applied tools: combining theory and data to evaluate policy
TRANSCRIPT
![Page 1: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/1.jpg)
Applied Tools:Combining Theory and Data to Evaluate Policy
Owen ZidarPrinceton
Spring 2020
Lecture 1b
This lecture draws from several lectures in applied economics (i.e., lectures by David Card, Raj Chetty, PatKline, Magne Mogstad, Kevin Murphy, Jesse Shaprio, Alex Torgovitsky, Chris Walters), parts of which arereproduced here.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 1 / 175
![Page 2: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/2.jpg)
Outline
1 Preliminaries: research designs, DGP, and applied modeling
2 Connecting theory and dataUsing supply and demandInterpreting regression results with optimizing agents
3 Potential Outcomes and SelectionSelection BiasExample: returns to selective colleges
4 The Roy Model and Selection Corrections
5 Decomposition methods and the Pay Gap
6 Review of 3 Applied Econometrics ToolsDifference-in-differencesEvent StudiesDiscrete Choice
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 2 / 175
![Page 3: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/3.jpg)
Preliminaries:Research designs, DGP, and applied modeling
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 3 / 175
![Page 4: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/4.jpg)
What is a research design (1/2)
Consider the effect of a treatment (e.g., tax) T on outcome y
yi = α + βTi + εi
Treatment is assigned based on “selection” model
Ti = αT + βTXi + ηi
Treatment may be non-random: cov(Xi , εi ) 6= 0, cov(ηi , εi ) 6= 0
Traditional approach to accounting for confounding factors or selection: control forobservables Xi when estimating treatment effect
yi = α + βTi + γXi + εi
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 4 / 175
![Page 5: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/5.jpg)
What is a research design (2/2)
Problem with this approach is that we don’t know the source of variation in Ti
There must be some reason that one person or location got treated and another did noteven if they are perfectly matched on observables (e.g., twins)
ηi must to correlated with Ti to have variation in Ti |Xi
But that same unobserved factor could also affect outcome: no way to know ifcov(ηi , εi ) = 0
A research design is a source of variation in ηi that is credibly unrelated to εiE.g., a reform that affects people above age 65 but not below. People at age 64 and 65 likelyto have similar outcomes ⇒ cov(ηi , εi ) = 0
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 5 / 175
![Page 6: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/6.jpg)
Interpreting regression results
It is easier to run a regression than to understand the results
How were the data generated? And what (part of the) data do we actually observe?
Is the effect big or small?
What substantively is in the error term εi and ηi?
How might outcomes change if we vary treatment?
Policy evaluation requires thinking about these questions. And connecting regressions withapplied statistical and economic models is very helpful for thinking through these issues
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 6 / 175
![Page 7: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/7.jpg)
How were the data generated?
How were the data generated? A Data Generating Process (DGP) is a completecharacterization of how the data were generated
Parametric example
Yi = β0 + β1Xi + εi
(Xi , εi ) ∼ N([µ1
µ2
],
[σ2
1 , σ12
σ12, σ22
])(β0, β1, µ1, µ2, σ1, σ12, σ2) ∈ R7
Everything there is to know about the data in 7 numbers!
Important to ask what (part of the) data do we actually observe? And what underlyingmodel led to what we are able to see?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 7 / 175
![Page 8: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/8.jpg)
Importance of DGPs for policy recommendations: WWII example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 8 / 175
![Page 9: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/9.jpg)
Importance of DGPs for policy recommendations: WWII example
WWII data: section of plane and bullet holes per square foot
Engine 1.11
Fuselage 1.73
Fuel system 1.55
Rest of plane 1.8
More bullet holes in the fuselage, not so many in the engines ⇒ protect fuselage?
Policy question – where should plane armor be allocated?
Wald’s insight: ask where are the missing holes. Don’t see planes that didn’t come back
Intuition: if you go to the recovery room at the hospital, you’ll see a lot more people withbullet holes in their legs than people with bullet holes in their chests; not because peopledon’t get shot in the chest; it’s because the people who get shot in the chest don’t recover
Source: https://medium.com/@penguinpress/
an-excerpt-from-how-not-to-be-wrong-by-jordan-ellenberg-664e708cfc3d
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 9 / 175
![Page 10: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/10.jpg)
Connecting theory and empirics
Real World → Model → Real World
(1) Make observations, (2) feed them into the model, and (3) use it to analyze keyquestions (e.g., what would happen if X happened?)
What data can we look at to understand what’s happening?
How are these data generated? E.g., did agents choose to be treated?
What predictions do different theories make?
How does the model rationalize what we are observing?
How should we change the model to better match the data?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 10 / 175
![Page 11: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/11.jpg)
The microeconomic approach to modelingA model is a simplified representation of reality that gets to the essence of what is going on
To understand how and why something happens, look at individual behavior
Who are the people making the choices?
What does each person want?
What decisions are optimal, given constraints?
How do the decisions of different players interact?
What adjusts if the choices aren’t mutually consistent?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 11 / 175
![Page 12: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/12.jpg)
Applying the microeconomic approach
Main application: Markets
Big picture: Supply and Demand
Consumers choose what to buy
Maximize utility subject to prices & budgets
Firms choose how much to produce, what prices to charge
Maximize profit subject to demand curves, costs
Equilibrium: prices and quantities adjust to clear market
What predictions do we make about the impact of supply/demand shifts, taxes, etc?
What can we say about consumer and producer welfare of different policies?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 12 / 175
![Page 13: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/13.jpg)
Connecting theory and data
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 13 / 175
![Page 14: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/14.jpg)
Overview
1 Connecting theory and data
Using supply and demandInequality: Katz and MurphyOptimization and the value of police services
2 Potential Outcomes and Selection
3 Roy model and college choice example
4 Applied Metrics Tools
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 14 / 175
![Page 15: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/15.jpg)
Connecting theory and data using supply and demand
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 15 / 175
![Page 16: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/16.jpg)
Using Supply and Demand Outline
1 Quantitative supply and demand framework
2 Using supply and demand to study inequality (Katz Murphy)
3 Using supply and demand – tax example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 16 / 175
![Page 17: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/17.jpg)
Recall the two ways the quantity demanded can change1. Moves along demand curve vs. 2. Shifts of demand curve
“Demand goes up” can mean one of two things.
