Public Economics - All Lecture Note PDF

Public Economics Lectures
Part 1: Introduction
Raj Chetty and Gregory A. Bruich
Harvard University
Fall 2012
Public Economics Lectures () Part 1: Introduction 1 / 49

What is Public Economics?
Public economics focuses on answering two types of questions
1 How do government policies a¤ect the economy?
2 How should policies be designed to maximize welfare?
Three motivations for studying these questions:
1 Practical Relevance
2 Academic Interest
3 Methodology

Motivation 1: Practical Relevance
Interest in improving economic welfare ! interest in public economics
Almost every economic intervention occurs through government

policy (i.e. involves public economics) via two channels:
Price intervention: taxes, welfare, social insurance, public goods
Regulation: min wages, FDA regulations (25% of products consumed),

zoning, labor laws, min education laws, environment
Government directly employs one sixth of U.S. workforce

Motivation 1: Practical Relevance
Stakes are extremely large because of broad scope of policies
Ex. Tax reforms immediately a¤ect millions
Contentious debate on the appropriate role of government in society
Romney: replacing Medicare with decentralized private insurance will

improve health outcomes and reduce costs
Obama: Romney’s proposal will worsen health outcomes and raise costs
Only one of these views can be correct
Injecting science into these debates has great practical value

Motivation 2: Academic Interest
Public economics is typically the end point for many other sub…elds
Macro, development, labor, and corporate …nance questions often

ultimately motivated by a public economics question
Ex 1: Macro studies on costs of business cycles and intertemporal

models of household behavior
Ex 2: Labor studies on employment e¤ects of the minimum wage
Natural to combine public …nance with another …eld
Understanding public …nance can help ensure that you work on

relevant topics

Motivation 3: Methodology
Public economics is at the frontier of a methodological transformation

in applied microeconomics
Data-driven approach to answering important policy questions
Combines a broad set of skills: applied theory, applied econometrics,

simulation methods
Useful skill set for many applied …elds in economics
Topics in the course re‡ect a broad set of methodological themes

Theme 1: Connecting Theory to Data
Modern public economics tightly integrates theory with empirical

evidence to derive quantitative predictions about policy
What is the optimal income tax rate?
What is the optimal unemployment bene…t level?
Traditional approach: theoretical models and numerical simulations
Theory often makes weak predictions: optimal tax rate between 0 and
100%
Numerical simulations rely on strong assumptions
Recent work derives robust formulas that can be implemented using

well-identi…ed empirical estimates
Theme 2: Quasi-Experimental Empirical Methods
Research in public economics exploits a variety of quasi-experimental
research designs to identify parameters of interest
Event studies, regression discontinuity, synthetic control
Good way to learn practical lessons in applied econometrics
What is “identi…cation by functional form” and why is it undesirable?
Is the LATE or ATE of greater interest in your problem?
When is propensity score reweighting credible?
When do weak instrument problems arise and how can they be …xed?
Emphasis on non-parametric graphical techniques rather than

parametric regression models
Anscombe (1973): Dataset 1
15
β=0.5 (0.12)
Earnings ($1000)
5 0 10
0 5 10 15 20
Years of Schooling

15
β=0.5 (0.12)
Earnings ($1000)
5 0 10
0 5 10 15 20
Years of Schooling

15
β=0.5 (0.12)
Earnings ($1000)
5 0 10
0 5 10 15 20
Years of Schooling

15
β=0.5 (0.12)
Earnings ($1000)
5 0 10
0 5 10 15 20
Years of Schooling

Theme 3: “Big Data”
Compelling implementation of quasi-experimental methods requires a

lot of data
Recent availability of very large datasets has transformed research in

applied microeconomics
Scanner data on consumer purchases
Tax and social security records
School district databases

Use of Pre-Existing Survey Data in Publications in Leading Journals, 1980-2010
Micro-data Based Articles using Survey Data (%)
100
80
60
40
20
1980 1990 2000 2010

Year
AER JPE QJE ECMA
Note: “Pre-existing survey”datasets refer to micro surveys such as the CPS or SIPP and do not
include surveys designed by researchers for their study. Sample excludes studies whose primary
data source is from developing countries.

Use of Administrative Data in Publications in Leading Journals, 1980-2010
Micro-data Based Articles using Admin. Data (%)
100
80
60
40
20
1980 1990 2000 2010

Year
AER JPE QJE ECMA
Note: “Administrative”datasets refer to any dataset that was collected without directly surveying
individuals (e.g., scanner data, stock prices, school district records, social security records).
Sample excludes studies whose primary data source is from developing countries.

United States Tax Data
7 billion tax records covering full pop. from 1996 to today
Includes a rich set of information on individuals
Earnings from W-2’s (covers non-…lers)
Employer ID
College attendance
Retirement savings, charitable contributions
Housing and mortgage interest
Geographical location
Birth, death, marriage, children, family structure
Analogous corporate databank contains information for 5 million …rms

per year, linked to workers
What are the Bene…ts of Administrative Data?
1 Higher quality information: virtually no missing data or attrition
Current Population Survey non-response rate now 31% for income
2 Longitudinal tracking over long periods
Match rates of 95% over 20+ years in studies of long-term impacts of

early childhood education [Chetty et al. 2011, Chetty, Friedman,
Rocko¤ 2012]

Earnings at Age 28 vs. Teacher Value-Added in Elementary School
$21,200
$21,000
Earnings at Age 28
$20,800
1 SD TVA = $182
(73)
$20,600
$20,400
-0.2 -0.1 0 0.1 0.2

Teacher Value-Added
Source: Chetty, Friedman, Rockoff 2012

What are the Bene…ts of Administrative Data?
1 Higher quality information: virtually no missing data or attrition
Current Population Survey non-response rate now 31% for income
2 Longitudinal tracking over long periods
Match rates of 95% over 20+ years in studies of long-term impacts of

early childhood education [Chetty et al. 2011, Chetty, Friedman,
Rocko¤ 2012]
3 Very large sample sizes: 2,000 times the size of the CPS
Can develop new research designs

Earned Income Tax Credit Schedule for Single Earners with One Child in 2008
4k
This income level maximizes tax refund
EITC Credit Amount ($1000)
3k
2k
1k
0k
$0k $5k $10k $15k $20k $25k $30k $35k
Family Earnings

U.S. Income Distributions for EITC-Eligible Individuals with Children in 2008
5%
4%
Percent of Tax Filers
3%
2%
1%
0%
$0 $10K $20K $30K $40K

Total Earnings (Real 2010 $)
One child Two children

Fraction of Tax Filers Who Report Income that Maximizes EITC Refund
in 1996
4.1 – 42.7%
2.8 – 4.1%
2.1 – 2.8%
1.8 – 2.1%
1.5 – 1.8%
1.2 – 1.5%
1.1 – 1.2%
0.9 – 1.1%
0.7 – 0.9%
0 – 0.7%
Source: Chetty, Friedman, Saez 2012

in 1999
4.1 – 42.7%
2.8 – 4.1%
2.1 – 2.8%
1.8 – 2.1%
1.5 – 1.8%
1.2 – 1.5%
1.1 – 1.2%
0.9 – 1.1%
0.7 – 0.9%
0 – 0.7%

in 2002
4.1 – 42.7%
2.8 – 4.1%
2.1 – 2.8%
1.8 – 2.1%
1.5 – 1.8%
1.2 – 1.5%
1.1 – 1.2%
0.9 – 1.1%
0.7 – 0.9%
0 – 0.7%

in 2005
4.1 – 42.7%
2.8 – 4.1%
2.1 – 2.8%
1.8 – 2.1%
1.5 – 1.8%
1.2 – 1.5%
1.1 – 1.2%
0.9 – 1.1%
0.7 – 0.9%
0 – 0.7%

in 2008
4.1 – 42.7%
2.8 – 4.1%
2.1 – 2.8%
1.8 – 2.1%
1.5 – 1.8%
1.2 – 1.5%
1.1 – 1.2%
0.9 – 1.1%
0.7 – 0.9%
0 – 0.7%

Theme 4: Behavioral Models
Recent work in public economics draws on insights from psychology

and economics literature
Strong evidence that individuals fail to optimize
Raises new policy questions
Suggests new policy instruments
E.g. information, social incentives

Correlation Between Response to EITC and EITC Filer Density by ZIP Code
Fraction Self-Emp Reporting Income that Maxes Refund
3.5%
3.0%
2.5%
2.0%
1.5%
R2 = 0.6
1.0%
-2 0 2 4 6
log (Number of EITC Filers Per Square Mile)

Background Facts: Size and Growth of Government
Government expenditures = 1/3 GDP in the U.S.
Higher than 50% of GDP in some European countries
Decentralization is a key feature of U.S. govt
One third of spending (10% of GDP) is done at state-local level (e.g.

schools)
Two thirds (20% of GDP) is federal

Federal Government Revenue and Expenditure 1930-2009
50
Revenue and spending (% of GDP)
40
30
20
10
0
1930 1940 1950 1960 1970 1980 1990 2000 2010

Year
Revenue Expenditure
Source: Office of Management and Budget, Historical Tables, FY 2011

Total Government Spending by Country, 1970-2007
60
Sweden
50
Percent of GDP
Canada
40
OECD Avg.
30
United States
20
1970 1975 1980 1985 1990 1995 2000 2005

Year
Source: OECD Economic Outlook (2009)

Federal Revenues (% of total revenue)
Other
Other 4.2% Excise
4.2%
2.7%
Excise
12.6%
Income Income
44% Payroll 45.4% Payroll
15.9% 37.5%
Corporate
23.2% Corporate
12.1%
1960 2008
Source: Office of Management and Budget, historical tables, government receipts by source

State/Local Revenues (% of total revenue)
Income Tax
5.9%
Federal Grants
9.4%
Property Income Tax
Tax 15.7% 14.3%
Property
Tax Federal
Other Grants
38.2% Sales Tax
17.7% 19.1%
17.9%
Sales Tax
Other
28.8%
33%
1960 2007
Source: U.S. Census Bureau, 2007 Summary of State & Local Government

Federal Spending (% of total spending)
Education, welfare, housing 4% Education, welfare, housing 9.7%
Other
Other 11.2%
Social 12.4%
Social
Security Security
13.5% 19.5% Health
Health, 2.9% 23.1%
UI and Net Interest Net
Disability 8.3%
Interest
8.9% 12.3%
UI and Disability, 6.3%
1960 2001
Source: Office of Management and Budget, historical tables, government outlays by function

International Tax Revenue by Type of Tax (2001, % of Total)
Mexico Norway OECD Average
Payroll Payroll Payroll

24.3% Consumption 20.5% Consumption 26.7%
31.3% 32.6%
Wealth,
2.2% Individual
Consumption Income
73.5% Individual
Corporate 24.2% Income
Wealth, 2.2% Income 21.7% Wealth, 5.5% 26%
Corporate Income, 9.3%
Source: OECD 2002

Government Intervention in the Economy
Organizing framework: “When is government intervention necessary

in a market economy?”
Market …rst, govt. second approach
Why? Private market outcome is e¢ cient under broad set of conditions

(1st Welfare Thm)
Course can be split into two parts:
1 How can govt. improve e¢ ciency when private market is ine¢ cient?
2 What can govt. do if private market outcome is undesirable due to

redistributional concerns?

E¢ cient Private Market Allocation of Goods
Amy’
s
Consumption
Bob’
s Consumption

First Role for Government: Improve E¢ ciency
Amy’
s
Consumption
Bob’
s Consumption

Second Role for Government: Improve Distribution
Amy’
s
Consumption
Bob’
s Consumption

First Welfare Theorem
Private market provides a Pareto e¢ cient outcome under three

conditions
1 No externalities
2 Perfect information
3 Perfect competition
Theorem tells us when the government should intervene

Failure 1: Externalities
Markets may be incomplete due to lack of prices (e.g. pollution)
Achieving e¢ cient Coasian solution requires an organization to

coordinate individuals – that is, a government
This is why govt. funds public goods (highways, education, defense)
Questions: What public goods to provide and how to correct

externalities?

Failure 2: Asymmetric Information and Incomplete Markets
When some agents have more information than others, markets fail
Ex. 1: Adverse selection in health insurance
Healthy people drop out of private market ! unraveling

Mandated coverage could make everyone better o¤
Ex. 2: capital markets (credit constraints) and subsidies for education
Ex. 3: Markets for intergenerational goods
Future generations’interests may not be fully re‡ected in market

outcomes

Failure 3: Imperfect Competition
When markets are not competitive, there is role for govt. regulation
Ex: natural monopolies such as electricity and telephones
This topic is traditionally left to courses on industrial organization

and is not covered in this course
But taking the methodological approach of public economics to

problems traditionally analyzed in IO is a very promising area

Individual Failures
If agents do not optimize, government intervention (e.g. by forcing

saving via social security) may be desirable
This is an “individual” failure rather than a market failure
Conceptual challenge: how to avoid paternalism critique
Why does govt. know better what’s desirable for you (e.g. wearing a
seatbelt, not smoking, saving more)
Di¢ cult but central issues for optimal policy design

Redistributional Concerns
Even when the private market outcome is e¢ cient, may not have
good distributional properties
E¢ cient markets generally seem to deliver very large rewards to small

set of people (top incomes)
Government can redistribute income through tax and transfer system

Why Limit Government Intervention?
One solution to these issues would be for the government to oversee

all production and allocation in the economy (socialism).
Serious problems with this solution
1 Information: how does government aggregate preferences and

technology to choose optimal production and allocation?
2 Government policies distort incentives (behavioral responses in private

sector)
In practice, there are sharp tradeo¤s between costs and bene…ts of

government intervention

Equity-E¢ ciency Tradeo¤
Amy’
s
Consumption
Bob’
s Consumption

Three Types of Questions in Public Economics
1 Positive analysis: What are the observed e¤ects of government

programs and interventions?
2 Normative analysis: What should the government do if we can choose

optimal policy?
3 Public choice/Political economy
Develops theories to explain why the government behaves the way it

does and identify optimal policy given political economy concerns
Criticism of normative analysis: fails to take political constraints into

account

Course Outline
1 Tax Incidence and E¢ ciency
2 Optimal Taxation
3 Income Taxation and Labor Supply
4 Corporate Taxation
5 Social Insurance
6 Public Goods and Externalities

Part 2: Incidence of Taxation
Harvard University
Fall 2012
Public Economics Lectures () Part 2: Tax Incidence 1 / 140

Outline
1 De…nition and Introduction
2 Partial Equilibrium Incidence
3 Partial Equilibrium Incidence with Salience E¤ects
4 Partial Equilibrium Incidence: Empirical Applications
5 General Equilibrium Incidence
6 Capitalization
7 Mandated Bene…ts

References on Tax Incidence
Kotliko¤ and Summers (1987) handbook chapter
Atkinson and Stiglitz text chapters 6 and 7
Chetty, Looney, and Kroft (2009)

De…nition
Tax incidence is the study of the e¤ects of tax policies on prices and
the distribution of utilities
What happens to market prices when a tax is introduced or changed?
Increase tax on cigarettes by $1 per pack
Introduction of Earned Income Tax Credit (EITC)
Food stamps program
E¤ect on price ! distributional e¤ects on smokers, pro…ts of

producers, shareholders, farmers, ...

Economic vs. Statutory Incidence
Equivalent when prices are constant but not in general
Consider the following argument:
Government should tax capital income b/c it is concentrated at the

high end of the income distribution
Neglects general equilibrium price e¤ects
Tax might be shifted onto workers
If capital taxes ! less savings and capital ‡ight, then capital stock
may decline, driving return to capital up and wages down
Some argue that capital taxes are paid by workers and therefore
increase income inequality (Hassett and Mathur 2009)
Overview
Tax incidence is an example of positive analysis
Typically the …rst step in policy evaluation
An input into thinking about policies that maximize social welfare
Theory is informative about signs and comparative statics but is

inconclusive about magnitudes
Incidence of cigarette tax: elasticity of demand w.r.t. price is crucial
Labor vs. capital taxation: mobility of labor, capital are critical

Overview
Ideally, we would characterize the e¤ect of a tax change on utility

levels of all agents in the economy
Useful simpli…cation in practice: aggregate economic agents into a

few groups
Incidence analyzed at a number of levels:
1 Producer vs. consumer (tax on cigarettes)

2 Source of income (labor vs. capital)
3 Income level (rich vs. poor)
4 Region or country (local property taxes)
5 Across generations (social security reform)

Partial Equilibrium Incidence: Key Assumptions
1 Two good economy
Only one relative price ! partial and general equilibrium are same
Can be viewed as an approx. of incidence in a multi-good model if

the market being taxed is “small”
there are no close substitutes/complements in the utility fn
2 Tax revenue is not spent on the taxed good
Tax revenue is used to buy untaxed good or thrown away
3 Perfect competition among producers
Relaxed in some studies of monopolistic or oligopolistic markets

Partial Equilibrium Model: Setup
Two goods: x and y
Government levies an excise tax on good x
Excise or speci…c tax: levied on a quantity (e.g. gallon, pack, ton)

Ad-valorem tax: fraction of prices (e.g. sales tax)
Let p denote the pretax price of x and q = p + t denote the tax

inclusive price of x
Good y , the numeraire, is untaxed

Partial Equilibrium Model: Demand
Consumer has wealth Z and has utility u (x, y )
∂D q ∂ log D
Let εD = ∂q D (q ) = ∂ log q denote the price elasticity of demand
Elasticity: % change in quantity when price changes by 1%
Widely used concept because elasticities are unit free

Partial Equilibrium Model: Supply
Price-taking …rms
Use c (S ) units of the numeraire y to produce S units of x
Cost of production is increasing and convex:
c 0 (S ) > 0 and c 00 (S ) 0
Pro…t at pretax price p and level of supply S is pS c (S )
With perfect optimization, the supply function for good x is implicitly

de…ned by the marginal condition p = c 0 (S (p ))
∂S p
Let εS = ∂p S (p ) denote the price elasticity of supply

Partial Equilibrium Model: Equilibrium
Equilibrium condition
Q = S (p ) = D (p + t )
de…nes an equation p (t )
dp
Goal: characterize dt , the e¤ect of a tax increase on price
First consider some graphical examples to build intuition, then

analytically derive formula

Tax Levied on Consumers
Price
C
$27.0
Consumer
Burden = $4.50 A
$22.5
Supplier
Burden = $3.00 D
$19.5
$15.0 B
$7.50
D+t D
1250 1500 Quantity

Tax Levied on Producers
Price
S+t
$7.50 S
B
$30.0
C
$27.0
Consumer
Burden = $4.50 A
$22.5
Supplier
Burden = $3.00 D
$19.5
1250 1500 Quantity

Perfectly Inelastic Demand
Price D S+t
S
$27.0
Consumer
burden
$22.5
$7.50
1500 Quantity

Perfectly Elastic Demand
Price S+t
S
$7.50
$22.5 D
Supplier
burden
$15.0
1500 Quantity

Formula for Tax Incidence
Implicitly di¤erentiate equilibrium condition
D (p + t ) = S (p )
to obtain:
dp ∂D 1
=
dt ∂p ( ∂S
∂p
∂D
∂p )
dp εD
) =
dt εS εD
Incidence on consumers:
dq dp εS
= 1+ =
dt dt εS εD

Formula for Tax Incidence
P
1 –excess supply of E
P1 1 created by imposition of tax
dp = E/Ý /S
/p
? /D
/p
Þ 2 2 –re-equilibriation of market
P2 through producer price cut
/D /S /D
ö dp/dt = /p
/Ý /p
? /p
Þ
D1
D2
Q
/D
E = dt × /p
Tax Incidence with Salience E¤ects
Central assumption of neoclassical model: taxes are equivalent to

prices ( dx dx
dt = dp )
In practice, are people fully aware of marginal tax rates?
Chetty, Looney, and Kroft (2009) test this assumption and generalize
theory to allow for salience e¤ects
Part 1: Test whether “salience” (visibility of tax-inclusive price)

a¤ects behavioral responses to commodity taxation
Does e¤ect of a tax on demand depend on whether it is included in

posted price?
Part 2: Develop formulas for incidence and e¢ ciency costs of

taxation that permit salience e¤ects and other optimization errors
Chetty et al.: Empirical Framework
Economy with two goods, x and y
Prices: Normalize the price of y to 1 and let p denote the (…xed)

pretax price of x.
Taxes: y untaxed, x subject to an ad valorem sales tax τ (not

included in posted price)
Tax-inclusive price of x is q = (1 + τ )p
Let demand for good x be denoted by x (p, τ )

Chetty et al.: Empirical Framework
If agents optimize fully, demand should only depend on the total
tax-inclusive price: x (p, τ ) = x ((1 + τ )p, 0)
Full optimization implies price elasticity equals gross-of-tax elasticity:

∂ log x ∂ log x
εx ,p = εx ,1 +τS
∂ log p ∂ log(1 + τ )
To test this hypothesis, log-linearize demand fn. x (p, τ ) to obtain
estimating equation:
log x (p, τ ) = α + β log p + θβ log(1 + τ )

θ measures degree to which agents under-react to the tax:
∂ log x ∂ log x εx ,1 +τ
θ= / =
∂ log(1 + τ ) ∂ log p εx ,p
Chetty et al.: Two Empirical Strategies
Two strategies to estimate θ:
1 Manipulate tax salience: make sales tax as visible as pre-tax price
E¤ect of intervention on demand:

v = log x ((1 + τ )p, 0) log x (p, τ )
Compare to e¤ect of equivalent price increase to estimate θ:

v
(1 θ ) =
εx ,p log(1 + τ )
2 Manipulate tax rate: compare εx ,p and εx ,1 +τ

θ = εx ,1 +τ /εx ,p
Chetty et al.: Strategy 1
Experiment manipulating salience of sales tax implemented at a

supermarket that belongs to a major grocery chain
30% of products sold in store are subject to sales tax
Posted tax-inclusive prices on shelf for subset of products subject to

sales tax (7.375% in this city)
Data: Scanner data on price and weekly quantity sold by product

Documenting Salience Mechanism
Concern with posting tax inclusive prices: may have in‡uenced

behavior through various channels besides salience
Common problem in …eld experiments termed “Hawthorne e¤ects”
Di¢ cult to rule out all mechanisms, but helpful to present evidence
that mechanism of interest is very powerful

TABLE 1
Evaluation of Tags: Classroom Survey
Mean Median SD
Original Price Tags:
Correct tax-inclusive price w/in $0.25 0.18 0.00 0.39
Experimental Price Tags:

Correct tax-inclusive price w/in $0.25 0.75 1.00 0.43
T-test for equality of means: p < 0.001

N=49
Students were asked to choose two items from image.

Asked to report “Total bill due at the register for these two items.”
Source: Chetty, Looney, and Kroft (2009)

Chetty et al.: Research Design
Quasi-experimental di¤erence-in-di¤erences
Treatment group:
Products: Cosmetics, Deodorants, and Hair Care Accessories
Store: One large store in Northern California
Time period: 3 weeks (February 22, 2006 – March 15, 2006)
Control groups:
Products: Other prods. in same aisle (toothpaste, skin care, shave)
Stores: Two nearby stores similar in demographic characteristics
Time period: Calendar year 2005 and …rst 6 weeks of 2006

Effect of Posting Tax-Inclusive Prices: Mean Quantity Sold
TREATMENT STORE
Period Control Categories Treated Categories Difference
Baseline 26.48 25.17 -1.31

(0.22) (0.37) (0.43)
Experiment 27.32 23.87 -3.45

(0.87) (1.02) (0.64)
Difference 0.84 -1.30 DDTS = -2.14

over time (0.75) (0.92) (0.64)

TREATMENT STORE
Baseline 26.48 25.17 -1.31

(0.22) (0.37) (0.43)
Experiment 27.32 23.87 -3.45

(0.87) (1.02) (0.64)

over time (0.75) (0.92) (0.64)
CONTROL STORES
Baseline 30.57 27.94 -2.63

(0.24) (0.30) (0.32)
Experiment 30.76 28.19 -2.57

(0.72) (1.06) (1.09)
Difference 0.19 0.25 DDCS = 0.06

over time (0.64) (0.92) (0.90)

TREATMENT STORE
Baseline 26.48 25.17 -1.31

(0.22) (0.37) (0.43)
Experiment 27.32 23.87 -3.45

(0.87) (1.02) (0.64)

over time (0.75) (0.92) (0.64)
CONTROL STORES
Baseline 30.57 27.94 -2.63

(0.24) (0.30) (0.32)
Experiment 30.76 28.19 -2.57

(0.72) (1.06) (1.09)
Difference 0.19 0.25 DDCS = 0.06

over time (0.64) (0.92) (0.90)
DDD Estimate -2.20

(0.58)

Statistical Signi…cance and Permutation Tests
Standard error computations always require speci…c assumptions

about error structure
Ex: allow for correlation in purchases across products within a

store-week-category cell
Standard parametric approach is to cluster standard errors by

store-week-category
But appropriate level of clustering is often unclear
Should we also allow for correlation across stores or categories?

Non-Parametric Permutation Tests
Useful technique for inference with correlated errors: permutation

(Fisher’s exact) test
Pretend intervention occurred in each of the other cells (store, week,

category) of the sample and recompute DDD estimate
Calculate where actual treatment e¤ect lies in empirical CDF of

placebo treatment e¤ects

Non-Parametric Permutation Tests
Key assumption underlying permutation test: treatment is truly

random
Probability of treatment cannot vary across cells
True by construction in experiment
But may not hold in settings where policy changes are endogenous
Ex: unemployment bene…ts increased during recessionary periods
Nevertheless, often a useful benchmark

Chetty et al.: Strategy 2
Compare e¤ects of price changes and tax changes
Alcohol subject to two state-level taxes in the U.S.:
Excise tax: included in price
Sales tax: added at register, not shown in posted price
Exploiting state-level changes in these two taxes, estimate θ
Addresses concern that experiment may have induced a Hawthorne

e¤ect

Change in Log Per Capita Beer Consumption Per Capita Beer Consumption and State Beer Excise Taxes
.1
.05
0
-.05
-.1
-.02 -.015 -.01 -.005 0 .005 .01 .015 .02
Change in Log(1+Beer Excise Rate)

Change in Log Per Capita Beer Consumption Per Capita Beer Consumption and State Sales Taxes
.1
.05
0
-.05
-.1
-.02 -.015 -.01 -.005 0 .005 .01 .015 .02
Change in Log(1+Sales Tax Rate)

Effect of Excise and Sales Taxes on Beer Consumption
Dependent Variable: ∆Log(per capita beer consumption)
Baseline Bus Cyc, 3-Year Diffs Food Exempt

Alc Regs.
(1) (2) (3) (4)
ΔLog(1+Excise Tax Rate) -0.87 -0.89 -1.11 -0.91
(0.17)*** (0.17)*** (0.46)** (0.22)***
ΔLog(1+Sales Tax Rate) -0.20 -0.02 -0.00 -0.14

(0.30) (0.30) (0.32) (0.30)
Business Cycle Controls x x x
Alcohol Regulation Controls x x x
Year Fixed Effects x x x x
F-Test for Equality of Coeffs. 0.05 0.01 0.05 0.04
Sample Size 1,607 1,487 1,389 937

Let fx (p, t, Z ), y (p, t, Z )g denote empirically observed demands
Place no structure on these demand functions except for feasibility:
(p + t )x (p, t, Z ) + y (p, t, Z ) = Z
Demand functions taken as empirically estimated objects rather than

optimized demand from utility maximization
Supply side model same as above
Market clearing price p satis…es
D (p, t, Z ) = S (p )
where D (p, t, z ) = x (p, t, z ) is market demand for x.

Pre-tax
price p
DÝp|t S = 0Þ
DÝp|t S Þ S ( p)
1 –excess supply of E
p0 1
created by imposition of tax
/S /D
dp = E/Ý /p
? /p
Þ 2
p1 2 –re-equilibriation of market
/D /S /D
ö dp/dt = S
/t S
/Ý /p
? /p
Þ through pre-tax price cut
S,D
E = tS /D//t S

Tax Incidence with Salience E¤ects: Formula
Incidence on producers of increasing t is
dp ∂D/∂t εD
= = θ
dt ∂S /∂p ∂D/∂p εS εD
1 Incidence on producers attenuated by θ
2 No tax neutrality: taxes on producers have greater incidence on

producers than non-salient taxes levied on consumers
Intuition: Producers need to cut pretax price less when consumers are
less responsive to tax

Tax Incidence: Empirical Applications
1 [Evans, Ringel, and Stech 1999]: Cigarette excise taxes
2 [Hastings and Washington 2010]: Food stamps
3 [Rothstein 2010]: Earned Income Tax Credit

Evaluating Empirical Studies
Consider ideal experimental design …rst
Then formulate a feasible design and analyze its ‡aws relative to ideal
design
Frontier for empirical papers: tradeo¤ between quality of research

design and importance/novelty

Developing Empirical Research Designs
All of the empirical studies we will consider here start by formulating

clear research designs
Why develop an explicit design rather than simply use all available
variation in tax rates?
Consider estimating e¤ect of a treatment (e.g., tax) T on outcome y
yi = α + βTi + εi
Treatment is assigned based on a “selection” model
Ti = αT + βT Xi + η i
Treatment may be non-random: cov (Xi , εi ) 6= 0, cov (η i , εi ) 6= 0

Traditional approach to accounting for confounding factors or
selection: control for observables Xi when estimating treatment e¤ect
yi = α + βTi + γXi + εi
Can be done using OLS regression, matching, propensity-score
reweighting, etc.
Problem with these approaches: don’t know source of variation in Ti
Must be some reason that one person got treated and another did not
even if they are perfectly matched on observables (e.g., twins)
η i must be correlated with Ti to have variation in Ti jXi
But that same unobserved factor could also a¤ect outcome: no way to
know if cov (η i , εi ) = 0
Cannot be sure that estimate of treatment e¤ect β is consistent

A “research design” is a source of variation in η i that is credibly

unrelated to εi
Ex: a reform that a¤ects people above age 65 but not below
People at age 64 and 65 likely to have similar outcomes !

cov (η i , εi ) = 0
General lesson: “controlling” for confounding factors using regression

or reweighting will rarely give you convincing estimates
However, reweighting can be a useful technique to obtain better

control groups when paired with a quasi-experimental research design
More on this below (Dinardo, Fortin, Lemieux 1996)

Evans, Ringel, and Stech (1999)
Question: How do cigarette tax increases a¤ect prices?
Do they take money from cigarette companies or smokers?
Partial equilibrium is a plausible approximation for cigarettes, so use

that framework here

Cigarette Taxation: Background
Cigarettes taxed at both federal and state levels in U.S.
Total revenue of about $35 billion per year, similar to estate taxation
Federal tax increased from $0.39 to $1.01 per pack in 2009
Variation among states: from 30 cents per pack in VA to $4.35 in NY

in 2012
Controversial commodity due to health and paternalism concerns

Since 1975, more than 200 state tax changes ! natural experiments
to investigate tax incidence
Exploit these state-level changes in excise tax rates using simple

di¤-in-di¤ research designs
Idea: Suppose federal govt. implements a tax change. Compare

cigarette prices before and after the change
D = [PA1 PA0 ]
Identi…cation assumption: absent the tax change, there would have

been no change in cigarette price

Di¤erence-in-Di¤erence
But what if price ‡uctuates because of climatic conditions or trends in

demand?
!First di¤erence (and time series) estimate biased
Can relax ID assumption using di¤-in-di¤
DD = [PA1 PA0 ] [ PB 1 PB 0 ]
State A: experienced a tax change (treatment)
State B: does not experience any tax change (control)
Identifying assumption for DD: “parallel trends:” absent the policy

change, P1 P0 would have been the same for A and B

Parallel Trends Assumption
Can use placebo DD to test parallel trends assumption
Analogous to permutation test: pretend reform occurred at other

points and replicate estimate
If DD in other periods is not zero, then DDt =1 likely biased
Useful to plot long time series of outcomes for treatment and control
Pattern should be parallel lines, with sharp change just after reform
Rest of U.S. a good control for MI in example above but not AZ

Triple Di¤erence
Some studies use a “triple di¤erence” (DDD)
Chetty, Looney, Kroft (2009): experiment using treatment/control

products, treatment/control stores
DDD = DDTS DDCS
DDTS : di¤erence of treat., cntrl products in treat. store
DDCS : di¤erence of treat., cntrl. products in cntrl. store
DDD is mainly useful as a robustness check:
DDCS 6= 0, unconvincing that DDD removes all bias
DDCS = 0, then DD = DDD but DD has smaller s.e.

Fixed E¤ects
ERS have data for 50 states, 30 years, and many tax changes
Want to pool all this data to obtain single incidence estimate
Fixed e¤ects generalize DD with S > 2 periods and J > 2 groups
Suppose that group j in year t experiences policy T of intensity Tjt
Want to identify e¤ect of T on price P. OLS regression:
Pjt = α + βTjt + ejt
With no …xed e¤ects, the estimate of β is biased if treatment Tjt is

correlated with ejt
Ex: states with higher taxes may have more anti-tobacco campaigns
Fixed E¤ects
Include time and state dummies to solve this problem:
Pjt = α + γt + δj + βTjt + ejt
Fixed e¤ect regression is equivalent to partial regression
P̂jt = βT̂jt + ejt
where P̂jt = Pjt Pj bjt is de…ned analogously

Pt and T
Identi…cation obtained from within-state variation over time
Note: common changes that apply to all groups (e.g. fed tax change)
captured by time dummy; not a source of variation that identi…es β

Fixed E¤ects vs. Di¤erence-in-Di¤erence
Advantage relative to DD: more precise estimates by pooling several

changes
Disadvantage: …xed e¤ects is a black-box regression, more di¢ cult to

check trends non-parametrically as with a single change
Combine with graphical, non-parametric evidence around certain policy

changes
Also useful to scatter residuals P̂jt vs. T̂jt
Same parallel trends identi…cation assumption as DD
Potential violation: policy reforms may respond to trends in outcomes
Ex: tobacco prices falling ! state decides to raise tax rate

Implement a …xed e¤ects model for prices
Regress price on state+year …xed e¤ects, covariates, and tax rate (in
cents)
Also estimate demand elasticities using …xed e¤ects estimator
Regress log quantity consumed on state/year …xed e¤ects, covariates,

and real tax rate (in cents)

Evans, Ringel, and Stech: Incidence Results
100% pass through implies supply elasticity of εS = ∞ at state level
Theory suggests that pass through would be lower at national level
Important to understand how the variation you are using determines

what parameter you are identifying

Evans, Ringel, and Stech: Demand Elasticity
Demand model estimate implies that: εD = 0.42
! 10% increase in price induces a 4.2% reduction in consumption
How to compute price elasticity of demand when using variation

arising from tax changes?
Tax passed 1-1 onto consumers, so we can substitute ∆P = ∆T here
Then compute εD from β̂ = (∆Q/Q )/∆T from regression coe¢ cient

of log demand on cigarette tax:
P ∆Q
εD = = β̂/P
Q ∆T
with P (price) and Q (quantity) are sample means
IV Estimation of Price Elasticities
How to estimate price elasticity of demand when tax and prices do
not move together 1-1?
Instrument for prices using taxes
First stage, taking note of F-stat:

Pjt = α0 + γt0 + δj0 + βTjt + ejt
Second stage:
bjt + ejt
Qjt = α + γt + δj + λP
Reduced form, using Tjt as an instrument for Pjt :
Qjt = α + γt + δj + µTjt + ejt
2SLS regression coe¤. is ratio of redued-form to …rst-stage coe¤.:
λ̂ = µ̂/ β̂
2SLS rescales reduced-form to account for ∆P/∆T 6= 1
Evans, Ringel, and Stech: Long Run Elasticity
DD before and after one year captures short term response: e¤ect of
current price Pjt on current consumption Qjt
F.E. also captures short term responses
What if full response takes more than one period? Especially

important considering nature of cigarette use
F.E. estimate biased. One solution: include lags (Tj ,t 1 , Tj ,t 2 , ...).
Are identi…cation assumptions still valid here? Tradeo¤ between LR

and validity of identi…cation assumptions

Evans, Ringel, and Stech: Distributional Incidence
Use individual data to see who smokes by education group and

income level
Spending per capita decreases with the income level
Tax is regressive on an absolute level (not only that share of taxes

relative to income goes down)
Conclusion: Taxes levied on cigarette companies lead to poor paying

more for same goods, with no impact on companies!

