Revisiting Event Study Designs,

with an Application to the Estimation of the Marginal Propensity to Consume∗

Kirill Borusyak, Harvard University

Xavier Jaravel, Stanford University

May 8, 2017

Abstract

A broad empirical literature uses “event study” research designs for treatment effect estimation, a setting
in which all units in the panel receive treatment but at random times. We make four novel points about
identification and estimation of causal effects in this setting and show their practical relevance. First, we show
that in the presence of unit and time fixed effects, it is impossible to identify the linear component of the path
of pre-trends and dynamic treatment effects. Second, we propose graphical and statistical tests for pre-trends.
Third, we consider commonly-used “static” regressions, with a treatment dummy instead of a full set of leads
and lags around the treatment event, and we show that OLS does not recover a reasonable weighted average
of the treatment effects: long-run effects are weighted negatively, and we introduce different estimators robust
to this issue. Fourth, we show that equivalent problems of under-identification and negative weighting arise
in difference-in-differences settings when the control group is allowed to be on a different time trend or in the
presence of unit-specific time trends.
We show the practical relevance of these issues in a series of examples from the existing literature. We
focus on the estimation of the marginal propensity to consume out of tax rebates: according to our preferred
specification, the marginal propensity to consume is much lower than (about half of) the main estimates in the
literature.
The main message for practitioners is that because of identification issues and negative weighting in event
study designs, results from common specifications are likely to seem non-robust. These problems can be alleviated
in a principled way by using parametric and semi-parametric estimators and tests.

∗ We thank Alberto Abadie, Isaiah Andrews, Raj Chetty, Itzik Fadlon, Ed Glaeser, Peter Hull, Guido Imbens, Larry Katz, Jack

Liebersohn and Jann Spiess for thoughtful conversations and comments. We are grateful to Jonathan Parker for his support in accessing
and working with the data. The results in the empirical part of this paper are calculated based on data from The Nielsen Company
(U.S.) LLC and provided by the Marketing Data Center and the University of Chicago Booth School of Business. The latest version of
the paper is available here.

1
1 Introduction
A broad empirical literature in labor economics, public finance, finance and empirical macroeconomics uses event
study research designs for treatment effect estimation, a setting in which all units in the panel receive treatment
but at random times. We make a series of novel points about identification and estimation of causal effects in such
a setting, which are closely related to the well-known age-cohort-time problem. We then establish the practical
relevance of these points in light of the existing literature and in a specific application, estimating the impulse
response function of consumption expenditures to tax rebate receipt.
We first show that in the presence of unit and time fixed effects, it is impossible to identify the linear component
of the path of pre-trends and dynamic treatment effects. Identification of the dynamic causal effects only up to
a linear trend is particularly problematic because researchers usually want to test for the absence of pre-trends
prior to the event and to document changes in the outcome variable after treatment in a non-parametric way.
Intuitively, the path of pre-trends and dynamic treatment effects is identified only up to a linear trend because,
“within a treatment unit”, one cannot disentangle the passing of absolute time from the passing of time relative to
the treatment event. We show formally that the collinearity problem that arises in this setting is effectively the
same as the age-cohort-time problem that others have studied.
We then propose two approaches to address this underidentification issue. Our first strategy is to restrict the pre-
trends in the fully dynamic specification, while keeping unit fixed effects. Our second strategy consists in replacing
unit fixed effects with unit random effects. We develop statistical tests for the validity of both approaches, as well
as a graphical test for the first approach.
Next, we turn to estimation of the average treatment effect, with a particular focus on specifications that are
meant to average all dynamic treatment effects post treatment. We show that the specification that is commonly
used in the literature estimates an average of treatment effects that severely overweighs short-run effects and weighs
long-run effects negatively. This issue can be serious, such that the estimated average treatment effect can be
outside of the convex hull of the true dynamic treatment effects. To address the problem, we provide alternative
parametric and semi-parametric estimation techniques, which always average dynamic effects in a convex way.
Intuitively, the short-run bias originates from a peculiar type of extrapolation implicitly performed by the OLS.
Recall that a difference-in-differences estimator is meant to use the outcome gap between some units A and B in
an early period s before they receive treatment for constructing counterfactuals for period t when B is treated but
A is not. However, OLS regression conflates the situations where neither unit is treated at time s, or both are—A
and B have the same treatment status in both cases, misleadingly opening room for comparison. Specifically, if
causal effects exhibit dynamics, the difference between post-treatment outcomes of B versus A reflects the wedge
between long- and short-run effects, since B was treated earlier. This difference is subtracted from an analogous one
for period t, thereby creating a short-run bias.
We demonstrate that the problem is particularly severe in event studies with no control group. While a control
group partially alleviates the problem, expanding the set of good comparisons, it has to be quite large for the
problem to disappear, and the short-time bias is present even then.
Beyond these core results, we show that OLS suffers from similar weighting problems in two related but different
designs. It happens in classical diff-in-diffs if the control group is allowed to be on a separate time trend—an approach
that applied studies sometimes take when they are not certain of the quality of their control group. It also applies
to regressions that include unit-specific time trends. In both cases, the problem arises even if the control group is
very large.
Finally, we establish the empirical relevance of these various points by describing a series of recent and influential
papers in the literature that run precisely the specifications whose undesirable properties we point out in this
paper. Moreover, we use extended data following Broda and Parker (2014) to estimate the impulse response