Move along a demand curve:Price falls, so quantity goes up
Q
P
D
A
B
Shift of a demand curve:Any P gives a higher Q
Q
P
D
D'
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 17 / 175
![Page 18: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/18.jpg)
Two ways the quantity demanded can change (Math)
The quantity demanded can change in two ways:
%∆QD = %∆D︸ ︷︷ ︸Shift
+ εD%∆P︸ ︷︷ ︸Movement Along
%∆QD is the percentage change in the quantity demanded
%∆D is the shift in demand in percentage terms
%∆P is the percentage change in price
εD is the elasticity of demand
Note that the shift and movement along are in terms of percent changes in quantities
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 18 / 175
![Page 19: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/19.jpg)
Two ways the quantity supplied can change (Math)
Similarly, the quantity supplied can change in two ways:
%∆QS = %∆S︸ ︷︷ ︸Shift
+ εS%∆P︸ ︷︷ ︸Movement Along
%∆QS is the percentage change in the quantity supplied
%∆S is the shift in supply in percentage terms
%∆P is the percentage change in price
εS is the elasticity of supply
Note that the shift and movement along are in terms of percent changes in quantities
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 19 / 175
![Page 20: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/20.jpg)
Unified Framework
What do we know?
1 %∆QD = %∆D + εD%∆P
2 %∆QS = %∆S + εS%∆P
In equilibrium, the change in quantity demanded and supplied have to be the same:
%∆QD = %∆QS
%∆D + εD%∆P = %∆S + εS%∆P
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 20 / 175
![Page 21: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/21.jpg)
Implications for Prices and Quantities
The magnitude of price changes reflect four forces:
%∆P =%∆D −%∆S
εS − εD
We can use this price change to determine the quantity change:
%∆Q = %∆S + εS(
%∆D −%∆S
εS − εD
)%∆Q =
−εD%∆S + εS%∆D
εS − εD
Bottom line: the quantity change is a an elasticity-weighted average of shifts in supply anddemand
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 21 / 175
![Page 22: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/22.jpg)
Application: Rise in Wage Inequality (from D. Autor)
From David Autor. Science 23 May 2014: Vol. 344 no. 6186 pp. 843-851
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 22 / 175
![Page 23: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/23.jpg)
College vs. High-school Gap in Median Earnings (D. Autor)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 23 / 175
![Page 24: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/24.jpg)
What do we actually observe (Katz-Murphy Example)
Quantity
Price
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 24 / 175
![Page 25: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/25.jpg)
Supply has increased, but outpaced by demand
There’s a “race between education and technology” (Goldin and Katz)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 25 / 175
![Page 26: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/26.jpg)
Katz Murphy
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 26 / 175
![Page 27: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/27.jpg)
Overview
logw1t
w2t= α− β log
L1t
L2t+ δt + et (1)
w1tw2t
is relative wages of college-educated workers
L1tL1t
is ratio of college to non-college workers
Goal: Show how to use micro theory to interpret the relationship between relative wagesand the supply and demand of college-educated workers
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 27 / 175
![Page 28: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/28.jpg)
Setup
The standard framework of production given different sets of labor skills assumes
1 Output y is given by y = f (K , h(L1, L2, ...)), where f (K , L) = ALαK 1−α
2 Return on capital r is exogenous3 h(L1, L2, ...) has a nested CES structure
Under the first two assumptions, we have
∂y
∂K= (1− α)ALαK−α = r
⇒ K = L
((1− α)A
r
)1/α
andy
K=
r
(1− α)(2)
Equation 2 shows that K adjusts to match the overall supply of total labor units L,keeping y/K constant and keeping K/L on a trend path that is driven by the rate ofgrowth of TFP.These assumptions are very plausible at a local level (or for “small open economies”) thatthat the price of capital as exogenous.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 28 / 175
![Page 29: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/29.jpg)
Simplifying the production function
Substituting for K , we get
y = ALαK 1−α = A1/α
(1− αr
) 1−αα
L (3)
which is linear in L. Equation 3 shows that under assumptions 1-3, we can ignore capital.
To analyze the effects of relative supply or relative technology changes (i.e., the part oftechnology embedded in h()), we need to specify the labor aggregator function.
A good starting point is a 2-group CES model:
L = h(L1, L2) = (θ1Lσ−1σ
1 + θ2Lσ−1σ
2 )σσ−1 (4)
where θ1 and θ2 are possibly trending over time.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 29 / 175
![Page 30: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/30.jpg)
Marginal Product of Each Group
The marginal product of group 1 is
h1(L1, L2) = θ1L−1σ
1 (θ1Lσ−1σ
1 + θ2Lσ−1σ
2 )1
σ−1 = θ1L−1σ
1 L1σ (5)
Likewise,
h2(L1, L2) = θ2L−1σ
2 L1σ (6)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 30 / 175
![Page 31: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/31.jpg)
Setting Wages equal to marginal products
Assuming w1/w2 = h1/h2 (i.e., MRTS=relative wage), we have:
logw1
w2= log
θ1
θ2− 1
σlog
L1
L2(7)
The slope of the relative demand curve is − 1σ , which is 0 if the two types are perfect
substitutes, and something larger otherwise.
This simple model is widely used to discuss “skill-biased technical change” (SBTC).
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 31 / 175
![Page 32: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/32.jpg)
Katz Murphy
In the traditional SBTC literature (e.g., Katz and Murphy, 1992), it is assumed that
logθ1t
θ2t= a + bt + et (8)
leading to a model for the relationship of relative wages to relative supplies:
logw1t
w2t= a + bt − 1
σlog
L1t
L2t+ et (9)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 32 / 175
![Page 33: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/33.jpg)
Implementation
Freeman (1976) and Katz and Murphy (1992) estimate models of this form, using 2“types” of labor - high-school equivalents and college equivalents.
Dropouts are assumed to be perfect substitutes for HS graduates with a relative efficiencyof (roughly) 70%.
Post-graduates are assumed to be perfect substitutes for college graduates with a relativeefficiency of (roughly) 125%.
People with 1-3 years of college are assumed to represent 1/2 unit of HS labor and 1/2unit of college labor.
(There are different conventions about whether supply should be based on the totalnumbers of adults in each education group, or total employees. There are also differentways to combine men and women).
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 33 / 175
![Page 34: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/34.jpg)
Estimates
The “magic number” is 1σ = 0.7, which implies σ = 1.4 (See KM, equation 19, page 69).
It has turned out to be hard to get a model like (9) to work as well as it did in KM’sstudy (and in Freeman, 1976) when the sample is extended to the 1990s and 2000’s.
Katz and Goldin (2008) present some estimates that have trend breaks in the last twodecades and manage to get estimates in the range of 1
σ = 0.7.
logw1t
w2t= a + bt − 1
σlog
L1t
L2t+ et (10)
See https://economics.mit.edu/files/15391 for more discussion and extensions.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 34 / 175
![Page 35: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/35.jpg)
The Model: Supply and Demand
Quantity traded and price are equilibrium outcomes from a system of simultaneousequations:
qSi = εSpi + ΓSXi + νSi
qDi = εDpi + ΓDXi + νDi
Where:
i indexes different markets, S indexes supply, D indexes demandq is log quantity, p is log priceX is a vector of (pre-determined) observable determinants of demand and supply (including aconstant term){νS , νD
}are unobservable determinants of supply and demand.