Cigarette Tax Incidence: Other Considerations
1 Lifetime vs. current incidence (Poterba 1989)
Finds cigarette, gasoline and alcohol taxation are less regressive (in
statutory terms) from a lifetime perspective
High corr. between income and cons share in cross-section; weaker
corr. with permanent income.
2 Behavioral models (Gruber and Koszegi 2004)
If agents have self control problems, incidence conc. on poor is

bene…cial to the extent that they smoke less
3 Intensive vs. extensive margin: Adda and Cornaglia (2006)
Use data on cotinine (biomarker) levels in lungs to measure inhalation

Higher taxes lead to fewer cigarettes smoked but no e¤ect on cotinine
in lungs, implying longer inhalation of each cigarette
Hastings and Washington 2010
Question: How does food stamps subsidy a¤ect grocery store pricing?
Food stamps typically arrive at the same time for a large group of
people, e.g. …rst of the month
Use this variation to study:
1 Whether demand changes at beginning of month (violating PIH)
2 How much of the food stamp bene…t is taken by …rms by increased

prices rather than consumers (intended recipients)

Hastings and Washington: Data
Scanner data from several grocery stores in Nevada
Data from stores in high-poverty areas (>15% food stamp recipients)

and in low-poverty areas (<3%)
Club card data on whether each individual used food stamps
Data from other states where food stamps are staggered across
month used as a control
Research design: use variation across stores, individuals, and time of

month to measure pricing responses

Hastings and Washington: Results
Demand increases by 30% in 1st week, prices by about 3%
Very compelling because of multiple dimensions of tests:

cross-individual, cross-store, cross-category, and cross-state
Interesting theoretical implication: subsidies in markets where

low-income recipients are pooled with others have better
distributional e¤ects
May favor food stamps as a way to transfer money to low incomes

relative to a subsidy such as the EITC

Rothstein 2010
How does EITC a¤ect wages?
EITC payments subsidize work and transfer money to low income

working individuals ($50 bil/year)
This subsidy could be taken by employers by shifting wage
Ex: inelastic demand for low-skilled labor and elastic supply ! wage
rate adjusts 1-1 with EITC
Policy question: are we actually transferring money to low incomes

through this program or are we just helping business owners?

Rothstein: Model
Rothstein considers a model of the labor market with three types of

agents
1 Employers
2 EITC-eligible workers
3 EITC-ineligible workers
Extends standard partial eq incidence model to allow for di¤erentiated

labor supply and di¤erent tax rates across demographic groups
Heterogeneity both complicates the analysis and permits identi…cation
Identi…cation strategy: compare wage changes across groups who

were a¤ected di¤erently by expansions of EITC program from 1992-94

Rothstein: Empirical Strategy
Two main challenges to identi…cation:

1 EITC 1992-1994 expansion when nation coming out of recession
! Compare to other workers (EITC ineligible, slightly higher incomes)
2 Violation of common trends assumption: technical change, more
demand for low-skilled workers in 1990s.
! Compare to trends in pre-period (essentially a DDD strategy)
Two dependent variables of interest:

1 [Prices] Measure how wages change for a worker of given skill
2 [Quantities] Measure how demand and supply for workers of each skill
type change because of EITC
Basic concept: use two moments – wage and quantity changes to

back out slopes of supply and demand curves

Rothstein: Empirical Strategy
Ideal test: measure how wage of a given individual changes when

EITC is introduced relative to a similar but ineligible individual
Problem: data is CPS repeated cross-sections. Cannot track “same

individual.”
Moreover, wage rigidities may prevent cuts for existing employees.
Solution: reweighting procedure to track “same skill” worker over

time (DiNardo, Fortin, and Lemieux 1996)

DFL Reweighting
Generalizes propensity score reweighting
Used to examine changes in distributions semi-parametrically,

conditioning on observables
Example: suppose wages are a function purely of height
When EITC is expanded, average observed height of workers falls

because less-skilled (shorter) people enter the labor force
We want to identify how wage distribution changes for people of

given height
Solution: hold “…xed” height semi-parametrically by reweighting the

distribution of wages ex-post to match heights ex-ante.
DFL Reweighting
Example: 100 short, 100 tall pre-reform and 200 short, 100 tall
post-reform
Then put 2/3 weight on tall and 1/3 on short when calculating wage
distribution after reform
Compare reweighted post-reform distribution to pre-reform

distribution to assess e¤ect of expansion on wages
Key assumption for causal interpretation of changes: selection on

observables
Here it is height; more generally, experience, age, demographics, etc.

Rothstein: Results
Basic DFL comparisons yield perverse result: groups that bene…ted

from EITC and started working more had more wage growth
Potential explanation: demand curve shifted di¤erentially – higher

demand for low skilled workers in 1990s.
To deal with this, repeats same analysis for 1989-1992 (no EITC
expansion) and takes di¤erences
Changes sign back to expected, but imprecisely estimated

Rothstein: Results
Ultimately uses quantity estimates and incidence formula to back out

predicted changes
Wage elasticity estimates: 0.7 for labor supply, 0.3 for labor demand
Implications using formulas from model:
EITC-eligible workers gain $0.70 per $1 EITC expansion
Employers gain about $0.70
EITC-ineligible low-skilled workers lose about $0.40
On net, achieve only $0.30 of redistribution toward low income

individuals for every $1 of EITC

Rothstein: Caveats
1 Identi…cation heavily complicated by recession, trends (SBTC); no

clean control group
2 Data limitations: no panel data; problems in measurement – no

annual income, cannot measure MTR
3 Short run vs. long run e¤ects; important due to evidence of nominal
wage rigidities.
4 Pure extensive-margin analysis. Intensive margin would go the other

way b/c EITC is not a marginal subsidy to wage for a very large
fraction of the population.

General Equilibrium Analysis
Now move beyond two-good partial equilibrium model to analyze

impacts on all prices
Typical goal: trace out full incidence of taxes back to original owners
of factors
Partial equilibrium: “producer” vs. consumer
General equilibrium: capital owners vs. labor vs. landlords, etc.

General Equilibrium Analysis
Two types of GE models:
1 Static: many sectors or many factors of production
Workhorse analytical model: Harberger (1962): 2 sector and 2 factors

of production
Computational General Equilibrium: many sectors, many factors of

production model
2 Dynamic
Characterize impacts over time or across generations
Asset price e¤ects: capitalization

Harberger 1962 Two Sector Model
1 Fixed total supply of labor L and capital K (short-run, closed

economy)
2 Constant returns to scale in both production sectors
3 Full employment of L and K
4 Firms are perfectly competitive
Implicit assumption: no adjustment costs for capital and labor

Harberger Model: Setup
Production in sectors 1 (bikes) and 2 (cars):
X1 = F1 (K1 , L1 ) = L1 f (k1 )
X2 = F2 (K2 , L2 ) = L2 f (k2 )
with full employment conditions K1 + K2 = K and L1 + L2 = L
Factors w and L fully mobile ! in eq., returns must be equal:

w = p1 F1L = p2 F2L
r = p1 F1K = p2 F2K
Demand functions for goods 1 and 2:
X1 = X1 (p1 /p2 ) and X2 = X2 (p1 /p2 )
Note: all consumers identical so redistribution of incomes via tax
system does not a¤ect demand via a feedback e¤ect
System of ten eq’ns and ten unknowns: Ki , Li , pi , Xi , w , r

Harberger Model: E¤ect of Tax Increase
Introduce small tax d τ on rental of capital in sector 2 (K2 )
All eqns the same as above except r = (1 d τ )p2 F2K
Linearize the 10 eq’ns around initial equilibrium to compute the e¤ect

of d τ on all 10 variables (dw , dr , dL1 , ...)
Labor income = wL with L …xed, rK = capital income with K …xed
Therefore change in prices dw /d τ and dr /d τ describes how tax is

shifted from capital to labor
Changes in prices dp1 /d τ, dp2 /d τ describe how tax is shifted from

sector 2 to sector 1
Kotliko¤ and Summers (Section 2.2) state linearized equations as a

fn. of substitution elasticities
Harberger Model: Main E¤ects
1. Substitution e¤ects: capital bears incidence
Tax on K2 shifts production in Sector 2 away from K so aggregate

demand for K goes down
Because total K is …xed, r falls ! K bears some of the burden

2. Output e¤ects: capital may not bear incidence
Tax on K2 implies that sector 2 output becomes more expensive

relative to sector one
Therefore demand shifts toward sector 1
Case 1: K1 /L1 < K2 /L2 (1: bikes, 2: cars)
Sector 1 is less capital intensive so aggregate demand for K goes down
Output e¤ect reinforces subst e¤ect: K bears the burden of the tax
Case 2: K1 /L1 > K2 /L2 (1: cars, 2: bikes)
Sector 1 is more capital intensive, aggregate demand for K increases
Subst. and output e¤ects have opposite signs; labor may bear some or
all the tax
3. Substitution + Output = Overshifting e¤ects
Case 1: K1 /L1 < K2 /L2

Can get overshifting of tax, dr < d τ and dw > 0
Capital bears more than 100% of the burden if output e¤ect su¢ ciently
strong
Taxing capital in sector 2 raises prices of cars ! more demand for

bikes, less demand for cars
With very elastic demand (two goods are highly substitutable), demand
for labor rises sharply and demand for capital falls sharply
Capital loses more than direct tax e¤ect and labor suppliers gain

3. Substitution + Output = Overshifting e¤ects
Case 2 : K1 /L1 > K2 /L2
Possible that capital is made better o¤ by capital tax
Labor forced to bear more than 100% of incidence of capital tax in

sector 2
Ex. Consider tax on capital in bike sector: demand for bikes falls,
demand for cars rises
Capital in greater demand than it was before ! price of labor falls

substantially, capital owners actually gain
Bottom line: taxed factor may bear less than 0 or more than 100% of
tax.
Harberger Two Sector Model
Theory not very informative: model mainly used to illustrate negative

result that “anything goes”
More interest now in developing methods to identify what actually

happens
Original application by Harberger: sectors = housing and corporations
Capital in these sectors taxed di¤erently because of corporate income

tax and many tax subsidies to housing
Ex: Deductions for mortgage interest about $80 bn total
Harberger made assumptions about elasticities and calculated

incidence of corporate tax given potential to substitute into housing
Computable General Equilibrium Models
Harberger analyzed two sectors; subsequent literature expanded

analysis to multiple sectors
Analytical methods infeasible in multi-sector models
Instead, use numerical simulations to investigate tax incidence e¤ects

after specifying full model
Pioneered by Shoven and Whalley (1972). See Kotliko¤ and

Summers section 2.3 for a review
Produced a voluminous body of research in PF, trade, and

development economics

CGE Models: General Structure
N intermediate production sectors
M …nal consumption goods
J groups of consumers who consume products and supply labor
Each industry has di¤erent substitution elasticities for capital and

labor
Each consumer group has Cobb-Douglas utility over M consumption

goods with di¤erent parameters
Specify all these parameters (calibrated to match some elasticities)

and then simulate e¤ects of tax changes

Criticism of CGE Models
Findings very sensitive to structural assumptions
Ex: assumption of perfect competition
Key behavioral elasticities and functional form assumptions
Modern econometric methods conceptually not suitable for GE

problems
The whole point is “spillover e¤ects” (contamination)
Need a new empirical paradigm to deal with these problems

Open Economy Application
Key assumption in Harberger model: both labor and capital perfectly

mobile across sectors
Now apply framework to analyze capital taxation in open economies,

where capital is more likely to be mobile than labor
See Kotliko¤ and Summers section 3.1 for a good exposition

Open Economy Application: Framework
One good, two-factor, two-sector model
Sector 1 : small open economy where L1 is …xed and K1 mobile
Sector 2 : rest of the world L2 …xed and K2 mobile
Total capital stock K = K1 + K2 is …xed

Open Economy Application: Framework
Small country introduces tax on capital income (K1 )
After-tax returns must be equal:
r = F2K = (1 τ )F1K
Capital ‡ows from 1 to 2 until returns are equalized; if 2 is large

relative to 1, no e¤ect on r
Wage rate w1 = F1L (K1 , L1 ) dec. when K1 dec. b/c L1 is …xed
Return of capitalists in small country is unchanged; workers in home

country bear the burden of the tax
Taxing capital is bad for workers!

Open Economy Application: Empirics
Mobility of K drives the previous result
Empirical question: is K actually mobile across countries?
Two strategies:
1 Test based on prices and equilibrium relationships [Macro-…nance]
2 Look at mobility directly [Feldstein and Horioka 1980]

Strategy One: Macro-Finance approach
Test based on prices and equilibrium relationships
Check whether net returns r are equal across countries
General …nding - covered interest parity: obligations that are

protected against ‡uctuations in in‡ation and exchange rates have the
same returns across countries
Di¢ culties in generalization: many assets yield di¤erent returns,

unexpected in‡ation, changes in currency exchange rates
Need models with uncertainty, risk aversion to deal with other assets
Di¢ cult to implement this test for risky assets

Feldstein and Horioka 1980
Second strategy: look at capital mobility directly
Feldstein and Horioka use data on OECD countries from 1960-74
Closed economy: S = I ; open economy: S I =X M
Motivates regression:
I /GDP = α + βS /GDP + ...
Find β = 0.89 (0.07)

Feldstein and Horioka 1980
In closed economy, β = 1
But do not know what β should be in an open economy
β may be close to 1 in open economy if
1 Policy objectives involving S I (trade de…cit balance)
2 Summing over all countries: S̄ = Ī as imports and exports cancel out
3 Data problem: S constructed from I in some countries

Open Economy Applications: Empirics
Large subsequent literature runs similar regressions and …nds mixed

results
Generally …nds more ‡ow of capital and increasing over time
General view: cannot extract money from capital in small open

economies
Ex. Europe: tax competition has led to lower capital tax rates
Could explain why state capital taxes are relatively low in the U.S.

General Equilibrium Incidence in Dynamic Models
Static analysis above assumes that all prices and quantities adjust
immediately
In practice, adjustment of capital stock and reallocation of labor takes

time
Dynamic CGE models incorporate these e¤ects; even more complex
Static model can be viewed as description of steady states
During transition path, measured ‡ow prices (r , w ) will not correspond

to steady state responses
How to measure incidence in dynamic models?

Capitalization and the Asset Price Approach
Asset prices can be used to infer incidence in dynamic models

(Summers 1983)
Study e¤ect of tax changes on asset prices
Asset prices adjust immediately in e¢ cient markets, incorporating the

full present-value of subsequent changes
E¢ cient asset markets incorporate all e¤ects on factor costs, output

prices, etc.
Limitation: can only be used to characterize incidence of policies on

capital owners
There are no markets for individuals

Simple Model of Capitalization E¤ects
Firms pay out pro…ts as dividends
Pro…ts determined by revenues net of factor payments:

Dt qt Xt wjt Ljt
V =∑ =∑
1+r 1+r
Change in valuation of …rm ( dV
dt ) re‡ects change in present value of
pro…ts
dV
dt is a su¢ cient statistic that incorporates changes in all prices
Empirical applications typically use “event study” methodology
Examine pattern of asset prices or returns over time, look for break at
time of announcement of policy change
Problem: clean shocks are rare; big reforms do not happen suddenly
and are always expected to some extent
Capitalization: Empirical Applications
Classic application: Cutler (1988) on e¤ect of Tax Reform Act of

1986 on corporations
Focus on two more recent studies here
1 [Friedman 2008] E¤ect of Medicare Part D on drug companies
2 [Linden and Rocko¤ 2008] E¤ect of a sex o¤ender moving into

neighborhood on home values

Friedman 2008
Medicare part D passed by Congress in 2003; enacted in 2006
Expanded Medicare coverage to include prescription drugs (provided

coverage for 10 mil additional people)
What is the incidence of Medicare part D? How much of the

expenditure is captured by drug companies through higher pro…ts?
Basic research design: event study
Plot excess (market adjusted) returns for drug companies around FDA
approval of drugs

Event Study Designs
Event studies are a powerful research design when treatments are

staggered in time across individuals
Use a group of treated individuals as counterfactuals for each other to

account for time series trends
Good for identifying sharp, short-run e¤ects but not longer-term

impacts
Methodology
1 De…ne "event time" as calendar time minus date of treatment for each
treated obs.
2 Plot means/medians, etc. of outcome variable by event time

Excess Returns Around Drug Approval Date

Friedman 2008
Test whether excess returns for high Medicare share drugs is higher
after Medicare Part D is passed
Let MMSi denote medicare market share drug class i. Second-stage

estimating equation:
Excessi = α + βMMSi + γPost2003t + λPost2003t MMSi

Distribution of Excess Returns around Drug Approval:
Pre-Reform (1999-2002)

Distribution of Excess Returns around Drug Approval:
Post-Reform (2004-2007)

Friedman: Results
Concludes that drug companies’pro…ts increased by $250 bn in

present value because of Medicare Part D
Rough calibration suggests that drug companies capture about 1/3 of

total surplus from program

Linden and Rocko¤ 2008
Another common application is to housing market to assess WTP for

amenities
Examples: pollution, schools, crime
Rocko¤ and Linden (2008) estimate costs of crime using

capitalization approach
Research design: examine how house prices change when a registered

sex o¤ender moves into a neighborhood
Data: public records on o¤ender’s addresses and property values in

North Carolina

Linden and Rocko¤: Results
Find house prices decline by about 4% ($5500) when a sex o¤ender is

located within 0.1 mile of the house
Implied cost of a sexual o¤ense given probabilities of a crime: $1.2

million
This is far above what is used by Dept of Justice
How to interpret evidence: true cost of crime or a behavioral e¤ect?
Why does price fall only within 0.1 mile radius?

Mandated Bene…ts
We have focused until now on incidence of price interventions (taxes,

subsidies)
Similar incidence/shifting issues arise in analyzing quantity

intervention (regulations)
Leading case: mandated bene…ts – requirement that employers pay

for health care, workers compensation bene…ts, child care, etc.
Mandates are attractive to government because they are “o¤

budget”; may re‡ect salience issues

Mandated Bene…ts
Tempting to view mandates as additional taxes on …rms and apply

same analysis as above
But mandated bene…ts have di¤erent e¤ects on equilibrium wages

and employment di¤erently than a tax (Summers 1989)
Key di¤erence: mandates are a bene…t for the worker, so e¤ect on

market equilibrium depends on bene…ts workers get from the program
Unlike a tax, may have no distortionary e¤ect on employment and

only an incidence e¤ect (lower wages)

Mandated Bene…ts: Simple Model
Labor demand (D) and labor supply (S) are functions of the wage, w
Initial equilibrium:
D (w0 ) = S (w0 )
Now, govt mandates employers provide a bene…t with cost t
Workers value the bene…t at αt dollars
Typically 0 < α < 1 but α > 1 possible with market failures
Labor cost is now w + t, e¤ective wage w + αt
New equilibrium:
D (w + t ) = S (w + αt )

Mandated Benefit
Wage S
Rate
w1 A
D1
L1 Labor Supply

Mandated Benefit
Wage S
Rate
w1 A
$1
D2 D1
L1 Labor Supply

Mandated Benefit
Wage S
Rate
$α
w1 A
B
C
w2
$1
D2 D1
L1 Labor Supply

Mandated Bene…ts: Incidence Formula
Analysis for a small t: linear expansion around initial equilibrium
(dw /dt + 1)D 0 = (dw /dt + α)S 0

dw /dt = (D 0 αS 0 )/(S 0 D0)
ηS
= 1 + (1 α)
ηS ηD
where
η D = wD 0 /D < 0
η S = wS 0 /S > 0
If α = 1, dw /dt = 1 and no e¤ect on employment
More generally: 0 < α < 1 equivalent to a tax 1 α with usual

incidence and e¢ ciency e¤ects
Mandated Bene…ts: Empirical Applications
Focus here on Acemoglu and Angrist (2001) study of Americans with

Disabilities Act
Other applications
1 Gruber and Krueger (1991) on workers compensation mandates

2 Gruber (1994) on mandated maternity bene…ts
3 Kolstad and Kowalski (2012) on health insurance mandates

Acemoglu and Angrist (2001)
Look at e¤ect of ADA regulations on wages and employment of the

disabled
The 1993 Americans with Disabilities Act requires employers to:
Make accommodations for disabled employees
Pay same wages to disabled employees as to non-disabled workers
Cost to accommodate disabled workers: $1000 per person on average
Theory is ambiguous on net employment e¤ect because of wage

discrimination clause

Mandated Benefit with Minimum Wage
Wage S
Rate
w1 A
B
minimum wage
w2
D2 D1
L1 Labor Supply

Acemoglu and Angrist 2001
Acemoglu and Angrist estimate the impact of act using data from the
Current Population Survey
Examine employment and wages of disabled workers before and after

the ADA went into e¤ect

Acemoglu and Angrist: Results
Employment of disabled workers fell after the reform:
About a 1.5-2 week drop in employment for males, roughly a 5-10%

decline in employment
Wages did not change
Results consistent w/ labor demand elasticity of about -1 or -2
Firms with fewer than 25 workers exempt from ADA regulations; no

employment reduction for disabled at these …rms
ADA intended to help those with disabilities but appears to have hurt
many of them because of wage discrimination clause
Underscores importance of considering incidence e¤ects before

implementing policies
Part 3: E¢ ciency Cost of Taxation
Harvard University
Fall 2012
Public Economics Lectures () Part 3: E¢ ciency 1 / 106

Outline
1 Marshallian surplus
2 Path dependence problem and income e¤ects
3 De…nitions of EV, CV, and excess burden with income e¤ects
4 Harberger Approximation
5 Exact Consumer Surplus (Hausman 1981)
6 Empirical Applications
7 Welfare Analysis in Behavioral Models

De…nition
Incidence: e¤ect of policies on distribution of economic pie
E¢ ciency or deadweight cost: e¤ect of policies on size of the pie
Focus in e¢ ciency analysis is on quantities, not prices

References
Auerbach (1985) handbook chapter
Chetty, Looney, Kroft (AER 2009)
Chetty (Annual Review 2009)
Mas-Colell, Whinston, Green Chapter 3 for background on price

theory concepts

E¢ ciency Cost: Introduction
Government raises taxes for one of two reasons:
1 To raise revenue to …nance public goods
2 To redistribute income
But to generate $1 of revenue, welfare of those taxed falls by more

than $1 because the tax distorts behavior
How to implement policies that minimize these e¢ ciency costs?
Start with positive analysis of how to measure e¢ ciency cost of a given

tax system

Marshallian Surplus: Assumptions
Simplest analysis of e¢ ciency costs: Marshallian surplus
Two assumptions:
1 Quasilinear utility: no income e¤ects, money metric
2 Competitive production

Partial Equilibrium Model: Setup
Two goods: x and y
Consumer has wealth Z , utility u (x ) + y , and solves
max u (x ) + y s.t. (p + τ )x (p + τ, Z ) + y (p + τ, Z ) = Z
x ,y
Firms use c (S ) units of the numeraire y to produce S units of x
Marginal cost of production is increasing and convex:
c 0 (S ) > 0 and c 00 (S ) 0
Firm’s pro…t at pretax price p and level of supply S is
pS c (S )

Model: Equilibrium
With perfect optimization, supply fn for x is implicitly de…ned by the

marginal condition
p = c 0 (S (p ))
0
Let η S = p SS denote the price elasticity of supply
Let Q denote equilibrium quantity sold of good x
Q satis…es:
Q ( τ ) = D (p + τ ) = S (p )
Consider e¤ect of introducing a small tax d τ > 0 on Q and surplus

Excess Burden of Taxation
Price
$30.0 A
1500 Quantity

Excess Burden of Taxation
Price
S+t
B
Excess Burden
$36.0
$30.0 A
$t
1350 1500 Quantity

E¢ ciency Cost: Qualitative Properties
1 Excess burden increases with square of tax rate
2 Excess burden increases with elasticities

EB Increases with Square of Tax Rate
P
S
P1
A
Q1 Q

P S+t1
S
B
P2
P1
A
$t1
D
Q2 Q1 Q

P S+t1+t2 S+t1
S
P3 E
B
P2
Change in EB
P1
A
$t2
C
Q3 Q2 Q1 Q

Comparative Statics
(a) Inelastic Demand (b) Elastic Demand
P P
S+t S+t
S S
B
P2
B
P2
P1 P1 A
A
C
D
$t $t C
Q 2 Q1 Q Q2 Q1 Q

Tax Policy Implications
With many goods, the most e¢ cient way to raise tax revenue is:
1 Tax inelastic goods more (e.g. medical drugs, food)
2 Spread taxes across all goods to keep tax rates relatively low on all
goods (broad tax base)
These are two countervailing forces; balancing them requires

quantitative measurement of excess burden

Measuring Excess Burden: Marshallian Surplus
How to measure excess burden? Three empirically implementable

methods:
1 In terms of supply and demand elasticities
2 In terms of total change in equilibrium quantity caused by tax
3 In terms of change in government revenue

Method 1: Supply and Demand Elasticities
1
EB = dQd τ
2
1 0 ηD
EB = S (p )dpd τ = (1/2)(pS 0 /S )(S /p ) d τ2
2 ηS ηD
1 ηS ηD dτ
EB = pQ ( )2
2 ηS ηD p
ηD
Note: second line uses incidence formula dp = ( η ηD )d τ
S
Tax revenue R = Qd τ
Useful expression is deadweight burden per dollar of tax revenue:
EB 1 ηS ηD d τ
=
R 2 ηS ηD p
Method 2: Distortions in Equilibrium Quantity
dQ p 0
De…ne η Q = dτ Q
η Q : e¤ect of a 1% increase in price via a tax change on equilibrium

quantity, taking into account the endogenous price change
This is the coe¢ cient β in a reduced-form regression:

τ
log Q = α + β +ε
p0
Identify β using exogenous variation in τ. Then:
dQ
EB = (1/2) d τd τ
dτ
dQ p Q
= (1/2) ( )( )d τd τ
dτ Q p
dτ
= (1/2)η Q pQ ( )2
p
Marginal Excess Burden of Tax Increase
Excess burden of a tax τ is
dQ 2
EB (τ ) = τ (1/2)
dτ
Consider EB from raising tax by ∆τ given pre-existing tax τ:
dQ
EB (∆τ ) = (1/2) [(τ + ∆τ )2 τ 2 ]
dτ
dQ
= (1/2) [2τ ∆τ + (∆τ )2 ]
dτ
dQ dQ
= τ ∆τ (1/2) (∆τ )2
dτ dτ
First term is …rst-order in ∆τ; second term is second-order ((∆τ )2 )
This is why taxing markets with pre-existing taxes generates larger

marginal EB
EB of ∆τ = 1% is 10 times larger if τ = 10% than if τ = 0.

First vs. Second-Order Approximations
Computing marginal excess burden by di¤erentiating formula for

excess burden gives:
dEB dQ
∆τ = τ ∆τ
dτ dτ
First derivative of EB (τ ) only includes …rst-order term in Taylor
expansion:
dEB 1 d 2 EB
EB (τ + ∆τ ) = EB (τ ) + ∆τ + (∆τ )2
dτ 2 d τ2
First-order approximation is accurate when τ large relative to ∆τ
Ex: τ = 20%, ∆τ = 5% implies …rst term accounts for 90% of EB
But introduction of new tax (τ = 0) generates EB only through

second-order term
Method 3: Leakage in government revenue
To …rst order, marginal excess burden of raising τ is:

∂EB dQ
= τ
∂τ dτ
Observe that tax revenue R (τ ) = Qτ
Mechanical revenue gain: ∂∂τR jQ = Q
Actual revenue gain: ∂∂τR = Q + τ dQ

dτ
MEB is the di¤erence between mechanical and actual revenue gain:

∂R dR dQ dQ ∂EB
jQ =Q [Q + τ ]= τ =
∂τ dτ dτ dτ ∂τ
First vs. Second-Order Approximations
Why does leakage in govt. revenue only capture …rst-order term?
Govt revenue loss: rectangle in Harberger trapezoid, proportional to ∆τ
Consumer and producer surplus loss: triangles in trapezoid

(proportional to ∆τ 2 )
Method 3 is accurate for measuring marginal excess burden given

pre-existing taxes but not introduction of new taxes

Excess Burden of a Tax Increase: Harberger Trapezoid
P S+τ+∆τ S+τ
Lost cons. S
surplus (2nd order)
B
Lost govt. revenue
(1st order)
τ A
∆τ
C
D Lost producer
surplus (2nd order)
Q2 Q1 Q

General Model with Income E¤ects
Drop quasilinearity assumption and consider an individual with utility
u (c1 , .., cN ) = u (c )
Individual’s problem:
max u (c ) s.t. q c Z
c
where q = p + τ denotes vector of tax-inclusive prices and Z is wealth
Labor can be viewed as commodity with price w and consumed in

negative quantity

Demand Functions and Indirect Utility
Let λ denote multiplier on budget constraint
First order condition in ci :
uci = λqi
These conditions implicitly de…ne:
ci (q, Z ): the Marshallian (“uncompensated”) demand function
v (q, Z ): the indirect utility function

Measuring Deadweight Loss with Income E¤ects
Question: how much utility is lost because of tax beyond revenue

transferred to government?
Marshallian surplus does not answer this question with income e¤ects
Problem: not derived from utility function or a welfare measure
Creates various problems such as “path dependence” with taxes on

multiple goods
∆CS (τ 0 ! τ̃ ) + ∆CS (τ̃ ! τ 1 ) 6= ∆CS (τ 0 ! τ 1 )
Need units to measure “utility loss”
Introduce expenditure function to translate the utility loss into dollars

(money metric)

Expenditure Function
Fix utility at U and prices at q
Find bundle that minimizes cost to reach U for q:
e (q, U ) = min q c s.t. u (c ) U
c
Let µ denote multiplier on utility constraint
First order conditions given by:
qi = µuci
These generate Hicksian (or compensated) demand fns:
ci = hi (q, u )
De…ne individual’s loss from tax increase as
e (q 1 , u ) e (q 0 , u )
Single-valued function ! coherent measure of welfare cost, no path
dependence
Compensating and Equivalent Variation
But where should u be measured?
Consider a price change from q 0 to q 1
Utility at initial price q 0 :
u 0 = v (q 0 , Z )
Utility at new price q 1 :

u 1 = v (q 1 , Z )
Two concepts: compensating (CV ) and equivalent variation (EV ) use
u 0 and u 1 as reference utility levels

Compensating Variation
Measures utility at initial price level (u 0 )
Amount agent must be compensated in order to be indi¤erent about

tax increase
CV = e (q 1 , u 0 ) e (q 0 , u 0 ) = e (q 1 , u 0 ) Z
How much compensation is needed to reach original utility level at

new prices?
CV is amount of ex-post cost that must be covered by government to

yield same ex-ante utility:
e (q 0 , u 0 ) = e (q 1 , u 0 ) CV

Equivalent Variation
Measures utility at new price level
Lump sum amount agent willing to pay to avoid tax (at pre-tax prices)
EV = e (q 1 , u 1 ) e (q 0 , u 1 ) = Z e (q 0 , u 1 )
EV is amount extra that can be taken from agent to leave him with
same ex-post utility:
e (q 0 , u 1 ) + EV = e (q 1 , u 1 )

E¢ ciency Cost with Income E¤ects
Goal: derive empirically implementable formula analogous to

Marshallian EB formula in general model with income e¤ects
Literature typically assumes either

1 Fixed producer prices and income e¤ects
2 Endogenous producer prices and quasilinear utility
With both endogenous prices and income e¤ects, e¢ ciency cost

depends on how pro…ts are returned to consumers
Formulas are very messy and fragile (Auerbach 1985, Section 3.2)

E¢ ciency Cost Formulas with Income E¤ects
Derive empirically implementable formulas using Hicksian demand

(EV and CV )
Assume p is …xed ! ‡at supply, constant returns to scale
The envelope thm implies that eqi (q, u ) = hi , and so:

Z q1
e (q 1 , u ) e (q 0 , u ) = h (q, u )dq
q0
If only one price is changing, this is the area under the Hicksian
demand curve for that good
Note that optimization implies that
h (q, v (q, Z )) = c (q, Z )

Compensating vs. Equivalent Variation
h(V(p1,Z)) h(V(p0,Z))
p
p1
p0
x(p1,Z) x(p0,Z) x

h(V(p1,Z)) h(V(p0,Z))
p
p1
EV
p0
x(p1,Z) x(p0,Z) x

h(V(p1,Z)) h(V(p0,Z))
p
p1
CV
p0
x(p1,Z) x(p0,Z) x

Marshallian Surplus
h(V(p1,Z)) h(V(p0,Z))
p
p1
Marshallian Surplus
p0
x(p1,Z) x(p0,Z) x

EV, CV, and Marshallian Surplus
With one price change:
EV < Marshallian Surplus < CV
But this is not true in general with multiple price changes because
Marshallian Surplus is ill-de…ned

Excess Burden
Deadweight burden: change in consumer surplus less tax paid
What is lost in excess of taxes paid?
Two measures, corresponding to EV and CV :
EB (u 1 ) = EV (q 1 q 0 )h (q 1 , u 1 ) [Mohring 1971]
EB (u 0 ) = CV (q 1 q 0 )h (q 1 , u 0 ) [Diamond and McFadden 1974]

h(V(p1,Z)) h(V(p0,Z))
p ~
p1
EBEV EBCV
p0
x(p1,Z) xC(p1,V(p0,Z)) x
~ ~ ~

h(V(p1,Z)) h(V(p0,Z))
p ~
p1
Marshallian
p0
x(p1,Z) xC(p1,V(p0,Z)) x
~ ~ ~

Excess Burden
In general, CV and EV measures of EB will di¤er
Marshallian measure overstates excess burden because it includes

income e¤ects
Income e¤ects are not a distortion in transactions
Buying less of a good due to having less income is not an e¢ ciency

loss; no surplus foregone b/c of transactions that do not occur
CV = EV = Marshallian DWL only with quasilinear utility (Chipman

and Moore 1980)

Implementable Excess Burden Formula
Consider increase in tax τ on good 1 to τ + ∆τ
No other taxes in the system
Recall the expression for EB:
EB (τ ) = [e (p + τ, U ) e (p, U )] τh1 (p + τ, U )
Second-order Taylor expansion:
MEB = EB (τ + ∆τ ) EB (τ )
dEB 1 d 2 EB
' ∆τ + (∆τ )2
dτ 2 d τ2

Harberger Trapezoid Formula
dEB dh1
= h1 (p + τ, U ) τ h1 (p + τ, U )
dτ dτ
dh1
= τ
dτ
d 2 EB dh1 d 2 h1
= τ
d τ2 dτ d τ2
2
Standard practice in literature: assume dd τh21 = 0 (linear Hicksian); not
necessarily well justi…ed b/c it does not vanish as ∆τ ! 0
dh1 1 dh1
) MEB = τ∆τ (∆τ )2
dτ 2 dτ
Formula equals area of “Harberger trapezoid” using Hicksian demands

Harberger Formula
Without pre-existing tax, obtain “standard” Harberger formula:

1 dh1
EB = (∆τ )2
2 dτ
General lesson: use compensated (substitution) elasticities to
compute EB, not uncompensated elasticities
To implement empirically, estimate Marshallian price elasticity and

income elasticity. Then apply Slutsky eqn:
∂hi ∂ci ∂ci

= + cj
∂qj ∂qj ∂Z
|{z} |{z} | {z }
Hicksian Slope Marshallian Slope Income E¤ect

Excess Burden with Taxes on Multiple Goods
Previous formulas apply to case with tax on one good
With multiple goods and …xed prices, excess burden of introducing a

tax τ k
1 2 dhk dhi
EB = τ ∑ τi τk
2 k d τ k i 6 =k d τk
Second-order e¤ect in own market, …rst-order e¤ect from other
markets with pre-existing taxes
Complementarity between goods important for excess burden

calculations
Ex: with an income tax, minimize total DWL tax by taxing goods
complementary to leisure (Corlett and Hague 1953)

Goulder and Williams 2003
Show that ignoring cross e¤ects by using one-good formula can be

very misleading
Di¤erentiate multiple-good Harberger formula w.r.t. τ k :

dEB dhk dhi
= τk ∑ τi
d τk d τk i 6 =k d τk
If τ k is small (e.g. gas tax), what matters is purely distortion in other

markets, e.g. labor supply
As τ k ! 0, error in single-market formula approaches ∞

Goulder and Williams 2003
Basic formula hard to implement because it requires estimates of all

cross-price elasticities
Goulder and Williams make formula empirically implementable by

making 3 assumptions:
1 No income e¤ects
2 Ignore interactions with commodities other than labor (other taxes are
small)
3 Assume good is of “average” substitutability with labor: cross partial

∂l
∂τ equals mean cross-partial across consumption goods
k

Goulder and Williams Formula
Obtain following formula for marginal excess burden of raising tax on
good k:
dEB τ Q τL L
= k k ηk η sk
d τk pk pk L
τ k , pk , and Qk are the tax, price, and quantity consumed of good k
η k and η L are own-price elasticity of good k and labor
sk = wlP(k1Q Kτ ) is budget share of good k
L
Only need estimates of own-price elasticities to implement this

formula
Why? Consumption tax and labor income tax have equivalent e¤ects
Price increase in all consumption goods has the same e¤ect on labor
supply as an increase in tax on labor:
(1 + t ) ∑ pk ck = wl
k

Goulder and Williams Formula: Intuition
Why do we only need to estimate η L ?
Step 1: consumption tax and labor income tax have equivalent e¤ects
Price increase in all consumption goods has the same e¤ect on labor
supply as an increase in tax on labor:
(1 + t ) ∑ pk ck = wl
k

Goulder and Williams Formula: Intuition
Step 2: Rank goods according to complementarity with labor (i.e.

cross-partial ddlτk )
dl
Find good at the mean level of d τk
A tax increase on this good has same e¤ect as an increase in sales tax
t on all consumption goods scaled down by sk
Therefore cross-elasticity of l w.r.t. τ k is equivalent to η L sk
Labor supply elasticity η L su¢ cient to calculate cross-elasticity for

good that has “average” level of substitutability

Goulder and Williams Results
Calibrate formula using existing elasticity estimates
Result: DWL of taxing goods such as gasoline is underestimated by a

factor of 10 in practice because of income tax
Caveat: is their approach and conclusion valid if there are salience

e¤ects?

