Chapter 6.

Difference in Differences

Joan Llull

Quantitative Statistical Methods II

Barcelona School of Economics
 Dierence in dierences setup

Randomized experiment ⇒ simple comparison of the mean outcome in treatment

and control groups (dierence estimator), unbiased and consistent estimate of the
ATE (Chapter 2).

The approach in this chapter: like in matching or sharp RD, adjusts somehow to
compensate confounders.

Key assumption: common trend.

Chapter 6. Dierence in Dierences 2

 Dierence in dierences

Chapter 6. Dierence in Dierences 3

 Formal discussion
Figure suggests to use trend observed for untreated to predict the counterfac-
tual trend for treated individuals in the absence of treatment:
E[Y0it |Di = 1, t = 1] = E[Yit |Di = 0, t = 1]
| {z }
level for controls at t=1
+ {E[Yit |Di = 1, t = 0] − E[Yit |Di = 0, t = 0]}.
| {z }
dierence in levels at t=0 dierence
Fundamental DD assumption: common trend:

E[Y0i1 − Y0i0 |Di = 1] = E[Y0i1 − Y0i0 |Di = 0].

Hence, the dierence in dierences coecient (which is an average treatment eect
on the treated) is:

β = {E[Yit |Di = 1, t = 1] − E[Yit |Di = 1, t = 0]}

− {E[Yit |Di = 0, t = 1] − E[Yit |Di = 0, t = 0]}.

Chapter 6. Dierence in Dierences 4

 Di-in-di and regression
The dierence in dierences coecient can be obtained as the β coecient in the
following regression:
Yit = β0 + βD Di + βT Tit + βDi Tit + Uit ,

where Tit = 1 if individual i is treatment period t = 1, and Tit = 0 otherwise.

With similar arguments as in previous chapters:

β0 = E[Yit |Di = 0, t = 0],

β0 + βD = E[Yit |Di = 1, t = 0],
β0 + βT = E[Yit |Di = 0, t = 1],
and β is the dierence in dierences coecient.

Chapter 6. Dierence in Dierences 5

 Di-in-di and regression
This regression model can be expanded in several ways:

Further periods: In such case, Tit is not a time dummy but, instead, a dummy that
equals one in the post-treatment period. One could additionally include time eects,
but the interaction term should be with the post dummy only.

Controls: the regression allows for controls, Xit (works as in regression vs matching).
Panel data: see below
Placebo analysis: a regression that simulates the dierence in dierences analysis
but for a point in time or group of individuals that resemble the treatment period or
group but that was actually not treated.

Chapter 6. Dierence in Dierences 6

 Triple dierences
Triple dierence: the dierence in dierences assumption does not hold, but the
change in trends is assumed to be the same across sub-groups, some of which should
be more aected than others.

Let Gi denote the (say sociodemographic) group to which individual i belongs. Then,
the triple-dierences model is:
Yit = β0 + βD Di + βT Tit + βG Gi + βGD Gi Di
+ βGT Gi Tit + βDT Di Tit + βGi Di Tit + Uit .

Example: Maternity leave policies combined with a tax reform that aects young
and old dierently.

Chapter 6. Dierence in Dierences 7

 Dierence in dierences
Male
Female

Chapter 6. Dierence in Dierences 8

 Synthetic Control Methods

Synthetic control methods: use longitudinal data to build the weighted average
of non-treated units that best reproduces the characteristics of the treated unit over
time prior to the treatment.

Thus, we build an articial control that has the best possible pre-trend possible,
and then we compute the dierence in dierences estimate using such synthetic
control group.

Chapter 6. Dierence in Dierences 9

 Panel Data
There is no need to have panel data to estimate a dierence in dierences model.

However, in the presence of panel data, one can estimate:

Yit = β0 + ηi + δt + βDit + Uit .

ηi provides an individual-specic estimate of  E[Y0it |i], the average Y0it for in-
dividual i across time periods t.

δt captures the average over-time trend in the absence of treatment.

β identies the average treatment eect on the treated (when the treatment
eect is constant), and it is identied from treated individuals, whose treatment
status Dit changes over the observation period.

Chapter 6. Dierence in Dierences 10

 Staggered Designs
Panel data regression allows us to identify the treatment eect from multiple treat-
ments implemented at dierent points in time.
Some observations contribute as control group in some treatments and as treat-
ment group in others.
⇒ the OLS estimate would provide a weighted average of all possible two-by-two
dierence in dierences estimators for each policy change (Godman and Bacon'21).

Specic weights depend on the absolute size of each subsample (treated and
control groups) and also about the timing of the treatments.
⇒ Changing spacing of treatments changes the weights and therefore the estimate.

If average treatment eects are constant across treatments ⇒ ATT.

Chapter 6. Dierence in Dierences 11
 Event Studies
Let T0i denote the timing in which individual i becomes treated.

Dene Sit ≡ t − T0i as the time period relative to the event (needs to be dened too
for untreated, by design or by assumption).

Yit = β0 + βD Di + βT Sit + βDi Sit 1{Sit ≥ 0} + Uit .

Child penalty example:

βD captured by gender dummies.

βT captured by dummies or trends of period in the event (birth of rst child).

β is identied from the change in wage for men and women in years after event.

This is often captured by time dummies instead of trends, in which case,

we can allow for the dummies for Sit < −1 to be dierent from zero, as we only
need to normalize the one for Sit = −1 to be equal to zero.
Chapter 6. Dierence in Dierences 12