Target parameters: εS and −εD
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 35 / 175
![Page 36: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/36.jpg)
We only observe the equilibrium, not supply/demand
Solid and dashed lines represent two different supply/demand systems with different elasticitiesεD1 6= εD2 and εS1 6= εS2 yet observed equilibrium can be rationalized by both systems
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 36 / 175
![Page 37: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/37.jpg)
Endogeneity
Endogeneity – equilibria across multiple markets i ∈ {1, 2, 3} do not trace out either supply ordemand
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 37 / 175
![Page 38: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/38.jpg)
Exclusion Restrictions - Supply shifter
Assume that we observe a variable (ZSi ) that enters the supply equation but is excluded
from the demand equation:
qSi = εSpi + ΓSXi + θSZSi + νSi
qDi = εDpi + ΓDXi + νDi
We further assume:
θS 6= 0 so that quantity supplied is a nontrivial function of ZSt
ZSi ⊥ νSi , νDi |Xi
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 38 / 175
![Page 39: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/39.jpg)
Exclusion Restrictions - Supply shifter
Using variation in ZSi identifies the elasticity of demand by shifting supply along the demand
curve.Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 39 / 175
![Page 40: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/40.jpg)
Exclusion Restrictions - Supply and Demand shifters
Assume that in addition to the supply shifter (ZSi ), we observe a variable (ZD
i ) thatenters the demand equation but is excluded from the supply equation:
qSi = εSpi + ΓSXi + θSZSi + νSi
qDi = εDpi + ΓDXi + θDZDi + νDi
We further assume:
θD 6= 0 so that quantity demanded is a nontrivial function of ZDt
ZDi ⊥ νSi , νDi |Xi
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 40 / 175
![Page 41: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/41.jpg)
Exclusion Restrictions - Supply and Demand shifters
Variation in ZDi (holding ZS
i constant) identifies the elasticity of supply.Variation in ZS
i (holding ZDi constant) identifies the elasticity of demand.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 41 / 175
![Page 42: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/42.jpg)
Supply and Demand shifters - Reduced Form
Solving equations for the equilibrium quantity and price on each market i , we obtain:
qi =εSΓD − εDΓS
εS − εDXi +
εSθDZDi − εDθSZS
i
εS − εD+εSνDi − εDνSiεS − εD
pi =ΓD − ΓS
εS − εDXi +
θDZDi − θSZS
i
εS − εD+νDi − νSiεS − εD
Denote by q∗ and p∗ the residual variation in q and p after partialling out variation in Xi .
Note: q∗i =εSθDZD
i −εDθSZS
i
εS−εD +εSνDi −ε
DνSiεS−εD and p∗i =
θDZDi −θ
SZSi
εS−εD +νDi −ν
Si
εS−εD
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 42 / 175
![Page 43: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/43.jpg)
IV estimates
βIV ,D =Cov
(q∗i ,Z
Si
)Cov
(p∗i ,Z
Si
) =−εDθS
−θS= εD
βIV ,S =Cov
(q∗i ,Z
Di
)Cov
(p∗i ,Z
Di
) =εSθD
−θD= εS
IV recovers the elasticities. In general, we need one instrument for each elasticity.
An interesting exception: When tax rate is an instrument ⇒ a single instrument (taxrate) recovers both elasticities (Gavrilova, Zoutman and Hopland 2018)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 43 / 175
![Page 44: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/44.jpg)
Using tax rates as an instrument
Assume there is an ad valorem tax rate ti imposed on producers. We defineτi = log (1 + ti ).
We also denote by pci the price paid by consumers and by psi = pci − τi the price receivedby suppliers.
We assume τi ⊥ νSi , νDi |Xi
Because the tax is on producers, it does not enter the demand equation ⇒ εD isidentified via standard exclusion restriction.
Economic theory generates an additional exclusion restriction: Ramsey ExclusionRestriction (see GZH 2018)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 44 / 175
![Page 45: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/45.jpg)
Identification of Demand
The tax is a “supply shifter” – it allows identification of εD
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 45 / 175
![Page 46: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/46.jpg)
Tax Rate as an Instrument
The system of equation becomes:
qDi = εDpci + ΓDXi + νDi
qSi = εSpci + θSZSi︸ ︷︷ ︸
=−εSτi︸ ︷︷ ︸=εS(pci −τi)
+ ΓSXi + νSi
Note: We impose an additional restriction – extremely common in public finance – thatsuppliers respond to the tax the same way they would respond to a cost shock(θS = −εS). This directly follows from an assumption of profit maximization.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 46 / 175
![Page 47: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/47.jpg)
Tax Rate as an Instrument - Reduced Form
Solving the previous system of equations for the equilibrium quantity and price on eachmarket i , we obtain:
qi =εSΓD − εDΓS
εS − εDXi +
εSεD
εS − εDτi +
εSνDi − εDνSiεS − εD
pci =ΓD − ΓS
εS − εDXi +
εS
εS − εDτi +
νDi − νSiεS − εD
Denote by q∗ and pc∗ the residual variation in q and pc after partialling out variation inXi .
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 47 / 175
![Page 48: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/48.jpg)
Tax Rate as an Instrument - IV estimate
βIV ,Dτ =Cov (q∗i , τi )
Cov(pc∗i , τi
) = εD
This directly follows from slide 43 and the fact that the tax is excluded from the demandequation (Standard Exclusion Restriction)
Can we identify more than just εD?
Yes, it is the role of the additional restriction that suppliers respond to the tax the sameway they would to an increase in marginal cost (θS = −εS). ⇒ Key implication is that
the passthrough of the tax (to consumers) is dpc
dτ = εS
εS−εD
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 48 / 175
![Page 49: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/49.jpg)
Tax Rate as an Instrument - Identifying εS
Because 1) εD is identified and 2) we can estimate the passthrough dpc
dτ , which is afunction of two elasticities, we can recover εS .
GZH 2018 recommend using the following IV estimator:
βIV ,Sτ =Cov (q∗i , τi )
Cov(ps∗i , τi
) = εS
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 49 / 175
![Page 50: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/50.jpg)
Interpreting regression results with optimizing agents
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 50 / 175
![Page 51: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/51.jpg)
Outline: Optimizing agents and regression results
1 Moneyball
2 The value of police services (example of “adding economics” to a nice reduced-formpaper)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 51 / 175
![Page 52: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/52.jpg)
Firm Optimization
Inputs X with prices w
Output Y = F (X ) with value V
Firm optimization gives us the following marginal condition:
VFi = wi
Value of output × Marginal product = marginal cost of input
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 52 / 175
![Page 53: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/53.jpg)
Implications
VFi = wi
V =wi
Fi
Cost per marginal unit of output must be the same for all factors
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 53 / 175
![Page 54: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/54.jpg)
Application #1: Moneyball
Baseball team
Output is wins
V is value of a win
X are player attributes
w are prices of player attributes
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 54 / 175
![Page 55: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/55.jpg)
Source: Hakes and Sauer (2006). Example from Jesse Shapiro.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 55 / 175
![Page 56: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/56.jpg)
Source: Hakes and Sauer (2006). Example from Jesse Shapiro.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 56 / 175
![Page 57: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/57.jpg)
#2 Optimization and the Value of Police Services
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 57 / 175
![Page 58: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/58.jpg)
Mello (2018) “More COPS, Less Crime”