Hausman 1981: Exact Consumer Surplus
Harberger formulas: empirically implementable but approximate
Alternative approach: structural estimation of demand model
Start by estimating Marshallian demand functions:
c (q, Z ) = γ + αq + δZ
Then integrate to recover underlying indirect utility function v (q, Z )
Invert to obtain expenditure function e (q, u ) and compute “exact” EB
Parametric approach: Hausman (AER 1981); non-parametric

approach: Hausman and Newey (ECMA 1995)

Harberger vs. Hausman Approach
Underscores broader di¤erence between structural and

quasi-experimental methodologies
Modern literature focuses on deriving “su¢ cient statistic” formulas

that can be implemented using quasi-experimental techniques
Now develop general distinction between structural and su¢ cient

statistic approaches to welfare analysis in a simple model of taxation
No income e¤ects (quasilinear utility)
Constant returns to production (…xed producer prices)
But permit multiple goods (GE)

Su¢ cient Statistics vs Structural Methods
N goods: x = (x1 , ..., xN ); prices (p1 , ...pN ); wealth Z
Normalize pN = 1 (xN is numeraire)
Government levies a tax t on good 1
Individual takes t as given and solves

N
max u (x1 , ..., xN 1 ) + xN s.t. (p1 + t )x1 + ∑ pi xi = Z
i =2
To measure EB of tax, de…ne social welfare as sum of individual’s

utility and tax revenue:
N 1
W (t ) = fmax u (x1 , ..., xN
x
1) + Z (p1 + t )x1 ∑ pi xi g + tx1
i =2
dW
Goal: measure dt = loss in social surplus caused by tax change
Primitives Sufficient Stats. Welfare Change
ω1
ω2
. K 1 ÝtÞ
.
dW
ÝtÞ
K 2 ÝtÞ dt
.
ωΝ
ω=preferences, β = f(ω,t) dW/dt used for

constraints y = β1X1 + β2X2 + ε policy analysis
ω not uniquely β identified using

identified program evaluation
Source: Chetty (2009)

Su¢ cient Statistics vs Structural Methods
Structural method: estimate N good demand system, recover u
Ex: use Stone-Geary or AIDS to recover preference parameters; then

calculate “exact consumer surplus” as in Hausman (1981)
Alternative: Harberger’s deadweight loss triangle formula
Private sector choices made to maximize term in red (private surplus)

N 1
W (t ) = fmax u (x1 , ..., xN 1 ) + Z
x
( p 1 + t ) x1 ∑ pi xi g + tx1
i =2
Envelope conditions for (x1 , ..., xN ) allow us to ignore behavioral

responses ( dx i
dt ) in term in red, yielding
dW dx1 dx
= x1 + x1 + t =t 1
dt dt dt
dx1
! dt is a “su¢ cient statistic” for calculating dW
dt
Heterogeneity
Bene…t of su¤ stat approach particularly evident with heterogeneity
K agents, each with utility uk (x1 , ..., xN 1 ) + xN
Social welfare function under utilitarian criterion:

K
W (t ) = fmax
x
∑ [uk (x1k , ..., xNk 1) + Z
k =1
N 1 K
(p1 + t )x1k ∑ pi xik ]g + ∑ tx1k
i =2 k =1
Structural method: estimate demand systems for all agents
Su¢ cient statistic formula is unchanged— still need only slope of
aggregate demand dx dt
1
dW K K
d ∑K k
k =1 x1 dx1
dt
= ∑ x1k + ∑ x1k + t
dt
=t
dt
k =1 k =1

Discrete Choice Model
Harberger su¢ cient statistic also works with discrete choice
Agents have value Vk for good 1; can either buy or not buy
Let F (V ) denote distribution of valuations
With 2 goods, utility of agent k is

Vk x1 + Z (p + t )x1
Social welfare:
Z
W (t ) = f max[Vk x1k + Z (p1 + t )x1k ]dF (Vk )g
Vk x1k
Z
+ tx1k dF (Vk )
Vk
This problem is not smooth at individual level, so cannot directly

apply envelope thm. as stated
Discrete Choice Model
Recast as planner’s problem choosing threshold above which agents

are allocated good 1:
( Z
)
∞
W (t ) = max
_ _ [ Vk (p1 + t )] dF (Vk ) + Z
V V
Z ∞
+t _ dF (Vk )
V
Again obtain Harberger formula as a fn of slope of aggregate demand

curve dx
dt :
1
R _∞
dW _ _ d V
dF (Vk )
= 1 F V + 1 F V +t
dt dt
dW dx1
) = t
dt dt

Economic Intuition for Robustness of Harberger Result
Deadweight loss is fully determined by di¤erence between marginal

willingness to pay for good x1 and its cost (p1 )
Recovering marginal willingness to pay requires an estimate of the

slope of the demand curve because it coincides with marginal utility:
p = u 0 (x (p ))
Slope of demand is therefore su¢ cient to infer e¢ ciency cost of a tax,

without identifying rest of the model

E¢ ciency Cost: Applications
1 [Income Taxation] Feldstein; Chetty; Gorodnichenko et al.
2 [Housing Subsidy] Poterba
3 [Diesel Fuel Taxation] Marion and Muehlegger

Feldstein 1995, 1999
Following Harberger, large literature in labor estimated e¤ect of taxes

on hours worked to assess e¢ ciency costs of taxation
Feldstein observed that labor supply involves multiple dimensions, not

just choice of hours: training, e¤ort, occupation
Taxes also induce ine¢ cient avoidance/evasion behavior
Structural approach: account for each of the potential responses to

taxation separately and then aggregate
Feldstein’s alternative: elasticity of taxable income with respect to

taxes is a su¢ cient statistic for calculating deadweight loss

Feldstein Model: Setup
Government levies linear tax t on reported taxable income
Agent makes N labor supply choices: l1 , ...lN
Each choice li has disutility ψi (li ) and wage wi
Agents can shelter $e of income from taxation by paying cost g (e )
Taxable Income (TI ) is

N
TI = ∑ wi li e
i =1
Consumption is given by taxed income plus untaxed income:
c = (1 t )TI + e

Feldstein Taxable Income Formula
Agent’s utility is quasi-linear in consumption:

N
u (c, e, l ) = c g (e ) ∑ ψi (li )
i =1
Social welfare:
N
W (t ) = f(1 t )TI + e g (e ) ∑ ψi (li )g + tTI
i =1
Di¤erentiating and applying envelope conditions for li

((1 t )wi = ψi0 (li )) and e (g 0 (e ) = t) implies
dW dTI dTI
= TI + TI + t =t
dt dt dt
Intuition: marginal social cost of reducing earnings through each
margin is equated at optimum ! irrelevant what causes change in TI
Taxable Income Formula
Simplicity of identi…cation in Feldstein’s formula has led to a large

literature estimating elasticity of taxable income
But since primitives are not estimated, assumptions of model used to

derive formula are never tested
Chetty (2009) questions validity of assumption that g 0 (e ) = t
Costs of some avoidance/evasion behaviors are transfers to other

agents in the economy, not real resource costs
Ex: cost of evasion is potential …ne imposed by government

Chetty Transfer Cost Model: Setup
Individual chooses e (evasion/shifting) and l (labor supply) to
max u (c, l, e ) = c ψ (l )
e,l
s.t. c = y + (1 t )(wl e) + e z (e )
Social welfare is now:
W (t ) = fy + (1 t )(wl e) + e
z (e ) ψ(l )g
+z (e ) + t (wl e)
Di¤erence: z (e ) now appears twice in SWF, with opposite signs

Excess Burden with Transfer Costs
Let LI = wl be the total (pretax) earned income and TI = wl e
denote taxable income
Exploit the envelope condition for term in curly brackets:

dW dz de d [wl e ]
= (wl e ) + (wl e) + +t
dt de dt dt
dTI dz de
= t+
dt de dt
dLI de dz de
= t t +
dt dt de dt
First-order condition for individual’s choice of e:
dz
t =
de
dW dLI
) =t (1)
dt dt
Intuition: MPB of raising e by $1 (saving $t) equals MPC
Chetty (2009) Formula
With both transfer cost z (e ) and resource cost g (e ) of evasion:
dW dLI de
= t g 0 (e )
dt dt dt
dTI dLI
= t fµ + (1 µ ) g
dt dt
t
= fµTI εTI + (1 µ)wl εLI g
1 t
EB depends on weighted average of taxable income (εTI ) and total

earned income elasticities (εLI )
Practical importance: even though reported taxable income is highly

sensitive to tax rates for rich, e¢ ciency cost may not be large!
Most di¢ cult parameter to identify: weight µ, which depends on

marginal resource cost of sheltering, g 0 (e )
Gorodnichenko, Martinez-Vazquez, and Peter 2009
Estimate εLI and εTI to implement formula that permits transfer costs
Insight: consumption data can be used to infer εLI
Estimate e¤ect of 2001 ‡at tax reform in Russia on gap between

taxable income and consumption, which they interpret as evasion

Gorodnichenko et al: Results
dTI
Taxable income elasticity dt is large, whereas labor income elasticity
dLI
dt is not
! Feldstein’s formula overestimates the e¢ ciency costs of taxation

relative to more general measure for “plausible” g 0 (e )
Question: could g 0 (e ) be estimated from consumption data itself?

Poterba 1992
Estimates e¢ ciency cost of subsidy for housing in the U.S. from

mortgage interest deduction
First need to de…ne “cost” of owning $1 of housing
De…nition: “user cost” – measures opportunity cost of living in home
Could rent the house to someone else at percentage rate

Rent
r=
Property Value
With marginal income tax rate τ and nominal interest i, net user cost
taking into account mortgage deduction is
c=r τ i

Poterba 1992
Poterba …rst calculates changes in user cost over 1980s
Tax reform in 1986 lowered tax rates for high income and raised user
cost of housing sharply
Prior to 1986: very high tax rates on high incomes (60%)
In 1990, only 28%
Nearly tripled the cost of housing

Poterba 1992
Calculates compensated elasticity using estimates in literature and

Slutsky eqn.
Rosen (1982): εH ,r = 1
Income elasticity: 0.75
Housing share: 0.25
3 1
) Compensated elasticity: 1+ ' 0.8
4 4
Intuition for large elasticity: broker calculates “how much house you
can a¤ord” if they spend 30% of income
Can “a¤ord” more with larger tax subsidy ! tax is e¤ectively salient
Calculates amount of overconsumption of housing and e¢ ciency cost

of housing subsidy
Poterba: Results
Tax reforms in 1980s reduced DWL from $12K to $2K for each
household earning $250K
Still have relatively large ine¢ ciency from subsidizing mortgages
This is why President Bush’s Tax Panel recommended cap or

elimination of subsidy for homeownership
But hard to implement politically

Marion and Muehlegger 2008
Study deadweight cost from taxing diesel fuels, focusing on evasion
Diesel fuel used for business purposes (e.g. trucking) is taxed, but
residential purposes (e.g. heating homes) is not
Substantial opportunity to evade tax
1993: government added red dye to residential diesel fuel
Easy to monitor cheating by opening gas tank of a truck
First document e¤ect of dye reform on evasion

Marion and Muehlegger: Excess Burden Calculations
Use reform to assess deadweight costs of evasion and taxation
Harder to evade ! elasticity of behavior with respect to tax is much

lower after reform
Estimate price and tax elasticities before and after reform
Use cross-state variation in tax rates and price variation from world
market
Note di¤erent interpretation of di¤erence between price and tax

elasticities in this study relative to tax salience papers

Price and Tax Elasticities By Year

Marion and Muehlegger: Results
Elasticities imply that 1% increase in tax rate raised revenue by

0.60% before dye reform vs. 0.71% after reform
Reform reduced deadweight cost of diesel taxation
MDWL = 40 cents per dollar of revenue raised before dye reform
MDWL = 30 cents per dollar after reform
Lesson: Deadweight cost depends not just on preferences but also on

enforcement technology
But again need to think carefully about marginal costs of evasion in

this context: social or transfer?

Welfare Analysis in Behavioral Models
Formulas derived thus far rely critically on full optimization by agents

in private sector
How to calculate e¢ ciency costs when agents do not optimize

perfectly?
Relates to broader …eld of behavioral welfare economics
Focus on two papers here:
1 Conceptual Issues: Bernheim and Rangel 2009
2 Applied Welfare Analysis: Chetty, Looney, Kroft 2009

Behavioral Welfare Economics
Abstractly, e¤ect of policies on welfare are calculated in two steps
1 E¤ect of policy on behavior
2 E¤ect of change in behavior on utility
Challenge: identifying (2) when agents do not optimize perfectly
How to measure objective function without tools of revealed

preference?
Danger of paternalism

Behavioral Welfare Economics: Two Approaches
Approach #1: Build a positive model of deviations from rationality
Ex: hyperbolic discounting, bounded rationality, reference dependence
Then calculate optimal policy within such models
Approach #2: Choice-theoretic welfare analysis (Bernheim and

Rangel 2009)
Do not specify a positive model to rationalize behavior
Instead map directly from observed choices to statements about welfare
Analogous to “su¢ cient statistic” approach

Consider three di¤erent medicare plans with di¤erent copays: L, M, H

and corresponding variation in premiums
We have data from two environments:
1 On red paper, H > M > L
2 On blue paper, M > H > L

Approach 1: build a model of why color a¤ects choice and use it to

predict which choice reveals “true” experienced utility
Approach 2: Yields bounds on optimal policy
L cannot be optimal given available data irrespective of positive model
Optimal copay bounded between M and H
Key insight: no theory of choice needed to make statements about

welfare
Do not need to understand why color a¤ects choice

Bernheim and Rangel 2009: Setup
Derive bounds on welfare based purely on choice data
In standard model, agents choose from a choice set x 2 X
Goal of policy is to identify optimal x
In behavioral models, agents choose from “generalized choice sets”

G = (X , d )
d is an “ancillary condition” – something that a¤ects choice behavior

but (by assumption) does not a¤ect experienced utility
Ex: color of paper, salience, framing, default option

Bernheim and Rangel 2009: Choice Sets
Let C (X , d ) denote choice made in a given GCS
Choice inconsistency if C (X , d ) 6= C (X , d 0 )
De…ne revealed preference relation P as xPy if x always chosen over

y for any d
Using P, can identify choice set that maximizes welfare instead of

single point
With continuous choices, e¤ectively obtain bounds on welfare

Bernheim and Rangel 2009: Compensating Variation
Consider a change in choice set from X to X 0 X
Compute CV as amount needed to make agent indi¤erent to restriction

of choice set for each d (standard calculation)
Lower bound on CV is minimum over all d’s
Upper bound on CV is maximum over all d’s

Bernheim and Rangel 2009: Compensating Variation
Ex: suppose insurance plans are restricted to drop M option
Under red paper condition, CV is 0 – no loss in welfare
Under blue paper condition, calculate price cut $z on H needed to

make agent indi¤erent between M and H.
Bounds on CV: (0, z )
If L option is dropped, bounds collapse to a singleton: CV = 0.

Bernheim and Rangel 2009: Re…nements
Problem: looseness of bounds
Bounds tight when ancillary conditions do not lead to vast changes in

choices
That is, bounds tight when behavioral problems are small
In cases where behavioral issues are important, this is not going to be

a very informative approach

Bernheim and Rangel 2009: Re…nements
Solution: “re…nements” – discard certain d’s as being

“contaminated” for welfare analysis
E.g. a neuroscience experiment shows that decisions made under red

paper condition are more rational
Or assume that choice rational when incentives are more salient
With fewer d’s, get tighter bounds on welfare and policy
Identifying “re…nements” typically requires some insight into positive

theory of behavior

Applied Welfare Analysis with Salience E¤ects
Chetty, Looney, and Kroft (2009) section 5
Derive partial-equilibrium formulas for incidence and e¢ ciency costs
Focus here on e¢ ciency cost analysis
Formulas do not rely on a speci…c positive theory, in the spirit of

Bernheim and Rangel (2009)

Welfare Analysis with Salience E¤ects: Setup
Two goods, x and y ; price of y is 1, pretax price of x is p.
Taxes: y untaxed. Unit sales tax on x at rate t S , which is not

included in the posted price
Tax-inclusive price of x: q = p + t S

Representative consumer has wealth Z and utility u (x ) + v (y )
Letfx (p, t S , Z ), y (p, t S , Z )g denote bundle chosen by a

fully-optimizing agent
Let fx (p, t S , Z ), y (p, t S , Z )g denote empirically observed demands
Place no structure on these demand functions except for feasibility:
(p + t S )x (p, t S , Z ) + y (p, t S , Z ) = Z

Price-taking …rms use y to produce x with cost fn. c
Firms optimize perfectly. Supply function S (p ) de…ned by:
p = c 0 (S (p ))
∂S p
Let εS = ∂p S (p )
denote the price elasticity of supply

E¢ ciency Cost with Salience E¤ects
De…ne excess burden using EV concept
Excess burden (EB) of introducing a revenue-generating sales tax t is:
EB (t S ) = Z e (p, 0, V (p, t S , Z )) R (p, t S , Z )

Preference Recovery Assumptions
A1 Taxes a¤ect utility only through the chosen consumption bundle.
Agent’s indirect utility given tax of t S is
V (p, t S , Z ) = u (x (p, t S , Z )) + v (y (p, t S , Z ))
A2 When tax inclusive prices are fully salient, the agent chooses the same
allocation as a fully-optimizing agent:
x (p, 0, Z ) = x (p, 0, Z ) = arg max u (x ) + v (Z px )
x
A1 speci…es ancillary condition: tax rate and salience does not enter
utility directly
A2 is a re…nement: behavior when tax is salient reveals true

preferences
E¢ ciency Cost with Salience E¤ects
Two demand curves: price-demand x (p, 0, Z ) and tax-demand

x ( p0 , t S , Z )
Two steps in e¢ ciency calculation:
1 Use price-demand x (p, 0, Z ) to recover utility as in standard model
2 Use tax-demand x (p, t S , Z )to calculate V (p, t S , Z ) and EB

∂x
Excess Burden with No Income E¤ect for Good x ( ∂Z = 0)
p, t S
C x( p,0) = u ' ( x)
xÝp 0 ,t S Þ
p0 + t S G D E
F /x//tS
/x//t S
EB p ? 12 Ýt S Þ 2 /x//p
/x//t S
tS /x//p
p0 A
B I H tS /x
/t S
x
x1* x1 x0

E¢ ciency Cost: No Income E¤ects
∂x
Without income e¤ects ( ∂Z = 0), excess burden of introducing a
S
small tax t is
1 S 2 ∂x /∂t S
EB (t S ) ' (t ) ∂x /∂t S
2 ∂x /∂p
1 S 2 εD
= (θt )
2 p + tS
Inattention reduces excess burden when dx /dZ = 0.
Intuition: tax t S induces behavioral response equivalent to a fully

perceived tax of θt S .
If θ = 0, tax is equivalent to a lump sum tax and EB = 0 because

agent continues to choose …rst-best allocation.

Same formula, but all elasticities are now compensated:
1 S 2 ∂x c /∂t S c
EB (t S ) ' (t ) ∂x /∂t S
2 ∂x c /∂p
1 c S 2 εcD
= (θ t )
2 p + tS
Compensated price demand: dx c /dp = dx /dp + xdx /dZ
Compensated tax demand: dx c /dt S = dx /dt S + xdx /dZ
Compensated tax demand does not necessarily satisfy Slutsky

condition dx c /dt S < 0 b/c it is not generated by utility maximization

1 S 2 ∂x c /∂t S c
EB (t S ) ' (t ) ∂x /∂t S
2 ∂x c /∂p
1 c S 2 εcD
= (θ t )
2 p + tS
With income e¤ects (dx /dZ > 0), making a tax less salient can raise
deadweight loss.
Tax can generate EB > 0 even if dx /dt S = 0
Example: consumption of food and cars; agent who ignores tax on

cars underconsumes food and has lower welfare.
Intuition: agent does not adjust consumption of x despite change in

net-of-tax income, leading to a positive compensated elasticity.
Directions for Further Work on Behavioral Welfare Analysis
1 Normative analysis of tax policy
Value of tax simpli…cation
Tax smoothing
2 Use similar approach to welfare analysis in other contexts
Design consumer protection laws and …nancial regulation in a less

paternalistic manner by studying behavior in domains where incentives
are clear

Part 4: Optimal Taxation
Harvard University
Fall 2012
Public Economics Lectures () Part 4: Optimal Taxation 1 / 80

Outline
1 Commodity Taxation: Ramsey Rule
2 Capital Income Taxation and Retirement Savings
3 Income Taxation: Mirrlees Model
4 Optimal Transfer Programs

Optimal Commodity Taxation: Introduction
Combine lessons on incidence and e¢ ciency costs to analyze optimal

design of commodity taxes
What is the best way to design taxes given equity and e¢ ciency
concerns?

Overview of Optimal Taxation
From an e¢ ciency perspective, would …nance government purely

through lump-sum taxation
With redistributional concerns, would ideally levy individual-speci…c

lump sum taxes
Tax higher-ability individuals a larger lump sum
Problem: cannot observe individuals’types
Therefore must tax economic outcomes such as income or

consumption, which leads to distortions

Ramsey vs. Mirrleesian Approaches
Two approaches to optimal taxation
1 Ramsey: restrict attention linear (t x) tax systems
2 Mirrleesian: Non-linear (t (x )) tax systems, with no restrictions on t (x )
Ramsey approach: rule out possibility of lump sum taxes by

assumption and consider linear taxes
Mirrleesian approach: permit lump sum taxes, but model their costs
in a model with heterogeneity in agents’skills

Four Central Results in Optimal Tax Theory
1 Ramsey (1927): inverse elasticity rule
2 Chamley (1985), Judd (1986): no capital taxation in in…nite horizon

Ramsey models
3 Diamond and Mirrlees (1971): production e¢ ciency
4 Atkinson and Stiglitz (1976): no consumption taxation with optimal

non-linear income taxation

Ramsey (1927) Tax Problem
Government sets taxes on uses of income in order to accomplish two

objectives:
1 Raise total revenue of amount E
2 Minimize utility loss for agents in economy

Ramsey Model: Key Assumptions
1 Lump sum taxation prohibited
2 Cannot tax all commodities (e.g., leisure untaxed)
3 Production prices …xed (and normalized to one):
pi = 1
) qi = 1 + τi

Ramsey Model: Setup
One individual (no redistributive concerns)
As in e¢ ciency analysis, assume that individual does not internalize

e¤ect of τ i on govt. budget
Captures idea that any one individual accounts for a small frac. of
economy
Individual maximizes utility
u (x1 , .., xN , l )
subject to budget constraint
q1 x1 + .. + qN xN wl + Z
Z = non wage income, w = wage rate

Ramsey Model: Consumer Behavior
Lagrangian for individual’s maximization problem:
L = u (x1 , .., xN , l ) + α(wl + Z (q1 x1 + .. + qN xN ))
First order condition:

uxi = αqi
Where α = ∂V /∂Z is marginal value of money for the individual
Yields demand functions xi (q, Z ) and indirect utility function V (q, Z )

where q = (w , q1 , .., qN )

Ramsey Model: Government’s Problem
Government solves either the maximization problem
max V (q, Z )
subject to the revenue requirement

N
τ x= ∑ τi xi (q, Z ) E
i =1
Or, equivalently, minimize excess burden of the tax system
min EB (q ) = e (q, V (q, Z )) e (p, V (q, Z )) E
subject to the same revenue requirement

Ramsey Model: Government’s Problem
For maximization problem, Lagrangian for government is:
LG = V (q, Z ) + λ[∑ τ i xi (q, Z ) E]

i
∂ LG ∂V
) = + λ[ xi + ∑ τ j ∂xj /∂qi ] = 0
∂qi ∂qi |{z}
|{z} |
j
{z }
Mechanical
Priv. Welfare Behavioral
E¤ect
Loss to Indiv. Response
∂V
Using Roy’s identity ( ∂q i
= αxi ):
(λ α)xi + λ ∑ τ j ∂xj /∂qi = 0

j
Note connection to marginal excess burden formula, where λ = 1 and

α=1
Ramsey Optimal Tax Formula
Optimal tax rates satisfy system of N equations and N unknowns:
∂xj xi
∑ τj ∂qi =
λ
(λ α)
j
Same formula can be derived using a perturbation argument, which is

more intuitive

Ramsey Formula: Perturbation Argument
Suppose government increases τ i by d τ i
E¤ect of tax increase on social welfare is sum of e¤ect on government

revenue and private surplus
Marginal e¤ect on government revenue:

dR = xi d τ i + ∑ τ j dxj
j
Marginal e¤ect on private surplus:

∂V
dU = d τi
∂qi
= αxi d τ i
Optimum characterized by balancing the two marginal e¤ects:
dU + λdR = 0
Ramsey Formula: Compensated Elasticity Representation
Rewrite in terms of Hicksian elasticities to obtain further intuition

using Slutsky equation:
∂xj /∂qi = ∂hj /∂qi xi ∂xj /∂Z
Substitution into formula above yields:
(λ α)xi + λ ∑ τ j [∂hj /∂qi xi ∂xj /∂Z ] = 0

j
1 ∂hi θ
)
xi ∑ τj ∂qj =
λ
j
where θ = λ α ∂
λ ∂Z (∑j τ j xj )

Ramsey Formula: Compensated Elasticity Representation
θ is independent of i and measures the value for the government of

introducing a $1 lump sum tax
θ=λ α λ∂(∑ τ j xj )/∂Z

j
Three e¤ects of introducing a $1 lump sum tax:
1 Direct value for the government of λ
2 Loss in welfare for individual of α
3 Behavioral e¤ect ! loss in tax revenue of ∂(∑j τ j xj )/∂Z

Intuition for Ramsey Formula: Index of Discouragement
1 ∂hi θ
xi ∑ τj ∂qj =
λ
j
Suppose revenue requirement E is small so that all taxes are also small
Then tax τ j on good j reduces consumption of good i (holding utility

constant) by approximately
∂hi
dhi = τ j
∂qj
Numerator of LHS: total reduction in consumption of good i
Dividing by xi yields % reduction in consumption of each good i =

“index of discouragement” of the tax system on good i
Ramsey tax formula says that the indexes of discouragements must be

equal across goods at the optimum
Inverse Elasticity Rule
Introducing elasticities, we can write Ramsey formula as:

N
τj θ
∑ 1 + τj εcij = λ
j =1
Consider special case where εij = 0 if i 6= j
Slutsky matrix is diagonal
Obtain classic inverse elasticity rule:
τi θ 1
=
1 + τi λ εii

Ramsey Formula: Limitations
Ramsey solution: tax inelastic goods to minimize e¢ ciency costs
But does not take into account redistributive motives
Necessities likely to be less elastic than luxuries
Therefore, optimal Ramsey tax system is likely regressive
Diamond (1975) extends Ramsey model to take redistributive motives

into account
Basic intuition: replace multiplier λ with average marginal utility for

consumers of that good

Application of Ramsey Approach to Taxation of Savings
Standard lifecycle model of consumption
max ∑ ut (ct ) s.t. qt ct W

t
where qt = (1 + τ t )pt
and τ 0 0
Consumption in each period isomorphic to consumption of di¤erent

goods
Can apply standard Ramsey formulas to calculate τ t
Capital income tax is a constant tax θ on interest rate:

1
qt =
(1 + (1 θ )r )t

Optimal Capital Income Tax Rate
For any θ > 0, implied tax τ t approaches ∞ as t ! ∞:

qt 1+r
= 1 + τt = ( )t
pt 1 + (1 θ )r
) lim τ t = ∞
t !∞
Ramsey formula implies that optimal τ t cannot be ∞ for any good
Therefore optimal capital income tax rate converges to 0 in long run

(Judd 1985, Chamley 1986)
Best policy is for govt to tax capital until it accumulates su¢ cient
assets to fund public goods and never tax capital again

Zero Capital Taxation in Ramsey Models
Fairly robust result in pure Ramsey framework (Bernheim 2002)
But not robust to:
Allowing for progressive income taxation (Golosov, Kocherlakota,

Tsyvinski 03)
Allowing for credit market imperfections (Aiyagari 95, Farhi and

Werning 11)
Finitely-lived agents with …nite bequest elasts. (Piketty and Saez 12)
More general issue: are agents this forward-looking when making

savings choices?
Return to this in the context of corrective taxes later in the course

Optimal Income Taxation: Outline
1 Optimal Static Income Taxation: Mirrlees (1971)
2 Empirical Implementation of Mirrlees Model: Saez (2001)
3 Income and Commodity Taxation: Atkinson and Stiglitz (1976)

[Spring Semester]
4 Optimal Transfer Programs: Saez (2002)

Key Concepts for Taxes/Transfers
Let T (z ) denote tax liability as a function of earnings z
1 Transfer bene…t with zero earnings T (0) [also called demogrant or

lumpsum grant]
2 Marginal tax rate T 0 (z ): individual keeps 1 T 0 (z ) for an additional

$1 of earnings (relevant for intensive margin labor supply responses)
3 Participation tax rate τ p = [T (z ) T (0)]/z: individual keeps

fraction 1 τ p of earnings when moving from zero earnings to
earnings z:
z T (z ) = T (0) + z [T (z ) T (0)] = T (0) + z (1 τp )
Relevant for extensive margin labor supply responses
4 Break-even earnings point z : point at which T (z ) = 0

US Tax/Transfer System, single parent with 2 children, 2009
$50,000 $50,000
$40,000 $40,000 Welfare:

TANF+SNAP
Tax credits:
Disposable Earnings
$30,000 $30,000
EITC+CTC
Earnings after
$20,000 $20,000 taxes
45 Degree Line
$10,000 $10,000
$0 $0
$0
$5,383
$10,765
$16,148
$21,530
$26,913
$32,295
$37,678
$43,060
$48,443
Gross Earnings (with employer payroll taxes)
Source: Saez 2010 AEA Clark Lecture

Optimal Income Tax with No Behavioral Responses
Utility u (c ) strictly increasing and concave
Same for everybody where c is after tax income
Income is z and is …xed for each individual, c = z T (z ) where

T (z ) is tax on z
Government maximizes Utilitarian objective:

Z ∞
u (z T (z ))h (z )dz
0
R
Subject to budget constraint T (z )h (z )dz E (multiplier λ)

Optimal Income Tax without Behavioral Responses
Lagrangian for this problem is:
L = [u (z T (z )) + λT (z )]h (z )
T (z ) : 0 = ∂L/∂T (z ) = [ u 0 (z T (z )) + λ]h (z )
0
) u (z T (z )) = λ
) z T (z ) = c constant for all z
) c = z̄ E
R
where z̄ = zh (z )dz average income
100% marginal tax rate; perfect equalization of after-tax income
Utilitarianism with diminishing marginal utility leads to egalitarianism

Mirrlees 1971: Incorporating Behavioral Responses
Standard labor supply model: Individual maximizes
u (c, l ) s.t. c = wl T (wl )
where c is consumption, l labor supply, w wage rate, T (.) income tax
Individuals di¤er in ability w distributed with density f (w )
Govt social welfare maximization: Govt maximizes

Z
SWF = G (u (c, l ))f (w )dw )
Z
s.t. resource constraint T (wl )f (w )dw E
and individual FOC w (1 T 0 ) uc + ul = 0
where G (.) is increasing and concave

Social Welfare Function
Mirrleesian approach: maximize weighted sum of utilities of ex-post

consumption
With equal weights G and diminishing marginal utility, we would

equate everyone’s income barring information constraints
But is this what people really want?
Ex: many would argue that it is ok for an entrepreneur to keep the

money he earned if he worked hard to get it
Is maximizing total ex-post utility the right objective function?

