• Download

• Fastfood.dta and do
• Injury.dta and do
• Kielmc.DTA

1
ECONOMETRÍA APLICADA
AVANZADA
Cristina Tello-Trillo

Semana 2
Lecture 2 - Introduccion

INTRODUCCION

3
 Bienvenidos a la Semana 2!
• Objetivo:

• En esta clase vamos a aprender uno de los

métodos mas importantes para el análisis
empírico: Diferencias-en-Diferencias

4
 Como vamos a aprender?
• La clase de hoy y manana sera divida entre
teoria/conceptos y aplicaciones empiricas del
‘real world’.

• Todos deberian tener su laptop con STATA.

5
 Conociéndonos
• Economista empirical aplicada
• Trabajo en el Centro de Estudios Economicos en el U.S.
Census Bureau – Ministerio de Comercio
• Profesora Adjunta en John Hopkins University y University of
Maryland.
• Pre-grado en la PUCP (alma mater!)
• M.A. y PhD en Economia en Yale University
• Trabajo en temas de comercio internacional y mercado de
trabajo:
– Como el comercio con China afecta los salarios, quienes
son los mas afectados, redistribucion de trabajo en
sectores.
– Como politicas ‘favorable a la familia’ puede afectar a las
decisiones laborales de la firma.
– Determinantes de la productividad de la firma. 6
 Evaluacion
• Controles de lectura previos a cada clase (30%)
• Propuesta de Investigación (40%)
• Participación en clase (20%)
• Ejercicios o Laboratorios calificados (10%)
– Problem set 1 (Fecha de entrega: Enero 8)
– Problem set 2 (Fecha de entrega: Enero 15)

7
 This Class
• Natural Experiments
• Difference-in-Differences
• Applications
• Class exercise

8
Lecture 2 – Natural Experiments & DD

NATURAL (QUASI)
EXPERIMENTS

9
 Why use Natural (Quasi)
Experiments?

• To solve for the endogeneity problem.

10
 OLS assumptions
• A1 Linearity in parameters:
• A2 No multi-collinearity
• A3 Exogeneity of the independent variables
– -> ->
– No information on X tell us something about the expected
value of the
– The causal interpretation of the coefficients are valid.
• A4 Random sampling of observations
• A5 Spherical errors: Homoscedasticity & no
autocorrelation:
–
–
• A6: Error terms should be normally distributed.
11
 Exogeneity
•
• The exogeneity assumption which
implies means that the
regressor is exogenous.
• A violation of this condition leads to an
endogeneity problem which precludes our
ability to make causal inferences (OLS is not
BLUE).

12
 Reasons for Endogeneity
• Measurement error
• Simultaneity bias: running quantity on prices
when estimating a demand equations.
• Omitted Variables: one or more variables in
X are not included in the regression because no
data on those variables are available –
estimation will be altered to the extent that the
missing variable and the included ones are
correlated.

13
Example (returns to schooling):
•
– X: observed factors that determine earnings, as
work experience, gender, family background
– U : unobserved factors that determine earnings,
as individual ability
•
– where
– A person level of education is (at least partially)
determined by a person’s ability
– High ability individuals do better in school and
therefore choose to attain a higher level of
education, and their high ability is the
fundamental reason for their high wages.
– wil not identify the causal effect of one
additional year on education on earnings
14
 OLS Assumptions:
Exogeneity
Violation of this assp:
Endogeneity; Causal effect
not properly identify

Solution to Endogeneity: Solution to Endogeneity:

Instrumental Variables, Panel Data Fixed Effects,
instrument for X include fixed effects
dummies

Solutions to Endogeneity:
Differences-in-Differences (changing
the framework of analysis)

15
• Another way to think about endogeneity is that it
indicates that the regressor is non-random/non-
independent.
– If a regressor X was completely random it shouldn’t be
correlated with the error term.
– Something that is random is not cause by anything and
has not predictability of correlation.