1 Policy relevant question: What is the effect of additional police on local criminalactivity?
2 Nice variation: ARRA police funding index increased COPS in some locations
3 Interesting Results:Average grant increased police by 0.7 per 10,000 residents (or 6% increase in police)Each officer reduces 4.3 violent crimes and 15.4 property crimesBenefit of $35 per resident vs $29 cost
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 58 / 175
![Page 59: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/59.jpg)
Economic Questions
This paper focuses on reduced-form effects, and provides a good example that shows how wecould think about optimization and the following questions:
1 What is the value of a marginal police officer?
2 How many police officers should we hire?
3 How should they be allocated? Should police focus more on violent crime?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 59 / 175
![Page 60: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/60.jpg)
Economic Framework
Local governments produce safety
y = f (L)
y are units of safety
L is number of police officers
Local governments maximize:
maxL
pf (L)− wL
p is the value of a unit of safety
w is wage of police officers
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 60 / 175
![Page 61: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/61.jpg)
Should we hire more police officers?
FOC
pf ′(L) = w
pf ′(L) is the marginal value of safety
w is the marginal cost of safety
Estimates suggest that pf ′(L) > w
Estimate of marginal benefit from Table 2 is $35.2 per 10K residents
Direct cost is roughly $29 per 10K residents
⇒ L < L∗
Keep hiring police until these are equal! (But also need to account for cost of raisingfunds).
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 61 / 175
![Page 62: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/62.jpg)
Economic framework with two types of crimes
Two types of safety y :
safety from violent crime y1
safety from property crime y2
Local governments maximize:
maxL1,L2
p1f (L1) + p2g(L2)− w(L1 + L2)
p1 is the value of a unit of safety from violent crime
p2 is the value of a unit of safety from property crime
L1 is number of police officers allocated to reducing violent crime
L2 is number of police officers allocated to reducing property crime
Note main outcome in paper is approx $68, 000× y1 + $4, 000× y2
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 62 / 175
![Page 63: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/63.jpg)
Optimal policing of violent crime?
FOC for violent crimes:
p1f′(L1) = w
p1 is approx $68,000
f ′(L1) = 4.3, i.e., hiring one more officer reduces # of violent crimes by 4.3
Marginal benefit is 4.3× $68, 000 ≈ $292, 400
If local governments are optimizing, then
f ′(L1)︸ ︷︷ ︸Marginal product
=w
$68, 000
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 63 / 175
![Page 64: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/64.jpg)
Optimal policing of property crime?
FOC for property crimes:
p2g′(L2) = w
p2 is approx $4,000
f ′(L2) = 15.4, i.e., hiring one more officer reduces # of property crimes by 15.4
Marginal benefit is 15.4× $4, 000 ≈ $61, 600
If local governments are optimizing, then
g ′(L2)︸ ︷︷ ︸Marginal product
=w
$4, 000
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 64 / 175
![Page 65: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/65.jpg)
Should police focus more on violent crime reduction?
FOCs for violent and property crimes:
p1f′(L1) = w
p2g′(L2) = w
But p1f′(L1) = $292, 000 > p2g
′(L1) = $62, 000
If local governments are optimizing, then
$68, 000︸ ︷︷ ︸Value of output
=w
f ′(L1)︸ ︷︷ ︸cost per marginal unit of output
$4, 000︸ ︷︷ ︸Value of output
=w
g ′(L2)︸ ︷︷ ︸cost per marginal unit of output
Seems like police should focus more on violent crime given p1 and p2Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 65 / 175
![Page 66: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/66.jpg)
Regional VariationShould the per capita size of the police force vary across locations?
FOC
pc f′(Lc) = wc
pc f′(Lc) is the marginal value of safety in location c
wc is the marginal cost of safety in location c
Would be interesting to analyze heterogeneity based on variation in
Initial force size Lc varies (so can trace out f ′(Lc))
Local cost of safety wc
Local value for safety pc can vary
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 66 / 175
![Page 67: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/67.jpg)
Demand for safetyWhere do the estimates of p1 and p2 come from?
Resident utility depends on level of safety and other consumption:
maxx ,yU(x , y) s.t. pyy + pxx = M
y is units of safety
x is a composite of other goods
M is income (and λ is MU of income)
FOC: ∂U∂y = λpy suggests that:
Marginal utility of safety depends on level of safety (so level of L)
Value of safety∂U∂y
λ is increasing in income (since λ is decreasing in M)
Thus, py should depend on level of L and local incomes
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 67 / 175
![Page 68: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/68.jpg)
Other thoughts
Estimates are interesting inputs for welfare analysis of an important non-traded good
1 Welfare analysis Could think about effective cost w that includes overhead and MCPFthat would rationalize current hiring levels
2 Time allocation Could weigh into debates about how police spend their time (violentcrime vs property crime)
3 Supply side Could learn more about production function of safety f (L) and g(L)
4 Demand side Could think more about value of unit of safety and the efficiency vs equityconsiderations of how police spending is allocated
5 Evaluating current police spending What social welfare function and/or cost of publicfunds are consistent with the level and allocation?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 68 / 175
![Page 69: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/69.jpg)
Potential Outcomes and Selection
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 69 / 175
![Page 70: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/70.jpg)
Causality, Potential Outcomes, and Selection
1 Potential Outcomes
2 Selection Bias
3 Example: return to selective colleges
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 70 / 175
![Page 71: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/71.jpg)
Motivation
Counterfactual questions are everywhere
What would happen if a job training program were expandedWhat would happen to prices/welfare if two firms marged?What would different monetary policy do to real output?What effect would this medication have on heart diseaseWhat will happen to global temps if emissions decrease?
Causal inference
Thinking about a counterfactual requires thinking about causalityTheory alone might (might) tell us the direction of causality
Even when it does, it will rarely tell us the magnitude
Causal inference uses data to address counterfactuals
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 71 / 175
![Page 72: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/72.jpg)
Potential Outcome Notation
Also known as the Neyman-Fisher-Roy-Quandt-Rubin causal model
D is a mutually exclusive and exhaustive set of states (“treatments”) e.g. training/notraining D = {0, 1}, prices D = [0,+∞), etc.