Social Welfare Function
Questions on appropriate social welfare function date to Rawls,

Nozick, Sen, and others
Notions of “equality of opportunity” and “just deserts” rather than

pure consequentalist perspective (Mankiw 2010, Weinzierl 2012)
But no widely applied tractable framework besides Mirrleesian

approach
One recent approach: Saez and Stantcheva’s (2012) “endogenous”

welfare weights G (c, T (wl ))
Fertile area to make a fundamental contribution

Mirrlees 1971: Results
Optimal income tax trades-o¤ redistribution and e¢ ciency
T (.) < 0 at bottom (transfer)
T (.) > 0 further up (tax) [full integration of taxes/transfers]
Mirrlees formulas are a complex fn. of primitives, with only a few

general results
1 0 T 0 (.) 1. T 0 (.) 0 is non-trivial and rules out EITC [Seade

1976]
2 Marginal tax rate T 0 (.) should be zero at the top if skill distribution
bounded [Sadka-Seade]

Mirrlees: Subsequent Work
Mirrlees model had a profound impact on information economics
Ex. models with asymmetric information in contract theory
But until late 1990s, had little impact on practical tax policy
Recently, Mirrlees model connected to empirical literature
Diamond (1998), Piketty (1997), and Saez (2001)
Su¢ cient statistic formulas in terms of labor supply elasticities instead

of primitives

Optimal Income Taxation: Su¢ cient Statistic Formulas
1 Revenue-maximizing linear tax (La¤er curve)
2 Top income tax rate (Saez 2001)
3 Full income tax schedule (Saez 2001)
See also section 4 of Chetty (Ann. Rev. 2009)

Revenue-Maximizing Tax Rate: La¤er Curve
With a constant tax rate τ, reported income z depends on 1 τ
(net-of-tax rate)
Tax Revenue R (τ ) = τ z (1 τ ) is inverse-U shaped:
R (τ = 0) = 0 (no taxes) and R(τ = 1) = 0 (nobody works)
Tax rate τ that maximizes R:

0
0 = R (τ ) = z τ dz /d (1 τ)
) τ MAX = 1/(1 + ε)
where ε = [(1 τ )/Z ]dz /d (1 τ ) is the uncompensated taxable
income elasticity w.r.t. 1 τ
Strictly ine¢ cient to have τ > τ
Technical note: why write ε as elasticity of z w.r.t. 1 τ instead of τ?

Optimal Top Income Tax Rate
Now consider constant mtr τ above …xed income threshold z̄
Derive optimal τ using perturbation argument
Assume away income e¤ects εc = εu = ε
Diamond (1998) shows this is a key theoretical simpli…cation
Assume that there are N individuals above z̄
Denote by z m (1 τ ) their average income, which depends on

net-of-tax rate 1 τ

Three e¤ects of small d τ > 0 reform above z̄
Mechanical increase in tax revenue:

dM = N [z m z̄ ]d τ
Behavioral response:
dz m
dB = Nτdz m = Nτ dτ
d (1 τ )
τ
= N ε̄ z m d τ
1 τ
Welfare e¤ect: money-metric utility loss is dM by envelope theorem:
If govt. values marginal consumption of rich at ḡ 2 (0, 1)

dW = ḡ dM
ḡ depends on curvature of u (c ) and SWF
τ
dM + dW + dB = Nd τ (1 ḡ )[z m z̄ ] ε̄ zm
1 τ
Optimal τ such that dM + dW + dB = 0 )
τ TOP (1 ḡ )(zm /z̄ 1)

=
1 τ TOP ε̄ zm /z̄
τ TOP decreases with ḡ [redistributive tastes]
τ TOP decreases with ε̄ [e¢ ciency]
τ TOP increases with zm /z̄ [thickness of top tail]
Note: this is not an explicit formula for top tax rate because zm /z̄ is
a fn. of τ
In US tax return data, z m /z̄ very stable above z̄ = $200K with

zm
z̄ = 2
Empirically, thickness parameter z m /z̄ unrelated to top tax rate τ

(Saez 1999)
How is this consistent with behavioral responses to taxation

(dz m /d (1 τ ) > 0)?
R
Increase in τ reduces both z >z̄ zh (z )dz and 1 H (z̄ )
R
Leaves zm = z >z̄ zh (z )dz /(1 H (z̄ ) constant
High taxes reduce number of people in tail, but could leave thickness of
tail unchanged

Diamond (1998) shows that with Pareto skill distribution, income

distribution is Pareto with parameter a invariant to τ
a zm
With Pareto distribution (f (z ) = a k a /z 1 +a ), a 1 = z̄ )a=2
1 ḡ
) τ TOP =
1 ḡ + 2ε̄
Ex: ε̄ = 0.5, ḡ = 0.5, a = 2 ) τ TOP = 33%

Zero Top Rate with Bounded Distribution
Suppose top earner earns z T , and second earner earns z S
Then z m = z T when z̄ > z S ) z m /z̄ ! 1 when z̄ ! z T )

τ
dM = Nd τ [z m z̄ ] ! 0 < dB = Nd τε̄ zm
1 τ
Optimal τ is zero for z̄ close to z T
Sadka-Seade zero top rate result
Result applies literally only to top earner: if z T = 2 z S then

z m /z̄ = 2 when z̄ = z S
Zero at top no longer considered to be of practical relevance

Connection to Revenue Maximizing Tax Rate
Revenue maximizing top tax rate can be calculated by putting 0

weight on welfare of top incomes
Utilitarian SWF ) ḡ = uc (z m ) ! 0 when z̄ ! ∞
Rawlsian SWF ) ḡ = 0 for any z̄ > min(z )
If ḡ = 0, we obtain τ TOP = τ MAX = 1/(1 + a ε̄)
Example: a = 2 and ε̄ = 0.5 ) τ = 50%
La¤er linear rate is a special case where z̄ = 0
) z m /z̄ = ∞ = a/(a 1) ) a = 1 ) τ MAX = 1/(1 + ε̄)

Optimal Non-Linear Income Tax
Now consider general problem of setting optimal T (z )
Let H (z ) = CDF of income [population normalized to 1] and h (z ) its

density [endogenous to T (.)]
Let g (z ) = social marginal value of consumption for taxpayers with

income z in terms of public funds
Let G (z ) be the average social marginalRvalue of consumption for

∞
taxpayers with income above z [G (z ) = z g (s )h (s )ds/(1 H (z ))]

General Non-Linear Income Tax
Consider small reform: increase T 0 by d τ in small band (z, z + dz )
Mechanical revenue e¤ect
dM = dzd τ (1 H (z ))
Mechanical welfare e¤ect
dW = dzd τ (1 H (z ))G (z )
Behavioral e¤ect: substitution e¤ect δz inside small band [z, z + dz ]:
dB = h (z )dz T 0 δz = h (z )dz T 0 d τ ε(z ) z/(1 T 0)
Optimum dM + dW + dB = 0

General Non-Linear Income Tax
Optimal tax schedule satis…es:
T 0 (z ) 1 1 H (z )
0
= [1 G (z )]
1 T (z ) ε (z ) zh (z )
T 0 (z ) decreasing in g (z 0 ) for z 0 > z [redistributive tastes]
T 0 (z ) decreasing in ε(z ) [e¢ ciency]
T 0 (z ) decreasing in h (z )/(1 H (z )) [density]
Connection to top tax rate: consider z ! ∞
G (z ) ! ḡ , (1 H (z ))/(zh (z )) ! 1/a
ε(z ) ! ε̄ ) T 0 (z ) = (1 ḡ )/(1 ḡ + a ε̄) = τ TOP

Negative Marginal Tax Rates Never Optimal
Suppose T 0 < 0 in band [z, z + dz ]
Increase T 0 by d τ > 0 in band [z, z + dz ]
dM + dW > 0 because G (z ) < 1 for any z > 0
Without income e¤ects, G (0) = 1

Value of lump sum grant to all equals value of public good
Concave SWF –> G 0 (z ) < 0
dB > 0 because T 0 (z ) < 0 [smaller e¢ ciency cost]
Therefore T 0 (z ) < 0 cannot be optimal
Marginal subsidies also distort local incentives to work
Better to redistribute using lump sum grant

Numerical Simulations of Optimal Tax Schedule
Formula above is a condition for optimality but not an explicit

formula for optimal tax schedule
Distribution of incomes H (z ) endogenous to T (.)
Therefore need to use structural approach (speci…cation of primitives)

to calculate optimal T (.)
Saez (2001) speci…es utility function (e.g. constant elasticity):

1
u (c, l ) = c (l )1 + ε
) l = [(1 T 0 )w ]ε
Calibrate the exogenous skill distribution F (w ) such that actual T (.)

yields empirical H (z )
Numerical Simulations
Use formula expressed in terms of F (w ) to solve for optimal T (z ):
∞ Z
T 0 (z (w )) 1 1 G 0 (u (s ))
= 1 + 1 f (s )ds,
1 T 0 (z (w )) ε wf (w ) w p
R
where p = G 0 (u (s ))f (s )ds is marginal value of public funds
Iterative …xed point method to solve for T (z ):
Start with initial MTR schedule T00 and compute incomes z 0 (w ) using
individual FOCs
Get T 0 (0) using govt budget constraint, compute utilities u 0 (w )

R
Compute p0 = G 0 (u 0 (s ))f (s )ds
Use formula to calculate T10 and iterate until convergence (Brewer,

Saez, Shephard 2009)
Commodity vs. Income Taxation
Atkinson and Stiglitz (1976) analyze optimal combination of income

and commodity taxes
K consumption goods c = (c1 , .., cK ) with pre-tax price

p = (p1 , .., pK )
Individual h has utility u (c1 , .., cK ) φh (z )
Assumes (1) separability between c and z and (2) homogeneity of

consumption sub-utility u
Main result: commodity taxation is super‡uous
max SWF = max SWF

t,T (.) t =0,T (.)

Atkinson-Stiglitz: Implications for Capital Taxation
Two period model: wage rate w in period 1, retired in period 2
Let δ = discount rate, ψ(.) disutility of e¤ort, and utility
u ( c2 )
u h ( c1 , c2 , z ) = u ( c1 ) + ψ(z/w )
1+δ
The budget constraint is
c1 + c2 / ( 1 + ( 1 θ )r ) z T (z )
Capital income tax θ is equivalent to tax on c2
Atkinson-Stiglitz implies that θ = 0 in the presence of an optimal

income tax

Atkinson-Stiglitz: Implications for Capital Taxation
If low ability people have higher δ then capital income tax tK > 0 is
desirable (Saez 2004)
Violates homogeneous utility assumption
Savings reveal information about type and relax IC constraint

Chamley-Judd vs. Atkinson-Stiglitz
Chamley-Judd: constrained policy instruments (linear taxes) but
dynamic
Atkinson-Stiglitz: full set of policy instruments (non linear income

tax) but static
New dynamic public …nance literature: full set of instruments in

dynamic model
In dynamic Mirrlees models, optimal capital tax is not zero (Golosov,

Kocherlekota, and Tsyvinski 2003)
Optimum features a wedge between MRS and MRTS
Intuition: payo¤ to distorting savings decisions relaxes IC constraints in

optimal income tax problem in next period
Does not emerge in Atkinson-Stiglitz because all income is earned in

…rst period
Optimal Transfer Programs
Several types of transfer programs are used in practice, each justi…ed

by a di¤erent theory and set of assumptions
Option 1: Negative Income Tax: TANF (Mirrlees 1971)
Bene…ts: no one omitted; low admin costs; no stigma
Costs: e¢ ciency loss from less work
Option 2: Work-for-welfare: EITC (Saez 2002)
Bene…ts: more incentive to work; low admin costs
Costs: e¢ ciency loss in phaseout range, no coverage of non-workers

Optimal Transfer Programs
Option 3: Categorical anti-poverty programs: assistance for blind

(Akerlof 1978)
Bene…ts: tagging relaxes incentive constraint by tying tax rate to

immutable qualities
Costs: not always feasible and limited coverage
Option 4: In-kind transfers: food stamps, public housing (Nichols

and Zeckhauser 1982)
Bene…ts: E¢ ciency gains from relaxing IC for high-types via ordeals
Costs: Paternalism (spend on the right things), ine¢ cient ordeal cost

Optimal Transfers: Mirrlees Model
Mirrlees model predicts that optimal transfer at bottom takes the

form of a Negative Income Tax
Lump sum grant T (0) for those with no earnings
High MTRs T 0 (z ) at the bottom to phase-out the lump sum grant

quickly
NIT optimal because it targets transfers to the most needy

Optimal Transfers: Participation Responses and EITC
Mirrlees result predicated on assumption that all individuals are at an

interior optimum in choice of labor supply
Rules out extensive-margin responses
But empirical literature shows that participation labor supply responses

are important, especially for low incomes
Diamond (1980), Saez (2002), Laroque (2005) incorporate such

extensive labor supply responses into optimal income tax model
Generate extensive margin by introducing …xed job packages (cannot

smoothly choose earnings)

Saez 2002: Participation Model
Model with discrete earnings outcomes: w0 = 0 < w1 < ... < wI
Tax/transfer Ti when earning wi , ci = wi Ti
Pure participation choice: skill i individual compares ci and c0 when

deciding to work
With participation tax rate τ i , ci c0 = wi (1 τi )
In aggregate, fraction hi of population earns wi , with ∑i hi = 1
Participation elasticity is
ei = ( ci c0 )/hi ∂hi /∂(ci c0 )

Saez 2002: Participation Model
Social Welfare function is summarized by social marginal welfare

weights at each earnings level gi
No income e¤ects ! ∑i gi hi = 1 = value of public good
Main result: work subsidies with T 0 (z ) < 0 (such as EITC) optimal
Key requirements in general model with intensive+extensive responses
Responses are concentrated primarily along extensive margin
Social marginal welfare weight on low skilled workers > 1 (not true
with Rawlsian SWF)

Saez 2002: Intuition for EITC
Two types: doctors (wage wh ) and plumbers (wl )
Both can choose whether to work, but doctors cannot become

plumbers
Transfer to 0 income individuals ! help plumbers but distort doctors’

incentives to work
Transfer to those with income of wl ! still help plumbers, but do not

distort doctors’incentives
Therefore better to have a larger transfer to wl than 0, i.e. have a

subsidy for work = EITC
In pure ext margin model, transfer T1 only distorts behavior of type 1
Higher types don’t move down
But transfer T0 distorts behavior of all types on extensive margin

Saez 2002: Optimal Tax Formula
Small tax cut dTi < 0 ) dci = dTi > 0. Three e¤ects:
1 Fiscal E¤ect: loss of tax revenue dM = hi dci
2 Welfare E¤ect: each worker in job i gains dTi so welfare gain

dW = hi gi dci
No …rst order welfare loss for switchers
3 Behavioral E¤ect: dhi = ei hi dci /(ci c0 )

!Tax loss: dB = (Ti T0 )dhi = ei hi dTi (Ti T0 ) / ( c i c0 )
FOC: dM + dB + dW = 0 )
τi Ti T0 1
= = (1 gi )
1 τi ci c0 ei
g1 > 1 ) T1 T0 < 0 ) work subsidy
Saez 2002: General Model
Model can be extended to allow both intensive and extensive

responses
Allow higher types to switch to lower jobs
General formula for optimal tax is a fn of both intensive and extensive

margin elasticity
Can be calibrated using empirical estimates of these elasticities

Tagging: Akerlof 1978
We have assumed that T (z ) depends only on earnings z
In reality, govt can observe many other characteristics X also

correlated with ability and set T (z, X )
Ex: gender, race, age, disability, family structure, height,...
Two major results:
1 If characteristic X is immutable then redistribution across the X

groups will be complete [until average social marginal welfare weights
are equated across X groups]
2 If characteristic X can be manipulated but X correlated with ability

then taxes will depend on both X and z

Mankiw and Weinzierl 2009
Tagging with Immutable Characteristics
Consider a binary immutable tag: Tall vs. Short
1 inch = 2% higher earnings on average (Postlewaite et al. 2004)
Average social marginal welfare weights ḡ T < ḡ S because tall earn

more
Lump sum transfer from Tall to Short is desirable
Optimal transfer should be up to the point where ḡ T = ḡ S
Calibrations show that average tall person (> 6ft) should pay $4500
more in tax
Problems with Tagging
Height taxes seem implausible, challenging validity of tagging model
What is the model missing?
1 Horizontal Equity concerns impose constraints on feasible policies:
Two people earning same amount but of di¤erent height should be

treated the same way
2 Height does not cause high earnings
In practice, tags used only when causally related to ability to earn

[disability status] or welfare [family structure, # kids, medical expenses]
Lesson: Mirrlees analysis [T (z )] may be most sensible even in an

environment with immutable tags
Nichols and Zeckhauser 1982: In-Kind Redistribution
In …rst-best full information model, no reason for in-kind transfers
In-kind transfer is tradeable at market price ! in-kind equivalent to

cash
In-kind transfer non-tradeable ! in-kind inferior to cash
Nichols and Zeckhauser: potential rationale for in-kind transfers

emerges in Mirrlees-type model with informational constraints
With heterogeneity in preferences, may be able to relax IC constraints

using in-kind transfers

Nichols and Zeckhauser: Simple Illustration
Consider a soup kitchen as an in-kind transfer policy
Let S = soup and W = wait in minutes
Two agents: poor (P) and rich (R)
Utility functions are increasing in S and decreasing in W :
Up = 2S .5W
Ur = S 1W
R has higher disutility from waiting and lower utility from soup
Social welfare
SWF = Up + Ur

Soup Kitchen without Wait: Cash Transfer
With a total of $100 in soup to give away and no wait times, the soup
will be split between the two agents
Both get some utility from soup, so both will claim it
Assume that they split it equally, resulting in
Up = 100
Ur = 50
SWF = 150
Equivalent to a cash-transfer program that pays each agent $50

Soup Kitchen with Wait Times: In-Kind Transfer
Now suppose we impose wait time of 51 minutes
R leaves - not worth it to him for $50 in food - gets Up = 0
P gets utility of 200 25.5 = 174.5
Social welfare with in-kind transfer (wait time) greater than cash
transfer (no wait time)
Targeting gains outweighing e¢ ciency losses from ordeal
Scope for such targeting depends upon degree of heterogeneity in

preferences

Income Taxation as Insurance (Varian 1980)
Important assumption in Mirrlees model: no ex-post uncertainty
Once skill type is revealed, agent controls income perfectly
In practice, there is considerable ex-post uncertainty in incomes (e.g.

unemployment shocks)
In this case, a progressive tax system could provide insurance
Do not want 100% insurance for moral hazard reasons
But some insurance desirable if individuals are risk averse

Varian: Taxation as Insurance
Income z = e + e where e is e¤ort and e is a random noise
Government observes only z and sets a tax schedule based on z
Individual utility
U = Eu (z T (z )) e
Chooses e = e to maximize this utility
E¤ort e low if tax schedule very redistributive
Government chooses T (.) to maximize indirect utility: trade-o¤

insurance vs incentives
Optimal tax system depends on parameters similar to those in

Mirrlees model
Varian Model: Private Insurance
Varian model has received less attention than Mirrlees model
One reason: government is not better than private market in

providing such insurance
In adverse selection (e.g. Mirrlees) models, only government can

improve redistributive outcomes once skills are revealed to agents
Agents cannot write contracts behind veil of ignorance
In pure moral hazard model with ex-post information revelation,

private markets should in principle reach optimum themselves
In practice, …rms o¤er wage contracts that provide some insurance

against bad luck
Ex: tenure system in universities, severance payments

Part 5: Income Taxation and Labor Supply
Harvard University
Fall 2012
Public Economics Lectures ()Part 5: Income Taxation and Labor Supply 1 / 220
Outline
1 Labor Supply Elasticity Estimation: Overview
2 Non-linear budget set methods
3 Summary of elasticity estimates in static models
4 Intertemporal Labor Supply Models
5 Elasticity of Taxable Income
6 Micro vs Macro Elasticities
7 Implications for Preference Parameters
References
Surveys in labor economics:
Pencavel (1986) Handbook of Labor Economics vol 1
Heckman and Killingsworth (1986) Handbook of Labor Econ vol 1
Blundell and MaCurdy (1999) Handbook of Labor Economics vol 3
Surveys in public economics:
Hausman (1985) Handbook of Public Economics vol 1
Mo¢ tt (2003) Handbook of Public Economics vol 4
Saez, Slemrod, and Giertz (JEL 2011)
Theoretical Issues in Estimation
Labor supply elasticity is a parameter of fundamental importance for

income tax policy
log l
Optimal tax rate depends inversely on εc = ∂∂log w U =U , the
compensated wage elasticity of labor supply
First discuss econometric issues that arise in estimating these (and

other) elasticities
Use a simple labor supply model to organize the empirical issues
Baseline Labor-Leisure Choice Model: Key Assumptions
1 One period
2 Intensive-margin, one dimensional choice
3 No frictions or adjustment costs
4 Linear tax system
Static Model: Setup
Let c denote consumption and l hours worked
Normalize price of c to one
1 +1/ε
Agent has utility u (c, l ) = c a l1+1/ε
Agent earns wage w per hour worked and has y in non-labor income
With tax rate τ on labor income, individual solves
max u (c, l ) s.t. c = (1 τ )wl + y
Labor Supply Behavior
First order condition

(1 τ )w = al 1/ε
Yields labor supply function
log l = α + ε log(1 τ )w
Here y does not matter because u is quasilinear
Log-linearization of …rst order condition for general utility u (c, l )

would yield a labor supply fn of the form:
log l = α + ε log(1 τ )w ηy
Can recover εc from ε and η using Slutsky equation
Problems with OLS Estimation of Labor Supply Equation
1 Econometric issues
Unobserved heterogeneity [tax instruments]
Measurement error in wages and division bias [tax instruments]
Selection into labor force [panel data]
2 Extensive vs. intensive margin responses [participation models]
3 Non-hours responses [taxable income]
4 Progressive taxes [non-linear budget set methods]
5 Frictions [macro comparisons, bounds]
Econometric Problem 1: Unobserved Heterogeneity
Early studies estimated elasticity using cross-sectional variation in

wage rates
Problem: unobserved heterogeneity
Those with high wages also have a high propensity to work
Cross-sectional correlation between w and h likely to yield an upward

biased estimate of ε
Solution: use taxes as instruments for (1 τ )w
Econometric Problem 2: Measurement Error/Division Bias
Wage w is typically not observed; backed out from dividing earnings

by reported hours
When hours are measured with noise, this can lead to “division bias”
Let l denote true hours, l observed hours
e
Compute w = l where e is earnings
) log l = log l + µ
) log w = log e log l = log e log l µ = log w µ
Measurement Error and Division Bias
Mis-measurement of hours causes a spurious link between hours and
wages
Estimate a regression of the following form:
log l = β1 + β2 log w + υ
Then
cov (log l, log w ) cov (log l + µ, log w µ)
Eb
β2 = =
var (log w ) var (log w ) + var (µ)
Problem: Eb
β2 6= ε because orthogonality restriction for OLS violated
Ex. workers with high mis-reported hours also have low imputed
wages, biasing elasticity estimate downward
Solution: tax instruments again

Econometric Problem 3: Selection into Labor Force
Consider model with …xed costs of working, where some individuals

choose not to work
Wages are unobserved for non-labor force participants
Thus, OLS regression on workers only includes observations with

li > 0
This can bias OLS estimates: low wage earners must have very high
unobserved propensity to work to …nd it worthwhile
In cross-sections, requires a parametric selection correction (e.g.

Heckman 1979)
Non-parametric approach: use panel data to estimate within-person

intensive-margin changes
Extensive vs. Intensive Margin
Related issue: want to characterize e¤ect of taxes on labor force

participation decision
With …xed costs of work, individuals may jump from non-participation

to part time or full time work (non-convex budget set)
Can be handled using a discrete choice model:
P = φ(α + ε log(1 τ) ηy )
where P 2 f0, 1g is an indicator for whether the individual works
Function φ typically speci…ed as probit or linear prob model
Note: here it is critical to have tax variation; regression cannot be run

with wage variation
Non-Hours Responses
Traditional literature focused purely on hours of work and labor force

participation
Problem: income taxes distort many margins beyond hours of work
More important responses may be on those margins
Hours very hard to measure (most ppl report 40 hours per week)
Two solutions in modern literature:
Focus on taxable income (wl) as a broader measure of labor supply

(Feldstein 1995)
Focus on subgroups of workers for whom hours are better measured,

e.g. taxi drivers
Progressive Taxes and Labor Supply
OLS regression speci…cation is derived from model with a single linear

tax rate
In practice, income tax systems are non-linear
Consider e¤ect of US income tax code on budget sets
Source: Congressional Budget O¢ ce 2005
Example 1: Progressive Income Tax
Example 2: EITC
Example 3: Social Security Payroll Tax Cap
Example 4: Negative Income Tax
Progressive Taxes and Labor Supply
Non-linear budget set creates two problems:
1 Model mis-speci…cation: OLS regression no longer recovers structural

elasticity parameter ε of interest
Two reasons: (1) underestimate response because people pile up at

kink and (2) mis-estimate income e¤ects
2 Econometric bias: τ i depends on income wi li and hence on li
Tastes for work are positively correlated with τ i ! downward bias in

OLS regression of hours worked on net-of-tax rates
Solution to problem #2: only use reform-based variation in tax rates
But problem #1 requires fundamentally di¤erent estimation method

Optimization Frictions
Standard methods assume that agents can costlessly adjust hours of

work
In practice, most hours changes occur with job switches (Altonji and
Paxson 1992)
And many individuals may be inattentive to change in tax rates
Implies that long-run impacts of policies may not be identi…ed from

short run variation
Optimization Frictions and Identi…cation
Chetty (2012) formalizes how frictions a¤ect identi…cation of

elasticities
Agents can choose any xt that generates a utility loss less than
exogenous threshold δ:
U (xi ) U (xi ) < δpxi
A given price p produces a choice set X (p, δ) instead of a single

point x (p )
Construction of Choice Set
151
150
Np t x D Ýp t Þ
149
148
147
Utility u( xt )
146
145
144
143
142
141
X Ý p t , NÞ
140
6 8 10 12 14 16 18 20 22
Source: Chetty 2009

Demand (xt)
Identification with Optimization Frictions
3.0
ε=1
2.8
log demand (log xt)
2.6
2.4
2.2
2.0
1.8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

log(pA) log(pB)
Source: Chetty 2009 log price (log pt)
3.0
ε=1
2.8
log demand (log xt)
2.6
2.4
2.2
2.0
1.8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

log(pA) log(pB)
3.0
ε=1
2.8
log demand (log xt)
2.6
2.4
2.2
2.0
1.8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

log(pA) log(pB)
3.0
ε=1
2.8
log demand (log xt)
2.6
2.4
2.2
2.0
1.8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

log(pA) log(pB)
Optimization Frictions and Identi…cation
Identi…cation problem: Multiple observed elasticities bε can be

generated by a model with a given structural elasticity when δ > 0
Conversely, multiple structural elasticities consistent with observed bε
Note that this is not a …nite-sample problem; does not disappear as

sample size approaches ∞
One focus of current research: how to deal with such frictions and
recover ε?
Estimating Elasticities Using Variation in Tax Rates
Now discuss literature estimating intensive-margin elasticities using

variation in tax rates for identi…cation
Begin by analyzing how to estimate elasticities with progressive taxes

in models without frictions
First discuss traditional NLBS estimation using maximum likelihood

methods
Then discuss recent literature on bunching estimators
Then discuss most recent work on frictions and bunching estimators
Non-Linear Budget Set Methods
Traditional approach to estimating elasticities with non-linear budget

sets pioneered by Hausman (1981)
Assume an uncompensated labor supply equation:
li = α + βwi (1 τ i ) + γyi + υi
Error term υi is normally distributed with variance σ2
Observed variables: wi , τ i , yi , and li
Technique: (1) construct likelihood function given observed labor

supply choices on NLBS, (2) …nd parameters (α, β, γ) that maximize
likelihood
Important insight: need to use “virtual incomes” in lieu of actual

unearned income with NLBS
Non-Linear Budget Set Estimation: Virtual Incomes
w3
y3
w2
y2
w1
y1
-L L2 l* L1
S ource: Haus man 1985
NLBS Likelihood Function
Consider a two-bracket tax system
Individual can locate on …rst bracket, on second bracket, or at the
kink lK
Likelihood = probability that we see individual i at labor supply li
given a parameter vector
Decompose likelihood into three components
Component 1: individual i on …rst bracket: 0 < li < lK
li = α + βwi (1 τ 1 ) + γy 1 + υi
Error υi = li (α + βwi (1 τ 1 ) + γy 1 ). Likelihood:
Li = φ((li (α + βwi (1 τ 1 ) + γy 1 )/σ)
Component 2: individual i on second bracket: lK < li . Likelihood:
Li = φ((li (α + βwi (1 τ 2 ) + γy 2 )/σ)
Likelihood Function: Located at the Kink
Now consider individual i located at the kink point
If tax rate is τ 1 and virtual income y 1 individual wants to work l > lK
If tax is τ 2 and virtual income y 2 individual wants to work l < lK
These inequalities imply:
α + βwi (1 τ 1 ) + γy 1 + υi > lK > α + βwi (1 τ 2 ) + γy 2 + υi

lK (α + βwi (1 τ 1 ) + γy 1 ) < υi < lK (α + βwi (1 τ 2 ) + γy 2 )
Contribution to likelihood is probability that error lies in this range:
Li = Φ[(lK (α + βwi (1 τ 2 ) + γy 2 ))/σ]

Φ[(lK (α + βwi (1 τ 1 ) + γy 1 ))/σ]
Maximum Likelihood Estimation
Log likelihood function is ` = ∑i log Li
Final step is solving

max `(α, β, γ, σ)
In practice, likelihood function much more complicated because of
more kinks, non-convexities, and covariates
But basic technique remains the same
Hausman (1981) Application
Hausman applies method to 1975 PSID cross-section
Finds signi…cant compensated elasticities and large income e¤ects
Elasticities larger for women than for men
Shortcomings of this implementation
1 Sensitivity to functional form choices
2 No tax reforms, so does not solve fundamental econometric problem

that tastes for work may be correlated with w
NLBS and Bunching at Kinks
Subsequent studies obtain di¤erent estimates (MaCurdy, Green, and

Paarsh 1990, Blomquist 1995)
Several studies …nd negative compensated wage elasticity estimates
Debate: impose requirement that compensated elasticity is positive or

conclude that data rejects model?
Fundamental source of problem: labor supply model predicts that

individuals should bunch at the kink points of the tax schedule
But we observe very little bunching at kinks, so model is rejected by

the data
Interest in NLBS models diminished despite their conceptual

advantages over OLS methods
Saez 2010: Bunching at Kinks
Saez observes that only non-parametric source of identi…cation for

elasticity in a cross-section is amount of bunching at kinks
Intuition: discontinuous reduction in wage rate at kink yields source of

non-parametric identi…cation
All other cross-sectional tax variation is contaminated by smooth

heterogeneity in tastes
Derives an estimator for the compensated taxable income elasticity

using amount of bunching at kinks
dz /z excess mass at kink
εc = =
dt/(1 t ) % change in NTR
Currently a popular approach (esp. when adapted to account for

frictions) because it yields highly credible estimates
Saez 2010: Bunching at Kinks
Saez implements this method using individual tax return micro data
(IRS public use …les) from 1960 to 2004
Advantage of dataset over PSID: very little measurement error
Finds sharp bunching around …rst kink point of the EITC for
self-employed
Later shown to be largely due to reporting e¤ects
However, no bunching observed at any kink for wage earners
Earnings Density and the EITC: Wage Earners vs. Self-Employed
Earnings Density and the EITC: Wage Earners vs. Self-Employed
Friedberg 2000: Social Security Earnings Test
Uses CPS data on labor supply of retirees receiving Social Security

bene…ts
Studies bunching based on responses to Social Security earnings test
Earnings test: phaseout of SS bene…ts above an exempt amount
Phaseout rate varies by age group - 50%, 33%, 0 (lower for older
workers)
Friedberg: Estimates
Estimates elasticities using Hausman method, …nds relatively large

compensated and uncompensated elasticities
Haider and Loughran (2008) replicate these results in admin. data

over mroe years
Find that degree of bunching and implied elasticities are 3 times larger
in data with less msmt error
Ironically, lost social security bene…ts are considered delayed

retirement with an actuarial adjustment of future bene…ts
!So the one kink where we do …nd real bunching is actually not real!
Why not more bunching at kinks?
1 Small structural elasticity
2 Noise in income generation process
3 Price misperceptions and salience e¤ects
Liebman and Zeckhauser (2009): “Schmeduling”
Ito (2012): empirical evidence that average price matters more than
marginal price
Ito (2012)
Ito (2012) presents evidence that individuals pay more attention to

average prices than marginal prices
Paper can be interpreted as estimating price-perception function

e(p + τ )
p
Studies electricity consumption in Orange County, CA with two

sources of price variation
Kinks in rate structure generate variation in marginal price
Spatial discontinuity coupled with sharp price increase in 2000 by one

utility changes average price
Consumption Density and Price Schedule in 2007: Bunching Around Kink Points
Source: Ito 2012
A Spatial Discontinuity in Electric Utility Service Areas in Orange County, California
Source: Ito 2012
Changes in Consumption from July 1999 to July 2000,
by Distance from the Utility Border
Source: Ito 2012
Changes in Consumption from August 1999 to August 2000,
by Distance from the Utility Border
Source: Ito 2012
3 Price misperceptions and salience e¤ects
4 Optimization frictions and rigidities in job o¤ers
Chetty, Friedman, Olsen, Pistaferri (2011)
Chetty, Friedman, Saez (2012)
Kleven and Waseem (2012)
Gelber, Jones, and Sacks (2012)

Bastani and Selin (2012)
Chetty, Friedman, Olsen, Pistaferri: Model
Firms post jobs with di¤erent hours o¤ers
Workers draw from this distribution and must pay search cost to
reoptimize
Firm cater to aggregate worker preferences: posted distribution …ts

aggregate tastes
Therefore not all workers locate at optimal hours
Bunching at kink and observed responses to tax reforms attenuated
Chetty et al. 2011: Testable Predictions
Model generates two key predictions:
1 [Size] Larger tax changes generate larger observed elasticities
Large tax changes are more likely to induce workers to search for a
di¤erent job
2 [Scope] Tax changes that apply to a larger group of workers generate

larger observed elasticities
Firms tailor jobs to preferences of common workers
30000
.44
25000
.42
15000 20000
.4
(sum) cnt
NTR
.38 .36
10000
.34
5000
220000 240000 260000 280000 300000 320000
pig
NTR (sum) cnt
Income Distribution for Wage Earners Around Top Tax Cutoff
100000
80000
Frequency
60000
40000
20000
-50 -40 -30 -20 -10 0 10 20 30 40 50
Taxable Income Relative to Top Bracket Cutoff (1000s DKr)
Source: Chetty et al. 2009
100000
80000
Frequency
60000
Excess mass = BÝAbÞ

40000
20000
-50 -40 -30 -20 -10 0 10 20 30 40 50
100000
Excess mass (b) = 0.81
Standard error = 0.05
80000
Frequency
60000
40000
20000
-50 -40 -30 -20 -10 0 10 20 30 40 50
Married Women vs. Single Men
30000
30000
Frequency (married women)
Frequency (single men)

20000
Married Women
Excess mass (b)= 1.79
20000
Single Men
10000
10000
0
-50 -40 -30 -20 -10 0 10 20 30 40 50

Teachers vs. Military
4000
8000
Teachers
Excess mass (b)= 3.54
3000
6000
Frequency (military)
Frequency (teachers)
2000
4000
1000
2000
Military
Excess mass (b) = -0.12
0
0
-50 -40 -30 -20 -10 0 10 20 30 40 50