2
function of consumption expenditures to tax rebate receipt, using the random timing of tax rebate receipt across
households. In this application, we find that underidentification can lead to seemingly unstable results when it is
not properly handled. We also find that the marginal propensity to consume estimated by the canonical regression
has a strong short-run bias (upwards) due to the weighting issues. At the same time, semi-parametric specifications
demonstrate that long-run effects are very difficult to identify in this type of natural experiment without undesirable
extrapolation.
Although our paper is aimed primarily for the applied reader, it contributes to three micro-econometric liter-
atures. Imbens (2015) notes that “possibly [...] because of lack of compelling evidence that simple methods (for
example, ordinary least squares regression) are not adequate, few of the estimators proposed in the recent literature
have been widely adopted”. We view our paper as providing such evidence for difference-in-difference settings with
differential event timing. Imbens’s paper and Sloczynski (2016) are similar in spirit. The seminal paper by Angrist
(1998), which argues in the other direction, provides a useful benchmark to highlight our points. We also speak
to the literature that noted, more informally and for specific applications, some problems with event studies and
unit-specific trends, such as Wolfers (2006), Fadlon and Nielsen (2015) and particularly Meer and West (2015).
Finally, in developing our semi-parametric estimator, we contribute to the large literature including Hirano et al.
(2003) and Abadie (2005).
The remainder of this paper is organized as follows. In Section 2, we describe our setting, the key specifications
we study, and how they are consistent with a causal model. Section 3 presents the underidentification problem and
our solutions, while Section 4 describes the negative weighting issue and how to address it. Section 5 considers a
variety of extensions. Finally, Section 6 relates our points to the existing literature and presents the estimation of
the marginal propensity to consume out of tax rebates as an application.

2 Setup
2.1 Data-generating Process
Consider a panel of i = 1, . . . , N units in which the outcome Yit is observed for t = 1, . . . , T periods (“calendar
time”), or possibly for a subset thereof. In our main setting, every unit receives treatment in some period Ei within
the sample and stays treated forever.1 Units with the same treatment period are referred to as a cohort. Let
Kit = t − Ei denote the “relative time”—the number of periods relative to the event. The indicator variable for
being treated can therefore be written as Dit = 1 {t ≥ Ei } = 1 {Kit ≥ 0}.
Empirical papers using this event study setup often pursue some of the following three goals. They first estimate
whether the treatment has an effect on average. Second, they test for pre-trends to lend credibility to the research
design. Finally, they may study in more depth the dynamics of the causal effect. With these goals in mind, we
choose a class of data-generating processes which is very flexible on the dynamics but abstracts away from a variety
of other specification issues:2

∞
X
Yit = α̃i + β̃t + γ̃k 1 {Kit = k} + ε̃it . (1)
k=−∞

Here {γ̃k } for k < 0 correspond to pre-trends, and for k ≥ 0—to dynamic effects k periods relative to the event.3
P∞
The average effect is k=0 ωk γ̃k for some weighting scheme {ωk }, but researchers rarely specify it based on their
1 We consider settings with a control group in Section 5.1.1.
2 We allow for more general models in the Extensions section, such as heterogeneity of treatment effects across units. All of our
results also directly extend to adding time-varying controls.
3 By k = ±∞ we mean the largest number possible given the sample.

3
economic question and instead rely on the regression to do something reasonable. Tildes indicate the parameters of
the “true model”, reflecting the data generating process, and later on we express the estimands of commonly-used
regression specification in terms of these parameters.4 α̃i and β̃t are unit and period fixed effects, respectively, and
ε̃it is random noise. We call equation (1) the fully dynamic specification.
This formulation is consistent with a causal model in which each unit i for each period t has a set of potential
(k)
outcomes Yit for each integer k, only one of which is realized. Treatment effects, expressed relative to one of them,
(−1) (k) (−1)
e.g. Yit , are homogenous across units and calendar time periods, and depend only on k: Yit − Yit = γ̃k .5
(−1)
Furthermore, Yit = α̃i + β̃t + ε̃it , which is a standard assumption necessary for the validity of difference-in-
difference approaches (Angrist and Pischke, 2008, p. 156). Together these assumptions deliver equation (1).
If the event is unpredictable, it is not known whether the current period corresponds to Kit = −1, −2, or any
(−1) (−2)
other negative number. As a consequence, Yit = Yit = . . . , so γ̃k = 0 for all k < 0. In that sense, random
timing of the event implies that there cannot be any pre-trends.6 Equation 1 reduces to the following specification,
which we call semi-dynamic and take to be true if the empirical design is valid:
∞
X
Yit = α̃i + β̃t + γ̃k 1 {Kit = k} + ε̃it . (2)
k=0

2.2 Current Practice

In the current practice, it is prevalent to estimate models similar to (1) and (2) using OLS with two-way (unit
and period) fixed effects. Different papers impose different restrictions on (1), but the following specification covers
most of them:
B−1
X
Yit = αi + βt + γk 1 {Kit = k} + γB+ 1 {Kit ≥ B} + εit , (3)
k=−A

where A ≥ 0 leads of treatment are included, together with B ≥ 0 terms for specific short-run effects and a single
last coefficient γB+ for all longer-run effects.7 Note the absence of tildes: unless A = 0 and B = ∞, this equation
does not coincide with the true model (2). We will study where its coefficients converge to in large samples as
functions of the true parameters. We will occasionally use hats to mark finite-sample objects.
The simplest and perhaps the most prevalent regression is (3) with A = B = 0, i.e.

