LSE – EC2C3, Gmeiner

Ungraded Practice Problems – Difference-in-Differences and Regression Discontinuity Practice

Comments are welcome if questions are unclear or the solutions seem incorrect.
1. A classic question in economics is the effect of the minimum wage on unemployment.
However, the minimum wage may have effects on the labour market in ways other than
unemployment. We are curious if, due to a minimum wage increase, workers increase
productivity. The idea is that it is more costly for a company to employ workers, so
employees must increase output to avoid being fired. We will measure productivity by the
quantity of goods produced by factory workers.

a. Suppose we could randomly assign a minimum wage increase at time 𝑡 to a set of

regions. Using data on 𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑖 that is gathered after time 𝑡, write a
regression equation that would allow us to estimate the effect of the minimum
wage on productivity.

b. Would the estimate from part a. be unbiased? Suppose that we did not randomly
assign the minimum wage increase. Would your answer change?

We can’t randomly assign minimum wage increases. We gather data from factory workers in
two regions, 𝑖 = 1 and 𝑖 = 2, for two time periods, 𝑡 = 1 and 𝑡 = 2. Region 𝑖 = 1 increased
their minimum wage at time 2. The minimum wage in region 𝑖 = 2 was unchanged.

𝑦 = Productivity
𝑡 = 1 is before the increase, 𝑡 = 2 is after the increase
𝜆𝑡 = effect (on productivity) of being time period 𝑡
𝑖 = 1 denotes the group that is treated. 𝑖 = 2 is control.
𝛾𝑖 = effect (on productivity) of being group 𝑖
𝛿 = causal effect of the minimum wage chnage on productivity (only applies in time 2 for group 1)

We assume that 𝐸[𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑖𝑡 ] = 𝛾𝑖 + 𝜆𝑡 + 𝛿 if the minimum wage increases and

𝐸[𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑖𝑡 ] = 𝛾𝑖 + 𝜆𝑡 if the minimum wage does not change.

As researchers, we would not know the following information, but as students doing practice
problems, suppose we know that 𝜆1 = 3.4, 𝜆2 = 4.2, 𝛿 = 2.8, 𝛾1 = 1.2, and 𝛾2 = 2.1.

c. We could use data from factories in region 1 and estimate the regression

𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑡 = 𝛽0 + 𝛽1 𝐴𝑓𝑡𝑒𝑟𝑡 + 𝑢𝑡

for which 𝐴𝑓𝑡𝑒𝑟𝑡 is a binary variable taking the value 1 if 𝑡 = 2. What is biggest concern
for why this regression might not give an unbiased estimate of the effect of the minimum
wage on productivity?

̂1 when estimating the equation in c?

d. What value would you expect to attain for 𝛽
LSE – EC2C3, Gmeiner

e. We could use data from both regions at time 𝑡 = 2 and estimate the regression

𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑣𝑖𝑡𝑦𝑖𝑡 = 𝛽0 + 𝛽1 𝑅𝑒𝑔𝑖𝑜𝑛 1𝑖 + 𝑢𝑖𝑡

for which 𝑅𝑒𝑔𝑖𝑜𝑛 1𝑖 is a binary variable taking the value 1 if 𝑖 is region 1. What is
biggest concern for why this regression may not give an unbiased estimate of the
effect of the minimum wage on productivity?

f. ̂1 when estimating the equation in e?

What value would you expect to attain for 𝛽

g. What is the difference-in-differences estimator of the effect of the minimum

wage on productivity in this setting?

h. What is the regression implementation of the difference-in-differences estimator?