17
 Randomized experiments
• In many of the “hard” sciences, the researcher
can simply design experiment to achieve the
necessary randomness.
– E.g. To determine the effect of a new drug, you
randomly give it to certain patients.
– E.g To determine the effect of modifying certain
gene, you modify it in a random sample of mice.

19
 But, it’s hard to do this in Economics
• Ethical, financial and practical reasons.
– E.g. we can’t randomly assign education to a random
group of people.
– E.g. we can’t randomly assign higher incarceration rate to
some states versus others just for research purposes.
• Therefore we need to rely on what we call “Natural (or
quasi) Experiments”
– Natural experiments is when some event causes a
random assignment of (or change in) a variable of
interest .
• Policy change, gov. randomization, climate event (e.g.
hurricanes)
– Quasi experiment: is when the intervention its on a
target population:
• An eligibility cutoff mark (students above a threshold in a
exam receive a grant).

20
 Quasi Experiments
Example:
• David Card (1990) use a quasi-experiment in which a
larger number of Cuban immigrants (125k) entered the
Miami, FL labor market in the “Mariel boatlift”.
• This resulted from a temporary lifting of restrictions on
emigration from Cuba in 1980.
• Half of the immigrants settled in Miami.
• Card estimate the causal effect on wages of an
exogenous increase in labor supply (caused by
immigration) by comparing the change in wages of low-
skilled workers in Miami to the change in wages of
similar workers in other comparable U.S. cities over the
same period.
• He conclude that the influx of immigrants had a 21
negligible effect on wages of low skilled workers.
See current academic debate about the impact of
immigration on labor market:

• https://www.wsj.com/video/the-mariel-
boatlift-immigration-impact-on-local-
workers/B196001A-12AE-44E1-8F00-
5D99918484A3.html

22
 Natural (Quasi) Experiment
• We can use such “natural/quasi” experiments
to ensure randomness of the regressor and
make causal inference.
– We use the randomness introduced into by the
natural experiment (climate disasters, sudden
supply shock, independent change in policy) to
uncover the causal effect of on .

23
 Natural Experiments
• Natural experiments can be used in many
ways:
– Use them to construct IV (quarter of birth,
proximity to college, seat-belt laws)
– Use the to construct regression discontinuity
(Policy evaluation class: cutoff for home loans at
credit score 620 and above)
– But admittedly, when most people refer to a
natural experiment, they are talking about DID
(Difference-in-Differences) regression.

24
Lecture 2 – Natural Experiments & DD

DIFFERENCES-IN-
DIFFERENCES

25
 Difference-in-Differences
• DID basically compares outcome for the
“treated” group to the outcome for the
“untreated” group where treatment is
randomly assigned by the natural experiment.

26
 Difference-in-Differences Like the Mariel
boatlift or wawa
wasi
• Let’s think about a simple evaluation policy.
• If we have data on a bunch of people right
before and right after a policy is enacted we
can try to identify the effect of a policy.
• Suppose we have two years of data, y0 and
y1, and that the policy is enacted in between.
• We could try to identify the effect by simply
looking at before and after the policy

27
 Difference-in-Differences

• The problem is that this attributes any

changes in time to the effect policy.
• Suppose that something else happened at
time other than just the program.
• We will attribute whatever that is (i.e
increasing trend) to the program.

28
 Difference-in-Differences
• To solve for this problem, suppose we have
two groups:
– People who are affected by the policy change
(treated)
– People who are not affected by the policy change
(control)
• We can use the control groups to pick up the
time changes:

• Then we can estimate our policy effect as a

)
29
Outcome=Y

Treatment
Group

𝑫𝑫 =intervention

effect
Unobserved
counterfactual
trend for
treatment group
Control
Group
Constant
difference
in
outcome

T=0 T=1
30
 Differences-in-Differences
• DID is usually implemented as an interaction
term between time and treatment group
dummy variables in a regression model:
•

• Post=1 for the second period (after the

intervention)
• Treatment=1 for the treatment group
• X = control variables

31
 Differences-in-Differences
•

Y T=0 T=1 Differenc

e
Treatment
Group
Control
Group
Difference in
Differences=

32
 𝑫𝑫

T=0 T=1 Differenc

e
Treatment
Group
Control
Group Difference in 33
Differences=
 DID Example
• Suppose you are interested in the effect of
minimum wages on employment (a classic
and controversial question in labor
economics).
• In a competitive labor market, increases in
the minimum wage would move us up a
downward-sloping labor demand curve.
– Employment would fall.