Taxable Income Distributions in 1994
3000
4000 6000 8000 10000 12000 14000

Frequency (all wage earners)
2000
Married Women
Excess Mass (b) = 1.03
1000
All Wage Earners
Excess Mass (b) = 0.61
0
210 220 230 240 250 260 270 280 290 300
Taxable Income (1000s DKR)
1995
3000
12000
2000
b = 1.25
s.e. = 0.16
b = 0.41
8000
s.e. = 0.08
1000
4000
0
210 220 230 240 250 260 270 280 290 300
1996
3000
12000
b = 1.55
2000
s.e. = 0.17
8000
b = 0.66
s.e. = 0.09
1000
4000
0
210 220 230 240 250 260 270 280 290 300
1997
3000
15000

b = 1.26
s.e. = 0.19
2000
10000
b = 0.58
s.e. = 0.01
1000
5000
0
210 220 230 240 250 260 270 280 290 300
1998
3000
12000
b = 1.71
s.e. = 0.18
2000
8000
b = 0.78
s.e. = 0.09
1000
4000
0
210 220 230 240 250 260 270 280 290 300
1999
4000
3000
12000
b = 1.49
s.e. = 0.16
2000
b = 0.62
8000
s.e. = 0.08
1000
4000
0
210 220 230 240 250 260 270 280 290 300
2000
4000
14000

3000
b = 1.50
s.e. = 0.21
2000
10000
b = 0.72
1000
s.e. = 0.09
6000
0
210 220 230 240 250 260 270 280 290 300
2001
4000
14000

3000
b = 1.44
s.e. = 0.20
10000
2000
b = 0.55
1000
s.e. = 0.10
6000
210 220 230 240 250 260 270 280 290 300
Married Women: Taxable Income Distribution at Middle Tax Cutoff
40000
30000
Frequency
20000

Predicted excess mass = 0.35

10000
-50 -40 -30 -20 -10 0 10 20 30 40 50

Taxable Income Relative to Middle Bracket Cutoff
Observed Elasticity vs. Size of Tax Change
All Wage Earners
0.01
Observed Elasticities
0.005
-0.005
0 5% 10% 15% 20% 25% 30%
Log Change in Net-of-Tax Rate
All Teachers
1500
1000
Frequency
500
0
-50 -40 -30 -20 -10 0 10 20 30 40 50
Wage Earnings Relative to Statutory Kink (1000s DKR)
Wage Earnings Distribution: Teachers with Deductions > DKr 20,000
10000
This group
starts paying
8000
top tax here

6000
Frequency
4000
2000
0
-50 -40 -30 -20 -10 0 10 20 30 40 50
Wage Earnings Relative to Statutory Kink (1000s DKR)
(a) Electricians,
Electricians (3114),2000
2000 (b) Salesmen, 1996
300
80
60
200
Frequency
Frequency
40
100
20
0
0
-100 -50 0 50 100 -100 -50 0 50 100
Wage Wage Earnings
Earnings Relative
Relative to Topto Top Bracket
Bracket Cutoff Cutoff
(1000s DKr) Wage Wage Earnings
Earnings Relative
(1000s DKr)
(c) Nurses and Midwifes, 2001 (d) Tellers and Clerks, 1998
150
400
300
100
Frequency
Frequency
200
50
100
0
0
-100 -50 0 50 100 -100 -50 0 50 100

Wage Wage Earnings
Earnings Relative
(1000s DKr) Wage Earnings Relative to Top Bracket Cutoff (1000s DKr)
Modes of Occupation-Level Wage Earnings Distributions
30
20
Frequency
10
0
-100 -50 0 50 100

Modes of Wage Earnings Distributions Relative to Top Bracket Cutoff (1000s DKr)
Chetty et al. 2011: Lessons
Frictions and coordinated …rm responses play a crucial role behavioral

responses to taxation
NLBS models may …t data better if these factors are incorporated
Standard methods of estimating elasticities that ignore these factors

likely to underestimate elasticities
Limitation: does not yield an estimate of structural elasticity ε or

actual impact of tax system on earnings distribn.
Chetty, Friedman, Saez (2012)
Identi…es impacts of EITC on earnings distribution given existence of

frictions
Use areas with no knowledge about the EITC schedule as a

counterfactual for earnings distribution in absence of EITC
Results suggest that earlier Danish study may have signi…cantly

understated impact of tax system on earnings distribution
Income Distribution For Single Wage Earners with One Child
Is the EITC having
an effect on this
3.5 distribution? 4k
3
Percent of Wage-Earners
3k
2.5
EITC Amount ($)

2
2k
1.5
1
1k
.5
0 0k
$0 $5K $10K $25K $20K $25K $30K $35K
W-2 Wage Earnings
Empirical Analysis: Two Parts
First, develop a proxy for local knowledge about EITC schedule based
on income manipulation by self-employed individuals
Self-employment income is self-reported ! easy to manipulate
Audit data reveal very high misreporting rates of SE income
Second, compare W-2 wage earnings distributions across areas to

uncover impacts of EITC of “real” earnings behavior
Wage earnings are directly reported to IRS by employers ! virtually no

scope for misreporting
Outline of Empirical Analysis
Step 1: Document variation across neighborhoods in sharp bunching

among self-employed
Earnings Distribution in Texas
5%
4%
Percent of Filers
3%
2%
1%
0%
-$10K $0 $10K $20K

Income Relative to 1st Kink
Earnings Distribution in Kansas
5%
4%
Percent of Filers
3%
2%
1%
0%
-$10K $0 $10K $20K

Income Relative to 1st Kink
Fraction of Tax Filers Who Report SE Income that Maximizes EITC Refund
in 2008
4.1 –42.7%
2.8 –4.1%
2.1 –2.8%
1.8 –2.1%
1.5 –1.8%
1.2 –1.5%
1.1 –1.2%
0.9 –1.1%
0.7 –0.9%
0 –0.7%

among self-employed
Step 2: Establish that variation in sharp bunching across

neighborhoods is driven by di¤erences in knowledge about EITC
schedule
Movers: Neighborhood Changes
Consider individuals who move across neighborhoods to isolate causal

impacts of neighborhoods on elasticities
54 million observations in panel data on cross-zip movers
Analyze how changes in neighborhood sharp bunching a¤ect movers’

behavior
Event Study of Sharp Bunching Around Moves
5% Effect of Moving to
10th Decile = 1.93
(0.13)
4%
Sharp Bunching for Movers
Effect of Moving to
1st Decile = -0.41
(0.11)
3%
2%
1%
0%
-4 -2 0 2 4
Event Year
Movers to Lowest Movers to Middle Movers to Highest
Bunching Decile Bunching Decile Bunching Decile
Learning and Memory
Knowledge model predicts asymmetric impact of moving:
Moving to a higher-bunching neighborhood should raise EITC refund
Moving to a lower-bunching should not a¤ect EITC refund
Change in EITC Refunds vs. Change in Sharp Bunching for Movers
120
Change in EITC Refund ($)
100
β= 59.7
(5.7)
80
60
β= 6.0
(6.2)
p-value for diff. in slopes: p < 0.0001
40
-1% -0.5% 0% 0.5% 1%
Change in ZIP-3 Sharp Bunching
Agglomeration: Sharp Bunching vs. EITC Filer Density by ZIP Code
3.5%
3.0%
Sharp Bunching
2.5%
2.0%
1.5%
R2 = 0.6
1.0%
-2 0 2 4 6
log (Number of EITC Filers Per Square Mile)

among self-employed

schedule
Step 3: Compare wage earnings distributions across low- and

high-knowledge neighborhoods to uncover impacts of EITC on
earnings
Is the EITC having

an effect on this
3.5 4k
distribution?
3
3k
2.5
EITC Amount ($)

2
2k
1.5
1
1k
.5
0 0k
$0 $5K $10K $25K $20K $25K $30K $35K

W-2 Wage Earnings
High vs. Low Bunching Areas
3.5 4k
3
3k
2.5
EITC Amount ($)

2
2k
1.5
1
1k
.5
0 0k
$0 $5K $10K $25K $20K $25K $30K $35K

W-2 Wage Earnings
Lowest Bunching Decile Highest Bunching Decile

among self-employed

schedule
Step 3: Compare wage earnings distributions across low- and

high-knowledge neighborhoods to uncover impacts of EITC on
earnings
Step 4: Compare impacts of changes in EITC subsidies on earnings

across low vs. high knowledge nbhds. to account for omitted variables
Child Birth Research Design
Individuals without children are essentially ineligible for the EITC
Birth of a child therefore generates sharp variation in marginal

incentives
Birth a¤ects labor supply directly, but cross-neighborhood

comparisons provide good counterfactuals
Earnings Distribution in the Year Before First Child Birth for Wage Earners
6%
Percent of Individuals
4%
2%
0%
$0 $10K $20K $30K $40K

Wage Earnings
Lowest Sharp Middle Sharp Highest Sharp
Earnings Distribution in the Year of First Child Birth for Wage Earners
6%
4%
2%
0%
$0 $10K $20K $30K $40K

Wage Earnings
Simulated EITC Credit Amount for Wage Earners Around First Child Birth
1900
Simulated One-Child EITC Amount ($)
1850
ß = 85.4
(7.2)
1800
1750
-4 -2 0 2 4
Age of Child
Change in Simulated One-Child EITC Refund ($) Changes in Simulated EITC around Births for Wage Earners
200
β= 26.5
100
(1.97)
-100
0% 2% 4% 6% 8%
ZIP-3 Self-Employed Sharp Bunching
0 to 1 Child
Change in Simulated One-Child EITC Refund ($) Changes in Simulated EITC around Births for Wage Earners
200
β= 26.5
100
(1.97)
0
β= -0.13
(1.08)
-100
0% 2% 4% 6% 8%
ZIP-3 Self-Employed Sharp Bunching
0 to 1 Child 2 to 3 Children
Composition of Wage Earnings Responses
Where is the increase in EITC refunds coming from?
By studying distributional shifts, we estimate that:
73% comes from increase in earnings in phase-in region
27% from reduction in earnings in phase-out region
Use neighborhoods with no self-emp. sharp bunching as

counterfactual to characterize causal impacts of EITC on earnings
distribution
Impact of EITC on Income Distribution
Percent of EITC-Eligible Households Below Threshold
50% of 100% of 150% of 200% of

Poverty Line Poverty Line Poverty Line Poverty Line
No EITC
Counterfactual
13.7% 31.9% 54.3% 77.3%
EITC, No
Behavioral 9.4% 22.0% 42.1% 71.1%
Response
EITC, with
Avg. Behavioral 8.2% 21.0% 42.0% 71.3%
Response
EITC with Top

Decile Behavioral 6.2% 19.6% 42.0% 71.7%
Response
Elasticity Estimates Based on Change in EITC Refunds Around Birth of First Child
Mean Phase-in Phase-out Extensive

Elasticity Elasticity Elasticity Elasticity
A. Wage Earnings
Elasticity in U.S. 2000-2005 0.10 0.14 0.06 0.10

(0.008) (0.011) (0.006) (0.009)
Elasticity in top decile ZIP-3's 0.46 0.58 0.30 0.59

(0.017) (0.021) (0.021) (0.033)
B. Total Earnings
Elasticity in U.S. 2000-2005 0.22 0.34 0.08 0.18

(0.013) (0.020) (0.004) (0.012)
Elasticity in top decile ZIP-3's 0.95 1.32 0.34 1.05

(0.026) (0.036) (0.012) (0.039)
Traditional Labor Supply Elasticity Estimates
Return to simple model with linear tax
Large literature in labor economics estimates e¤ects of taxes and

wages on hours worked and participation
Now discuss some estimates from this older literature
Negative Income Tax
Best way to resolve identi…cation problems: exogenously increase the

marginal tax rate
NIT experiment conducted in 1960s/70s in Denver, Seattle, and other

cities
First major social experiment in U.S.
Provided lump-sum welfare grants G combined with a steep phaseout

rate τ (50%-70%)
Analysis by Rees (1974), Ashenfelter and Plant (1990), and others
Several groups, with randomization within each; approx. N = 75

households in each group
NIT Experiments: Ashenfelter and Plant 1990
Present non-parametric evidence of labor supply e¤ects
Compare implied bene…t payments to treated vs control households
Di¤erence in bene…t payments aggregates hours and participation

responses
This is the relevant parameter for expenditure calculations and

potentially for welfare analysis (revenue method of calculating DWL)
But does not decompose estimates into income and substitution

e¤ects
Hard to identify the elasticities relevant to predict labor supply e¤ects

of other programs
NIT Experiments: Findings
Signi…cant labor supply response but small overall
Implied earnings elasticity for males around 0.1
Implied earnings elasticity for women around 0.5
Response of women is concentrated along the extensive margin
Problems with Experimental Design
Estimates from NIT not considered very credible today for two reasons:
1 Self reported earnings
Treatments had …nancial incentives to under-report earnings.

Reported earnings not well correlated with actual payments
!Lesson: need to match with administrative records
2 Selective attrition
After initial year, data was collected based on voluntary income reports
by families to qualify for the grant
Those in less generous groups/far above breakeven point had much less
incentive to report
Consequently attrition rates were much higher in these groups
!No longer a random sample of treatment + controls

Tax Reform Variation
Modern studies use tax changes as natural experiments
Representative example: Eissa (1995)
Uses the Tax Reform Act of 1986 to identify the e¤ect of MTRs on
labor force participation and hours of married women
TRA 1986 cut top income MTR from 50% to 28% from 1986 to 1988
But did not signi…cantly change tax rates for the middle class
Substantially increased incentives to work of wives of high income

husbands relative wives of middle income husbands
DD strategy: compare women in top 1% households (treatment) with

women in 90th percentile and 75th percentile (controls)
Data: CPS, 1983-85 and 1989-91

Eissa 1995: Results
Participation elasticity around 0.4 but large standard errors
Hours elasticity of 0.6
Total elasticity (unconditional hours) is 0.4 + 0.6 = 1
Eissa 1995: Caveats
Does the common trends assumption hold?
Potential story biasing the result:

Trend toward “power couples” and thus DD might not be due to taxes
In the 1980s, professionals had non-working spouses
In the 1990s, professionals married to professionals
While for middle class, always married to working middle class wives
Problem: starting from very di¤erent levels for T and C groups
Liebman and Saez (2006) show that Eissa’s results are not robust
using admin data (SSA matched to SIPP)
Bianchi, Gudmundsson, and Zoega 2001
Use 1987 “no tax year” in Iceland as a natural experiment
In 1987-88, Iceland switched to a withholding-based tax system
Workers paid taxes on 1986 income in 1987; paid taxes on 1988

income in 1988; 1987 earnings never taxed
Data: individual tax returns matched with data on weeks worked from
insurance database
Random sample of 9,274 individuals who …led income tax-returns in

1986-88
Bianchi, Gudmundsson, and Zoega 2001
Large, salient change: ∆ log(1 MTR ) 49%, much bigger than

most studies
Note that elasticities reported in paper are w.r.t. average tax rates:
∑(L87 LA )/LA
εL,T /E =
∑ T86 /E86
∑ 87 EA )/EA
( E
εE ,T /E =
∑ T86 /E86
Estimates imply earnings elasticity w.r.t. marginal tax rate of roughly
0.37 (Chetty 2012)
Is this a Frisch or Hicksian elasticity?
Responses to Low-Income Transfer Programs
Particular interest in treatment of low incomes in a progressive tax
system: are they responsive to incentives?
Recent literature has focused on welfare reform in mid-1990’s
Reform modi…ed AFDC cash welfare program to provide more
incentives to work by
1 Requiring recipients to go to job training
2 Limiting the duration for which families able to receive welfare
3 Reducing phase out to 66 cents of bene…ts per $1 earnings instead of
100% cli¤
Variation across states because Fed govt. gave block grants with
guidelines
EITC also expanded during this period: general shift from welfare to
“workfare”
Monthly Welfare Case Loads: 1963-2000
Welfare Reform: Two Empirical Questions
1 Incentives: did welfare reform actually increase labor supply
Test whether EITC expansions and changes in welfare policies a¤ect

labor supply
2 Bene…ts: did removing many people from transfer system reduce their
welfare?
How did consumption change?
Focus on single mothers, who were most impacted by reform
Meyer and Rosenbaum 2001
Study the …rst question: impact of welfare reforms and EITC on labor
supply
Document dramatic (6 pp, 10%) increase in LFP for single women

with children around EITC expansion
No change for women without children
Employment Rates for Single Women with and without Children
1
Employment Rate
.9
.8
.7
1984 1986 1988 1990 1992 1994 1996

Year
Children No Children
Source: Meyer and Rosenbaum 2001
Problem: EITC expansion took place at same time as welfare reform
Try to disentangle e¤ects of welfare waivers, changes in AFDC and

state taxes, etc. using state-level variation
Fit a regression model with following policy variables at state level

(interacted with # of kids, etc.):
EITC
AFDC bene…ts
Medicaid
Waivers
Training
Child Care
From 1984-1996, the extra increase in single mom’s relative to single

women without kids is explained by:
1 EITC expansion (60%)
2 Welfare max bene…t reduction (AFDC and food stamps) (25%)
3 Medicaid if work (-10%) (insigni…cant and wrong sign)
4 Welfare waivers (time limits) 15%
5 Child care and training: 15%
Meyer and Sullivan 2004
Examine the consumption patterns of single mothers and their

families from 1984–2000 using CEX data
Question: did single mothers’consumption fall because they lost

welfare bene…ts and were forced to work?
Total Consumption: Single Mothers 1984-2000
Relative Consumption: single women with/without children
Meyer and Sullivan: Results
Material conditions of single mothers did not decline in recent years,

either in absolute terms or relative to single childless women or
married mothers
In most cases, evidence suggests that the material conditions of single

mothers have improved slightly
Question: is this because economy was booming in 1990s?
Is workfare approach more problematic in current economy?
Changing Elasticities: Blau and Kahn 2007
Identify elasticities from 1980-2000 using “grouping instrument”
1 De…ne cells (year/age/gender/education) and compute mean wages

2 Instrument for actual wage with mean wage
Identify purely from group-level variation, which is less contaminated

by individual endogenous choice
Result: total hours elasticity (including int + ext margin) shrank from
0.4 in 1980 to 0.2 today
Interpretation: elasticities shrink as women become more attached to

the labor force
Summary of Static Labor Supply Literature
1 Small elasticities for prime-age males
Probably institutional restrictions, need for one income, etc. prevent a

short-run response
2 Larger responses for workers who are less attached to labor force
Married women, low incomes, retirees
3 Responses driven by extensive margin in short-run
Ext margin (participation) elasticity around 0.2
Int margin (hours) elasticity close 0
Intertemporal Models and the MaCurdy Critique
What parameter do reduced-form regressions of labor supply on

wages or taxes identify?
MaCurdy critique: reduced-form studies did not identify any

parameter of interest in a dynamic model
Instead, estimate a mix of income e¤ects, intertemporal substitution

e¤ects, and compensated wage elasticities
MaCurdy (1981) develops a tractable method (two stage budgeting)

to identify preference parameters in a life-cycle model of labor supply
See Chetty (2006) for a simple exposition of two-stage budgeting
Life Cycle Model of Labor Supply
With time-separable utility, agents maximize

T
U= ∑ β t u ( ct , lt )
t =0
First order conditions
lt : βt ult + λwt /(1 + r )t = 0

ct : βt uct + λ/(1 + r )t = 0
Combining yields: ul (lt ) = wt uc
Intratemporal f.o.c. same as in static model
Intertemporal f.o.c.: uct /uct +1 = β(1 + r )
Dynamic Life Cycle Model: Policy Rules
λ = uc0 is the marginal utility of initial consumption
The two …rst order conditions imply that

lt = l (wt , λ/( β(1 + r ))t )
ct = c (wt , λ/( β(1 + r ))t )
Current labor and consumption choice depends on current wt
All other wage rates and initial wealth enter only through the budget
constraint multiplier λ (MaCurdy 1981)
Easy to see for separable utility:

u (c, l ) = u (c ) v (l )
0
) v (lt ) = λwt /[ β(1 + r )]t
) lt = v 0 1 (λwt /[ β(1 + r )]t )
Su¢ ciency of λ greatly simpli…es solution to ITLS model
Dynamic Life Cycle Model: Frisch Elasticity
Frisch intertemporal labor supply elasticity de…ned as:
wt ∂l
δ=( ) jλ
lt ∂wt
Experiment: change wage rate in one period only, holding all other
wages, and consumption pro…le constant
Can show that δ > 0: work more today to take advantage of

temporarily higher wage
In separable case:
lt = v 0 1 (λwt /[ β(1 + r )]t )

∂l λ
) jλ = >0
∂wt β(1 + r )t v 00 (lt )
Dynamic Life Cycle Model: Three Types of Wage Changes
1 Evolutionary wage change: movements along pro…le
2 Parametric change: temporary tax cut
3 Pro…le shift: changing the wage rate in all periods
Equivalent to a permanent parametric change
Implicitly the elasticity that static studies estimate with unanticipated

tax changes
Frisch vs. Compensated vs. Uncompensated Elasticities
Frisch elasticity Compensated static elasticity

Compensated static elasticity Uncompensated static elasticity
Compensated static elasticity: changing wages in all periods but

keeping utility constant
Uncompensated static elasticity: changing wages in each period with

no compensation
First inequality is due to inter-temporal substitution:
When wage increases only in 1 period, substitute labor from other

periods toward this period
When it increases in all periods, do not have this motive
Second inequality is due to income e¤ects (as in static model)

Frisch vs. Compensated vs. Uncompensated Elasticities
Frisch elasticity Compensated static elasticity

Compensated static elasticity Uncompensated static elasticity
Without income e¤ects, all three elasticities are equal
Otherwise inequalities are strict
Di¤erence in elasticities related to anticipated vs. unanticipated

changes
Frisch elasticity is of central interest for calibration of macro business

cycle models
Frisch Elasticities Implied by Hicksian Elasticity of 0.33
d [wli ,t ] 2 Ai ,t
εF = ε + ρ ( )
dYi ,t wli ,t
Income Effect: -d[wl*]/dY

0.00 0.11 0.22 0.33 0.44 0.55 0.66 0.35
0.00 0.33 0.33 0.33 0.33 0.33 0.33 0.33 0.50
0.20 0.33 0.34 0.35 0.36 0.38 0.41 0.44 0.55
0.40 0.33 0.34 0.36 0.39 0.43 0.49 0.55 0.60
EIS 0.60 0.33 0.34 0.37 0.42 0.48 0.56 0.66 0.65
(ρ) 0.80 0.33 0.35 0.38 0.44 0.53 0.64 0.77 0.70
1.00 0.33 0.35 0.39 0.47 0.58 0.71 0.88 0.75
1.20 0.33 0.35 0.41 0.50 0.63 0.79 0.99 0.79
1.40 0.33 0.35 0.42 0.53 0.67 0.87 1.10 0.84
Source: Chetty 2011
Structural Estimates: MaCurdy 1983 and Pencavel 2002
MaCurdy (1983)
Structural estimate using panel data for men and within-person wage
variation
Find both Frisch and compensated wage elasticity of around 0.15
But wage variation is not exogenous
Pencavel (2002)
Instruments with trade balance interacted with schooling and age
Frisch elasticity: 0.2
Uncompensated wage elasticity: 0-0.2
Card (1991) Critique of ITLS models
Critiques value of ITLS model
Fails to explain most variation in hours over lifecycle
Sheds little light on pro…le-shift elasticities that we care about
Di¢ cult to identify key parameters
Blundell, Duncan, and Meghir 1998
Use quasi-experimental variation coupled with ITLS model to identify

Frisch elasticity
Group individuals by birth cohort (decade) interacted with education

(e.g. high school or more)
Tax reforms in the UK in 1980s a¤ected groups very di¤erently
Key innovation over pure reduced-form studies: use consumption data

as a control for permanent income
Yields a structurally interpretable (λ constant) elasticity estimate
Blundell, Duncan, and Meghir: Results
Compensated wage elasticities: 0.15-0.3, depending on number of kids
Virtually no income e¤ects
Identi…cation assumption is common trends across cohort/ed groups
Reforms in 80s went in opposite directions at di¤erent times
!Secular trends cannot explain everything
Manoli and Weber 2011
Use variation in retirement bene…ts as a function of job tenure in

Austria to estimate intertemp subst. elasticity
Administrative panel for full population of Austria, 1980-2005
Question: how much do people delay retirement in order to get higher

(anticipated) bene…ts?
Rough estimate of intertemp subst. elasticity on extensive margin of

0.2 at annual level
Lump-Sum Severance Payments at Retirement
1
.75
Fraction of Last Year's Salary
.5
.333
0
0 5 10 15 20 25 30
Years of Tenure at Retirement
Source: Manoli and Weber 2009
Distribution of Tenure at Retirement
Quarterly Frequency
6000
Number of Individuals
4000
2000
0
10 15 20 25
Years of Tenure at Retirement
Source: Manoli and Weber 2009
Taxable Income Elasticities
Modern public …nance literature focuses on taxable income elasticities

instead of hours/participation elasticities
Two main reasons
1 Convenient su¢ cient statistic for all distortions created by income tax
system (Feldstein 1999)
2 Data availability: taxable income is precisely measured in tax return

data
Good overview of this literature: Saez et al. 2010 (JEL)
US Income Taxation: Trends
The biggest changes in MTRs are at the top
1 [Kennedy tax cuts]: 91% to 70% in ’63-65
2 [Reagan I, ERTA 81]: 70% to 50% in ’81-82
3 [Reagan II, TRA 86]: 50% to 28% in ’86-88
4 [Bush I tax increase]: 28% to 31% in ’91
5 [Clinton tax increase]: 31% to 39.6% in ’93
6 [Bush Tax cuts]: 39.6% to 35% in ’01-03
Feldstein 1995
Feldstein (1995) estimates the e¤ect of TRA86 on taxable income for

top earners
Constructs three income groups based on income in 1985
Looks at how incomes and MTR evolve from 1985 to 1988 for
individuals in each group using panel
Feldstein: Results
Feldstein obtains very high elasticities (above 1) for top earners
Implication: we were on the wrong side of the La¤er curve for the rich
Cutting tax rates would raise revenue
Feldstein: Econometric Criticisms
DD can give very biased results when elasticity di¤er by groups
Suppose that the middle class has a zero elasticity so that
∆ log(z M ) = 0
Suppose high income individuals have an elasticity of e so that
∆ log(z H ) = e∆ log(1 τH )
Suppose tax change for high is twice as large:
∆ log(1 τ M ) = 10% and ∆ log(1 τ H ) = 20%

e 20% 0
Estimated elasticity ê = 20% 10% = 2e
Feldstein: Econometric Criticisms
Sample size: results driven by very few observations (Slemrod 1996)
Auten-Carroll (1999) replicate results on larger Treasury dataset
Find a smaller elasticity: 0.65
Di¤erent trends across income groups (Goolsbee 1998)
Triple di¤erence that nets out di¤erential prior trends yields elasticity
<0.4 for top earners
Slemrod: Shifting vs. “Real” Responses
Slemrod (1996) studies “anatomy” of the behavioral response

underlying change in taxable income
Shows that large part of increase is driven by shift between C corp

income to S corp income
Looks like a supply side story but government is actually losing revenue
at the corporate tax level
Shifting across tax bases not taken into account in Feldstein e¢ ciency
cost calculations
18%
Wages S-Corp. Partner. Sole P. Dividends Interest Other
16%
14%
12%
10%
8%
6%
4%
2%
0%
1960
1962
1964
1966
1968
1970
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
FIGURE 7
The Top 1% Income Share and Composition, 1960-2000
Source: Saez 2004
Goolsbee: Intertemporal Shifting
Goolsbee (2000) hypothesizes that top earners’ability to retime

income drives much of observed responses
Analogous to identi…cation of Frisch elasticity instead of compensated

elasticitiy
Regression speci…cation:
TLI = α + β1 log(1 taxt ) + β2 log(1 taxt +1 )
Long run e¤ect is β1 + β2
Uses ExecuComp data to study e¤ects of the 1993 Clinton tax

increase on executive pay
Goolsbee 2000
Most a¤ected groups (income>$250K) had a surge in income in 1992

(when reform was announced) relative to 1991 followed by a sharp
drop in 1993
Simple DD estimate would …nd a large e¤ect here, but it would be

picking up pure re-timing
Concludes that long run e¤ect is 20x smaller than substitution e¤ect
E¤ects driven almost entirely by retiming exercise of options
Long run elasticity <0.4 and likely close to 0
Gruber and Saez 2002
First study to examine taxable income responses for general
population, not just top earners
Use data from 1979-1991 on all tax changes available rather than a
single reform
Simulated instruments methodology
Step 1: Simulate tax rates based on period t income and characteristics

MTRtP+3 = ft +3 (yt , Xt )
MTRt +3 = ft +3 (yt +3 , Xt +3 )
Step 2 […rst stage]: Regress log(1 MTRt +3 ) log(1 MTRt ) on
log(1 MTRtP+3 ) log(1 MTRt )
Step 3 [second stage]: Regress ∆ log TI on predicted value from …rst
stage
Isolates changes in laws (ft ) as the only source of variation in tax rates
Gruber and Saez: Results
Find an elasticity of roughly 0.1-0.4 with splines
But this is very fragile (Giertz 2008)
Sensitive to exclusion of low incomes (min income threshold)
Sensitive to controls for mean reversion
Subsequent studies …nd smaller elasticities using data from other

countries
Evidence from Danish Tax Reforms
Observed Earnings Responses to Small Tax Reforms
1
% Residual Change in Wage Earnings
0.5
0
-0.5
-1
-1.5
-4 -2 0 2 4 6
% Residual Change in Net-of-Tax Wage Rate ∆log(1-t)
Imbens et al. 2001: Income E¤ects
Estimate income e¤ects using lottery winnings
Survey responses matched with administrative data on earnings from

Social Security Administration
Divide sample into three subgroups:
1 Losers [N = 259]: “season ticket holders” who won $100-$5K
2 Winners [N = 237]: anyone who won prizes of $22K to $9.7 mil
3 Big Winners [N = 43]: winners of prizes >$2 mil total ($100K/yr)
Estimate marginal propensity to earn out of unearned income of

d [wl ]/dy = 0.1
Taxable Income Literature: Summary
Large responses for the rich, mostly intertemporal substitution and

shifting
Responses among lower incomes small in short run
Pattern con…rmed in many settings (e.g. Kopczuk 2009 - Polish ‡at

tax reform)
But many methdological problems remain to be resolved
Econometric issues: mean reversion, appropriate counterfactuals
Which elasticity is being identi…ed?
Macro Evidence
Macroeconomists estimate/calibrate elasticities by examining

long-term trends/cross-country comparisons
Identi…cation more tenuous but estimates perhaps more relevant to

long-run policy questions of interest
Use aggregate hours data and aggregate measures of taxes (average

tax rates)
But highly in‡uential in calibration of macroeconomic models
Macro models require high elasticities to …t both business cycle and

cross-country data
Prescott 2004
Uses data on hours worked by country in 1970 and 1995 for 7 OECD
countries
Technique to identify elasticity: calibration
Rough logic: posit a utility function u (c, l; ε)
Hicksian elasticity of ε = 0.8 best matches cross-country evidence
Note that this is analogous to a regression without controls for other

variables abstracting from GE e¤ects
Aggregate Hours vs. Net-of-Tax Rates Across Countries (Prescott Data)
7.4
εPrescott=0.7 Japan 1970-74
Log Hours Worked Per Adult
Japan 1993-96
US.1970-74
UK 1993-96
7.2
Germany.1970-74
France 1970-74
UK 1970-74
Canada 1993-96 US 1993-96
Canada 1970-74
Germany 1993-96 Italy 1970-74
France 1993-96
6.8
Italy 1993-96
-1 -.8 -.6 -.4 -.2

Log (1-Tax Rate)
Business Cycle Fluctuations in Employment Rates in the U.S.
Log Deviation of Employment from HP Filtered Trend
.03
.02
.01
-.01
-.02
-.03
1950 1960 1970 1980 1990 2000 2010

Year
Employment Employment - Men (25-54)
Reconciling Micro and Macro Estimates
Recent interest in reconciling micro and macro elasticity estimates
Three potential explanations
1 Statistical Bias: regulations, culture di¤ers in countries with higher tax

rates [Alesina, Glaeser, Sacerdote 2005]
2 Extensive vs. Intensive margin: “Indivisible Labor” [Rogerson 1988;

Rogerson and Wallenius 2008]
L = Nh
d log L d log N d log h d log h
= + >
d (1 τ ) d (1 τ ) d (1 τ ) d (1 τ )
3 Optimization frictions: short run vs. long run [Chetty 2012]
Chetty (2012) asks two questions
1 Can frictions quantitatively explain micro-macro puzzle and other

puzzles in labor supply literature?
2 Given frictions, what can we say about the “structural” elasticity?
Structural elasticity controls long run responses (e.g. Europe vs US)
To illustrate potential importance of frictions, …rst calculate utility

loss from ignoring tax changes under neoclassical model with ε = 0.5
Setup
Consider a static demand model; results hold in dynamic model
N individuals with quasilinear utility over two goods:

x 1 1/ε
ui (x, y ) = y + ai
1 1/ε
Agent i’s optimal demand for good x:
ai ε
)
xi (p ) = (
p
) log xi (p ) = α ε log p + vi
where vi = αi α denotes i’s deviation from mean demand
Under orthogonality condition Evi jp = 0,

E log x1 E log x0
ε=
log p1 log p0
!Observed response to price increase (p0 to p1 ) identi…es ε.
Optimization Frictions: Examples
Agent pays adjustment cost ki to change consumption
Demand set optimally at initial price p0
Let x (p ) denote observed demand at price p
De…ne observed elasticity estimated from price increase as

E log x1 E log x0
bε =
log p1 log p0
Observed elasticity confounds structural elasticity ε with adjustment
cost distribution:
bε = P (∆ui > ki )ε
e(p )
Behavioral example: price misperception p
E log p
e ( p1 ) E log pe ( p0 )
bε = ε
log p1 log p0
Restrict size of frictions by requiring that utility loss is less than

exogenous threshold δ:
U (xi ) U (xi ) < δpxi
This restriction generates a class of models around nominal model
Includes adjustment cost models, inattention, etc.
A δ class of models maps price to a choice set X (p, δ) instead of a

single point x (p )
Construction of Choice Set
151
150
Np t x D Ýp t Þ
149
148
147
Utility u( xt )
146
145
144
143
142
141
X Ý p t , NÞ
140
6 8 10 12 14 16 18 20 22
Source: Chetty 2009

Demand (xt)
3.0
ε=1
2.8
log demand (log xt)
2.6
2.4
2.2
2.0
1.8
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

log(pA) log(pB)
Bounds and Partial Identi…cation
Chetty (2012) derives bounds on ε given estimates of ε̂
Approach is closely related to modern econometrics literature on

partial identi…cation
Also called “set identi…cation” or “moment inequalities” in IO
Pioneered by Manski (1993)
See Tamer (2010 Annual Review) for a good summary
Partial Identi…cation
Main concept: avoid making strong assumptions about data

generating process
Instead, derive non-parametric bounds on parameters under

worst-case scenarios
Classic example: missing data with a binary outcome

Traditionally, assume data is missing at random
But easy to derive bounds by assuming missing data is all either 0 or 1
Chetty (2012) applies similar logic to handle model uncertainty rather

than missing data
Calculation of Bounds on Structural Elasticity
3.6
3.4
3.2
3
log demand (log xt)
2.8
2.6
2.4
2.2
1.8
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2

log(pA) log(pB)
a) Upper Bound on Structural Elasticity
3.6
3.4
3.2
3
log demand (log xt)
2.8
2.6
2.4
2.2
1.8
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2

log(pA) log(pB)
b) Lower Bound on Structural Elasticity
3.6
3.4
3.2
3
log demand (log xt)
2.8
2.6
2.4
2.2
1.8
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2

log(pA) log(pB)
Bounds on Elasticity with Optimization Frictions
For small δ, the range of structural elasticities consistent with an

observed elasticity bε in a δ class of models is approximately
4δ 4δ
[bε + (1 ρ),bε + (1 + ρ)]
(∆ log p )2 (∆ log p )2
1 bε
where ρ = (1 + (∆ log p )2 )1/2
2δ
Maps an observed elasticity bε, size of price change ∆ log p, and degree
of optimization frictions δ to bounds on ε.
Bounds shrink with (∆ log p )2
Extensive Margin Responses
Now consider bounds on extensive margin elasticities
Assume that x 2 f0, 1g and ‡ow utility is
ui (x, y ) = y + bi x
Let F (bi ) denote distribution of tastes for x
Agents optimally buy x if taste bi > p ! θ = 1 F (p )
Let structural extensive elasticity be denoted by
log θ A (pA ) log θ B (pB )