For each d ∈ D, there is a potential outcome Yd (a random variable)
Yd is what would have happened if the state were d
We observe the actual state, a random variable D ∈ DWe also observe an outcome Y , related to potential outcomes as
Y =∑d∈D
Yd1 [D = d ] = YD
binary case: Y = DY1 + (1− D)Y0(”switching regression”)
Y = YD is observed, but Yd for d 6= D are unobserved
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 72 / 175
![Page 73: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/73.jpg)
The fundamental problem of causal inference (Holland, 1986)Y = DY 1 + (1 − D)Y 0
Causal inference requires assumptions/restrictions on the “missing” potential outcomes
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 73 / 175
![Page 74: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/74.jpg)
What Do We Want to Measure?
We are interested in counterfactuals, Yd for d 6= D
These variables capture the “what if” aspect of causality
Since they are random variables, they can be summarized in many ways
That is, there are many possible parameters of interest
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 74 / 175
![Page 75: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/75.jpg)
Example: Program Evaluation
Suppose d ∈ {0, 1} indicates participation in a job training program
Y is a scalar labor market outcome such as earnings
If D = 1 we observe Y1 (but not Y0) and if D = 0 we observe Y0
There are many possible questions one could ask:
What would be average earnings if everyone were trained, i.e. E [Y1]?What is the average effect of the program, i.e. E [Y1 − Y0]?What about only for those who are trained, i.e. E [Y1 − Y0|D = 1]?
What is useful depends on what question we want to answer!
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 75 / 175
![Page 76: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/76.jpg)
Latent Variable Notation
Many empirical models in economics look like a special case of:
Y = g (D,U) ,
where g is a function and U are unobservable variables
A causal interpretation of this model is implicitly saying:
Yd = g (d ,U) for every d ∈ D
This could impose assumptions, depending on what g and U are
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 76 / 175
![Page 77: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/77.jpg)
Going from potential outcomes to parameters of interest
Three considerations/ pillars of econometrics:
Identification (what can be learned?)
Estimation (how best to learn it?)
Inference (how certain am I?)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 77 / 175
![Page 78: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/78.jpg)
Identification
The parameter of interest is a function of the unobservables. In the PO model, this issome function of {Yd}d∈DWhat could we learn about this function from the observables, (Y ,D)?
This is the question of identification
Identification can be seen as the link between data and theory
By theory here I mean the way in which we believe the world works
This can include what we think of as “economic theory”
But it need not be so formal – a synonym would be assumptions
Theory is often ambiguous – otherwise we wouldn’t need data
Identification always involves a (possibly simple) model
This model encodes our assumptions (theory)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 78 / 175
![Page 79: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/79.jpg)
Identification vs. Statistical Inference
In practice, we only see a finite sample of the observables {{Yi ,Di}}ni=1
From this we know the sample distribution
However, we don’t know the population distribution of (Y ,D)
Statistical inference is using the sample to learn about the population
It is useful to separate identification from statistical inference:
sample →︸︷︷︸statistical inference
population →︸︷︷︸identification
unobserved parameters
The second arrow is logically the first thing to consider
Can’t recover a parameter when we know the population distribution? Then you alsocouldn’t recover it with the sample distribution!
Informally, identification is sometimes seen as having “infinite data”
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 79 / 175
![Page 80: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/80.jpg)
Identification is Prior to Statistical Inference
Return to the example of job training and earnings
Suppose we care about the average effect of the program on participants:
ATT = E [Y1 − Y0|D = 1] = E [Y |D = 1]︸ ︷︷ ︸fnc. of pop. dist.
− E [Y0|D = 1]︸ ︷︷ ︸fnc. of unobs.
An important ingredient in a decision to continue or end the program
The first term is a function of the population distribution
Using the sample to understand this from data is the domain of statistics
The question of identification is about the second term
What can we say about E [Y0|D = 1].
under different assumptions?
Must answer this question before we can construct an estimate of ATT
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 80 / 175
![Page 81: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/81.jpg)
Selection Bias
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 81 / 175
![Page 82: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/82.jpg)
Selection
There is selection into the treatment state D if
Yd |D = d︸ ︷︷ ︸observable
is distributed differently from Yd |D = d ′︸ ︷︷ ︸unobserved
for d ′ 6= d
This is not the case under the random assignment assumption
Expected to occur if agents choose D with knowledge of {Yd}d∈DSelection is common
Particularly concerning if you are trained in neoclassical economicsAgents choose a job training program (D ∈ {0, 1}) to max utilityUtility will incorporate expected future earnings (Y0,Y1)Agents who choose job training might do so because of low Y0
Data typically supports this story (“Ashenfelter’s (1978) dip”)Alternatively, might choose D = 0 because of high Y0
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 82 / 175
![Page 83: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/83.jpg)
Selection Bias
Consider the simple treatment/control mean contrast under selection
This contrast would be the ATE under random assignment
Decompose the contrast into a causal effect and selection bias:
E [Y |D = 1]− E [Y |D = 0]
= E [Y1|D = 1]− E [Y0|D = 1]︸ ︷︷ ︸ATT
+ E [Y0|D = 1]− E [Y0|D = 0]︸ ︷︷ ︸selection bias
= E [Y1|D = 0]− E [Y0|D = 0]︸ ︷︷ ︸ATU
+ E [Y1|D = 1]− E [Y1|D = 0]︸ ︷︷ ︸selection bias
First term is causal effect for those who were treated/untreatedUnder random assignment would have ATT = ATU = ATE
Second term is how the treated would have been different anywayUnder random assignment this would be 0
The first expression is more natural if thinking of D = 0 as baseline
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 83 / 175
![Page 84: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/84.jpg)
Selection on Observables
A simple relaxation of random assignment is selection on observables
Suppose that we observe (Y ,D,X ) where X are covariates
The selection on observables assumption is that
{Yd}d∈D ⊥ D|X
Says: Conditional on X , treatment is as-good-as randomly assigned