Yit = αi + βt + γ0+ Dit + εit . (4)

We refer to this specification as static or, following Allegretto et al. (2013), canonical, and will discuss it at great
length later in the paper. Compared to the fully dynamic one, it imposes no pre-trends and constant treatment
effects for all k. The other extreme is, of course, A = B = ∞, which is just the fully dynamic specification with no
restrictions.
Often regressions are run using all available data, but sometimes the sample is balanced around the event
time: only observations with Kit ∈ [k, k̄], k < 0 ≤ k̄, are included and only for units which are observed for all
corresponding periods. We discuss pros and cons of this approach later on.
4 The notation for these estimands does not use tildes.
5 Note that one cannot hope to estimate treatment effects relative to the situation in which the event never happens, simply because
this circumstance is not observed in the data. Picking k = −1 as the omitted category is an innocent and standard normalization.
6 In some settings anticipation of the event is possible but limited to a fixed number of A periods. In that case γ̃ = 0 for k < −A,
k
and any k < −A can be chosen as the omitted category.
7 One term, e.g. γ
−1 1 {kit = −1}, can be omitted as a normalization.

4
3 Underidentification of the Fully Dynamic Specification
3.1 Problem
In this section, we show that the fully dynamic specification, given by equation (1), suffers from a fundamental
underidentification problem. The goal of such a specification is to recover the dynamic path of causal effects
∞
{γ̃k }k=−∞ . We show that a linear trend in this path is not identified. One can start from any set of points
∞
estimates {γk }k=−∞ , add a linear trend (in k) and adjust the sets of point estimates for the year fixed effects βt and
∞
the unit fixed effects αi to keep the same predicted value. Identification of the dynamic causal effects {γ̃k }k=−∞ up
to a linear trend is particularly problematic because researchers usually want to test for the absence of “pre-trends”
prior to the event,8 and more generally are hoping to document changes in the outcome variable after treatment in
a non-parametric way.9 In this section, we first illustrate the underidentification issue graphically to gain intuition.
We then show mathematically where it stems from.
The intuition for why the fully dynamic specification is underidentified can easily be grasped with a simple
example. We intentionally make it extreme, to isolate the fundamental problem from any additional effects or
noise, which exist on top of it in real data; the mechanism is completely general. Consider Fig. 1, which plots the
outcomes for a simulated dataset, which covers two cohorts. One is treated early at t = 2 (solid line), and the other
one later at t = 4 (dashed line). Both groups are observed for all periods t = 1, . . . , 7, and the outcomes exhibit
linear growth with the same slope of one, albeit starting from different levels. There are two interpretations of what
could cause such dynamics. On one hand, treatment could have no impact on the outcome, in which case the level
difference corresponds to the cohort effect and trends are just a common feature of the environment, formalized
by the time fixed effects. On the other hand, one could note that the outcome equals the number of periods since
treated for both groups and all time periods: at the moment of treatment the outcome equals zero, it is negative
before and positive after. So a possible interpretation is that the outcome is entirely driven by causal effects of
treatment and anticipation of treatment. Note that in the former case, the solid line is a vertical shift up (a level
shift) from the dashed line, while in the latter case it is a horizontal shift to the left that is due to the differential
timing. With straight lines, these are observationally equivalent. One cannot hope to distinguish between dynamic
causal effects and a combination of cohort effects with time trends, or more generally, of unit and time effects.
8 Incontrast, underidentification of the set of point estimates {γ̃k }∞k=−∞ up to a constant, which is a well-known fact, is not
problematic because it does not in itself prevent the study of dynamic effects and pre-trends.
9 Alternatively, researchers could focus on testing for specific changes in the outcome variable after treatment that are invariant up

to a linear transformation of the path of the dynamic causal effects {γk }∞ k=−∞ , i.e. they are identified. A change in the slope of the
outcome variables or a sudden jump in the outcome variable after treatment are examples of parametric specifications that could be
tested despite the underidentification problem.

5
 Figure 1: Underidentification of Fully Dynamic Specification

We can now turn to a formal investigation of these issues. Observe that for any constant h, the following two
representations are equivalent:
∞
X
Yˆit ≡ αi + βt + γk 1 {Kit = k}
k=−∞
∞
X
= (αi + h · Ei ) + (βt − h · t) + (γk + h · k) 1 {Kit = k} (5)
k=−∞

because by definition t − Ei = Kit . As a result, the dynamic causal effects {γk + h · k} fit the data just as well as
the original {γk } path, although these two sets of coefficients paint vastly different pictures about the causal effects.
To gain further intuition about the nature of the underidentification, we show that the empirical model above
nests a specification with collinear terms. Specifically, replace the set of unit fixed effects {αi } with a linear predictor
in initial treatment period λ+αEi (i.e. the outcomes of different “cohorts” of units experiencing the event at different
times are allowed to differ in a linear way), the set of year fixed effects {βt } with a time trend βt, and the set of
fully dynamic causal effects {γk } with a trend in relative time γKit . The fundamental underidentification problem
described above can be seen immediately in the following regression:

Yit = λ + αEi + βt + γKit + uit

given that t − Ei = Kit . In other words, the presence of a linear term in the initial treatment period is necessary
for the identification problem to arise. Unit fixed effects subsume such effects. In the presence of a control group,
cohort fixed effects or unit fixed effects do not cause any identification problem because the control group pins down
the year effects. The problem is effectively the same as the (well-known) age-cohort-time problem in the regression

Yit = αE + βt + γt−Ei +uit ,

|{z}i |{z} | {z }
Cohort FE Time FE Age FE

where Ei is the date of birth.