Write your equation such that 𝛽3 represents the answer from g.

i. What is the critical assumption of the difference-in-differences estimator from g

and h? Answer (1) in English and (2) in math.

j. Shown below is the plot of productivity over time for regions 1 and 2 prior to
time 1. Does this graph make it seem likely the assumption from question (i) is
true?

k. Suppose instead the graph looked as shown below. Does this graph make it seem
likely the assumption from question (i) is true?

l. Write a regression equation that allows us to formally perform the type of test we
performed in parts j and k. Explain.
 LSE – EC2C3, Gmeiner

Note that even if the graph makes it seem like the assumption is true, the assumption might not be
true. If the graph makes it seem like the assumption is not true, it might be. The assumption that is the
answer to question (i) is an assumption, and we can’t prove it.

m. Regardless of your answers to 𝑗 and 𝑘, suppose the assumption from 𝑖 is true and
the graph from 𝑘 shows the trends in unemployment prior to time 1. Which graph
shows the assumption from 𝑖 is true? The dashed line shows the change in
unemployment that would have occurred if there had been no change in the
minimum wage.

i. ii. iii. iv.

For n, assume the following:

𝑦12 = 8.2, ̅̅̅̅
̅̅̅̅ 𝑦11 = 4.6, ̅̅̅̅
𝑦22 = 6.3, and ̅̅̅̅
𝑦21 = 5.5.

n. What is the difference-in-differences estimate of the effect of the minimum

wage?

o. Suppose parallel trends is not true. Suppose that in the absence of a minimum
wage change, region 1 would have increased productivity by 2.3, and region 2’s
change was 0.8.
What is the bias between the difference-in-differences estimate and the true value
of the effect of the minimum wage? (If it helps, you may define the notation 𝜆𝑖𝑡
to represent the effect of time 𝑡 for region 𝑖).
LSE – EC2C3, Gmeiner

2. A government has a program to provide extra funding to cities if the population is above a
threshold, 𝐷. We want to estimate the effect of extra funds on scores of students, 𝑆𝑐𝑜𝑟𝑒𝑠𝑖 .
We define 𝐿𝑎𝑟𝑔𝑒𝑖 to be a binary variable that takes the value 1 if 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖 > 𝐷.

a. Write a regression equation that implements a regression discontinuity.

b. What is the necessary assumption of this regression discontinuity? Does it seem

true?

c. Use the table of summary statistics that is below to evaluate the assumption from
b. Means are presented above standard deviations, which are in parentheses.

Population > D Population < D

% Immigrants 23.3 5.8
(8.4) (3.4)
Household Income 65,930 43,850
(45,920) (23,430)
Teacher – Student Ratio 19.3 17.5
(3.4) (2.8)
N 43 632

d. Regression discontinuities are typically not estimated while using all data, but
instead estimated using data for which the running variable is in a small window
around the cutoff. Why? Answer in context.

e. What are the pros and cons of increasing the size of the window? Answer in
general and in context.
LSE – EC2C3, Gmeiner

Regardless of the model you wrote in part a, suppose we estimate the equation below,

𝑆𝑐𝑜𝑟𝑒𝑠𝑖 = 𝛽0 + 𝛽1 𝐿𝑎𝑟𝑔𝑒𝑖 + 𝛽2 (𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖 − 𝐷) + 𝛽3 (𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖 − 𝐷) ⋅ 𝐿𝑎𝑟𝑔𝑒𝑖 + 𝑢𝑖

In which we have restricted to observations for which 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖 ∈ [𝐷 − ℎ, 𝐷 + ℎ] for

̂0 = 62, 𝛽
some ℎ. We attain estimates: 𝛽 ̂1 = 2.3, 𝛽
̂2 = 4.5, and 𝛽
̂3 = −1.1.

f. Draw a figure that shows these coefficient estimates.

g. Write English sentences that interpret the coefficient estimates.

Suppose the standard errors are respectively 8.8, 3.4, 2.4, and 0.8.
h. We originally restricted to a window of observations that satisfy,
𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖 ∈ [𝐷 − ℎ, 𝐷 + ℎ]. If we were to make ℎ larger, how would you
expect the standard errors to change? How would you expect coefficient estimates
to change? (Can we make any reasonable guess at all?)