Other terms: unconfoundedness, ignorable treatment assignment
Underlies causal interpretations of linear regression
We will look into this connection more later
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 84 / 175
![Page 85: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/85.jpg)
Pre-Determined Covariates
Advice - only condition on predetermined observables
For selection on observables to be plausible, X should be predetermined
In particular, D should not have a causal effect on X
Usually this really is a temporal issue (measured before vs. after D)
Intuition is clear – we want to condition on selection into treatment
Simple but trivial example
Suppose we accidentally included Y as part of X
Then clearly we aren’t going to have (Y0,Y1) ⊥ D|XLess trivial examples
Don’t include earnings 1 year after the program in X
Don’t include employment after the program in X
Don’t include marital status after the program in X
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 85 / 175
![Page 86: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/86.jpg)
Imputation Estimator for the ATE
Suppose D ∈ {0, 1} and recall the first point identification result:
ATE = E︸︷︷︸over X
[E [Y |D = 1,X ]− E [Y |D = 0,X ]]
Suppose we have an i.i.d. sample of data {(Yi ,Di ,Xi )}Ni=1
A natural imputation estimator is given by:
ATE =1
N
N∑i=1
µ1 (Xi )− µ0 (Xi )
where:µ1 (x) is an estimator of µ1 (x) ≡ E [Y |D = 1,X = x ]µ0 (x) is an estimator of µ0 (x) ≡ E [Y |D = 0,X = x ]
Estimate conditional means, then take the sample analog
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 86 / 175
![Page 87: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/87.jpg)
Imputation Estimator for the ATT
The ATT is actually easier to estimate, since
ATT = E [Y1 − Y0|D = 1]
= E [Y |D = 1]− E [µ0 (X ) |D = 1] = E [Y − µ0 (X ) |D = 1]
An imputation estimator here would be
ATT =1
N1
∑i :Di=1
Yi − µ0 (Xi ) with N1 =N∑i=1
Di
Imputation of the first term is simplified (sample average)
Using control group to impute control outcomes for treated groupNo need to do the opposite, so don’t need to estimate µ1
Notice that the sample average weights X according to the treated groupSo, same issue remains (how to estimate µ0), but a bit simpler
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 87 / 175
![Page 88: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/88.jpg)
A Slightly More Complicated Imputation Estimator
Using insight from the ATT imputation, we could estimate ATE with
ATE =1
N
N∑i=1
Di (Yi − µ0 (Xi )) + (1 − Di ) (µ1 (Xi ) − Yi )
=N1
N︸︷︷︸≈P[D=1]
1
N1
∑i :Di =1
Yi − µ0 (Xi )
︸ ︷︷ ︸
≡ATT
+N0
N︸︷︷︸≈P[D=0]
1
N0
∑i :Di =0
µ1 (Xi ) − Yi
︸ ︷︷ ︸
≡ATU
Makes it clear that we know Yi = Ydi for Di = d
No need to impute Y1i for Di = 1 or Y0i for Di = 0
The two forms of imputation will be numerically identical if:
1
Nd
∑i :Di =d
µd (Xi ) =1
Nd
∑i :Di =d
Yi – often the case, e.g. linear regression
In either form, primary problem is estimating µ0 and/or µ1
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 88 / 175
![Page 89: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/89.jpg)
Parametric Estimators
We could also just use the good ol’ fashioned parametric model, e.g.,
µd (x) = αd + β′dx
Estimate αd , βd by regressing Y on X among D = d subpopulations then
ATE =1
N
N∑i=1
[α1 + β1
′Xi − α0 − β0
′Xi
]= Y1 − Y0︸ ︷︷ ︸
Naive contrast
+
(N1
Nβ0 +
N0
Nβ1
)′(X0 − X1
)︸ ︷︷ ︸
regression adjustment
which can be shown after noting αd = Yd − β′d Xd and rearranging terms
Widely used approach, but considered poor taste by many
Concern about functional forms driving results via extrapolation
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 89 / 175
![Page 90: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/90.jpg)
Connection to Linear Regression
Nevertheless, linear regression is the most widely used approach
In fact, by far the most widely used specification looks like this:
µd (x) = αd + x ′β ≡ α0 + (α1 − α0) d + x ′β
Under selection on observables, this implies that
E [Y |D,X ] = Dµ1 (X ) + (1− D)µ0 (X ) = α0 + (α1 − α0)D + X ′β
This is restrictive and implies constant treatment effects:
E [Y1 − Y0|X = x ] = α1 − α0 does not depend on x
In contrast, the specification on the previous slide had
E [Y1 − Y0|X = x ] = α1 − α0 + (β1 − β0)′x still depends on x
Can be implemented as a single regression of Y on 1, D, X and DX
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 90 / 175
![Page 91: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/91.jpg)
Example: returns to selective colleges
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 91 / 175
![Page 92: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/92.jpg)
Dale and Krueger (2002)
Question, empirical setting
The impact of selective colleges on labor market outcomes
Compare outcomes of those who attend selective vs. non-selective?
Clearly unlikely to be a causal effect
Even after conditioning on observables such as GPA
Unobservables play a large role in college admissions
Methodology
Match (condition) on the set of colleges to which amitted
Key variable which they use to argue causality
Assumption is selection on a (carefully-chosen) observable
Use linear regression (without interactions)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 92 / 175
![Page 93: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/93.jpg)
The Empirical Strategy
Stylized idea
UPenn (selective) vs. Penn State (less selective)
Condition on students with similar observables
GPA, sex, race, athlete, high school rank
Find students who were only admitted to UPenn and Penn State
Compare log wages of those who attended UPenn vs. Penn State
Practical challenges
Defining selectivity → they use average SAT
Implementing the grouping above more generally
Main estimates use average SAT instead of school identity...
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 93 / 175
![Page 94: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/94.jpg)
Grouping Scheme
Also report estimates that group using school identities directlyAnd estimates that group using Barron’s selectivity ranking (1-5)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 94 / 175
![Page 95: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/95.jpg)
Main Results
Columns 1-2: Raw coefficient is positive and large
Columns 3-5: Match on schools admitted to
Column 6: Match on schools applied to – they argue upward bias here
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 95 / 175
![Page 96: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/96.jpg)
Possible Explanations and Critiques
Explanations1 More likely to enter academia/public sector at a more selective college
They check this by including controls for occupationNot a good strategy – occupation is an outcome (not predetermined)
2 “Big fish, small pond” phenomenon
Supported by changing outcome variable to college class rankHeterogeneity by parental income – find benefis for low income
Critiques
Not interacting groups with SAT imposes constant effects
Selectivity has the same labor market impact for any choice set?