We want to stress that typically, but not always, only a linear component of the {γk } path is not identified.10
10 There are some exception to these rules, e.g. when treatment is staggered but happens at periodic intervals.

6
It is not possible to reproduce a nonlinear {γk } path perfectly with unit and period fixed effects. The reason is
that by definition, such a path is a nonlinear function of γ(t − Ei ), and it cannot be represented as a sum of any
functions α(Ei ) and β(t).11
∞
In sum, the linear trend in the dynamic schedule of causal effects {γ̃k }k=−∞ is not identified, because one
cannot disentangle the effects of passing of absolute time t and relative time k when there is no control group and
in presence of unit fixed effects. More specifically, unit fixed effects create an identification problem because they
subsume “linear cohort effects” (i.e. linear predictors of the form λ + αEi ). The calendar year (t) is equal to the
year in which the event happens for unit i (Ei ) plus the “relative time” (Kit ): there is a perfect linear relationship
between these effects and it is therefore impossible to observe independent variation in these variables.

3.2 Solutions
3.2.1 Overview

As we explained above, calendar time, relative time, and a linear term in the initial treatment period cannot
be included together in the regression. To avoid this, additional restrictions on the statistical model have to be
imposed. Dropping unit fixed effects is an immediate fix, with the advantage of being very easy to implement, but
it suffers from two important drawbacks: it requires strengthening the identification assumption, and it reduces
power. Although dropping unit fixed effects may be a reasonable approach in some settings, we develop two other
approaches that address underidentification and do not suffer from these drawbacks. Our first strategy is to restrict
the pre-trends in the fully dynamic specification, while keeping unit fixed effects, and we show how to test this
restriction. Our second strategy is to replace unit fixed effects with unit random effects, which is also testable.
Both of these strategies can be justified by reasonable assumptions about the nature of the variation in the timing
of the event across units. The restriction of pre-trends is justified when the event is unpredictable conditional on
unit characteristics, while the random effects model is warranted when the timing of the event is randomly assigned
across units.12 Consider two examples to clarify this distinction. When the effect of health shocks on income
is of interest (e.g. Dobkin et al., 2015), it is plausible that low-income individuals may be less healthy and get
hospitalized earlier on average. Yet, conditional on permanent income level, individuals may not be able to predict
when the health shock will happen. In contrast, in the case of tax rebates studied by Parker et al. (2013), people
could know when they are getting the rebate. However, the date was assigned based on the last two digits of the
Social Security Number, so it was uncorrelated with any relevant individual characteristics.
The following subsections develop each of these approaches and discuss two related issues: the role of unit fixed
effects in balanced samples, and the possibility of adding a control group as a solution to the identification problem.

3.2.2 Restricting Pre-Trends

We begin with the situation where event timing is supposed to be randomly assigned conditionally on the fixed
effect α̃i , and unpredictable. The former assumption justifies the use of difference-in-differences type approaches.
And latter one means that the outcome cannot be adjusted based on anticipation of the event, so there can be no
pre-trends, γ̃k = 0 for k < 0. This assumption can be tested statistically and graphically, and then imposed for
efficient estimation of causal effects. We discuss these issues in order.
Under the no pre-trends null hypothesis, the true model is semi-dynamic, which has no identification issues. The
alternative allows for the fully dynamic specification, which is only set identified. Despite this, the F-test works and
can be implemented in the following straightforward way. Start from the fully dynamic regression and drop any
11 To see why this is the case, imagine that t and E are continuous and take a cross-partial derivative ∂ 2 γ(t − E )/∂t∂E =
i i i
−γ 00 (t − Ei ) 6= 0, whenever γ is nonlinear. In contrast, α(Ei ) + β(t) always has zero cross-partial.
12 These two senses of “randomness” of the event timing appear to have been conflated in the existing literature.