η=
log pA log pB
Let θ = observed participation rate and b
η = observed extensive
elasticity
Bounds on Structural Elasticities: δ = 1%, ∆ log p = 20%
4.0
Intensive Margin
3.5 Bounds
3.0
Elasticity (ε)
2.5
2.0
1.5
1.0 Extensive Margin

Bounds
0.5
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Observed Elasticity ( ε )
^
Source: Chetty 2011
Extensive vs. Intensive Margin Bounds
Bounds on extensive margin

p elasticities shrink linearly with δ rather
than in proportion to δ
Intuition: agents are not near optima to begin with on extensive

margin ! …rst-order utility losses from failing to reoptimize
Marginal agent loses bene…t of price cut if he doesn’t enter market
Application to Taxation and Labor Supply
What can be learned about structural elasticity from existing

estimates?
Collect estimates from a broad range of studies that estimate

intensive margin Hicksian elasticities
Calculate bounds on the intensive margin structural elasticity with

frictions of δ = 1% of net earnings
Bounds on Intensive-Margin Hicksian Elasticities with δ = 1% Frictions
å å
se( P ) ∆log(1-τ) εL εU
Study Identification P
(1) (2) (3) (4) (5) (6) (7)
A. Hours Elasticities
1. MaCurdy (1981) Lifecycle wage variation, 1967-1976 0.15 0.15 0.39 0.03 0.80
2. Eissa and Hoynes (1998) U.S. EITC, 1984-1996, Men 0.20 0.07 0.07 0.00 15.29
3. Eissa and Hoynes (1998) U.S. EITC, 1984-1996, Women 0.09 0.07 0.07 0.00 15.07
4. Blundell et al. (1998) U.K. Tax Reforms, 1978-1992 0.14 0.09 0.23 0.01 1.78
5. Ziliak and Kniesner (1999) Lifecycle wage, tax variation 1978-1987 0.15 0.07 0.39 0.03 0.80
Mean observed elasticity 0.15
B. Taxable Income Elasticities
6. Bianchi et al. (2001) Iceland 1987 Zero Tax Year 0.37 0.05 0.49 0.15 0.92
7. Gruber and Saez (2002) U.S. Tax Reforms 1979-1991 0.14 0.14 0.14 0.00 4.42
8. Saez (2004) U.S. Tax Reforms 1960-2000 0.09 0.04 0.15 0.00 3.51
9. Jacob and Ludwig (2008) Chicago Housing Voucher Lottery 0.12 0.03 0.36 0.02 0.84
10. Gelber (2010) Sweden, 1991 Tax Reform, Women 0.49 0.02 0.71 0.28 0.86
11. Gelber (2010) Sweden, 1991 Tax Reform, Men 0.25 0.02 0.71 0.12 0.54
12. Saez (2010) U.S., 1st EITC Kink, 1995-2004 0.00 0.02 0.34 0.00 0.70
13. Chetty et al. (2011a) Denmark, Top Kinks, 1994-2001 0.02 0.00 0.30 0.00 0.93
14. Chetty et al. (2011a) Denmark, Middle Kinks, 1994-2001 0.00 0.00 0.11 0.00 6.62
15. Chetty et al. (2011a) Denmark Tax Reforms, 1994-2001 0.00 0.00 0.09 0.00 9.88
Bounds on Intensive-Margin Hicksian Elasticities with δ = 1% Frictions
å å
se( P ) ∆log(1-τ) εL εU
Study Identification P
(1) (2) (3) (4) (5) (6) (7)
C. Top Income Elasticities

16. Feldstein (1995) U.S. Tax Reform Act of 1986 1.04 0.26 0.37 2.89
17. Auten and Carroll (1999) U.S. Tax Reform Act of 1986 0.57 0.12 0.37 0.21 1.53
18. Goolsbee (1999) U.S. Tax Reform Act of 1986 1.00 0.15 0.37 0.47 2.14
19. Saez (2004) U.S. Tax Reforms 1960-2000 0.50 0.18 0.30 0.14 1.77
20. Kopczuk (2010) Poland, 2002 Tax Reform 1.07 0.22 0.30 0.44 2.58
D. Macro/Cross-Sectional
21. Prescott (2004) Cross-country Tax Variation, 1970-96 0.46 0.09 0.42 0.18 1.20
22. Davis and Henrekson (2005) Cross-country Tax Variation, 1995 0.20 0.08 0.58 0.07 0.57
23. Blau and Kahn (2007) U.S. wage variation, 1980-2000 0.31 0.004 1.00 0.19 0.51
Bounds on Intensive-Margin Hicksian Elasticities with δ=1%
2.5
2
Elasticity
MaCurdy (1981)
1.5
0.5
0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Percentage Change in Net of Tax Rate ∆ log (1 –τ)
Source: Chetty 2011
3 Feldstein (1995) No disjoint sets: δ = 1% reconciles all estimates
2.5
Goolsbee TRA86
2 Saez (2004)
Elasticity
MaCurdy (1981)
1.5
Prescott (2004)
1 Gelber (2010)
Davis and Henrekson Blau and Kahn
(2005) (2007)
0.5
0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Source: Chetty 2011
3 Feldstein (1995) Unified Bounds Using All Studies: (0.47, 0.51)
2.5
Goolsbee TRA86
2 Saez (2004)
Elasticity
MaCurdy (1981)
1.5
Prescott (2004)
1 Gelber (2010)
(2005) (2007)
0.5
0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Source: Chetty 2011
3 Feldstein (1995)
Unified bounds excluding macro+top income: (0.28, 0.54)

2.5
Goolsbee TRA86
2 Saez (2004)
Elasticity
MaCurdy (1981)
1.5
Prescott (2004)
1 Gelber (2010)
(2005) (2007)
0.5
0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Source: Chetty 2011
Unified Bounds on Intensive Margin Elasticity vs. Degree of Frictions
1.2
1
Elasticity (ε)
.8
.6
.4
εδ−min=0.33
.2
0
δmin= 0.5% 1% 2% 3% 4% 5%
Optimization Frictions as a Fraction of Net Earnings (δ)
Source: Chetty 2011
Bounds on Extensive-Margin Hicksian Elasticities with δ=1% Frictions
0.5
Meyer and Rosenbaum (2004)
Blau and
Kahn (2007)
Extensive Margin Elasticity
0.4
Blundell et al. (2011)

0.3
Graversen (1998) Prescott (2004)
0.2 Jacob and Ludwig (2008)

Eissa and Hoynes (2004)
Nickell (2003)
Davis and Henrekson (2005)
0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Percentage Change in Net of Average Tax Wage ∆ log (1 –τ)
Source: Chetty 2011

Micro vs. Macro Labor Supply Elasticities
Chetty, Guren, Manoli, and Weber (2012): can frictions explain gap
between micro and macro elasticities?
Collect estimates of intensive and extensive margin elasticities

adjusted for frictions and evaluate macro predictions
Aggregate Hours vs. Net-of-Tax Rates Across Countries (Prescott Data)
7.4
εPrescott=0.7 Japan 1970-74
εmicro=0.58
Log Hours Worked Per Adult
Japan 1993-96
US.1970-74
UK 1993-96
7.2
Germany.1970-74
France 1970-74
UK 1970-74
Canada 1993-96 US 1993-96
Canada 1970-74
Germany 1993-96 Italy 1970-74
France 1993-96
6.8
Italy 1993-96
-1 -.8 -.6 -.4 -.2

Log (1-Tax Rate)
Prescott (2004) Prediction Based on Micro Elasticity
Table 2: Micro vs. Macro Labor Supply Elasticities
Extensive Intensive Aggregate

Margin Margin Hours
micro 0.25 0.33 0.58

Steady State (Hicksian)
macro 0.17 0.33 0.50
micro
Intertemporal Substitution
(Frisch)
macro
à Indivisible labor + frictions reconcile micro and macro steady-state elasticities
Business Cycle Fluctuations in Employment Rates in the U.S.
Log Deviation of Employment from HP Filtered Trend
.03
.02
.01
-.01
-.02
-.03
1950 1960 1970 1980 1990 2000 2010

Year
Employment Real Wages × Micro Extensive Frisch
Table 2: Micro vs. Macro Labor Supply Elasticities
Extensive Intensive Aggregate

Margin Margin Hours
micro 0.25 0.33 0.58

Steady State (Hicksian)
macro 0.17 0.33 0.50
Intertemporal Substitution
micro 0.32 0.54 0.86
(Frisch)
macro [2.77] [0.54] 3.31
à Even with indivisible labor, Frisch elasticity of aggregate hours >1 is

inconsistent with micro evidence
à Challenge: matching employment flucs. with extensive Frisch of 0.3
• Search/labor wedge models provide one solution
Labor Supply Elasticities: Implications for Preferences
Labor supply elasticities central for tax policy because they determine
e¢ ciency costs
But optimal income tax policy also depends on bene…ts of

redistribution (curvature of utility fn.)
u (c ) ψ (l )
Curvature of u (c ): γ = uuccc c determines how much more low

income individuals value $1 relative to higher income individuals
Risk aversion parameter γ also central for social insurance literature

and macro models
Evidence on labor supply elasticities also contains information about

γ (King, Plosser, Rebelo 1988; Basu and Kimball 2002; Chetty 2006)
Chetty 2006
Suppose marginal utility of consumption declines quickly, i.e. γ large
Then as wages rise, individuals should quickly become sated with

goods
Therefore, they should opt to consume much more leisure when

wages rise
But this would imply εl ,w << 0
Ex: if marginal utility of consumption drops to zero, agent reduces

labor supply 1-1 as wage rises
But we know that increases in wages do not cause sharp reductions in

labor supply (εl ,w > 0.1)
Places an upper bound on size of γ

Formula for Risk Aversion
Let y = unearned inc, w = wage, l = labor supply and u (c, l ) =

utility
At an interior optimum, l must satisfy
wuc (y + wl, l ) = ul (y + wl, l )
Work until point where marginal utility of an additional dollar is o¤set

by marginal disutility of work required to earn that dollar
Comparative statics of this condition implies (if ucl = 0):
wl εl ,y
γ= (1 + )
y εl c ,w
Risk aversion directly related to ratio of income e¤ect to substitution

e¤ect
Labor Supply and Risk Aversion: Intuition
Assume y = 0. At initial wage w0 , agent works l0 hours
Consider e¤ect of increasing w by 1% to w1

Shifts wuc curve up by 1% (substitution e¤ect)
Shifts wuc curve down by ∂∂ log uc
log w = γ% because γ is elasticity of MU
w.r.t. c (income e¤ect)
Therefore, γ < 1 () εl ,w > 0
If ucl 6= 0, then ul curve shifts when w changes
But the shift is ul relatively small, so change in l can still be used

to get a bound on γ
uc ,ul
Case A: γ < 1 -ul(w0l,l)
w0uc(w0l,l) -ul range with
w1uc(w1l,l)
complementarity
w1uc(w1l,l)
Case B: γ > 1
lB l0 lA l
Source: Chetty 2006
Labor Supply Elasticities and Implied Coefficients of Relative Risk Aversion
Income Compensated γ γ
Study Sample Identification Elasticity Wage Elasticity Additive ∆c/c=0.15
(1) (2) (3) (4) (5) (6) (7)
A. Hours
MaCurdy (1981) Married Men Panel -0.020 0.130 0.46 0.60

Blundell and MaCurdy (1999) Men Various -0.120 0.567 0.63 0.82
MaCurdy, Green, Paarsch (1990) Married Men Cross Section -0.010 0.035 1.47 1.81
Eissa and Hoynes (1998) Married Men, Inc < 30K EITC Expansions -0.030 0.192 0.88 1.08
Married Women, Inc < 30K EITC Expansions -0.040 0.088 0.64 1.34
Friedberg (2000) Older Men (63-71) Soc. Sec. Earnings Test -0.297 0.545 0.93 1.46
Blundell, Duncan, Meghir (1998) Women, UK Tax Reforms -0.185 0.301 0.93 1.66
Average 0.69 0.94
B. Participation
Eissa and Hoynes (1998) Married Men, Inc < 30K EITC Expansions -0.008 0.033 0.44 0.48
Married Women, Inc < 30K EITC Expansions -0.038 0.288 0.15 0.30
Average 0.29 0.39
C. Earned Income
Imbens, Rubin, Sacerdote (2001) Lottery Players in MA Lottery Winnings -0.110

Feldstein (1995) Married, Inc > 30K TRA 1986 1.040 0.32 0.41
Auten and Carroll (1997) Single and Married, Inc>15K TRA 1986 0.660 0.50 0.65
Average 0.41 0.53
D. Macroeconomic/Trend Evidence
Blau and Kahn (2005) Women Cohort Trends -0.278 0.646 0.60 1.29
Davis and Henrekson (2004) Europe/US aggregate stats Cross-Section of countries -0.251 0.432 1.74 2.25
Prescott (2004) Europe/US aggregate stats Cross-Country time series -0.222 0.375 1.78 2.30
Average 1.37 1.95
Overall Average 0.71 0.97
Source: Chetty 2006
Chetty 2006: Results
Labor supply evidence justi…es use of u (c ) = log c
wl εl ,y
Formula γ = (1 + y ) εl c ,w useful in tax, insurance, and other
applications
Income Distribution
We have covered evidence on two of the three elements critical for

optimal income taxation
1 Labor supply elasticities

2 Measurement of preferences/social welfare weights
3 Measurement of income distribution
Third piece can be well measured using tax data, even for high
incomes (Piketty and Saez 2004)
Saez 2004: Long-Run Evidence
Compares top 1% relative to the bottom 99%
Bottom 99% real income increases up to early 1970s and stagnates

since then
Top 1% increases slowly up to the early 1980s and then increases

dramatically up to year 2000.
Corresponds to the decrease in MTRs
Pattern exempli…es general theme of this literature: large responses

for top earners, no response for rest of the population
Bottom 99% Tax Units
40% $40,000
35% $35,000
30% $30,000
Marginal Tax Rate
25% $25,000
20% $20,000
15% $15,000
10% $10,000
5% Marginal Tax Rate

5% Average Income $5,000
0% $0
1970
1972
1974
1976
1978
1980
1984
1986
1988
1990
2000
1982
1998
1960
1962
1964
1968
1992
1994
1996
1966
Source: Saez 2004
Top 1% Tax Units
80% $800,000
70% $700,000
60% $600,000
Marginal Tax Rate
50% $500,000
40% $400,000
30% $300,000
20% $200,000
10% Marginal Tax Rate

5% Average Income $100,000
0% $0
1970
1972
1974
1976
1978
1980
1984
1986
1988
1990
2000
1982
1998
1960
1962
1964
1968
1992
1994
1996
1966
Source: Saez 2004
Part 6: Social Insurance
Harvard University
Fall 2012
Public Economics Lectures () Part 6: Social Insurance 1 / 178

Outline
1 Motivations for Social Insurance
2 Unemployment Insurance
3 Workers’Compensation
4 Disability Insurance
5 Health Insurance

De…nition of Social Insurance
Transfers based on events such as unemployment, disability, or age
Contrasts with welfare: means-tested transfers
SI is the biggest and most rapidly growing part of government

expenditure today

Growth of Social Insurance in the U.S.
Social
Health Security Income
0.4% 3.6% Security
5%
Health
National 9.4%
Defense
20.7% Social
Other Security
21.6% 20.7%
National Defense
69.4% Other Income
34.8% Security
14.5%
1953 2008
Source: Office of Management and Budget, historical tables, government outlays by function

Social Insurance Spending, 1996
% of Central % of Total
Government Government
% of GDP Expenditures Expenditures
Sweden 32.47% 86.60% 49.58%
Germany 28.05% 82.91% 49.44%
Mexico 1.36% 8.82% 6.39%
Columbia 6.61% 43.33% N/A
United Kingdom 17.53% 43.13% 33.77%
United States 12.22% 59.76% 30.02%
Japan 2.50% 19.44% 16%
Czech Republic 11.89% 38.90% 25.75%
Source: Krueger and Meyer 2002

Unemployment Benefit Systems in Developed Countries
120
100
Net replacement rate (%)
80
60
40
20
0
0 5 10 15 20 25 30 35 40 45 50 55 60
Time (months)
Belgium Hungary Spain Sweden USA

Source: OECD Benefits and Wages 2002

Main Questions in Social Insurance
1 Why have social (as opposed to private, or any) insurance?
2 What type of SI system maximizes social welfare?
Tradeo¤ between two forces:

Bene…ts – reducing risk (‡uctuations in consumption)
Distortion – changes in incentives for workers and …rms –> ine¢ cient
behavior and DWL
Generate new distortions as you …x the problem you set out to solve
–> second-best solution
Identify optimal policy by combining theoretical models of social

insurance with empirical evidence on program e¤ects

References
1 Institutional details: see handout posted on course website
2 Expected utility theory: See MWG or other graduate texts
3 Survival analysis: Kiefer (1988 JEL)
4 Surveys: Krueger and Meyer (2002) Handbook chapter
5 Chetty and Finkelstein (2012) Handbook chapter

Why have social insurance?
Motivation for insurance: reduction in risk for risk-averse individuals
Unemp Ins: risk of involuntary unemployment

Workers’comp and DI: risk of injuries/disabilities
Social Security annuity: risk of living too long
But why is government intervention needed to provide this

insurance?
Possible sources of market failure here:
1 Informational problems (adverse selection)

2 Individual optimization failures (myopia/improper planning)
3 Macroeconomic shocks

Adverse Selection as a Motivation for SI
Key paper: Rothschild and Stiglitz (1976); see MWG Ch. 13 for a
good review
More recent “su¢ cient statistic” version that can be connected to

data: Einav, Finkelstein, and Cullen (QJE 2010)
Consider an environment with asymmetric information, e.g.

individuals know risk of losing job but insurer does not
Main result: can lead to market failure where no equilibrium supports

provision of insurance
Government intervention through mandated insurance can increase

welfare

Rothschild-Stiglitz model
Economy with two types, low-risk (L) and high-risk (H)
A fraction f of the individuals are high-risk
Type L has a chance pL of becoming unemployed in a given year
Type H has a chance pH > pL of becoming unemployed.
In good state (state 1), income is E1 for both types; in bad state,
income is E2 < E1 .

Rothschild-Stiglitz: Key Assumptions
1 Static model: individuals arrive in the period either employed or

unemployed; no savings/dynamics.
2 No moral hazard: agents choose insurance contract but make no

choices after signing a contract.
3 Insurance market is perfectly competitive, so …rms earn zero pro…ts

in equilibrium.

Rothschild-Stiglitz: Contracts
An insurance contract is described by a vector α = (α1 , α2 )
Consumption in the two states: (E1 α1 , E2 + α2 )
Type i’s expected utility is
Vi ( α ) = ( 1 p i ) u ( E1 α 1 ) + p i u ( E2 + α 2 )
Any contract that earns non-negative pro…ts is feasible
Zero-pro…t condition ) …rms price insurance s.t.

1 p
α2 = α1
p
where p is risk rate of those who purchase contract.

Rothschild-Stiglitz: Equilibrium
De…nition
An equilibrium is de…ned by a set of insurance contracts such that
(1) individuals optimize: both types cannot …nd a better contract than the
ones they chose
(2) …rms optimize: all …rms earn zero pro…ts
Two types of equilibrium:
1 Pooling: both types are o¤ered the same contract α.
2 Separating: high-risk types choose a contract αH while low-risk types

choose a di¤erent contract αL .

Rothschild-Stiglitz: First Best Solution
In …rst best, insurer can distinguish types (perfect information)
In this case, equilibrium is separating
1 pi
Plugging in α2 = pi α1 , each type solves
1 pi
max(1 pi ) u ( w α1 ) + pi u (w + α1 ).
α1 pi
Solution
1 pi
Set MRS12 = pi , i.e. u 0 (c1 ) = u 0 (c2 ), i.e. full insurance
Both types are perfectly insured: earn their expected income

(1 pi )w regardless of the state.
Rothschild-Stiglitz: Second Best Problem
Firms cannot distinguish types in practice, because they cannot

determine true layo¤ risks, illness history, etc.
With contracts above, all the high risk types buy the low risk’s
contract and insurer goes out of business
Hence optimal contracts di¤er when information is asymmetric

Rothschild-Stiglitz: Second Best Solution
Result #1: no pooling equilibrium exists
If H and L types are pooled in a contract α,low-risk types lose money

in expectation.
1 p
Zero-pro…t condition requires α2 = p α1 but p > pL .
Low-risk type gets fewer dollars in state 2 than he should if the

insurance were fair for him.
Creates an opportunity for a new insurer to enter and “pick o¤” low
risk types by o¤ering slightly less insurance at a better price: higher
c1 , lower c2
Only low risk types switch, because they value c1 more.

Result #2: in a separating eq, Type H obtains full insurance and
Type L is under-insured
Intuition: in any sep. eq., both types are getting actuarially fair
insurance because of the zero-pro…ts condition
For H, no cost to …rm in providing full ins. (worst that can happen is
that L will join the pool, raising pro…ts)
But for L, full ins. would create an incentive for H to buy this
(cheaper) policy, forcing …rm into negative pro…ts
Incentive constraints always bind downward – “no distortion at the

top” result in standard asymmetric info. models
In eq., L gets as much ins as possible without inducing H to deviate

and pretend to be low-risk
Rothschild-Stiglitz: Gains from Government Mandate
There can be gains from government intervention through mandated

insurance
Consider an example where
E1 = 100, E2 = 0
p 1 3
u (c ) = c, pL = , pH = , f = 10%
4 4
In candidate separating eq., type H gets perfect insurance:
r
1
EUH = u (100(1 pH )) = 100 =5
4

Type L gets as much ins. as possible without making H want to

deviate at actuarially fair rate for L:
s
q
1 3 1 pL L
5= 100 αL1 + α1
4 4 pL
Solving gives αL1 = $3.85, αL2 = $11.55 – nowhere near full insurance
for low risk type.
Note that expected utility for low risk type is

3p 1p
EUL = 100 3.85 + 3 3.85 = 8.2.
4 4

Now suppose govt. comes in and mandates pooled insurance at

actuarial rate. Everyone gets an income of
9 3 1 1 7
( + )100 = 100 = 70.
10 4 10 4 10
H bene…ts from this: now pooling with less risky people

p
But L bene…ts too! Expected utility is 70 > 8.2

Because there are relatively few high risk types, L types bene…t from
pooling with them and getting full insurance coverage.
Note: pooled contract of 70 could be o¤ered by a private …rm,

destroying separating eq. proposed above
Hence there is actually no equilibrium in this example

Adverse Selection as a Motivation for SI
More generally, consider an economy in which people di¤er in their

risks of becoming unemployed
Adverse selection can destabilize the market:
Firm provides UI but lowest-risk (tenured people) drop out ) rates

have to rise
But then even moderate-risk types opt out ) rates rise further, more
drop out, ...
Could cause unraveling to the point where virtually no one is insured by
private market
UI program that pools everyone can lead to (ex-ante) welfare
improvements
What tool does the govt. have that private sector does not? Ability
to mandate
Adverse Selection: Empirical Evidence
Empirical evidence shows that adverse selection is a real source of

market failures in practice
Standard test: “positive correlation” property in equilibrium

(Chiappori and Salanie 2000)
Are those who buy more insurance more likely to …le claims?
Could be driven by both moral hazard + AS but not in certain contexts

such as death
Example: Finkelstein and Poterba (2004): adverse selection in U.K.

annuity market.
Annuities = ins. against the risk of living too long.

Finkelstein and Poterba 2004
Study two types of annuity markets: compulsory vs. voluntary.
Examine two features of annuity contracts
degree of backloading (in‡ation indexing and escalation of payments

over time)
payments to estate in event of death (guarantees and capital
protection).
Test for positive correlation in two ways
1 In eq., those who purchase backloaded annuities have lower mortality

rates
2 In eq., those who purchase annuities with payment to estate have
higher mortality rates
Both e¤ects should be stronger in voluntary markets

Limitations of Positive Correlation Test
1 Does not account for other dimensions of heterogeneity that may
confound the correlation
Literature on “advantageous selection” (e.g., Finkelstein and McGarry

2006)
2 Correlation does not clearly map into parameters that control welfare
costs of selection
Einav, Finkelstein, and Cullen (2010) develop “cost curve” tests that
map to measures of welfare costs
3 Only applicable in markets that exist, i.e. those that have not totally
unravelled
Hendren (2012) uses subjective expectations data to bound welfare

costs in markets that have unraveled
Individual Optimization Failures as a Motivation for SI
Given adverse selection, expect individuals to “self-insure” against

temp. shocks by building up savings
With such bu¤er stocks, still no need for large social safety nets to
insure against temporary shocks such as unemployment
In practice, individuals appear to be very liquidity constrained when

hit by shocks: median job loser has <$200 in assets
Suggests 1st Welfare thm also does not hold due to individual failures
to optimize
Individuals may misperceive the probability of a layo¤
Firms may not be able to debias people in equilibrium, leading to role

for govt. (Spinnewijn 2009)
Aggregate Shocks as a Motivation for SI
Private ins. (cross-sectional pooling) relies on idiosyncratic risks so

those who are well o¤ can pay those who are poor
Government is the only entity able to coordinate risk-sharing across

di¤erent groups that are all a¤ected by negative shocks
Inter-generational risk sharing required if everyone is poor at the same

time
Particularly relevant for UI and maybe social security
Less so for health-related shocks

Optimal Social Insurance
Now turn to question of optimal design of SI policies
Take as given that market provides no insurance for some reason
In the simple Rothschild-Stiglitz model, perfect insurance is optimal
But this abstracts from moral hazard
Individuals will not work if they have perfect unemp insurance
Must take this distortion into account to …nd optimal level of social
insurance

Unemployment Insurance
Potential bene…ts
1 Smoother path of consumption
2 Better job matches
Potential distortions
1 Less job search, higher unemployment rate
2 Workers’preferences distorted toward unstable jobs
3 Shirking on the job

Optimal UI: Outline
1 Optimal level of UI bene…ts ignoring …rm responses [Baily-Chetty

model]
Theory applies to all income security programs discussed later
2 Distortions to …rms’layo¤ decisions due to imperfect exp rating

[Feldstein model]
3 Other issues: Post-unemployment outcomes, general equilibrium

e¤ects

Replacement Rate
Common measure of program’s size is its “replacement rate”
net bene…t
r=
net wage
UI reduces agents’e¤ective wage rate from …nding a new job to
w (1 r )
Feldstein (1978): UI makes e¤ective wages very low because of

interaction with tax system:
(0.5 )w
1970: No tax ) r = (1 .18 .05 .07 )w = 72%
Incentives worse for some subgroups: secondary income earner faces

MTR of 50% ) r = 1.3
Today, federal income taxes paid on UI bene…ts, so rep. rate is

50-60%
Unemployment Insurance Benefit Schedule in Michigan, 2009
362 400
300
Weekly Benefits ($)
200
100
$0 if highest total quarterly earnings < $2,871 ($220/wk)

0
0 200 400 600 800 1000

Weekly Wage Earnings in Highest Quarter ($)
Source: Michigan Department of Energy, Labor, and Economic Growth 2009

Baily-Chetty model
Canonical analysis of optimal level of UI bene…ts: Baily (1978)
Shows that the optimal bene…t level can be expressed as a fn of a

small set of parameters in a static model.
Once viewed as being of limited practical relevance because of strong

assumptions
Chetty (2006) shows formula actually applies with arbitrary choice

variables and constraints.
Parameters identi…ed by Baily are su¢ cient statistics for welfare

analysis ) robust yet simple guide for optimal policy.

Baily-Chetty model: Assumptions
1 Fixed wages – no GE e¤ects
2 No distortions to …rm behavior (temporary layo¤s); implicitly assume

perfect experience rating
3 No externalities such as spillovers to search

Baily-Chetty model: Setup
Static model with two states: high (employed) and low (unemployed)
Let wh denote the individual’s income in the high state and wl < wh
income in the low state
Let A denote wealth, ch consumption in the high state, and cl

consumption in the low state
Agent is initially unemployed. Controls probability of being in the bad

state by exerting search e¤ort e at a cost ψ(e )
Choose units of e so that the probability of being in the high state is

given by p (e ) = e

Baily-Chetty model: Setup
UI system that pays constant bene…t b to unemployed agents
Bene…ts …nanced by lump sum tax t (b ) in the high state
Govt’s balanced-budget constraint:
e t (b ) = (1 e) b
Let u (c ) denote utility over consumption (strictly concave)
Agent’s expected utility is
eu (A + wh t (b )) + (1 e )u (A + wl + b ) ψ (e )

First Best Problem
In …rst best, there is no moral hazard problem
To solve for FB, suppose government chooses b and e jointly to

maximize agent’s welfare:
max e (A + wh t ) + (1 e )u (A + wl + b ) ψ (e )
b,e
1 e
s.t. t = b
e
Solution to this problem is u 0 (ce ) = u 0 (cu ) ) full insurance

Second Best Problem
In second best, cannot eliminate moral hazard problem because e¤ort

is unobserved by govt.
Problem: Agents only consider private marginal costs and bene…ts

when choosing e
Social marginal product of work is wh wl
Private marginal product is wh wl b t
Agents therefore search too little from a social perspective, leading to

e¢ ciency losses

Second Best Problem
Agents maximize expected utility, taking b and t (b ) as given
max eu (A + wh t ) + (1 e )u (A + wl + b ) ψ (e )
e
Let indirect expected utility be denoted by V (b, t )
Government’s problem is to maximize agent’s expected utility, taking

into account agent’s behavioral responses:
max V (b, t )
b,t
s.t. e (b )t = (1 e (b ))b

Second Best Problem
Problem
Optimal Social Insurance
max V (b, t (b ))
b
s.t. e (b )t (b ) = (1 e (b ))b
e (b ) = arg max e u (A + wh t ) + (1 e ) u (A + wl + b ) ψ (e )
e
Formally equivalent to an optimal Ramsey tax problem with

state-contingent taxes

Two Approaches to Optimal Social Insurance
1 Structural: specify complete models of economic behavior and

estimate the primitives
Identify b as a fn. of discount rates, nature of borrowing constraints,

informal ins. arrangements.
2 Su¢ cient Statistic: derive formulas for b as a fn. of reduced-form

elasticities
Baily-Chetty formula is one example

Chetty (2006) Su¢ cient Statistic Formula
At an interior optimum, the optimal bene…t rate must satisfy
dV /db (b ) = 0
To calculate this derivative, write V (b ) as
V (b ) = max eu (A + wh t (b )) + (1 e )u (A + wl + b ) ψ (e )
e
Since fn has been optimized over e, Envelope Thm. implies:
dV (b ) dt 0
= (1 e ) u 0 ( cl ) eu (ch )
db db
∂e
Can ignore ∂b terms because of agent optimization

Kaplan 2009
Exploiting f.o.c.’s from agent optimization particularly useful in more

complex models
Kaplan (2009): unemployed youth move back in with their parents.
How does this a¤ect optimal UI?
Kaplan takes a structural approach and estimates a dynamic model of

the decision to move back home

Su¢ cient Statistic Approach to Kaplan 2009
Suppose moving home raises consumption by H and has a cost g (H ):
V (b ) = max eu (A + wh t (b ))
e,H
+(1 e )[u (A + wl + b + H ) g (H )] ψ (e )
Variable H drops out, as did e, because of agent optimization
dV (b )
Formula derived for db is una¤ected by ability to move home:
dV (b ) dt 0
= (1 e ) u 0 ( cl ) eu (ch )
db db
where cl is measured in the data as including home consumption (H)

Chetty (2006) Su¢ cient Statistic Formula
The government’s UI budget constraint implies

dt 1 e
b de 1 e ε
= 2
= (1 + 1 e,b )
db e e db e e
dV (b ) ε
=) = (1 e )fu 0 (cl ) (1 + 1 e,b )u 0 (ch )g
db e
Setting dV (b )/db = 0 yields the optimality condition
u 0 ( cl ) u 0 ( c h ) ε
= 1 e,b
u 0 ( ch ) e
LHS: bene…t of transferring $1 from high to low state
RHS: cost of transferring $1 due to behavioral responses

Baily-Chetty Formula
u 0 ( cl ) u 0 ( ch ) ε
0
= 1 e,b
u ( ch ) e
This equation provides an exact formula for the optimal bene…t rate
u 0 (c l ) u 0 (c h )
Implementation requires identi…cation of u 0 (c h )
u 0 (c l ) u 0 (c h )
Three ways to identify u 0 (c h )
empirically
1 Baily (1978), Gruber (1997), Chetty (2006): cons-based approach
2 Shimer and Werning (2007): reservation wages
3 Chetty (2008): moral hazard vs liquidity

Consumption-Based Formula
Write marginal utility gap using a Taylor expansion
u 0 ( cl ) u 0 ( ch ) u 00 (ch )(cl ch )
u 00 (c )c
De…ning coe¢ cient of relative risk aversion γ = u 0 (c )
, we can write
u 0 ( c l ) u 0 ( ch ) u 00 ∆c
ch (1)
u 0 ( ch ) u0 c
∆c
= γ
c
Gap in marginal utilities is a function of curvature of utility (risk
aversion) and consumption drop from high to low states

Theorem
The optimal unemployment bene…t level b satis…es
∆c ε1 e,b
γ (b )
c e
where
∆c ch cl
= = consumption drop during unemployment
c ch
u 00 (ch )
γ = ch = coe¢ cient of relative risk aversion
u 0 ( ch )
d log 1 e
ε1 e,b = = elast. of probability of unemp. w.r.t. bene…ts
d log b

∆c ε1 e,b
γ (b )
c e
Intuition for formula: LHS is marginal social bene…t of UI, RHS is

marginal social cost of UI
Extends to model where agent chooses N other behaviors and faces M

other constraints, subject to some regularity conditions (Chetty 2006).
Envelope conditions used above still hold
Empirical work on UI provides estimates of the three key parameters

(γ, ∆c
c , ε).

Empirical Estimates: Duration Elasticity
Early literature used cross-sectional variation in replacement rates
Problem: comparisons of high and low wage earners confounded by

other factors.
Modern studies use exogenous variation from policy changes (e.g.