SAT is too coarse a measure of school quality
Replacing SAT by dummies for schools they strongly reject 0 effect
Students might still be sorting on unobservables within groups
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 96 / 175
![Page 97: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/97.jpg)
Roy Model and Selection Corrections
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 97 / 175
![Page 98: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/98.jpg)
Roy Model of Self-Selection
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 98 / 175
![Page 99: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/99.jpg)
Roy Model of Self-Selection
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 99 / 175
![Page 100: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/100.jpg)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 100 / 175
![Page 101: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/101.jpg)
Roy Model of Self-Selection
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 101 / 175
![Page 102: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/102.jpg)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 102 / 175
![Page 103: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/103.jpg)
Roy Model of Self-Selection
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 103 / 175
![Page 104: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/104.jpg)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 104 / 175
![Page 105: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/105.jpg)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 105 / 175
![Page 106: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/106.jpg)
Roy Model: Parametric example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 106 / 175
![Page 107: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/107.jpg)
Roy Model: Parametric example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 107 / 175
![Page 108: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/108.jpg)
The basic Roy model and selectionModel of College Education
Suppose you are interested in the benefit of College Education (D = 1) relative to nothaving College Education (D = 0)
For each individual you observe realized wage:
Y = DY1 + (1− D)Y0
Where:
Y1 = Xβ1 + U1
Y0 = Xβ0 + U0
D = 1 (Y1 > Y0)(U1
U0
)∼ N
((00
),
(σ2 ρσρσ 1
))Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 108 / 175
![Page 109: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/109.jpg)
Using Roy model to think about returns to college
Note:
U0 − U1 ∼ N(0, σ2 + 1− 2ρσ
)Cov (U1,U0 − U1) = ρσ − σ2
Cov (U0,U0 − U1) = 1− 2ρσ
Decision rule:
D = 1 (Y1 > Y0)
= 1 (Xβ1 + U1 > Xβ0 + U0)
= 1 (X (β1 − β0) > U0 − U1)
Implies:
P (D = 1|X ) = P (X (β1 − β0) > U0 − U1) = Φ
(X (β1 − β0)√σ2 + 1− 2ρσ
)Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 109 / 175
![Page 110: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/110.jpg)
The basic Roy model, selection and target parametersCollege Education: Treatment parameters conditional on X
ATE = E (Y1 − Y0|X ) = X (β1 − β0)
ATT = E (Y1 − Y0|X ,D = 1)
= E (Xβ1 + U1 − Xβ0 − U0|X ,X (β1 − β0) > U0 − U1)
= X (β1 − β0)− E (U0 − U1|X ,U0 − U1 < X (β1 − β0))
= X (β1 − β0) +√σ2 + 1− 2ρσ
φ
(X (β1−β0)√σ2+1−2ρσ
)Φ
(X (β1−β0)√σ2+1−2ρσ
)︸ ︷︷ ︸
>0
Intuition: Those who select into college benefit from it
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 110 / 175
![Page 111: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/111.jpg)
The basic Roy model, selection and target parameters
ATU = X (β1 − β0)− E (U0 − U1|X ,U0 − U1 ≥ X (β1 − β0))
= X (β1 − β0)−√σ2 + 1− 2ρσ
φ
(X (β1−β0)√σ2+1−2ρσ
)1− Φ
(X (β1−β0)√σ2+1−2ρσ
)︸ ︷︷ ︸
<0
Intuition: individuals do not select into college because they do not benefit from it
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 111 / 175
![Page 112: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/112.jpg)
The basic Roy model, selection and target parameters
Graphical intuition for the sign of the selection bias (the expectation of truncated normal):
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 112 / 175
![Page 113: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/113.jpg)
What is a “Policy Relevant” Parameter?
The MTE framework partitions all agents in a clear way
Provides a foundation for thinking about “ideal” treatment effects
The “ideal” treatment effect clearly depends on the question
The ATE receives a lot of attention in the literature
But not very useful for policy – can agents still choose D?
The ATT is somewhat clearer in this regard
Loss in benefit to treated group from discontinuing D = 1
Perhaps more relevant is changing the agent’s choice problem
For example, D ∈ {0, 1} is attending a four-year college
Average effect of forcing college/no college (ATE) is not interesting
Nor is the effect on college-goers of shutting down college (ATT)
More interesting are the effects via D of adjusting tuition Z
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 113 / 175
![Page 114: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/114.jpg)
Decomposition methods and the Pay Gap
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 114 / 175
![Page 115: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/115.jpg)
Decomposition Methods
Decomposition methods are traditionally used to separate differences between groups intoa component explained by observables and an unexplained component
This is conceptually similar to decomposing a treatment/control difference into acomponent explained by controls (bias) and an unexplained component (treatment effect)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 115 / 175
![Page 116: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/116.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 116 / 175
![Page 117: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/117.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 117 / 175
![Page 118: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/118.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 118 / 175
![Page 119: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/119.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 119 / 175
![Page 120: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/120.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 120 / 175
![Page 121: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/121.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 121 / 175
![Page 122: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/122.jpg)
Oaxaca-Blinder Decompositions
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 122 / 175
![Page 123: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/123.jpg)
APPENDIX:Example of pay-gap between natives and immigrants
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 123 / 175
![Page 124: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/124.jpg)
Oaxaca-Blinder Decomposition Example from David Card
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 124 / 175
![Page 125: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/125.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 125 / 175
![Page 126: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/126.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 126 / 175
![Page 127: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/127.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 127 / 175
![Page 128: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/128.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 128 / 175
![Page 129: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/129.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 129 / 175
![Page 130: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/130.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 130 / 175
![Page 131: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/131.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 131 / 175
![Page 132: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/132.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 132 / 175
![Page 133: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/133.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 133 / 175
![Page 134: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/134.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 134 / 175
![Page 135: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/135.jpg)
Oaxaca-Blinder Decomposition Example
Graduate Public Finance (Econ 524) Applied Tools Lecture 1b 135 / 175
![Page 136: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/136.jpg)
Outline
1 Preliminaries: research designs, DGP, and applied modeling
2 Connecting theory and dataUsing supply and demandInterpreting regression results with optimizing agents
3 Potential Outcomes and SelectionSelection BiasExample: returns to selective colleges
4 The Roy Model and Selection Corrections
5 Decomposition methods and the Pay Gap
6 Review of 3 Applied Econometrics ToolsDifference-in-differencesEvent StudiesDiscrete Choice
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 136 / 175
![Page 137: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/137.jpg)
Difference-in-differences
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 137 / 175
![Page 138: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/138.jpg)
Difference-in-DifferencesIdea is to use population unaffected by a program as a longitudinal control group for a treated population
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 138 / 175
![Page 139: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/139.jpg)
Difference-in-Differences
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 139 / 175
![Page 140: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/140.jpg)
Difference-in-Differences
The identifying assumption is that the change in the control group outcomes serves as avalid proxy for the change in the treatment group outcomes
For this reason, the most convincing DDs exploit data from many periods
Consider an example of Davis (2004) studying the effects of the emergence of a cancerCluster in Churchill County in 2010
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 140 / 175
![Page 141: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/141.jpg)
Treatment versus Control (Davis 2010)Control groups (Lyon County and NV) rarely deviated from treatment group
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 141 / 175
![Page 142: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/142.jpg)
Difference-in-Differences (Davis 2010)Compelling design because T-C difference has same long run mean and stable pre-treatment behavior
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 142 / 175
![Page 143: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/143.jpg)
Difference-in-Differences: Issues with common trendsWe worry about source of the trend (which is not observable) and whether it is stable
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 143 / 175
![Page 144: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/144.jpg)
Difference-in-Differences: formalizing common trend assumptionWe worry about source of the trend (which is not observable) and whether it is stable
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 144 / 175
![Page 145: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/145.jpg)
Difference-in-Differences: Potential outcome assumption
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 145 / 175
![Page 146: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/146.jpg)
Difference-in-Differences
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 146 / 175
![Page 147: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/147.jpg)
Interpretation in Functional Form
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 147 / 175
![Page 148: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/148.jpg)
Regression Implementation
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 148 / 175
![Page 149: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/149.jpg)
Event studies
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 149 / 175
![Page 150: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/150.jpg)
Event studies
Event studies are generalizations of the DD design where different units are treated atdifferent times
Started in finance. Typically looked at excess returns, which are deviations of stockreturns over a level implied by some stock market index (in practice, one usually runs aregression of individual returns on a market average and takes a residual)
Simply examined what happens to the mean value of excess returns in the neighborhoodof a financial event
The basic idea — of re-ordering a panel into event time — spilled over to evaluationliterature
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 150 / 175
![Page 151: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/151.jpg)
Event studies
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 151 / 175
![Page 152: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/152.jpg)
Event studies
The βk can then be plotted over time and provide estimates of mean outcomes in “eventtime” after having taken out the individual and year specific effects.