7
two terms corresponding to k1 , k2 < 0. This is the minimum number of restrictions for point identification, to pin
down a constant and a linear term in Kit . Then use the F-test on the pre-trends remaining in the model.13 The
F-test compares the residual sums of squares under the restricted and unrestricted specifications, where the former
is always semi-dynamic, and the latter is fully dynamic with two restrictions. Precisely due to underidentification,
the fully dynamic specification with two restrictions is effectively unrestricted and its fit is identical for any k1 and
k2 , so the F-statistic will be invariant to k1 and k2 even in finite samples.
This test has power only against non-linear pre-trends. Indeed, nothing in the data can point to a linear pre-
trend—that is the essence of underidentification. However, if the empirical design is actually flawed, i.e. event
timing is correlated with unobservables, there is no reason for pre-trends to be exactly linear, and the test will
detect them.
While we are not aware of any empirical papers implementing this F-test, a common way to check for pre-trends
is to plot the path of γ̂k before and after treatment. Sometimes this is called the event study approach. It originates
from the standard difference-in-differences setup, where only an additive constant in γ̃k is not identified, and it is
irrelevant for visual inspection; γ̂−1 is typically set to zero. In contrast, in the staggered design two restrictions
have to be made. Different choices of the two restrictions, γ̂k1 = γ̂k2 = 0, will matter a lot for the whole picture:
the whole path of estimated γ̂k gets rotated by adding ĥ · k for some constant ĥ. If there are no pre-trends in
the data-generating process, ĥ asymptotically converges to zero for any k1 and k2 as the number of units grows.
However, in finite samples difference may be large, particularly in longer panels, since ĥ is multiplied by k.
So how does one pick the two omitted categories? While choosing k1 = −1 and k2 = −2 might seem natural, we
propose setting the omitted categories far apart. Under the null hypothesis, this greatly reduces standard errors for
most individual coefficients on the graph. To understand why, imagine that a line on a plane is drawn through two
points with fixed x-coordinates x1 6= x2 , but stochastic y-coordinates, for simplicity with mean zero. The position
of the line will be much more stable when x1 and x2 are far from each other. This is true both for the slope of
the line (the analog of ĥ) and its value at a typical x (the analog of the γ̂k ). The fully-dynamic specification with
two restriction effectively draws a line and evaluates all dynamic effects relative to it. When k1 is far from k2 , e.g.
k1 = −1 and k2 close to the most negative value of K in the sample, it will be much less likely that a linear pre-trend
(although perhaps statistically insignificant) will be visible. Remember that linear pre-trends are never possible to
detect in the data, so choosing k2 = −2 would only reduce the usefulness of the graph, distracting attention from
nonlinearities in the pre-trends.
Even if the two restrictions are chosen well, this graph should only be used to evaluate pre-trends; it does not
estimate the treatment effects efficiently. Once the researcher is comfortable with the assumption of no pre-trends,
all γk , k < 0, should be set to zero. The semi-dynamic specification should be estimated and its coefficients plotted
to provide a graphical illustration of the dynamics of causal effects.14

3.2.3 Unit Random Effects

We now consider a second sense in which the timing of the event is random: the treatment period Ei is independent
of the relevant unit characteristics—in our model (1), the time-invariant unit intercept α̃i . In such a setting, the
estimation can be carried out without unit fixed effects, which are no longer necessary for the research design.
Dropping unit fixed effects immediately addresses the underidentification problem.15 However, doing so reduces
efficiency: instead, we propose to carry out estimation in a random effects model. In addition to increasing efficiency,
another advantage of using a random effects model is that we can test the hypothesis that the treatment period is
13 Both the restricted and unrestricted specifications are identified now, so standard results about the F-test behavior apply.
14 Note that in case there is truly a linear trend in the set of {γ̃k }∞
−∞ , then the results from the fully dynamic specification can be
interpreted as a test for any change relative to this linear trend around the time of the event.
15 Recall that unit fixed effects create an identification problem because they subsume “linear cohort effects”

8
independent of the unit fixed effects. As in the case of the no pre-trends assumption, the random effects assumption
can be tested against some, although not all, alternatives, and then imposed to regain identification.
When we discussed underidentification of the fully dynamic specification, we emphasized that for any path of
{γk }, identical fit of the model can be produced with the path {γk + h · k}. But the same result holds for the unit
fixed effects: {αi − h · Ei } and {αi } fit the data equally well, as long as other coefficients in the model are adjusted
appropriately (see equation (5)). As a consequence, it is impossible to test whether α̃i is uncorrelated with Ei —the
16
estimates can always be made uncorrelated by choosing h.
Yet, independence is testable; consider the following simple F-test. The restricted specification allows for
arbitrary time and causal effects, but no cohort (or unit) effects. The unrestricted one is

X ∞
X
Yit = αe 1 {Ei = e} + βt + γk 1 {Kit = k} + noise.
e k=−∞

It is almost equivalent to the fully dynamic specification, except that cohort effects are included instead of unit
effects. As in the F-test of Section 3.2.2, two normalizations are required, but this time two arbitrary cohort terms
αe 1 {Ei = e} should be dropped.17 A joint restriction αe ≡ 0 is then tested, but because of underidentification, the
test is only powerful against nonlinear functions E [αi | Ei ].
When the researcher is comfortable to impose the independence assumption, they should use the random effects
estimator as their preferred one for the full path of {γ̃k }. Remember that in general, the setup does not imply there
are no pre-trends. If the units’ outcomes can adjust to the randomized, yet known in advance event timing, the
pre-trends are part of the treatment effect.

3.2.4 Related Issues

Using a Balanced Sample: As previously discussed, without unit fixed effects there is no underidentification
problem, but in some settings the research design requires the inclusion of unit fixed effects, for instance if treatment
is unpredictable only conditional on some time-invariant characteristics. In such settings, a potential easy fix for
the underidentification problem would be to drop fixed effects and balance the sample around the initial treatment
period (i.e. restrict the sample such that each unit appears for the same number of periods before and after the
initial treatment period). Indeed, there is a view that unit fixed effects are essentially irrelevant in event studies
on a balanced panel.
The intuition underlying this view is that balancing the sample addresses one key issue that is handled by unit
fixed effect in unbalanced panels: the changing composition of the sample. Omitting unit fixed effects when working
with an unbalanced panel may be a big assumption because of selection into treatment. For instance, units that are
treated earlier in the sample may be different from other units and because the sample is unbalanced they spend a
bigger share of the sample under treated status. Therefore, in the absence of unit fixed effect one would worry that
the estimated coefficient on the treatment dummy may partly reflect selection.
In fact, balancing the sample does not help address such selection effects. We show in Appendix D that omitting
unit fixed effects when working with balanced panels is just as worrisome as when working with unbalanced panels
(for the purpose of dealing with selection effects). A balanced panel appears to solve the “selection issue” that is
salient in the case of unbalanced panels because each unit gets treated for the same number of years during the
panel. However, in practice year fixed effects absorb all of the variation for years at the beginning (where all units
are untreated) or end (where all units are treated) of the sample. For this reason, the point estimate we obtain for
16 An example of a conceptually different moment restriction which could be tested but would not provide much intuition is

Cov α2i , Ei = 0.