Meyer 1990)

Weekly
Benefit
Amount
WBAAmax
After Benefit Increase
WBABmax
Before Benefit Increase
WBAmin
E1 E2 E3 Previous Earnings
Low Earnings Group High Earnings Group
Source: Krueger and Meyer 2002

Hazard Models
De…ne hazard rate ht = number that …nd a job at time t divided by

number unemployed at time t
This is an estimate of the probability of …nding a job at time t

conditional on being unemployed for at least t weeks
Standard speci…cation of hazard model: Cox “proportional hazards”
ht = αt exp( βX )
Here αt is the non-parametric “baseline” hazard rate in each period t

and X is a set of covariates
Semi-parametric speci…cation – allow hazards to vary freely across

weeks and only identify coe¢ cients o¤ of variation across spells
Hazard Models
Useful to rewrite expression as:
log ht = log αt + βX
Key assumption: e¤ect of covariates proportional across all weeks

d log ht d log hs
=β= 8t, s
dX dX
If a change in a covariate doubles hazard in week 1, it is forced to
double hazard in week 2 as well
Restrictive but a good starting point; can be relaxed by allowing for

time varying covariates Xt

Meyer 1990
Meyer includes log UI bene…t level as a covariate:
log ht = log αt + β1 log b + β2 X
In this speci…cation,
d log ht
= β1 = εht ,b
d log b
Note: in exponential survival (constant-hazard) models,

εht ,b = ε1 e,b
Meyer estimates εht ,b = 0.9 using administrative data for UI

claimants
Subsequent studies get smaller estimates; consensus: εht ,b = 0.5

(Krueger and Meyer 2002)
Consumption Smoothing Bene…ts of UI
Gruber (1997) takes the Baily formula to the data by estimating

consumption smoothing response.
Same methodology as Meyer
Uses cross-state and time variation and uses drop in food consumption
as the LHS variable.
Data: PSID food consumption

Gruber 1997
Gruber estimates
∆c b
= β1 + β2
c w
Finds β1 = 0.24, β2 = 0.28
Without UI, cons drop would be about 24%
Mean drop with current bene…t level (b = 0.5) is about 10%
Implies a 10 pp increase in UI replacement rate causes 2.8 pp

reduction in cons. drop
Suggests that ins. markets are not perfect and UI does play a
consumption smoothing role, but estimates are imprecise
Key area for future work: admin. consumption data

Consumption Smoothing Bene…ts of UI
What is substituting for/getting crowded out by UI?
Cullen and Gruber (2000) emphasize spousal labor supply
Study wives of unemployed husbands
Examine wives’labor supply as a fn of level of husbands’UI bene…ts
For a $100/wk increase in UI bene…t, wives work 22 hrs less per month
In the absence of UI, wives would work 30% more during the spell than
they do now
Engen and Gruber (1995) document that higher UI bene…ts lower

ex-ante savings, another crowdout channel

Calibrating the Model
Gruber calibrates Baily’s model using his and Meyer’s estimates:

∆c ε1 e,b
γ
c e
b ε1 e,b
γ ( β1 + β2 ) =
w e
Solving for the optimal replacement rate yields:
b ε /e 1 β1
= 1 e,b ( )
w β2 γ β2
Plugging in ε1 e,b = .43 as in Gruber (1997) and e = .95 (5%

unemployment rate) yields:
b .43/.95 1 ( .24)
= ( )
w .28 γ .28

Calibrating the Model
b
Results: w varies considerably with γ
γ 1 2 3 4 5 10
b
w 0 0.05 0.31 0.45 0.53 0.7
Gruber: introspection and existing evidence suggests γ < 2
Implies optimal bene…t level is much lower than observed

Measurement of Risk Preferences
Parameter that is most poorly identi…ed: γ
Barseghyan et al. (2012) and Einav et al. (2012) estimate risk

preferences across di¤erent domains
Do individuals who choose more insurance for health also choose more
insurance for long-term disability?
Find positive correlation in risk preferences but substantial

heterogeneity across domains
Suggests that appropriate value of γ is highly context-dependent
Could be due to behavioral factors and framing but also due to

neoclassical factors

Chetty and Szeidl (2007): Consumption Commitments
Standard expected utility model: one composite consumption good c
Composite commodity assumes that people can cut back on all

consumption goods at all times freely.
E.g. when unemployed, cut consumption of food, housing, cars,

furniture, etc.
In practice, di¢ cult to adjust many elements of consumption in short

run because of …xed adjustment costs

Homeowners’Consumption around Unemployment Shocks
.05
Food and Housing Growth Rates
.025
0
-.025
-.05
-.075
-4 -2 0 2 4
Year relative to unemployment
Housing (Home Value) Food

Source: Chetty and Szeidl 2007

Renters’Consumption around Unemployment Shocks
.05
Food and Housing Growth Rates
.025
0
-.025
-.05
-.075
-4 -2 0 2 4
Year relative to unemployment
Housing (Rent) Food

Source: Chetty and Szeidl 2007

Commitments and Risk Aversion
How do commitments a¤ect risk aversion?
Utility over two goods, food and housing:
U (f , h ) = u (f ) + v (h ).
Adjusting h requires payment of a …xed cost k
Agent follows an (S, s ) policy

Commitments Model: Implications for UI
Commitments amplify risk aversion
Ex: 50% food, 50% housing
Suppose unemployed agent forced to cut expenditure by 10%
Then have to cut food cons by 20%, leading to larger welfare loss
Model of commitments suggests that γ might actually exceed 4 for

unemployment shocks
γ 1 2 3 4 5 10
b
w 0 0.05 0.31 0.45 0.53 0.7
Problem: γ hard to estimate precisely by context

Alternative Formulas for Optimal UI
Since γ and ∆cc are hard to identify, recent work has sought
alternative ways of calculating optimal bene…t.
Two approaches
1 Moral hazard vs. liquidity (Chetty 2008)
2 Reservation wage response (Shimer Werning 2007)
Note that any formula is only one representation of optimal bene…t

Chetty 2008: Moral Hazard vs. Liquidity
Discrete time dynamic search model
Individual lives for T periods
Interest rate and discount rate equal to 0
Individual loses job in period t = 0
Let u (ct ) denote ‡ow utility over cons.
Dynamic budget constraint:
At +1 = At + yt ct
Asset limit: At L

Chetty 2008: Baseline Assumptions
1 Assets prior to job loss exogenous
2 No heterogeneity
3 Fixed wages: choose only search intensity, not reservation wage
Each of these is relaxed in paper, so model nests search models used

in structural literature (e.g. Wolpin 1987)

Chetty 2008: Job Search Technology
If unemployed in period t, worker …rst chooses search intensity st
Finds a job that begins immediately in period t with probability st
If job found, consumes cte . Jobs are permanent, pay wage wt τ.

Chetty 2008: Job Search Technology
If no job found: receives bene…t bt , consumes ctu , enters t + 1

unemployed
Cost of job search: ψ(st )
cte = ct+1e = …
st
Period t
st+1 ct+1e
1-st
ctu
1-st+1
ct+1u

Chetty 2008: Value Functions
Value function for agent who …nds a job in period t:
Vt (At ) = max u (At At + 1 + w τ ) + Vt + 1 ( At + 1 )

A t +1 L
Value function for agent who does not …nd a job in period t:
Ut (At ) = max u (At At +1 + bt ) + Jt +1 (At +1 )

A t +1 L
where Jt +1 (At +1 ) is value of entering next period unemployed.
Agent chooses st to maximize expected utility
Jt (At ) = max st Vt (At ) + (1 s t ) U t ( At ) ψ(st )

st

Chetty 2008: Optimal Search Behavior
First order condition for optimal search intensity:
ψ0 (st ) = Vt (At ) U t ( At )
Intuitively, st is chosen to equate the marginal cost of search e¤ort

with the marginal value of search e¤ort.
E¤ect of bene…ts on durations:
∂st /∂bt = u 0 (ctu )/ψ00 (st )

Chetty 2008: Moral Hazard vs. Liquidity Decomposition
Bene…t e¤ect can be decomposed into two terms:
∂st /∂At = fu 0 (cte ) u 0 (ctu )g/ψ00 (st ) < 0

∂st /∂wt = u 0 (cte )/ψ00 (st ) > 0
) ∂st /∂bt = ∂st /∂At ∂st /∂wt
∂st /∂At is “liquidity e¤ect”
∂st /∂wt is “moral hazard” or price e¤ect
Liquidity and total bene…t e¤ects smaller for agents with better
consumption smoothing capacity

Source: Chetty 2008
Chetty 2008: Formula for Optimal UI
∂st /∂At = fu 0 (cte ) u 0 (ctu )g/ψ00 (st ) 0

∂st /∂wt = u 0 (cte )/ψ00 (st ) > 0
∂st /∂At LIQ u 0 (ctu ) u 0 (cte )
) = =
∂st /∂wt MH u 0 (cte )
Can show that the Baily formula holds in this model:

u 0 (ctu ) u 0 (cte ) ε
0 e = 1 e,b
u ( ct ) e
Combining yields formula that depends solely on duration elasticities:
∂st /∂At ε1 e,b
=
∂st /∂bt ∂st /∂At e
ε1 e,A ε1 e,b
=
ε1 e,b Ab ε1 e,A e

Intuition for Moral Hazard vs. Liquidity Formula
Formula is a “revealed preference” approach to valuing insurance
Infer value of UI to agent by observing what he would do if money

given as a cash-grant without distorted incentives
If agent would not use money to extend duration, infer that only takes
longer because of price subsidy (moral hazard)
But if he uses cash grant to extend duration, indicates that UI

facilitates a choice he would make if markets were complete
Same strategy can be used in valuing other types of insurance
Make inferences from agent’s choices instead of directly computing

costs and bene…ts of the policy
Key assumption: perfect agent optimization

Moral Hazard vs. Liquidity: Evidence
Two empirical strategies
1 Divide agents into liquidity constrained and unconstrained groups and

estimate e¤ect of bene…ts on durations using changes in UI laws.
2 Look at lump-sum severance payments to estimate liquidity e¤ect.

TABLE 1
Summary Statistics by Wealth Quartile for SIPP Sample
Net Liquid Wealth Quartile

1 2 3 4
(< -$1,115) (-$1,115-$128) ($128-$13,430) (>$13,430)
Median Liq. Wealth $466 $0 $4,273 $53,009

Median Debt $5,659 $0 $353 $835
Median Home Equity $2,510 $0 $11,584 $48,900
Median Annual Wage $17,188 $14,374 $18,573 $23,866
Mean Years of Education 12.21 11.23 12.17 13.12

Mean Age 35.48 35.18 36.64 41.74
Fraction Renters 0.43 0.61 0.35 0.16

Fraction Married 0.64 0.59 0.60 0.63
All monetary variables in real 1990 dollars
Source: Chetty 2008

Figure 3a
Effect of UI Benefits on Durations: Lowest Quartile of Net Wealth
1 .8
Fraction Unemployed
Mean rep. rate = .53

.6

.4 .2
Wilcoxon Test for Equality: p = 0.01
0 10 20 30 40 50
Weeks Unemployed
Avg. UI benefit below mean Avg. UI benefit above mean
Source: Chetty 2008

Figure 3b
Effect of UI Benefits on Durations: Second Quartile of Net Wealth
1 .8
Fraction Unemployed

.6

.4 .2
0 10 20 30 40 50
Weeks Unemployed
Source: Chetty 2008

Figure 3c
Effect of UI Benefits on Durations: Third Quartile of Net Wealth
1 .8
Fraction Unemployed

.6

.4 .2
0 10 20 30 40 50
Weeks Unemployed
Source: Chetty 2008

Figure 3d
Effect of UI Benefits on Durations: Highest Quartile of Net Wealth
1 .8
Fraction Unemployed

.6

.4

.2
0 10 20 30 40 50
Weeks Unemployed
Source: Chetty 2008

TABLE 2
Effect of UI Benefits: Cox Hazard Model Estimates
(1) (2) (3) (4) (5)
Pooled Stratified Stratified with Full Controls
Full cntrls No cntrls Avg WBA Max WBA Ind. WBA
log UI ben -0.527
(0.267)
Q1 x log UI ben -0.721 -0.978 -0.727 -0.642
(0.304) (0.398) (0.302) (0.241)
Q2 x log UI ben -0.699 -0.725 -0.388 -0.765
(0.484) (0.420) (0.303) (0.219)
Q3 x log UI ben -0.368 -0.476 -0.091 -0.561
(0.309) (0.358) (0.370) (0.156)
Q4 x log UI ben 0.234 0.103 0.304 0.016
(0.369) (0.470) (0.339) (0.259)
Q1=Q4 p-val 0.039 0.013 0.001 0.090

Q1+Q2=Q3+Q4 p-val 0.012 0.008 0.002 0.062
Number of Spells 4529 4337 4054 4054 4054
Source: Chetty 2008

TABLE 3
Summary Statistics for Mathematica Data
Pooled No Severance Severance

(0.83) (0.17)
Percent dropouts 14% 15% 6%
Percent college grads 17% 13% 34%
Percent married 58% 56% 68%
Mean age 36.2 35.2 40.6
Median pre-unemp annual wage $20,848 $19,347 $30,693

Median job tenure (years) 1.9 1.5 4.8
Source: Chetty 2008

Figure 5
Effect of Severance Pay on Durations
1 .9
Fraction Unemployed
.6 .7 .5.8
0 5 10 15 20
Weeks Unemployed
No Severance Received Severance
Source: Chetty 2008

Figure 6a
Effect of Severance Pay on Durations: Below Median Net Wealth
1
Fraction Unemployed
.6 .4 .8
0 5 10 15 20
Weeks Unemployed
Source: Chetty 2008

Figure 6b
Effect of Severance Pay on Durations: Above Median Net Wealth
1
Fraction Unemployed
.6 .4 .8
0 5 10 15 20
Weeks Unemployed
Source: Chetty 2008

TABLE 4
Effect of Severance Pay: Cox Hazard Model Estimates
Pooled By Liquid Wealth By Sev. Amt.
Severance Pay -0.233

(0.071)
(Netliq < Median) x Sev Pay -0.457
(0.099)
(Netliq > Median) x Sev Pay -0.088
(0.081)
(Tenure < Median) x Sev Pay -0.143
(0.055)
(Tenure > Median) x Sev Pay -0.340

(0.119)
Equality of coeffs p-val <0.01 0.03
N=2428; all specs. include full controls.
Source: Chetty 2008

Chetty 2008: Implications for Optimal UI
Plug reduced-form estimates of de/dA and de/db into formula to

calculate dW /db
Welfare gain from raising bene…t level by 10% from current level in
U.S. (50% wage replacement) is $5.9 bil = 0.05% of GDP
Small but positive
In structural models calibrated to match su¢ cient statistics, dW /db

falls rapidly with b
Small dW /db suggests we are currently near optimal bene…t level

Card, Chetty, and Weber 2007
Use discontinuities in Austria’s unemployment bene…t system to

estimate liquidity e¤ects
Severance payment is made by …rms out of their own funds
Formula for sev. pay amount for all non-construction workers:

Severance Amt.
(months of pay)
0
0 36 60
Job Tenure

Figure 3
Frequency of Layoffs by Job Tenure
40000
30000
Number of Layoffs
20000
10000
0
12 18 24 30 36 42 48 54 60
Previous Job Tenure (Months)
Source: Card, Chetty, and Weber 2007

Age by Job Tenure
34
33
Mean Age
32
31
30
12 18 24 30 36 42 48 54 60

Figure 4
Selection on Observables
.95
Mean Predicted Hazard Ratios
.9
.85
.8
12 18 24 30 36 42 48 54 60

Figure 5a
Mean Nonemployment Duration (days) Effect of Severance Pay on Nonemployment Durations
165
160
155
150
145
12 18 24 30 36 42 48 54 60

TABLE 3a
Effects of Severance Pay and EB on Durations: Hazard Model Estimates
(1) (2) (3)

Restricted Restricted Full
Sample Sample Sample
Severance pay -0.122 -0.125

(0.019) (0.017)
Extended benefits -0.084 -0.093

(0.018) (0.016)
Sample size 512,767 512,767 650,922
NOTE--All specs are Cox hazard models that include cubic polynomials with
interactions with sevpay and/or extended benefit dummy.

Shimer and Werning 2007: Reservation-Wage Model
Reservation wage model: probability of …nding job (e) determined by

decision to accept or reject a wage o¤er, not search e¤ort
Wage o¤ers drawn from distribution w F (x )
Agent rejects o¤er if net wage w t is less than outside option b,

implying that probability of …nding a job is e = 1 F (b + t )
Agent’s expected value prior to job search:
W (b ) = (1 F (b + t ))E [u (w t )jw t > b ] + F (b + t )u (b )
Reservation wage prior to job search satis…es
u (w̄0 t ) = W (b )

Shimer and Werning 2007: Reservation-Wage Formula
Government’s problem is
max W (b ) = max u (w̄0 t ) = max w̄0 t
It follows that
dW d w̄0 dt
=
db db db
d w̄0 1 e 1
= (1 + ε1 e,b )
db e e

Shimer and Werning 2007: Reservation-Wage Formula
d w̄ 0
Implement formula using estimates of db reported by Feldstein and
Poterba (1984)
Find gains from raising UI bene…ts 5 times larger than Chetty (2008)
But reservation wage elasticity estimates questionable
Do greater bene…ts ! longer durations ! better outcomes later on?

No.
Ex: evidence from Austrian discontinuity (Card, Chetty, Weber 2007)
Note: all the formulas above take such match quality gains into
account via envelope conditions

Figure 5a
Mean Nonemployment Duration (days) Effect of Severance Pay on Nonemployment Durations
165
160
155
150
145
12 18 24 30 36 42 48 54 60

Figure 10a
Effect of Severance Pay on Subsequent Wages
0
-.02
Wage Growth
-.04
-.06
-.08
-.1
12 18 24 30 36 42 48 54 60

Figure 10b
Average Monthly Job Ending Hazard in Next Job Effect of Severance Pay on Subsequent Job Duration
.2
.15
.1
.05
0
-.05
12 18 24 30 36 42 48 54 60

Figure 9a
Effect of Benefit Extension on Nonemployment Durations
165
Mean Nonemployment Duration (days)
160
155
150
145
140
135
12 18 24 30 36 42 48 54 60
Months Employed in Past Five Years

Effect of Extended Benefits on Subsequent Wages
.1
.05
Wage Growth
0
-.05
-.1
12 18 24 30 36 42 48 54 60
Months Worked in Past Five Years

Effect of Extended Benefits on Subsequent Job Duration
Average Monthly Job Ending Hazard in Next Job
.05
0
-.05
-.1
-.15
12 18 24 30 36 42 48 54 60
Months Worked in Past Five Years

Spike at Bene…t Exhaustion
Most striking evidence for distortionary e¤ects of social insurance:

“spike” in hazard rate at bene…t exhaustion
Katz and Meyer (1990), Meyer (1990), ...
Traditional measure of hazard: exiting UI system
Preferred measure based on theory: …nding a job
The two could di¤er if workers transit o¤ of UI but are still jobless
Ex. may not go to pick up last unemployment check

Particularly important in European context, where you can remain
registered on UI inde…nitely

Time Until Benefits Lapse Empirical Hazard
.2
Unemployment Exit Hazard
.15
.1
.05
0
20 15 10 5 0
Weeks of Eligibility Left
Source: Meyer 1990

Job Finding vs. Unemployment Exit Hazards: 20 Week UI
.2 .15
Weekly Hazard Rate
.05 .1 0
0 10 20 30 40 50
Weeks Elapsed Since Job Loss
Job Finding Hazards Unemp Exit Hazards
Source: Card, Chetty, Weber 2007b (AER P&P)

Job Finding vs. Unemployment Exit Hazards: 30 Week UI
.2 .15
Weekly Hazard Rate
.05 .1 0
0 10 20 30 40 50
Job Finding Hazards Unemp Exit Hazards

Effect of Benefit Expiration on Hazard Rates
.1
Difference in Weekly Hazard UI20-UI30
.05
0
-.05
-.1
0 10 20 30 40 50
Unemployment Exit Hazards Job Finding Hazards

UI and Firm Behavior
Preceding discussion assumed perfect experience rating of UI
Firms’layo¤ incentives are not distorted
But in practice, UI is not perfectly experience rated
Feldstein (1976, 1978) shows:
Theoretically that imperfect experience rating e¤ect can raise rate of

temporary layo¤s
Empirically that this e¤ect is large in practice

Experience Rating in Washington, 2005
10
8
UI Tax Rate (%)
6
4
2
0
0 2 4 6 8 10
Benefit Ratio (100*UI Benefits Paid/Payroll)
Washington’
s UI Tax Schedule Perfect Experience Rating
Source: Washington State Joint Legislative Task Force on Unemployment Insurance Benefit Equity 2005

UI and Firm Behavior: Feldstein 1976 model
Firms o¤er workers stochastic contracts, with wage and probability of

temporary layo¤
Two states: high demand and low demand
In equilibrium, competitive …rms will o¤er contract that pays worker

his marginal product in expectation over two states at cheapest cost
to …rm
Firm pro…ts by laying o¤ workers with imperfect exp rating
Layo¤s generate …rst-order gain in pro…ts at a second-order cost from

added risk to worker
In an imperfectly experience-rated economy, …rms choose a positive

rate of layo¤s in low output state
Feldstein 1978: Empirical Results
First observation: more than half of …rms are above the max rate or
below the min rate
No marginal incentive for these …rms to reduce layo¤s.
Uses cross-state/time variation in UI bene…ts
10% increase in UI bene…ts causes a 7% increase in temp layo¤

unemployment
E¤ect is twice as large for union members as non-union, suggesting

worker-…rm coordination.

Topel 1983
Feldstein does not directly show that imperfect exp rating is to blame
for more temp layo¤s b/c not using variation in experience rating itself
Topel (1983) uses state/industry variation in …nancing of UI
Variation in tax rate on …rms from min/max thresholds for exp rating
Finds that imperfect subsidization accounts for 31% of all temp layo¤
unemployment, a very large e¤ect
See Krueger and Meyer (2002) for review of more recent studies,
which …nd similar results but smaller magnitudes

UI Savings Accounts
Alternative to UI transfer-based system (Feldstein and Altman 2007)
Instead of paying UI tax to government, pay into a UI savings account.

If unemployed, deplete this savings account according to current
bene…t schedule
If savings exhausted, government pays bene…t as in current system
(…nanced using a tax).
Idea: people internalize loss of money from staying unemp longer.
Reduces distortion from UI while providing bene…ts as in current

system.
But modelling this formally is di¢ cult: to internalize incentives at
retirement, people must be forward looking, but then no need to force
them to save.

Feldstein and Altman 2007
Address feasibility: How many people hit negative balance on UI

account and just go back to old system?
Simulate how UI savings accounts would evolve using actual earnings

histories from PSID.
Calculations imply that only 1/3 of spells will occur with negative
balances, so most people still have good incentives while unemployed.
Total tax payments are less than half what they are in current system.
In their simulation, bene…ts are identical; only question is how costs

change.

Feldstein and Altman 2007
Calculation of changes in present value of lifetime wealth from switch

to UISA by income quintile:
Q1 Q2 Q3 Q4 Q5
Present Value Gain: -$95 +$22 -$67 +$94 +$468
Net PVG is positive
Without change in behavior, how is the pie larger?
Reason: discounting at 2% but earning 5.5% interest

Takeup
Mean takeup rate is very low – a major puzzle in this literature

(Currie 2004)
Why leave money on the table?
Andersen and Meyer (1997) show that after-tax UI replacement rate

a¤ects level of takeup.
So at least some seem to be optimizing at the margin.
Takeup low in many govt. programs. (UI, food stamps, EITC, etc.)
Possible explanations: myopia, stigma, hassle, lack of info.

Black, Smith, Berger, and Noel 2003
Experiment in KY where some UI claimants were randomly assigned

to receive re-employment services
E.g., assisted job search, employment counseling, job search workshops,

retraining programs
Treatment [N = 1236] required to receive services in order to get UI

bene…ts
Control [N = 745]: exempt from services

Black, Smith, Berger, and Noel 2003: Results
Treatment group exit UI system earlier, receiving 2.2 fewer weeks of

bene…ts on average
Most signi…cant increase in exits in wks 2-3, when noti…ed of

mandatory services

General Equilibrium: Acemoglu and Shimer 1999
UI can be e¢ ciency-enhancing in equilibrium.
Standard models focus only on distortionary costs, and assume that

total output always lower when UI is provided.
But this ignores potentially important GE e¤ect: more risky jobs

provided in eq. if workers are insured.
Provision of UI raises availability of risky jobs (e.g. tech jobs) and can
raise e¢ ciency in equilibrium
So if workers are risk averse, tradeo¤ may not be very hard – both
raise output and insure them better.

Dynamics: Path of UI Bene…ts
Classic reference is Shavell and Weiss (1979), who solved for optimal
path of bene…ts in a 3 period model.
Tradeo¤: upward sloping path ! more moral hazard but more

consumption-smoothing bene…ts.
Recent literature that is very active in this area: “new dynamic public
…nance” – optimal path of unemployment and disability programs.
Hopenhayn and Nicolini (1997) – numerical simulations for case where

govt can control consumption
Shimer and Werning (2008) – with perfect liquidity and CARA utility,
optimal bene…t path is ‡at

Optimal Insurance in Behavioral Models
We do not have a model consistent with the data that can explain
both savings behavior pre-unemployment and search behavior
post-unemployment
Evidence that unemployment is indeed costly and bene…ts can improve

welfare a lot for certain liquidity-constrained groups
Simple rational model cannot rationalize level of savings that people

have when they get unemployed
Interesting direction for future research: optimal SI with behavioral

considerations (see e.g., Spinnewijn 2009)

Workers Compensation
Insurance against injury at work
Covers both lost wages and medical bene…ts
Rationales for govt. intervention:
Market may fail due to adverse selection
Workers may be unaware of risks on the job
Litigation costs (origin of system in 1920s)
Substantial variation in bene…ts across states for di¤erent injuries

Maximum Indemnity Benefits in 2003
Type of permanent impairment

State Arm Hand Index finger Leg Foot Temporary Injury
(10 weeks)
California $108,445 $64,056 $4,440 $118,795 $49,256 $6,020

Hawaii 180,960 141,520 26,800 167,040 118,900 5,800
Illinois 301,323 190,838 40,176 276,213 155,684 10,044
Indiana 86,500 62,500 10,400 74,500 50,500 5,880
Michigan 175,657 140,395 24,814 140,395 105,786 6,530

Missouri 78,908 59,521 15,305 70,405 52,719 6,493
New Jersey 154,440 92,365 8,500 147,420 78,200 6,380
New York 124,800 97,600 18,400 115,200 82,000 4,000
Source: Gruber 2007

Theory of Workers’Compensation
Formally very similar to that of unemployment insurance
If prob of injury cannot be controlled, model same as Baily-Chetty
If prob of injury can be controlled, that distortion must be taken into

account in calculation
Leisure now includes bene…ts of having more time to heal
Similar formal theory, so literature is mostly empirical

Outline of Empirical Evidence
1 Monday e¤ects and impact on worker behavior
2 Firm side responses
3 E¤ect on equilibrium wage

Day of the Week E¤ect
Intertemporal distortions, moral hazard e¤ect of workers’comp.
Card & McCall (1994): test if weekend injuries lead to Monday e¤ect.
Look at uninsured workers, who should have bigger Monday e¤ect.
Find no di¤erence in e¤ect between insured and uninsured.
Other explantations:
Gaming system for more days o¤.
Pure reporting e¤ect if pain does not go away.
Suggests that incentives matter a lot.

E¤ects of Bene…ts on Injuries
Potential incentive e¤ects to look for on worker’s side:
Number of claims of injury
Duration of injuries
Meyer, Viscusi, and Durbin (1995):
Implement DD analysis for workers’comp durations
Find large e¤ects on duration using reforms in MI and KY

Firm Side Responses
Purchasing insurance leads to imperfect experience rating and moral

hazard
Self-insured …rms: stronger incentives to improve safety
Also, have incentive to ensure that workers return to work quickly
Krueger (1990): compares behavior of self-insured …rms with others
Finds self-insured have 10% shorter durations
But could be biased by selection

E¤ect on Equilibrium Wage
Workers’compensation is a mandated bene…t
When …rms hire, should adjust wage downwards if workers value

bene…t (Summers 1989)
Gruber-Krueger (1991) test this using changes in WC laws
85% of WC cost is shifted to workers, no signi…cant employment e¤ect
Fishback-Kantor (1995) study initial implementation of program
Find 100% shift to workers’wages
Both studies suggest that bene…ts valued close to cost

Directions for Further Research on WC
Decomposition into liquidity vs. moral hazard e¤ects
Better evidence on …rm side responses
Consumption smoothing bene…ts

Disability Insurance
See Bound et. al (HLE 1999) for an overview
Insures against long-term shocks that a¤ect individuals at home or

work
Federal program that is part of social security
Eligible if unable to “engage in substantial gainful activity” b/c of

physical/mental impairment for at least one (expected) year
Main focus of literature is sharp rise in the size of the program

Source: Mullen, Maestas, Strand (2012)
Two Views on the Rise in DI
One perspective: moral hazard from a lenient system that leads to

ine¢ ciency
Another perspective: program is now helping more needy people who

have high disutilities of work
Empirical work attempts to distentangle these two views

Theory of Disability Insurance
Key additional element relative to UI models is screening and waiting

periods
Less relevant for unemployment because it is easy to identify who has

a job and who does not
Diamond-Sheshinski (1995) build a model that incorporates screening
Characterize optimal properties of solution but do not derive an

empirically implementable formula for optimal screening rule or
bene…t level

Diamond and Sheshinski 1995
Individuals have di¤erent disutilities of working ψi
To max social welfare, not desirable for those with high ψi to work.
First best: Individual i works i¤
Marginal product > ψi
But govt observes only an imperfect signal of ψi ! sets a higher

threshold for disability
Result: lower bene…t rate if screening mechanism has higher noise to

signal ratio

Empirical Evidence: Bound-Parsons Debate
Question: Did increase in DI bene…ts cause decline in labor supply?
Well-known debate between Bound & Parsons in 1980’s is of

methodological interest
Parsons (1980)
Uses cross-sectional variation in replacement rates
Data on men aged 45-59 in 1966-69 NLSY
OLS regression:
LFPi = α + βDIrepratei + εi
where DIreprate is calculated using wage in 1966
Finds elasticity of 0.6
Simulations using this elasticity imply that increase in DI can

completely explain decline in elderly labor force participation
Bound highlights key econometric problem in Parson’s speci…cation
DIreprate variation correlated with wage
Identi…cation assumption: LFP rates equal across wage groups
Parson’s solution: “control” for wage rate
LFPi = α + βDIrepratei + f (wagei ) + εi
Does this resolve the problem?

Identi…cation by Functional Form
LFPi = α + βDIrepratei + f (wagei ) + εi

This is an example of identi…cation by “functional form”
As f is made increasingly ‡exible, standard error on β goes to in…nity
Problem is that only source of variation is due to wages
To illustrate practice importance, Bound replicates Parson’s regression
on sample that never applied to DI and obtains a similar elasticity
Motivates literature that focuses on quasi-experiments other sources
of non-parametric identi…cation
Key idea of non-parametric identi…cation: with su¢ ciently large

samples, estimate is identi…ed without parametric assumptions on f
Impose functional forms only for computational convenience and
precision in …nite samples
Bound (1990) proposes a technique to bound e¤ect of DI on LFP rate
Uses data on LFP of rejected applicants as a counterfactual
Idea: if rejected applicants do not work, then surely DI recipients

would not have worked
Rejected applicants’LFP rate is an upper bound for LFP rate of DI

recipients absent DI
Illustrate using better data from Mullen, Maestas, Strand (2012)

Gruber 2000
Exploits di¤erential law change in Quebec and rest of Canada as a

natural experiment
In 1987, 36% inc. in bene…ts in rest of Canada; in Quebec, no change
Estimates e¤ect of law change on labor force participation of men

aged 45-59
Uses DD method on NLFP rates of men aged 45-59

Gruber 2000
Implied elasticity of non-employment rate w.r.t. DI bene…t level:

0.25-0.3
Agrees more with Bound than Parsons
But estimates are imprecise and only capture short-run e¤ects

Maestas, Mullen, Strand 2012
Maestas, Mullen, and Strand (2012) use random variation in examiner
assignment to identify e¤ects of DI
Disability cases randomly assigned by computer to examiners at state

board
Substantial discretion generates signi…cant variation across examiners

in allowance rates
Instrument for receipt of DI w/ examiner’s conditional allowance

propensity
Another approach: use set of examiner f.e.’s as instruments
Use administrative data on DI decisions and earnings from SSA
1 million observations

Estimate causal e¤ects of DI on employment rates using IV regressions
First stage:
DIi = a + φi examineri + νi
Second stage:
yi = α + βDIi + εi
Note that …rst stage coe¤s. φi are average DI allowance rates by
examiner
Therefore IV regression is equivalent to examiner-level OLS regression
y e = a + βDI e + εe
Weighting this OLS regression by number of individuals per examiner

will yield identical estimate of β

Conclude that DI receipt reduces probability of employment by 28%

for marginal applicants
Important to recognize that this is a LATE for people who are at the
margin of getting DI
Severely disabled individuals would be granted DI by all examiners

and are not captured in this LATE
Hence should be interpreted as an upper bound on ATE
Maestas et al. con…rm this by studying heterogeneity in treatment

e¤ects by disease severity


Results highlight importance of estimating relevant LATE for policy
If policy question is raising DI bene…t, relevant treatment e¤ect is for

people on margin w.r.t. bene…t change
May not be the same people as those who are on the margin with
respect to examiner decision
This is the advantage of directly studying the policy of interest
Does not require extrapolations from estimated LATE to the

policy-relevant su¢ cient statistic
Also important to note that paper estimates uncompensated

elasticities
Critical to distinguish moral hazard vs. liquidity for normative purposes
Might still have people with disutility ϕi > wi working when rejected
Methodological Note: Weak Instruments
Random assignment instrument to judges, classes, etc. now popular

in many applications
Common problem: weak instruments
Arises when each examiner has few cases
Not an issue in Maestas et al. but common e.g. in education with 20

kids per class
In this case, IV estimate of β will be biased toward OLS

Weak Instruments Problem
Recall examiner-level OLS regression
y e = a + βDI e + εe
Same observation appears on LHS and RHS of this regression
With few individuals, this biases β toward OLS
Extreme case: one individual per examiner equivalent to OLS
If an examiner gets a draw of particularly sick people, they will both

get DI and have low employment rates even if DI has no causal e¤ect

Weak Instruments Problem
Problem vanishes as number of individuals per examiner grows large
Only remaining variation in DI e is due to examiner e¤ects
Key question: does number of instruments grow at same rate as

sample size?
If sample gets bigger by adding more people per examiner, then

instruments are asymptotically strong
If sample gets bigger by adding more examiners, then instruments and

asymptotically weak

Weak Instruments: Solutions
How to obtain unbiased estimates with weak instruments?
Traditional recommendation: LIML
Currently preferred alternative: Jackknife IV
Leave out own observation for each i when estimating DI e in …rst stage
Second stage regression becomes
yi = a + bDI j 6=i + εi
Directly …xes own-observation problem and is more robust than LIML

(Kolesar et al. 2012, Kolesar 2012)

Autor and Duggan 2003
Focus on interaction between DI and UI systems
Observe that DI claims rise in recessions, may reduce measured

unemployment rate
Idea: consider a worker laid o¤ in current recession
Given generosity of DI program, instead of claiming UI and searching

for a job, he applies for DI
One less unemployed person –> unemployment rate lower
But economic situation is the same: one less person working
Test this hypothesis using cross-state variation in employment shocks

Autor and Duggan 2003: Bartik Shocks
Standard technique to construct state-level employment shocks over a

…ve year window:
Calculate industry shares in a given state in base year
Calculate employment changes over …ve year period by industry using

data on national employment (excluding state in question)
Project changes in each state’s employment using national changes
Ex: if car industry declines over a …ve year period, assign a negative
employment shock to Michigan
Then correlate state employment shocks with DI applications

Employment Shocks and DI Applications: 1979-1984
6
MS
4
E[DI Apps/Pop | X]
AL AR
WV GA LA
2 KY FL
MT MO AZ
NCTN ME
MA
MISC
OK
0 OHRI
PA VT CT IL TXDE NY CA NV
NJ VA MD
IN IA OR CO NM
MN KS WA
WI NE
SD ND
-2 UT
NH
WY ID HI
AK
-4
-6
Coefficient = -0.094, se = 0.062, t = -1.51
-8
-8 -6 -4 -2 0 2 4 6 8
E[Change in Employment/Pop | X]
Source: Autor and Duggan 2003