34
 DID Example: Card & Krueger
(1994)
• Card & Krueger (1994) analyze the effect of a minimum
wage increase in New Jersey using a differences-in-
differences methodology.
• In April 1992 NJ increased the state minimum wage from
$4.25 to $5.05. Pennsylvania's minimum wage stayed at
$4.25.

35
• They surveyed about 410 fast food stores
(Burger King, Roy Rogers, Wendy's and
KFC) both in NJ and in PA both before
(Feb) and after (Nov) the minimum wage
increase in NJ.
• The survey included questions on
employment, starting wages, prices, and
other store characteristics (company-
owned)

36
Question:
• Why fast food restaurants?
– Fast food restaurants are the leading
employer of low-wage (min-wage) workers.
– Job requirements in fast food restaurants are
homogenous: easier to obtain reliable
measures of employment, wages and product
prices.
– No tips

37
38
 DID Example: Card & Krueger
(1994)
• They write the following specification:

• is the observed employment at restaurant

i, in state s, at time t
• is a dummy=1 if the obs is from NJ (high
min wage state)
• is a dummy=1 if the obs is from Nov (post
policy change)
• Diff-in-diff estimate:
• (Y_NJ post – Y_NJ pre) – (Y_PA post – Y_PA
pre) =
39
 DID Strategy
• The differences-in-differences strategy amounts
to comparing the change in employment in NJ to
the change in employment in PA.
• The population differences-in-differences are:
•

40
 DID Example

• Surprisingly, employment rose in NJ relative

to PA after the minimum wage change
41
 DID Example
• We can also estimate

By doing first differences:

42
43
 DID Example
Advantages of regression:
• It is easy to calculate standard errors.
• We can control for other variables which
may reduce the residual variance (lead to
smaller standard errors).
• It is easy to include multiple periods.
• We can study treatments with different
treatment intensity. (e.g. varying increases
in the minimum wage for different firms)
44
45
(Clase domingo…)

46
47
 Intensity of the treatment

• The value of GAP, is a

strong predictor of the
actual proportional wage
change between waves 1
and 2 or the increase in
wages at store i necessary
to meet the new minimum
wage rate.
• The mean value of GAP
among New Jersey stores
is 0.10. Thus the estimate
in column (ii) implies a 1.44
increase in FTE
employment in New Jersey
relative to Pennsylvania.

48
 DID Example: Card & Krueger
(1994)
• How convincing is this evidence against
the labor demand story?
– The key identifying assumption in DD is that
employment trends would be the same in both
states in the absence of the treatment.

49
50
PA

NJ

51
 Card & Krueger (1994): Credible results?

• The key assumption for any DD strategy is that the

outcome in treatment and control group would follow
the same time trend in the absence of the treatment.
• Common trend assumption is difficult to verify but
one often uses pre-treatment data to show that the
trends are the same.
• Even if pre-trends are the same one still has to
worry about other policies changing at the same
time.
• Card and Krueger (2000) obtained administrative
payroll data for restaurants in New Jersey and
Pennsylvania for a number of years.
52
• The data reveals fairly substantial year-to-year employment
variation in other periods.
• Employment trends in the two areas did not significantly differ,
particularly for the 14 counties sample. 53
How can we explain these results?
An alternative to the conventional competitive
model:
• NJ Stores were supply constrained
(monopsony). Stores were in need of more
workers, but no worker wanted to work at the
old minimum wage.
• Price effects. Stores in NJ simply pass-
through the minimum wage with a price
increase. They generated more revenue, thus
increase the number of employees.
54
Card & Krueger (1994): Other Critiques
• Min wage increase decided in early 1990 and
implemented in April 92. The “Before” survey is
conducted in Feb 1992.
• How could this announcement invalidate the
identification strategy?
– It might be the case that some restaurants start
implementing the min wage policy earlier.
• Also, Nov might be too early to capture the treatment
effect on employment (labor market frictions)
• Note: We don’t know which biases actually happened,
and how big the effects were.
– A good paper will check & rule out as many possible biases
given the data.
– Card and Krueger (1994,2000) present a wide variety of
alternative specifications to probe the robustness of their initial
conclusions.
55
 Regression DID Including Leads and Lags