17 The unrestricted specification can be estimated by OLS or random effects—the results will be identical.

9
the treatment effect by running a regression in a balanced sample is exactly the same as the one obtained by running
the same regression in this balanced sample further restricted to years in the middle of the sample (which means we
are effectively running a regression on an unbalanced panel). Appendix D discusses this more formally and shows
that endogenous selection effects in the data generating process affects the consistency of the point estimate in the
same way under balanced and unbalanced panels.

Adding a Control Group: A number of empirical papers using random timing research designs start with a
sample which includes units that are never treated, or could conceivably collect such a sample. Therefore, it would
be possible to include a control group of units that never experience treatment in the estimation sample. In many
instances, it is difficult to construct a control group that plausibly provides valid counterfactuals for the treatment
group. But assuming that such a control group is available, then including it in the estimation sample solves the
underidentification problem because the control group can be used to estimate the year effects independently of the
causal effect of treatment. However, the strategy of adding a control group has two important limitations. First,
if the control group is small relative to the treatment group, important finite-sample issues can arise. Second, one
cannot allow the control group to be on its own time trend, otherwise the underidentification problem is left intact.
We discuss both of these issues in greater depth in Section 5.

4 Negative Weighting in Canonical Regression

4.1 Problem
In this section we show that fundamental underidentification discussed above creates problems for more restricted
specifications. Our flagship case will be the canonical regression (4), but the argument extends to all specifications
with two-way fixed effects, which do not allow for flexible dynamic treatment effects. We show that these regressions
estimate an average of treatment effects that severely overweighs short-run effects and weighs long-run effects
negatively. That is, if programs P1 and P2 have similar short-run effects but P1 is uniformly more efficient in the
long-run, the canonical regression will show that P1 has lower average effect.
Assume that the design is valid in the sense that there are no pre-trends, so the true model is semi-dynamic (2).
People often summarize treatment effects by the γ coefficient from the canonical regression,

Yit = αi + βt + γDit + εit .

This specification is valid under a restriction that γ̃k are equal for all k ≥ 0, i.e. that treatment leads to an
immediate and permanent jump in the outcome variable and no further effects. This restriction should not hold in
most applications, in our view. Quite often treatment effects are growing or decaying over time, may kick in with
a delay, etc. However, there is a perception that γ should estimate average treatment effects with some reasonable
weights. While true in some other contexts (e.g. Angrist 1998), we show that this logic does not apply to the
canonical regression estimand-–the weights are not even necessarily positive.18
The first step in the argument is that γ is some weighted average of {γ̃k } with weights which can be easily
estimated from the data and solely depend on the grid —the distribution of treatment periods Ei and sample
selection.

Lemma 1. The canonical regression OLS estimand can be expressed as a weighted average of dynamic treatment
18 Imbens (2015) is one of the few papers that shows negative weighting by OLS, in a different context.

10
effects,
∞
X
γ= ωk γ̃k
k=0

with weights ωk that sum up to one and equal the coefficients for Dit from in the following regressions:

1 {Kit = k} = F Ei + F Et + ωk Dit + noise, k ≥ 0. (6)

where F Ei denotes unit fixed effects and F Et time fixed effects.

To gain intuition for this lemma, note the following: by linearity of OLS, one can recover the coefficient for Dit in
the canonical regression with Yit as the outcome by instead running two canonical regressions with subcomponents
of Yit as the outcome variable and then summing up the coefficients
 on Dit obtained in each of these regressions.19
Consider first running a canonical regression with α̃i + β̃t + ε̃it as the outcome (one could do this in theory if the
P∞
true parameters of the model were known), and then another canonical regression with k=0 γ̃k 1 {Kit = k} as the
outcome, and finally sum up the coefficients for Dit obtained in each of these two regressions to recover γ. The first
regression will load on the fixed effects and not on Dit , so γ comes solely from the second regression. By the same
logic, one further notes that γ can be recovered by running a series of canonical regressions with γ̃k 1 {Kit = k} as
the outcomes (repeating for all k ≥ 0) and summing up the coefficients. Since γ̃k is a multiplicative constant in
γ˜k 1 {Kit = k}, each of these canonical regressions generates coefficients for Dit that can be written γ̃k · ωk , for ωk
determined by specification (6). Importantly, the only variables required by (6) are the unit identifier, calendar
time, and relative time—what we call the grid.20
Now the question is whether these ωk weights are “reasonable” in some sense, and our answer is strongly negative.
The canonical regression estimand suffers from a severe short-run bias, and weights long-run effects negatively.
Although a general characterization of ωk does not seem feasible, we demonstrate our result and intuitions for it in
four ways. First, we show that the short-run bias originates from a peculiar type of extrapolation performed by the
OLS when treatment timing is heterogenous across units. Second, we connect negative weighting to the propensity
score regression to show why long-run effects are more likely to be negatively weighted. Third, we solve for an
approximation of the weights in the general case, where the short-run bias always arises. Finally, we get a striking
closed-form solution for ωk in a simplified special case.

