Title: Difference%20in%20Difference%20and%20Regression%20Discontinuity
1Difference in Difference and Regression
Discontinuity
2Review
- From Lecture I
- Causality is difficult to Show from cross
sectional observational studies - What caused what?
- X caused Y, Y caused X
- Omitted Variable Bias/Confounding
- In some cases you can say whether the estimate is
an upper-bound or lower bound estimate - Other times impossible to sign bias since omitted
variables bias the coefficient of interest
positively and negatively. Net effect impossible
to determine a-priori.
3Review (cont.)
- Discussed Randomized Control Trials as a simple
(but not necessarily practical) way to solve the
causality problem - Randomization works because we can be sure about
temporal precedence - Randomization works because treatment and control
groups are balanced on observables and
un-observables
4Review (cont.)
- Also quickly presented some other commonly used
research designs - X 01 - Observe only data from post treatment (X)
- 01 X 02 Observe data from pre and post
treatment periods - 01 02 X 03 Observe data from pre and post
treatment observe a longer pre period - Common Feature of all these designs is that there
is NO CONTROL GROUP
5Difference in Difference I
- 01 X 02
03 04 - 01 is the pre-period treatment group data
- 02 is the post intervention treatment group data
- 03 is the pre-period control group data
- 04 is the post intervention control group data
6Difference in Difference I (cont.)
- Lets Compare to 01 X 02 design
- No control group, leads to the strong assumption
that over time, without an intervention,
Dependent variable of interest would not have
changed - 01 X 02
03 04 - Diff. in Diff. treatment design accounts for the
fact that dependent variable might change even if
there were no intervention - Similar to the RCT framework the control group
provides the counterfactual
7Difference in Difference I (cont.)
- Simplest representation is 2 X 2 Table
- Diff. in Diff. Estimate
- E(YT1) E(YT0) E(YC1) E(YC0)
- Same result even if you calculate
- E(YT1) E(YC1) E(YT0) E(CT0)
E(YT0) E(YT1)
E(YC0) E(YC1)
8Difference in Difference I (cont.)
- In words you are subtracting out the change in
the control group from the change in the
treatment group - If the treatment had not effect then this is
tantamount to saying that the two differences are
equal - If the treatment had an effect then either the
first term is bigger than the second term
(positive effect) or the second term is bigger
than the first term (negative effect)
9Difference in Difference II
- Why is Diff. and Diff. powerful?
- MAIN REASON We have a control group
- Another problem with cross-sectional studies is
that we worry about unobserved and hard to
measure differences between the treatment and
control group - In the difference in difference estimate,
Unobserved differences across treatment and
control that stay constant over time are
differenced out - Another way of saying this is that these
unobserved unchanging characteristics effect the
Level but not the changes
10Difference in Difference III
- Problems with Difference in Difference Estimation
- Lets remember What made RCT powerful
- We knew the assignment mechanism RANDOMIZATION
- Note that there is no randomization in Diff. in
Diff - Unit of observation (for ex. state) still chooses
whether or not to get treatment - Choice leads to the potential problem that
treatment and control groups are different - Consequently we are still concerned with some of
the usual problems from cross-sectional studies
11Difference in Difference III (cont.)
- Main Concern is History
- How can we be sure that other interventions are
also not simultaneously occurring with treatment? - For ex. Some states in an effort to reduce smoke
might enact anti-smoking laws in public spaces - Very possible that the states that enact
anti-smoking laws simultaneously enact other
anti-smoking measures as well (increase
advertising, increase taxes etc.) - For these changes not to bias the diff. in diff.
estimate we would have to argue that the control
group also enacted these other changes at the
same time.
12Difference in Difference III (cont.)
- Specification Checks
- Plot pre intervention trends over time for
dependent variables for treatment and control
groups separately. - IF trends are parallel in treatment and control
groups and see sudden change after intervention
then you are potentially safe - If trends are not parallel then possible bias
from other sources - Create False Treatments and Redo estimation
- For ex. If intervention happened in 1990, assign
intervention in treatment group to 1989 and see
if you still find an effect - If you find an effect then likely that something
else is driving your findings
13Difference in Difference III (cont.)
- Use an outcome that shouldnt be affected by the
intervention and redo estimation
14Difference in Difference IV
- Still Other Concerns
- Policy intervention is tied to outcome
- Difference in Difference will overstate true
effect - Mean reversion is again a potential problem
- My sense is that this is only a problem for some
outcomes (wages is a good ex.) - Long term effect might be difficult to estimate
- Estimate is most reliable right after
intervention - Long term effects likely confounded by other
variables - Functional Form
- Means or Logs
15Card Krueger - An Example
- What is the Effect of a Minimum Wage increase on
employment? - Theory says rise in wages should lead to less
employment - Firms are profit-maximizing already, taxing one
input (labor) should lead to a decrease in its
use
16Card Krueger (cont.)