• Including leads into the DD model is an easy way

to analyze pre-trends.
• Lags can be included to analyze whether the
treatment effect changes over time after treatment.
• The estimated regression would be:

Includes q lags (pre-treatment effects) and m leads

(post-treatment effects). Treatment occurs in year 0.

56
 Study Including Leads and Lags - Autor (2003)

• Autor (2003) includes both leads and lags in a DD

model analyzing the effect of increased
employment protection on the firm’s use of
temporary help workers.
• In the US employers can usually hire and fire
workers at will.
• Some states courts have made some exceptions to
this employment at will rule and have thus
increased employment protection.
• Different states have passed these exceptions at
different points in time.
• The standard thing to do is to normalize the
adoption year to 0.
• Autor then analyzes the effect of these exceptions
on the use of temporary help workers. 57
Temporary workers

• The lags are very close to 0. No evidence for anticipatory

effects (good news for the common trends assumption).
• The leads show that the effect increases during the first years
of the treatment and then remains relatively constant.

58
 Example: Meyer et al. (1995)
• Meyer, Viscusi, and Durbin evaluate effect of increase in
disability payments on time out of work.
• In Kentucky and Michigan the benefit amount for high-
earnings individuals was raised by approximately 50
percent, while low-earnings individuals, who were
unaffected by the maximum benefit, did not experience a
change in their incentives.
• They will compare time out of work for individuals injured
before and after the increase in the maximum weekly
benefit amount.
• Treatment:
– Change in disability payment
• Treatment Group:
– High earners
• Control Group:
– Low earners
67
• Within a given state at a point in time, the weekly
benefit for temporary total disability is a piecewise
linear increasing function of previous earnings. 68
• The group of workers with previous earnings of at least E3 experience
the full effect of the benefit increase (treatment group)
• The comparison group those with earnings between E1 and E2 (the
low-earnings group). The benefits these individuals receive are
unaffected by the increase in the maximum weekly benefit.
69
• Workers' compensation programs are run by the individual states
and differ widely in their coverage, types of benefits, levels of
benefits, and available methods of insurance underwriting.
• Workers' compensation provides both payments for medical care
and indemnity (cash) benefits for work-related injuries.
• There were only three large increases in the temporary total
maximum benefit levels in the states and time periods included in
the NCCI data base used in this study.
• These increases occurred in Florida, Kentucky, and Michigan. The
Florida increase coincided with a major overhaul of the workers'
compensation law, so that the before versus after comparisons
reflect multiple aspects of the change in benefit structure.
– The Kentucky increase of July 15, 1980, raised the maximum
benefit from $131 to $217 per week, a 66-percent increase.
– The Michigan increase on January 1,1982, raised the maximum
benefit from $181 to $307 per week, a 70-percent increase

70
• The data source for this study is the Detailed
Claim Information (DCI) data base collected
by the National Council on Compensation
Insurance (NCCI).
• The key variables in the data set that we use
are: date injured, duration of temporary total
benefits, total medical costs, previous
earnings, weekly benefit amount, benefit type
(i.e., temporary total or permanent partial),
type of injury (body part affected and the type
of damage).