4.1.1 Intuition: Forbidden Extrapolations

In a nutshell, the intuition for the short-run bias is related to “forbidden extrapolations” performed by OLS, where
post-treatment periods are used to provide counterfactuals for the earlier ones, rather than the other way round.21
As a benchmark with no bias, we first remind the reader how classical diff-in-diff estimators perform extrapo-
lations using double-differences. Consider the simplest setting which involves two groups of homogenous units, A
19 To be explicit, if Y = A + B, then the specifications
Y = βY X + 
A = βAX + η
B = βB X + ζ
yield
Cov(Y, X) Cov(A, X) + Cov(B, X)
βY = = = βA + βB .
V ar(X) V ar(X)
The result holds for multivariate X.
20 To show that weights always add up to one, imagine that γ̃ = γ̃ 6= 0 for all k ≥ 0. The canonical regression is then correctly
P k P
specified and provides consistent estimate γ = γ̃. But γ = k≥0 ωk γ̃, so k≥0 ωk = 1. Since weights do not depend on the true
outcomes, this result holds generally.
21 This intuition derives from Meer and West (2015, Sec. 2.1). We are very grateful to Jonathan Meer for pointing us at their insightful

paper.

11
 Table 1: Treatment Status in the Minimal Examples
Panel A Panel B
Period \ Group A B
Period \ Group A B
0 0 0
0 0 0
1 0 1
1 0 1
2 1 1

(control) and B (treatment), and two time periods 0 and 1. Units in A are never treated, and units in B receive
treatment at t = 1. Table 1A illustrates how the treatment status depends on the group and period. Then in the
canonical regression
yit = αi + βt + γDit + noise,

the estimate of the treatment effect, γ, is determined by a double difference

γclassic = (yB1 − yA1 ) − (yB0 − yA0 ) .

The difference between period-0 outcomes is extrapolated to period 1 to serve as a no-treatment counterfactual. It
is permitted by the conventional parallel trends assumption, so we call it an “admissible extrapolation”.
We now proceed to the case of a staggered treatment diff-in-diff, where a different type of a double-difference also
provides information on treatment effects. Extend the previous example by adding period 2 when unit A also gets
treated (see Table 1B). Then under the canonical specification, the difference yB2 − yA2 identifies the permanent
level difference between A and B, αB − αA , in the same way as yB0 − yA0 does. As a result,

γstaggered = (yB1 − yA1 ) − (yB2 − yA2 )

is also a valid estimator of the treatment effects, provided they are homogenous.
This double-difference becomes problematic when short- and long-run effects differ. Indeed, yB1 and yA2 both
reflect the immediate causal effect, denote it γS , whereas yB2 reflects the long-run effect γL , and the remaining
outcome observations are pre-treatment. It is then evident that γclassic = γS and γstaggered = 2γS − γL . Instead of
estimating the average, the second number puts double weight on the short-run effect and a negative weight on the
long-run one (in earlier notation, ω0 = 2 and ω1 = −1). We call that a forbidden extrapolation.
A typical panel, like the one from our three-period example, will lend itself for both admissible and forbidden
extrapolations, and both types will be used by OLS for maximal efficiency under treatment effect homogeneity. But
robustness to treatment effect dynamics is, in our view, more important for empiricists than efficiency, so forbidden
extrapolations should not be used.

4.1.2 Formal Results

We now turn to a more formal discussion of how weights ωk are determined and when they can be negative. We
show in Appendix C that the weighting scheme implicitly used by OLS is proportionate to the residuals in the
linear propensity score regression. That is, suppose treatment indicator Dit is regressed on all other right-hand side
variables—unit and period fixed effects in our case. Then observations it which have fitted values above Dit (and
hence negative residuals) will have a negative weight in the original canonical regression—larger Yit produces smaller
γ. Although Dit is a dummy variable, the linear probability model can easily generate fitted values D̂it > Dit = 1
for some post-treatment observations. The outcome variable Yit contains treatment effects for these observations,
and those will be weighted negatively.

12
 Which observations could suffer from this problem? As once-treated units stay treated forever, the probability
of being treated increases over time, so time fixed effects in the probability score regression should be increasing
in t. Similarly, units that are treated earlier (with small Ei ) are treated for a larger fraction of the periods.
Therefore, fitted values will be particularly large for observations corresponding to early treated units at the end
of the sample—precisely those which identify long-run treatment effects. They will be negatively weighted, or at
least underweighted, by the canonical regression.
This argument also implies that negative weighting will not happen in models without individual fixed effects,
such as
Yit = F Et + γDit + noise.

Fitted values from the simple propensity score regression of Dit on all calendar time dummies are fractions of
treated units in each year, which always lie between zero and one. This result is consistent with Angrist (1998)
who show that when the regression is saturated, i.e. solely includes dummies for all levels of a single categorical
variable, OLS weights treatment effects by the variance of treatment conditional on the controls (see also Angrist
and Pischke, 2008, sec. 3.3.1). Such variance is of course non-negative.
Appendix B.2 provides another treatment of the problem. Instead of solving for ωk , we look at the weighting
scheme implied by a simplified specification that is nested in the canonical specification. More precisely, unit and
time fixed effects are replaced by corresponding linear terms (we used this trick in discussing the underidentification
problem as well). We show in Proposition 2 that values of k that are larger than average (among treated observations)
are always underweighted relative to their sample weights.