Jacobson, LaLonde, and Sullivan (1993) is a famous example. Look at effects of job losson earnings
JLS embellish the model to include individual-specific trends and allow for interactionsbetween individual level characteristics and the event dummies
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 152 / 175
![Page 153: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/153.jpg)
Event studies: Jacobson, LaLonde, and Sullivan (1993)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 153 / 175
![Page 154: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/154.jpg)
Event studies
Note that with a common event data ei = e, then the event study specification simplifiesto a standard difference in differences model and the coefficients βk merely plot thebehavior of outcomes in the treatment group relative to the control group before andafter treatment
When treatment dates vary, then the ES approach is potentially a more efficient means ofpooling together several different DDs, even in cases where all the units get treated.
ES compares changes in outcomes of treated groups to both units that have not yet beentreated and units that will never be treated – good to check if those two sets of controlsare in fact exchangeable. (one can do this by re-estimating the model without the nevertreated units and seeing how estimates change
Good practice to start simple and compare means of treatment and control groups inevent time. Then add controls
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 154 / 175
![Page 155: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/155.jpg)
Event studies: other tips
Units that are treated early will have fewer pre-treatment obs and units that are treatedlate will have few post treatment obs, so that’s why binning end points can make sense
Alternatively, sometimes people focus on a balanced panel
Note if you do not have any never treated controls, you will not be able to include alldummies so will have to normalize one of the event coefficients to zero, which JLS do(and -1 in event time is standard)
It is common to find one thing in levels and another in logs because (in part) these areparametric models and linearity assumptions are not innocuous. Important to try differentspecifications of the time effects γt
Greenstone Hornbeck and Moretti (2010) compute ES estimates on a treatment sample(which wins MDP) and a control sample (which narrowly lose them) and then take thedifference to obtain more credible estimates of impacts. This extra difference is roughlyequivalent to a DDD model
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 155 / 175
![Page 156: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/156.jpg)
Discrete choice
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 156 / 175
![Page 157: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/157.jpg)
Consumers decide whether or not to buy
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 157 / 175
![Page 158: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/158.jpg)
Consumers decide whether or not to buy
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 158 / 175
![Page 159: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/159.jpg)
Consumers decide whether or not to buy
The first graph shows the share of consumers buying a product is 50% when it’s price is $5
The second graph shows the share of consumers buying a product is 30% when it’s priceis $6
How can we think about how responsive demand will be to changes in price whenconsumers are making discrete (i.e., buy or not) choices?
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 159 / 175
![Page 160: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/160.jpg)
Analytical Setup
Suppose that individual i buys if her value exceeds the price, i.e., buy if vi > P
This value can be a function of common things (e.g., income, credit conditions, etc) oridiosyncratic tastes but at this stage, specifying what is in vi doesn’t matter. The fractionof people who buy is:
Prob(Q = 1) = P(vi > P) (11)
= 1− F (P) (12)
where F (x) is the c.d.f. of vi . Note this is why the demand curve looks like a CDFrotated clockwise 90 degrees
A c.d.f. describes the probability that a real-valued random variable X with a givenprobability distribution will be found to have a value less than or equal to x
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 160 / 175
![Page 161: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/161.jpg)
Elasticity of Demand
What is the elasticity of this curve?
Q(P) = N(1− F (P)) (13)
where N is the size of the population (e.g., number of potential consumers in your market)
εD =dQ(P)
dP
Q
P(14)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 161 / 175
![Page 162: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/162.jpg)
Elasticity of Demand
What is the derivative?
dQ(P)
dP= −Nf (P) (15)
where N is the size of the population (e.g., first time home buyers in an area)
f(x) is the probability density function (p.d.f.)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 162 / 175
![Page 163: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/163.jpg)
Elasticity of Demand
εD =dQ(P)
dP
P
Q(16)
= −Nf (P)P
N(1− F (P))(17)
=−f (P)
1− F (P)P (18)
What matters for responsiveness?
Fraction of people at the margin f (P)
Fraction of people already buying 1− F (P)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 163 / 175
![Page 164: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/164.jpg)
From $5, a $1 dollar increase in price ⇓ demand by 20%
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 164 / 175
![Page 165: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/165.jpg)
From $8, a $1 dollar increase in price ⇓ demand by 2%
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 165 / 175
![Page 166: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/166.jpg)
Elasticity of Demand: In words
Takeaways:
For very homogeneous populations, you’ll have very elastic demand
If tastes are more spread out, you’ll see smaller responses
At the extreme in which everyone is the same, demand will be a step function, so there issome price above which no one will buy and below which everyone will buy.
In this case, things will be very inelastic at high prices, but very elastic near the price, andthen unresponsive at very low prices
Thinking about consumer choice in this way will be helpful for evaluating how effectivesales can be
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 166 / 175
![Page 167: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/167.jpg)
Demand if V ∼ N(µ, σ)
N
P
D(P)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 167 / 175
![Page 168: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/168.jpg)
Demand if V ∼ U(A,B)
A
B
•
•
P
D(P)
→ At B, no one buys
→ At A, everyone buys
Back
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 168 / 175
![Page 169: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/169.jpg)
Discrete Choice more generally: Di = argmaxjUij
See Discrete Choice Methods with Simulation by Ken Trainhttps://eml.berkeley.edu/books/choice2.html.
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 169 / 175
![Page 170: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/170.jpg)
Discrete Choice more generally: Di = argmaxjUij
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 170 / 175
![Page 171: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/171.jpg)
Discrete Choice more generally:
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 171 / 175
![Page 172: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/172.jpg)
Discrete Choice: probit (normal distribution)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 172 / 175
![Page 173: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/173.jpg)
Discrete Choice: logit (logistic distribution)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 173 / 175
![Page 174: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/174.jpg)
Discrete Choice: logit (logistic distribution)
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 174 / 175
![Page 175: Applied Tools: Combining Theory and Data to Evaluate Policy](https://reader033.vdocuments.mx/reader033/viewer/2022060603/629690281421e73e74197b85/html5/thumbnails/175.jpg)
Ken Train has an excellent book on discrete choice
For more detail, see Discrete Choice Methods with Simulation by Ken Train
The book is very readable and is freely available online:https://eml.berkeley.edu/books/choice2.html
Graduate Public Finance (Econ 524) Applied Tools Lecture 1c 175 / 175