- NJ enacted a state law that increased the minimum
wage from 4.25 to 5.05 - Effective April 1, 1992
- Card and Krueger(1994) use a Diff. in Diff.
research design to examine whether this change
led to lower employment - Control group is Pennsylvania where the minimum
wage did not change over this time period
17Card Krueger (cont.)
- Card Krueger Look at the effects in Fast Food
Industry, Why? - Leading employer of low-wage workers
- Easier to measure prices, employment and wages in
this industry - Survey Burger King, KFC, Wendys and Roy Rogers
chains - Exclude McDonalds because McDonalds had a poor
response rates to surveys in previous work - Initial survey conducted in late February and
early March 1992, - A month before the NJ minimum wage increase
- Secondary Survey conducted in November and
December 1992
18Card Krueger (cont.)
- Around 80 response rate in pre-period
- 90 of these 80 responded in post-period
- One Key question Is the wage increase in N.J.
meaningful? - Yes, average starting wage in New Jersey
restaurants increased by 10 (4.61 to 5.05) - In wave 1 31 had a starting wage of 4.25
- In Pennsylvania, In wave 1, average starting wage
in Pennsylvania was 4.63 and - In wage two there was no change
19Card Krueger Results
Avg. Full Time Employees Before
Avg. Full Time Employees After
- Diff in Diff Estimate
- 21.03-20.44 21.17-23.33
- .59--2.162.75
- Standard Error on estimate is 1.36
- Conclusion Estimate is positive but not
statistically significant at the 5 level
20.44 21.03
23.33 21.17
NJ
PA
20CK Results (cont.)
- Lets compare to the 01X02 design
- Question Given the CK data what would you have
concluded about the effect of the increase in
minimum wage if you used this design? - This simpler design would have said that the
effect of the minimum wage hike is positive and
the magnitude.59 - The Diff. in Diff. estimate also says the effect
of the minimum wage hike is positive but the
magnitude is now 2.7 - Including a control group increases the 01X02
estimate by a factor of close to 5
21CK Results (cont.)
- Regression Framework
- Each observation in the data is a store
- Dependent variable is Change in employment
- Independent variables include region, chain
dummies (burger king etc.) - State Dummy for whether or not in New Jersey
- Regression coefficient on State Dummy 2.33
- On average the law leads to an increase of 2.33
employees - But standard error on the estimate is 1.33 so not
statistically different from zero
22CK-Other Specifications (cont.)
- Some stores not affected if they are already
above the minimum wage - Create a GAP variable
- 0 for stores in Pennsylvania
- 0 for stores in NJ whose wage is already above
5.05 - (5.05-initial wage)/initial wage for other NJ
states - increase in wages
- Again find a positive effect but not
statistically significant
23CK-Other Specifications (cont.)
- change in employment in the dependent variable
- Exclude management employees
- Include part time workers in employment
- Exclude stores along the coast of NJ
- These stores might have a different seasonal
pattern - Finally surveyors called some stores in NJ more
often to get data. Exclude these stores from
sample - NONE OF THESE CHANGES AFFECT THE BASIC RESULTS
24CK - Other Specifications (cont.)
- Non Wage-Offsets
- Offset raise in minimum wage by reducing non-wage
compensation (fringe benefits) - Main fringe benefit is free and reduced price
meals - Do not find any changes in this measure
- Future wage offsets
- Reduce the rate at which salaries increase
- Examined the average time to first pay raise
25Regression Discontinuity
- Arbitrary Threshold determines whether or not a
unit gets assigned to treatment or Control group - Anti-Discrimination law only applies to firms
with at least 15 employees - Rabbinic Scholar Maimonides says Class size
cannot exceed 40, if so must group student into
smaller classes - For ex. 42 students means average class size is
20.5 - 80 students means two classes of size 40 but 81
students means average class size of 27
26Regression Discontinuity (cont.)
- In this framework, for most examples, being above
threshold implies you are in the treatment group - In this framework, for most examples, being below
threshold implies you are in the control group - Look for a Change in magnitude of the outcome
variable right around this threshold
27(No Transcript)
28Regression Discontinuity (cont.)-
- This research design might make you think of 01 X
02 - But its not? Why is that?
- For one thing there is no time component
- Second 01 is NOT A VALID control group
29Regression Discontinuity
- Two types of regression discontinuity
- Sharp Regression Discontinuity
- W_I 1 if X_I gt C
- All units with X gt C are assigned the treatment
- All units with Xlt C are assigned to control
30Probability of Assignment
31Regression Discontinuity
- Sharp Regression Discontinuity
- We assume that effect without treatment is linear
- There is no way to verify this since treatment is
assigned to individuals above the cutoff
32Potential and Observed Outcome
33Regression Discontinuity
- Fuzzy Regression Discontinuity Design
- Probability of receiving does not have to be 1 at
the threshold - For ex. Individuals above some threshold could be
offered a treatment - The offer does not lead all individuals to take
up treatment - As an example think of a voucher scheme that
allows people to move neighborhoods. - For some individuals voucher amount is not enough
to get them to comply
34(No Transcript)