71
• The fraction of previous earnings replaced by workers'
disability compensation rises dramatically for the high-
earnings group which received the benefit increase,
but remains constant for the low earnings group.
72
 Example: Meyer et al. (1995)
• Open INJURY.DTA
• Outcome= Mean Duration
• sort highearn afchnge
Before After Difference
Change Change
Treatment
Group =
High earn
Control
Group =
Low earn
Difference in 73
Differences=
74
tab highearn afchng, sum(durat)

75
Kentucky Before After Differenc
and Change Change e
Michigan
Treatment 11.76 13.93 =2.17
Group =
High earn
Control 7.47 8.61 =1.14
Group =
Low earn
Difference in =1.03
Differences=

76
• We emphasize the mean of the log of duration because this statistic is likely
to be more precisely measured and less susceptible to the influence of a
few large obs.
• This issue of robustness (log) is important because the distribution of claims
lengths has some large values, but most values are small. 77
• Statistically, if your variables are right-
skew (that is, they have a long tail at the
high end) then a measure such as
correlation or regression can be influenced
a lot by one or a few cases at the high end
(outliers, leverage points, influential
points).
• Taking the log can aid by reducing or
eliminating skew.

78
• Substantively, some concepts are better thought of in terms of
ratios than differences. Take the two volume measures you
discuss. Now, compare two companies:
• One a small company trading on NASDAQ that few people
have heard of, the other a mega-corporation. The former will
get very few tweets per day. The latter will get many; similarly
for trading volume. Suppose (just to pick numbers) that
company A typically gets 100 tweets a day and the latter gets
100,000.
• If company A's tweets go up from 100 to 500 (a difference of
400, a ratio of 5) that's huge news - something must be going
on. But if company B's go up from 100,000 to 100,400 (a
difference of 400, a ratio very close to 1) no one cares. The
rough equivalent would be if it went from 100,000 to 500,000.

79
Regression estimates of DID

80
 Regression estimates of DID
• Duration of the spell of time-out-of-work (time
that you receive workers' compensation
disability benefits)
• period after the reform
• = treated or high earnings group dummy
(Income > E3).

• Diff-in-Diff =

81
Kentuck Before After Differenc
y and Change Change e
Michigan
Treatmen 11.76 13.93 =2.17
t Group =
High earn
Control 7.47 8.61 =1.14
Group =
Low earn Difference in
4.29 =1.03 Differences
82
83
 // logs
global controls "manuf construc head neck upextr trunk lowback lowextr occdis"

reg ldurat afchnge highearn treated male age prewage $controls, robust

reg ldurat afchnge highearn treated male age prewage $controls ///
if ky==1, robust
est store regky

reg ldurat afchnge highearn treated male age prewage $controls ///
if mi==1, robust
est store regmi

esttab regky regmi, se mtitle("Kentucky" "Michigan") star(* 0.1 ** 0.05 *** 0.01)

84
85
The reform increase around
16-18% the duration in
weeks of Disability
Compensation.

86
• The outcome variable (duration in weeks of Disability compensation) is in logs.
• The reform increase around 15-20% the duration in weeks of Disability Compensation.
• Michigan estimates are not significantly different from zero: in Michigan benefits depended partly on
tax filing status and the number of dependents, besides pre-earnings.

87
• The results of this study suggest a substantial effect of
the level of temporary total benefits on the duration of
workers' compensation claims.
• The KY estimates are generally significantly different
from zero, while the MI estimates are similar in
magnitude to the KY estimates but are less precisely
measured and generally not significantly different from
zero.
• Should we increase workers' compensation
disability benefits?
– The longer durations that we find after benefit increases
may not indicate a loss in social welfare, as longer
recovery times may improve subsequent health.
– Higher benefits may enable injured workers to complete
their recovery before returning to work.
– To examine this question, we would like to be able to
examine health status after an individual returns to work.

88
 Class Exercise
• Kiel and McClain (1995) studied the effect that a new
garbage incinerator had on housing values in North
Andover, MA.
• The rumor that a new incinerator would be built in North
Andover began after 1978, and construction began in
1981. The incinerator was expected to be in operation
soon after the start of construction; the incinerator
actually began operating in 1985.
• We will use data on prices of houses that sold in 1978
and another sample on those that sold in 1981.
• The hypothesis is that the price of houses located near
the incinerator would fall relative to the price of more
distant houses.
89