6
MS
4
E[DI Apps/Pop | X]
KY AR LA
WV
2 GA
OK MI MT
ME TX
MO SC AL NM
IN NC FL
SD KS TN WA RI
0 OH
PA DE IL VA CO
MA
NY
OR
CAAZ NV
WY IA VT WI MD
ND NE MN CT NJ ID HI
NH AK
-2 UT
-4
-6
Coefficient = -0.262, se = 0.067, t = -3.90
-8
-8 -6 -4 -2 0 2 4 6 8

6 MS
WV KY
4 AR
E[DI Apps/Pop | X]
AL LA
2 SC TNNC MO
IA
ME MA NM
INGA TX CO FL
WA OR MI MT
0 NH OKAZ NV
DE NY
SD
OH WIIL MN
PA CA ID
VT VA UT
NJ RI MD
KS WY NE ND
-2 CT HI
AK
-4
-6
Coefficient = -0.343, se = 0.130, t = -2.64
-8
-8 -6 -4 -2 0 2 4 6 8

8 MS
6
AR
WV
4 AL
KY
E[DI Apps/Pop | X]
SC NC
ME TN
2 OK MO DE FL RI
GA NM
LA MT IN
KS
PA VTNY MI MA
0 WY SD
NV
NEVA AZ
IDOR OH
CTNH
TX IL
IAHI CACO
WA MDNJ
WI
AK MN
-2 ND UT
-4
-6
Coefficient = -0.849, se = 0.164 t = -5.18
-8
-8 -6 -4 -2 0 2 4 6 8

Autor and Duggan 2003
Unemployment would be 0.65% higher if not for post-‘84 trends in DI

participation
Trace decline in LFP to the rise in DI over the past two decades via:
The 1984 inclusion of mental illness in DI eligibility
Rising wage inequality (combined with the progressivity of system)
Bottom line: DI applications are clearly sensitive to incentives
But evidence is insu¢ cient to make welfare statements
Essential to decompose bene…t e¤ects into income and price elasticities

to make normative judgment

Part 7: Public Goods and Externalities
Harvard University
Fall 2012
Public Economics Lectures () Part 7: Public Goods and Externalities 1 / 111

Externalities: Outline
1 De…nition and Basic Model
2 Correcting Externalities
3 Prices vs. Quantities (Weitzman 1974)
4 2nd Best Taxation with Externalities (Sandmo 1975)
5 Empirical Applications

De…nition
An externality arises whenever the utility or production possibility of

an agent depends directly on the actions of another agent.
Important distinction between “pecuniary” vs. “non-pecuniary”

externalities
Consuming an apple vs. consuming loud music
Not a technological distinction; depends on market in place
Coasian view: can convert all externalities into pecuniary externalities

with appropriate markets, property rights.
Only non-pecuniary externalities justify policy intervention

Externalities: Main Questions
1 Theoretical: what is the best way to correct externalities and move

closer to the social optimum?
2 Empirical: how to measure the size of externalities?
Key di¤erence: cannot use revealed-preference

Model of Externalities
Firms produce x cars using c (x ) units of the numeraire y
Generates x units of pollution: P (x ) = x
Consumers have wealth Z and quasilinear utility:
u (x ) + y d P (x )
where d = marginal damage (MD) of pollution
Social welfare is
W = u (x ) + Z c (x ) d x
Let p denote the market price of cars

Model of Externalities: Equilibrium
Firms max pro…ts:

max px c (x )
Consumers max utility, taking level of pollution as …xed:
max u (x ) + Z px
Demand satis…es
u 0 (x D ) = p
Supply satis…es
c 0 (x S ) = p
PMB equals PMC in equilibrium:
u 0 (x D ) = c 0 (x S )
But this is not Pareto e¢ cient

Negative Production Externalities: Pollution
SMC=PMC+MD
Price
S=PMC
P*
PM
MD
D = PMB = SMB
0 Q* QM Quantity

Model of Externalities: Deadweight Loss
Perturbation argument: can increase social welfare by reducing

production by ∆x:
dW = u 0 (x )∆x c 0 (x )∆x d ∆x
= d ∆x > 0 if ∆x < 0
First Welfare Theorem does not hold
Analogous result for consumption externalities

Negative Consumption Externalities
Price
S=PMC=SMC
PM
MD
P*
D = PMB
SMB=PMB-MD
0 Q* QM Quantity

Remedies for Externalities
1 Coasian bargaining solution
2 Pigouvian corrective taxation
3 Regulation
4 Permits (cap-and-trade)

Coasian Solution
Externalities emerge because property rights are not well de…ned.
Establish property rights to create markets for pollution.
Consider example of pollution in a river.
If consumer owns river, in competitive equilibrium, …rms pay d for

every unit of pollution emitted.
Marginal cost of production is now c 0 (x ) + d, leading to 1st best.
Symmetric solution when …rm owns river.
Assignment of property rights a¤ects distribution but not e¢ ciency

Coasian Solution: Limitations
1 Cost of bargaining
Ex: air pollution – would require millions of agents to coordinate and
bargain
To reduce transactions costs, need an association to represent agents
This “association” is the government
2 Asymmetric information: competitive equilibrium can break down
Often hard to identify precise source of damage
E.g. atmospheric pollution very di¤use, marginal damages unclear

Pigouvian Taxation
Impose tax t = MD (Q )
Restores Pareto e¢ ciency and maximizes social welfare
Practical limitations:
Must know marginal damage function to set t
Di¢ cult to measure the marginal damage in practice

Pigouvian Tax
SMC=PMC+MD
Price S=PMC+t
S=PMC
$t
P*
P2
P1
D = PMB = SMB
0 Q* Q2 Q1 Quantity

Regulation
Quantity-based restriction: reduce pollution to …xed level or face legal

sanctions
Same outcome as Pigouvian taxation: move people to x2
Disadvantages: no marginal incentives
Allocative ine¢ ciency with heterogeneity in costs of pollution reduction
Dynamic ine¢ ciency: no incentive to innovate
These problems can be solved by cap and trade system

Permits: Cap-and-Trade
Cap total amount of pollution and auction permits to …rms
Then allow …rms to trade permits to pollute
Hybrid of regulation and Coasian solution: create the market
In eq., …rms with highest MC of reducing pollution will buy permits;

those that can easily reduce pollution will do so
If total number of permits is set to achieve the social optimum, both

allocative and productive e¢ ciency will be achieved
Also have dynamic incentives to innovate because each …rm is bearing

a marginal cost of pollution

Weitzman 1974: Prices vs. Quantities
Price mechanism (taxes) identical to quantity mechanism (permits) in

simple model above. How to choose?
Weitzman (1974): with uncertainty re. shape of MB and MC curves,

price and quantity no longer equivalent
Now the standard method of choosing between regulation and taxes

Weitzman 1974: Prices vs. Quantities
Let q denote pollution reduction starting from private market eq.,

where q = 0.
Let B (Q ) denote social bene…ts of pollution reduction
Let C (Q ) denote social costs.
In simple model above:
MB of pollution reduction is constant, B 0 (Q ) = d.
MC given by loss in surplus from producing one less car: u 0 (x ) c 0 (x ).
More generally, MC should be interpreted as cost of reducing pollution

through cheapest method (e.g. cleaner plants)

Market for Pollution Reduction
Price
PMCQ=SMCQ
SMBQ
Q* Pollution Reduction

Optimal Policy without Uncertainty
In eq’m, PMB of pollution reduction is 0 ) level of pollution

reduction is Q = 0.
Social optimum:
max B (Q ) C (Q )
C 0 (Q ) = B 0 (Q )
With no uncertainty, can obtain optimum with either quantity or price
policy.
Quantity: require amount Q .
Price: set price for pollution reduction of p = C 0 (Q ).

Optimal Policy with Uncertainty
Now suppose that there is uncertainty about the marginal costs of

reducing pollution.
Cost is now C (Q, θ ) with θ unknown.
Marginal cost lies between MCLB and MCUB , with mean value given
by MCmean .
Objective: maximize expected social welfare:
Eθ [ B (Q ) C (Q , θ )] >? Eθ [B (Q (p )) C (Q (p ), θ )]
Optimal choice depends on B 00 (Q )/C 00 (Q )
Quantity regulation preferred if MB steep relative to MC

MB steep, Quantity regulation

MB Steep, Price Regulation

Quantity Regulation Price Regulation

Price Band vs. Quantity Band with Steep MB

MB Flat, Quantity
MB Flat, Price Regulation
regulation

Weitzman: Uncertainty about Bene…ts
Now suppose that there is uncertainty about the marginal bene…ts of

reducing pollution but that the costs are known
Price and quantity policies are again equivalent
For a given p, the government knows the Q that will result exactly
since p = C 0 (Q )
More generally, uncertainty matters only when it is about the

cost/bene…t schedule for the agent who chooses level of pollution
reduction
If consumer chooses level of pollution reduction, then only uncertainty

about marginal bene…ts matters

Optimal Second-Best Taxation with Externalities
In general, cannot restore 1st best b/c externality is one of many

deviations from …rst best.
Most important other deviation: govt also uses distortionary taxes to

…nance public goods and redistribute income.
Sandmo (1975): optimal tax policy with externalities and a revenue

requirement.
Combination of Ramsey and Pigou problems

Sandmo 1975: Setup
Denote by d (xN ) the externality cost of consumption of good N
Let w be the wage rate and qi = pi + τ i denote post-tax prices.
Let Z denote non wage income.
Producer prices …xed; all pre tax prices normalized to 1.
Individuals have utility functions of the following form:
u (x1 , .., xN , l ) d (xN )
Utility is maximized subject to:
q1 x1 + .. + qN xN wl + Z

Sandmo 1975: Setup
Individual maximization program
L = u (x1 , .., xN , l ) + λ(wl + Z (q1 x1 + .. + qN xN ))
Maximization yields indirect utility v (q ).
Government maximization program:
max W (q ) = v (q ) d (q )
q
s.t. ∑ τi xi R
Analogous to Ramsey tax problem, but here SWF di¤ers from private
sector objective

Sandmo 1975
Let θ = marginal social welfare gain from $1 of a lump sum tax and
λ = marginal value of relaxing agent’s budget constraint
τ ip = optimal Pigouvian tax rate (when R = 0)
τ ip = 0 for goods 1 to N 1 and τ ip = d 0 (xN ) for good N
τ ir = optimal Ramsey tax rate (when d (xn ) = 0)
Let τ i denote optimal tax rate in Sandmo model

Sandmo 1975: Additivity Result
Main result: can express optimal tax rate as Ramsey rate plus
Pigouvian correction.
Consider case where Slutsky matrix is diagonal (zero cross-price

elasticities)
Then optimal tax on good i, τ i satis…es

τ i τ ip
= (θ/λ)/ecii
1 + τi
θxic dxic
) τi = / + τ ip
λ dpi
= τ ip + τ ir

Sandmo 1975: Additivity Result
Useful analytic representation but not an explicit formula for the

optimal tax rate
Ramsey tax will a¤ect level of cons, which a¤ects optimal Pigouvian tax
Conversely, Pigouvian tax will a¤ect optimal Ramsey tax rate
Qualitative lesson: no justi…cation to tax goods that are

complementary to those that produce negative externalities
Just tax fuel, not cars

Double Dividend Debate
Claim: gas tax has two “dividends”
1 discourages pollution, raising social welfare
2 allows govt. to reduce other distortionary taxes, improving e¢ ciency
True if we are at a corner where revenue req. is below level what is

generated by optimal Pigouvian taxes
More realistic case: already at a Ramsey-tax interior optimum

Double Dividend Debate
Suppose we discover that production of computers generates negative
externality.
Is there a “double dividend” from taxing computers?
No. Already at Ramsey optimum ! no e¢ ciency gain from raising

taxes on PC’s and reducing taxes on other goods
Only get single dividend of improving environment
Obtain double dividend only if taxes on polluting good were initially

too low from a Ramsey perspective.
General lesson: separate externality and optimal second-best tax

problems.
Measure externalities and identify optimal corrective taxes without

worrying about other aspects of tax system
Externalities: Empirical Measurement
Two approaches
Indirect market-based methods
Contingent valuation

Edlin and Karaca-Mandic 2006
Accident externalities from driving automobiles.
If I drive, I increase probability you will get into an accident !

externality cost imposed on you
How to estimate this externality cost and appropriate Pigouvian tax

on driving?
Examine relationship between tra¢ c density and per-capita insurance

costs and premiums
Look at slope to infer size of externality cost
Identi…cation assumption: variation in tra¢ c density at state level not

correlated with other determinants of premiums (e.g. types of cars,
etc.)
Edlin and Karaca-Mandic 2006
Tra¢ c density substantially increases marginal insurer costs
Insurers set premia to cover average costs in market equilibrium
Individual’s marginal cost on other driver’s not internalized
Externality is convex
Increase in tra¢ c density from average driver has external cost of

$2,000 per year in California
Comparable …gure in $10 per year in North Dakota
Suggests that insurance premiums should be doubled in CA to achieve

social optimum

Brookshire et al. 1982
Infer willingness to pay for clean air using e¤ect of pollution on

property prices (capitalization)
Compare prices of houses in polluted vs non-polluted areas.
Pi = α + Pollutioni + Xi β + ei
Econometric problems
Omitted variables: polluted neighborhoods worse on many dimensions
Deeper problem: sorting
Recover marginal WTP rather than average WTP
People with allergies avoid polluted areas

Chay and Greenstone 2005
Also study home prices but use Clean Air Act as an exogenous change
in pollution.
Clean Air Act: imposed ceilings on pollution levels by county in mid

1970s.
High pollution counties experience sharp reductions in pollution levels

relative to low pollution counties

Chay and Greenstone 2005
Conclusion: 1% increase in pollution ! 0.25% decline in house values
Clean air act increased house values by $45 bil (5%) in treated
counties
Conceptual concern with short-run market-based methods: people

may not be fully aware of changes in pollution

Glaeser and Luttmer 2003
Without tradeable permits, e¢ ciency costs of regulation can be very

high because of allocation distortions
Study allocation of apartments under rent control
Standard model assumes that with price controls, still have allocative
e¢ ciency
Those who value the apartments most get them
But regulation will generally lead to allocative ine¢ ciency that

generates …rst-order welfare losses
For small price caps, allocation ine¢ ciency dwarfs undersupply

ine¢ ciency

Glaeser and Luttmer 2003
Quantify welfare losses from misallocation by comparing consumption

patterns in rent-controlled (NYC) and free-market places across
demographic groups.
Predict apartment size using number in family, income, education,

age, etc. using 105 large MSAs
Test if actual apartment allocations in NYC match predictions
Identifying assumption: preferences stable across MSAs
Check: placebo tests using Chicago and Hartford

Contingent Valuation
For some outcomes, it is impossible to have a market value
Ex: protecting endangered species
Common solution: “contingent valuation” surveys
How much would you be willing to pay to avoid extinction of whales?

Diamond and Hausman 1994
Describe problems with contingent valuation using surveys
No resource cost to respondents
Lack of consistency in responses
Framing E¤ects: whales then seals vs. seals then whales
WTP to clean one lake = WTP to clean 5 lakes
Diamond and Hausman: let experts decide based on a budget voted

on by individuals for the environment instead of relying on valuation

Behavioral Economics Applications: Internalities
Sin taxes intended to correct “internalities.”
Internal costs of smoking cigarettes dwarf the external costs.
Suggests that conventional Pigouvian taxation should be small

(relative to actual taxes observed on e.g. cigarettes and alcohol).
Q: Does addictive nature of cigarettes motivate taxation?
A: Highly sensitive to positive model of addiction
Challenge: di¢ cult to determine which model is right empirically

Becker and Murphy 1988
Show that addictive goods can be modeled in perfectly rational

framework
Dynamic model with habit formation
Current consumption of the addictive good decreases utility in future

periods but increases marginal utility of consumption tomorrow
If discount rate high enough, rationally choose to become addicted
Implication: no reason for special taxes on these goods; set taxes

according to Ramsey rules

Gruber and Koszegi 2004
Hyperbolic discounting preferences for smokers
U0 = u (c0 ) + β( ∑ γt u (ct )) with β < 1.

t 1
U1 = u (c1 ) + β( ∑ γt u (ct ))
t 2
Planner maximizes U0 with β = 1 (true utility).
Individuals overconsume c: fail to take full account of harm to future

selves.
Taxes reduce demand for each self; can partly correct the internality.
Calibration implies corrective tax should be very large.

Bernheim and Rangel 2004
Model of “cue-triggered” addiction. Two selves:
Cognitive self with rational preferences
Visceral brain triggered by random cues in which addictive good is

consumed at any cost.
Probability of trigger increases with past consumption levels.
Ideal policy: only allow rational consumption, eliminate consumption

in hot mode.
Corrective taxation may not be desirable: only distorts consumption in

rational state, not visceral state.
Better solution: regulated dispensation – must place orders one

period in advance
O’Donoghue and Rabin 2006
Studies optimal sin taxes in a model with two types of consumers:

rational and those who overconsume (e.g., because of self-control
problems)
Can be thought of as a hybrid of Becker and Gruber-Koszegi models
Key result: irrationality among a few consumers leads to substantial

role for corrective taxation/subsides.
For rational individuals, excess burden due to taxation is second-order

(Harberger triangle).
For irrational individuals, welfare gains from correction of internality is

…rst-order (Harberger trapezoid)
Therefore always optimal to have a positive tax; calibrations suggest

fairly large corrective taxes
Application: Retirement Savings
Many believe that people do not save enough for retirement because
of myopia, self-control problems, etc.
What are the best corrective policies to increase savings?
Price subsidies: 401(k)’s, IRA’s (Du‡o et al. 2006)
Nudges: defaults and automatic enrollment in pension plans (Madrian

and Shea 2001)
Commitment devices (Bernartzi and Thaler 2004; Ashraf, Karlan, and

Yin 2006)
Information provision and …nancial literacy (Lusardi and Mitchell 2011)
Focus on the …rst two here

Du‡o et al. (2006): Price Subsidies
Du‡o et al. conduct a randomized experiment providing matches for

IRA contributions
Subject pool: H&R Block tax …lers
Main …nding: provision of a non-zero subsidy signi…cantly increase

IRA participation rates
Equivalent government Saver’s Credit program has no impact,

suggesting that salience matters

Madrian and Shea (2001): Defaults
Madrian and Shea show that defaults have powerful e¤ects on savings
behavior even though they do not change budget set
Clearly violates neoclassical model
Utility consequences of changing retirement savings rate are large
Therefore di¢ cult to explain with optimization costs
Carroll et al. (2009) propose a model with optimization costs and

hyperbolic discounting to explain the pattern
Hyperbolic discounters keep postponing plans to set up retirement

account because cost of delaying by one period is small
Interesting implication: short deadlines may help improve decision

making
Chetty et al. (2012): Crowdout in Retirement Savings
Do defaults increase total savings or just lead to shifting of assets

from non-retirement to retirement accounts?
Even inattentive individuals still have to satisfy budget constraint by

cutting consumption or savings in non-retirement accounts
Do price subsidies raise total savings or induce shifting across

accounts?
Large literature on “crowdout” in retirement savings accounts (Engen,

Gale, Scholz 1996; Poterba, Venti, Wise 1996)
Data in U.S. on wealth outside retirement accounts very limited
Chetty et al. (2012) analyze this question using third-party reported

data on all …nancial wealth for population of Denmark
Chetty et al. (2012): Price Subsidies
Denmark has two types of pension accounts: capital pensions and

annuity pensions
Reform in 1999 in Denmark lowered subsidy for saving in capital

pensions by 12 cents per DKr
Question 1: how did this a¤ect contributions to capital pensions?
Question 2: how much money was shifted to annuity pensions and to

non-retirement taxable accounts?

Impact of Capital Pension Subsidy Reduction On Capital Pension Contributions
3.5
Total Capital Pension Contribution Rate
Subsidy for Capital

Pension Reduced
1.5 2 2.5 3
1995 2000 2005 2010

Year
25-75K Below Top Tax Cutoff 25-75K Above Top Tax Cutoff
Impact of 1999 Capital Pension Subsidy Reduction on Distribution of
Capital Pension Contributions for Prior Contributors
60
Extensive-margin responders account for

100% of reduction in capital pension
40
20
0
-100 -80 -60 -40 -20 0 20 40 60 80 100
Percentage Change in Capital Pension Contributions (Pt –Pt-1)/ Pt-1

1997 to 1998 1998 to 1999

Difference in Fraction of High Savers Before vs. After Subsidy Change,
Pre-Reform (1996-1998) minus Post-Reform (1999-2001)
0
Change in % with Taxable Savings Above Mean
-0.5
-1.0
-1.5
Change in Slope at Cutoff = -0.8% (0.1%)

-2.0
Crowd-Out of Pension Contribution

ΔTaxable Saving / Δ Pension Contrib.:
β = -0.99 (0.19)
-2.5
-75000 -50000 -25000 0 25000 50000 75000

Income Relative to Top Tax Cutoff (DKr)
Chetty et al. (2012): Automatic Contributions
Next, study impacts of automatic contributions
Employers make pension contributions on workers behalf

automatically
Research design: event study when workers switch …rms
Retirement savings rate can change sharply when workers switch …rms
Do workers o¤set these changes in private savings as neoclassical

model predicts?

Event Study around Switches to Firm with >3% Increase in Employer Pension
Contrib.: Switchers with Positive Individual Pensions and Savings in
Year Prior to Switch
Contribution or Savings Rate (% of income)
18
14
10
6
Δ Employer Pensions = 5.65

Δ Total Pensions = 4.86
Δ Total Savings = 4.44
2
-5 0 5
Year Relative to Firm Switch
Employer Pensions Total Pensions Total Savings

Changes in Total Savings Rates vs. Changes in Employer Pension Rates
for Firm Switchers, Cond. on Lagged Savings
10
Percent Change in Total Savings Rate
0 5
Total Savings
Pass-Through Rate: β = 90%
-5
(0.9%)
-5 0 5 10
Percent Change in Employer Pension Contributions
Chetty et al. (2012): Active vs. Passive Savers
Why do automatic contributions have much larger impacts on total

savings than price subsidies?
Hypothesis: active vs. passive choice, as in Carrol et al. (2009)
Analyze heterogeneity of responses to test key predictions of this

model
1 Price subsidies a¤ect active savers
2 Automatic contributions a¤ect passive savers
3 Active savers save more for retirement to begin with

Percent Responding to Capital Pension Subsidy Change in 1999
by Frequency of Active Changes in Other Years
25
% with Sharp Response in 1999
20
15
10
5
0
0 20 40 60 80 100
Percentage of Other Years with Change in Individual Pension Contributions
Pensions Pass-Through of Employer Pension Changes for Firm-Switchers
by Frequency of Active Changes in Other Years
Pass-Through of Emp. Pensions to Total Pensions
.98
.96
.94
.92
.9
.88
0 20 40 60 80 100
Percentage of Other Years with Change in Individual Pension Contributions
Heterogeneity in Response to Capital Pension Subsidy by Wealth/Income Ratio
25
% with Sharp Response in 1999
20
15
10
0 .5 1 1.5
Wealth/Income Ratio in 1998

Heterogeneity in Pass-Through of Employer Pensions by Wealth/Income Ratio
Pass-Through of Employer Pensions to Total Savings
40 60 80 100 120
0 .5 1 1.5 2
Wealth/Income Ratio in Year Prior to Switch

Chetty et al. (2012): Correcting “Internalities”
Tax subsidies tend to in‡uence the behavior of those who are already
saving
Need to be an active saver to pay attention and respond to subsidies
More general lesson: economic tools (prices) may not be the best way
to change the behavior of non-optimizing agents
Non-traditional tools such as “nudges” may be more e¤ective

Public Goods: Outline
1 De…nitions and Baseline Model
2 Samuelson Rule
3 Public Goods with Endogenous Private Provision
4 Public Goods with Distortionary Taxation
5 Alternative Instruments

Public vs. Private Goods
Private goods bene…t one individual h
∑ Xh X
h
Public goods bene…t several individuals simultaneously
Xh X 8h
Ex: can of coke vs. teaching a class
Pure: can accommodate any number of users.
Impure: subject to congestion
radio vs. roads

Private Good
Person 1’
s
Consumption
Person 2’
s Consumption

Public Good
Person 1’
s
Consumption
Person 2’
s Consumption

Public vs. Private Goods
Rival vs. non-rival.
Pure are non-rival
Excludable vs. non-excludable.
National Radio: impossible to exclude. Teaching: possible to exclude
Most economic analysis focuses on pure public goods
Public goods ) equilibrium outcome ine¢ cient (large scale

production externalities)

Public Goods Model: Setup
Economy with H households, indexed by h = 1, .., H
Two goods X and G
X is always private, individual h consumes quantity X h
Denote by X = ∑h X h the total quantity of good X in the economy
Denote by G h consumption of good G by h, with G = ∑h G h
Utility of h is U h = U h (X h , G )

Public Goods Model: Setup
Social welfare = weighted sum of utilities, βh weight on h
βh 0 and at least one βh > 0
Production possibility F (X , G ) = 0
Assume that U h is increasing in X and G

First Best if G is Private
To identify Pareto e¢ cient outcomes, solve:
max ∑ βh U h (X h , G h )
h
s.t. F (∑ X h , ∑ G h ) 0 [λ]
h h
Lagrangian:
L= ∑ βh U h λF
First order conditions
[X h ] : βh UXh = λFX
[G h ] : βh UGh = λFG

First Best if G is Private
Taking ratios of FOCs yields
UGh F
h
= G
UX FX
Set of Pareto e¢ cient allocations is set of allocations that satisfy:
h
MRSGX = MRTGX 8h
Decentralized market equilibrium will implement such an allocation

(1st Welfare Thm).

First Best if G is a Pure Public Good
To identify Pareto e¢ cient outcomes, now solve:
max ∑ βh U h (X h , G )
h
s.t. F (∑ X h , ∑ G h ) 0 [λ]
h h
FOC’s:
[X h ] : βh UXh = λFX
[G ] : ∑ βh UGh = λFG
h
Using βh = λFX /UXh from f.o.c. for X h we obtain:
Uh FG
∑[ UGh ] = FX
h X

Samuelson (1954) Rule
Condition for Pareto e¢ ciency: sum of MRS is equal to MRT:
∑ MRSGX
h
= MRTGX
h
Intuition: an additional unit of G increases the utility of all

households in the public good case
With G a private good, an additional unit only increases one

individual’s utility

Decentralized Private Provision is Suboptimal
Private good X and a pure public good G as above
Price of each good is normalized to 1
Each household starts with an endowment Y h of good X .
Individual h contributes G h to public good funding.
Consumption of public good is G = ∑h G h for everyone.
Consumption of the private good is X h = Y h G h for individual h.

Decentralized Private Provision is Suboptimal
Individual h solves
max U h (X h , G 1 + .. + G h + .. + G H )
s.t. X h + G h = Y h .
Nash equilibrium outcome is UXh = UGh
Samuelson Rule not satis…ed
Pareto improvement if each person invested 1/H more dollars in the

public good:
∆W = UXh (1/H ) + UGh = UGh (1 1/H ) > 0.
Market outcome is ine¢ cient: underprovision of G

Optimal Second Best Provision of PG’s
Now suppose government provides public good to rectify

under-provision in market equilibrium
Two complications arise
1 Crowdout of private sector provision
Private contributions to charity exceed $250 bn. per year
Key model: Bergstrom, Blume, and Varian (1986)
2 Government cannot …nance PGs through lump sum taxation
Need to modify Samuelson rule to account for distortionary taxation?
Related to Sandmo (1975) analysis of externalities

Bergstrom, Blume, Varian (1986): Setup
Individual h solves:
max U h (Xh , Gh + G h )
X h ,G h
s.t. Xh + Gh = Yh
FOC is UXh = UGh
Nash equilibrium exists and is unique
G s.t. all individuals optimize given others’behavior
Let G denote private equilibrium outcome

Bergstrom-Blume-Varian Model: Crowd-out
Now suppose government introduces lump sum taxes t h on each

individual h
Revenue used to …nance expenditure on public good T = ∑ t h
Individual’s optimization problem is now:
max U (X h , Gh + G h + T)
h h h
s.t. X + G = Y th

Bergstrom-Blume-Varian Model: Crowd-out
Let Zh = Gh + th denote total contribution of individual h.
Can rewrite this as:
max U (X h , Zh + Z h)
h h h
s.t. X + Z = Y
This is isomorphic to original problem ) Z = G
Total public good provision is unchanged!
Each person simply reduces voluntary provision by th

BBV Model: Key Assumptions
1 No corners: assumed the set of contributors are the same in both
situations.
With corners, transfer neutrality breaks down: tax increase T results in

no private contribution from individuals with G h < T , but
contributions increase on net.
2 Ignores direct utility from giving: U (X h , G h , G ).
Andreoni’s (1990) “warm glow” model.
Stigler and Becker (1977) critique: should not simply modify

preferences to explain patterns
3 Ignores prestige/signalling motives
Glazer and Konrad (1996)

Empirical Evidence on Crowd-Out
Two empirical questions motivated by theory
1 How large is the degree of crowd-out in practice?
2 What are the income and price e¤ects on charitable giving?
Two strands of empirical literature
1 Field evidence (observational studies)
2 Lab experiments
Traditionally, lab experiments have been more in‡uential but recent

…eld studies may change this
Lab experiments may not capture important motives for giving: warm
glow, prestige
Hungerman 2005
Studies crowdout of church-provided welfare (soup kitchens, etc.) by

government welfare
Uses 1996 Clinton welfare reform act as an instrument for welfare

spending
One aspect of reform: reduced/eliminated welfare for non-citizens
Motivates a di¤-in-di¤ strategy: compare churches in high non-citizen

areas with low non-citizen areas before/after 1996 reform
Estimates imply that total church expenditures in a state go up by 40

cents when welfare spending is cut by $1

Andreoni and Payne (2003, 2008): Fundraising Mechanism
Government spending crowds-out private donations through two
channels: willingness to donate + fundraising
Use tax return data on arts and social service organizations
Instrument for government spending using changes in state budget

due to federal grants
Key …ndings:
$1000 increase in government grant leads to $250 reduction in private

fundraising
$1 more of government grant to a charity leads to 56 cents less private

contributions
70 percent ($0.40) due to the fundraising channel
Suggests that individuals are relatively passive actors

Marwell and Ames 1981
Early lab experiments testing free-rider behavior.
Groups of 5 subjects, each given 10 tokens.
Can invest tokens in either an individual or group account.
Individual: 1 token = $1 for me; Group: 1 token = 50 cents for

everyone
Nash equilibrium is 100% individual but Pareto e¢ cient outcome is

100% group.
Compute fraction invested in group account under various treatments
Marwell and Ames 1981
Finding: 40 to 60% of tokens were still invested in the public good.
Experiment run on various groups of high school and college students.
Only one group free-rode a lot: 1st year econ graduate students (20%
donation rate).
“Economists Free Ride, Does Anyone Else?”
Andreoni (1988, 1993) implements experiments with repeated

contributions
Shows that contributions to public goods fall over time but remain
positive
Expanding the Policy Set: Social Prices
Traditional public goods and externalities literature focuses on

economic incentives
Induce public goods provision by changing relative prices of goods
Another potential policy tool: manipulation of social prices
Exploit concerns for perception by peers to encourage pro-social

behavior
E.g. have researchers compete on publications and help society in the

process
Social Pressure: Existing Evidence
Recent examples from psychology and political science
Cialdini (2003) on energy conservation: telling people how their energy

use compares with averages reduces energy use
Gerber et al. (2008): using social pressure to increase voter turnout
Social Prices as a Policy Tool: Research in Progress
1 [Perez-Truglia and Cruces 2012] Social incentives in campaign

contributions and political polarization
2 [Chetty, Saez, Sandor 2012] Comparing cash and social incentives to

reduce referee times at JPubE
3 [Chetty, Mobarak, Singhal] Increasing tax revenue in Bangladesh

using social incentives
Theoretical question: optimal social prices and policy design
Figure 2b: Turnaround times, DFL reweighted on pre-experiment turnaround
Percentage of Reports Still Pending 100%
75%
50% 27.3 33.7 44.8 47.8
25%
Control = 4 week: p = 0.00
Control = Social: p = 0.01
4 week = Cash: p = 0.00
0%
0 20 40 60
Days since invitation
Control Social 4 week Cash
Source: Chetty, Saez, Sandor 2010


Public Economics - All Lecture Note PDF

Uploaded by

Copyright:

Available Formats

Public Economics - All Lecture Note PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Public Economics - All Lecture Note PDF

Uploaded by

Copyright:

Available Formats

Public Economics Lectures

Raj Chetty and Gregory A. Bruich

Public Economics Lectures () Part 1: Introduction 1 / 49

Public economics focuses on answering two types of questions

1 How do government policies a¤ect the economy?

2 How should policies be designed to maximize welfare?

Three motivations for studying these questions:

Public Economics Lectures () Part 1: Introduction 2 / 49

Interest in improving economic welfare ! interest in public economics

Almost every economic intervention occurs through government

Price intervention: taxes, welfare, social insurance, public goods

Regulation: min wages, FDA regulations (25% of products consumed),

Government directly employs one sixth of U.S. workforce

Public Economics Lectures () Part 1: Introduction 3 / 49

Stakes are extremely large because of broad scope of policies

Ex. Tax reforms immediately a¤ect millions

Contentious debate on the appropriate role of government in society

Romney: replacing Medicare with decentralized private insurance will

Only one of these views can be correct

Injecting science into these debates has great practical value

Public Economics Lectures () Part 1: Introduction 4 / 49

Macro, development, labor, and corporate …nance questions often

Ex 1: Macro studies on costs of business cycles and intertemporal

Ex 2: Labor studies on employment e¤ects of the minimum wage

Natural to combine public …nance with another …eld

Understanding public …nance can help ensure that you work on

Public Economics Lectures () Part 1: Introduction 5 / 49

Public economics is at the frontier of a methodological transformation

Data-driven approach to answering important policy questions

Combines a broad set of skills: applied theory, applied econometrics,

Useful skill set for many applied …elds in economics

Topics in the course re‡ect a broad set of methodological themes

Public Economics Lectures () Part 1: Introduction 6 / 49

Modern public economics tightly integrates theory with empirical

What is the optimal income tax rate?

What is the optimal unemployment bene…t level?

Traditional approach: theoretical models and numerical simulations

Numerical simulations rely on strong assumptions

Recent work derives robust formulas that can be implemented using

Event studies, regression discontinuity, synthetic control

Good way to learn practical lessons in applied econometrics

What is “identi…cation by functional form” and why is it undesirable?

Is the LATE or ATE of greater interest in your problem?

When is propensity score reweighting credible?

Emphasis on non-parametric graphical techniques rather than

Public Economics Lectures () Part 1: Introduction 9 / 49

Public Economics Lectures () Part 1: Introduction 10 / 49

Public Economics Lectures () Part 1: Introduction 11 / 49

Public Economics Lectures () Part 1: Introduction 12 / 49

Compelling implementation of quasi-experimental methods requires a

Recent availability of very large datasets has transformed research in

Scanner data on consumer purchases

Tax and social security records

School district databases

Public Economics Lectures () Part 1: Introduction 13 / 49

1980 1990 2000 2010

Public Economics Lectures () Part 1: Introduction 14 / 49

1980 1990 2000 2010

Public Economics Lectures () Part 1: Introduction 15 / 49