4.1.3 Example of Negative Weighting

Now we consider a very simple grid that may approximate the structures of the data in some applications.

Proposition 1. Suppose there are T ≥ 2 time periods, Ei is distributed uniformly among them, and for each unit
i, the outcome is observed for all periods. Then,

(T − k) (T − 2k − 1)
ωk = , k = 0, . . . , T − 1.
T (T 2 − 1) /6

Strikingly, ωk < 0 for k > (T − 1) /2.22

Figure 2 illustrates this proposition by plotting the weights ωk for T = 20. For comparison, it also shows the
fraction sk of observations with each k in the post-treatment sample, which could be a possible definition of a
reasonable weighting scheme. It is clear that short-run effects are severely overweighted, whereas long-run effects
enter negatively. That means that a program that produces uniformly larger effects may look worse in the canonical
regression.
22 All proofs are given in Appendix B.1.

13
 Figure 2: Weights in Canonical Regressions

When treatment effects have strong dynamics, there will be a wide discrepancy between the sample size-weighted
average treatment effects and the canonical regression estimand. Figure 3 shows two examples of this. Panel A
corresponds to the case treatment permanently changes the slope (growth rate) rather than the level of the outcome,
i.e. γ̃k = k + 1 for k ≥ 0. Canonical regression completely misses the effects, estimating γ to be zero! The following
corollary formalizes the result:
Corollary 1. Suppose the assumptions of Proposition 1 hold. Then, when treatment changes the slope of the
outcome’s growth, i.e. γ̃k = ϕ(k + 1) for k ≥ 0 and some constant ϕ, then the canonical regression OLS estimand
γ = 0, regardless of ϕ and the panel length T .
The canonical regression coefficient in this case lies outside of the convex hull of the effects at each time horizon.
We show later that this is not just a theoretical possibility but happens in applications. Although negative weights
in Figure 2 are not very large, they are multiplied by the large treatment effects at longer horizons in this example.

Figure 3: Biases of Canonical Regressions

A: Treatment Affects the Slope B: Temporary Effects

Panel B of Figure 3 considers a situation when treatment effects are temporary, and the outcome gradually (in
seven periods) reverts back to the original trajectory. Canonical regression will produce a large coefficient close to
the largest, very short-run effect, and does not characterize the average well.

14
 The weighting problem extends to more flexible specifications with two-way fixed effects. Consider for example,
a “capped” specification

B−1
X
Yit = αi + βt + γk 1 {Kit = k} + γB+ 1 {Kit ≥ B} + εit .
k=0

If treatment effects after Kit = B are not constant, as assumed by this specification, this will make the estimate
of γB+ unreliable. But also, through the wrong choice of individual and time fixed effects, short-run effects will be
biased. Figure 4 provides an illustration for B = 6 when the true effects are growing linearly, as in Figure 3A. The
long-run effect is outside of the convex hull of the true effects for k ≥ 6, and short-run effects are downward biased.

4.2 Solutions
Unlike with the underidentification issue, the solution for the weighting problem is quite straightforward. Re-
searchers should never run regressions which impose any restrictions on the dynamics of treatment effects post-
treatment. They should estimate them flexibly and average the coefficients manually with some weights (e.g. pro-
portionately to the sample size). When pre-trends can be assumed away, that amounts to fitting the semi-dynamic
regression (2). If some anticipation effects are possible, a more general specification (3) should be employed with
the number of leads A reflecting the horizon of anticipation. Here A < ∞ is required for identification, and B = ∞
is necessary to avoid weighting problems.
Another solution is to find a valid control group that never experiences treatment, yet faces the same time effects
β̃t . The control group helps identify the time effects, alleviating the problem. Importantly, the control group should
not be allowed to be on a separate time trend (see Section 5.1.1).
Note, however, that if the control group is small relative to the treatment group, it would not help. For the
problem to disappear, time effects should be identified solely from the control group. Figure 5 shows, in the context
of Proposition 1, how weights ωk vary with the fraction of units (and, equivalently, observations) in the control
group. Having a 10% control group does almost nothing, and even with equally sized groups, the weighting scheme
is still quite far from sk . Running a flexible specification and averaging the effects manually seems worthwhile even
in presence of a control group.23
Although this approach solves the problem when the only type of treatment effect heterogeneity is across the
time horizon, averaging of the effects across individuals or calendar times, that is implied by the semi-dynamic
specification, may not be perfect. We return to this problem in Extensions, proposing a matching procedure that
works more generally.

5 Extensions
It should by now be clear to the reader that both problems—underidentification and negative weighting—stem from
the ability of two-way fixed effects to recover the relative time variable Kit in our staggered design without the
control group. Here we demonstrate that equivalent problems may arise in a variety of different empirical designs.
We also extend our setting to the case with more general treatment effect heterogeneity and show that OLS for
the semi-dynamic specification has undesirable extrapolation properties. We propose a matching scheme robust to
this type of heterogeneity.
23 A lazy man’s alternative is to “boost” the control group: use weighted OLS where all control group observations are weighted by

a very large number, e.g. 1,000 times more than the treatment group. This may increase standard errors but from the identification
point of view, it is equivalent to having a very large control group. This weighted OLS can be viewed as a two-stage procedure: time
effects are first identified from a regression Yit = αi + βt + noise on the control group only, and then used in the canonical regression
for the treatment group.

15
5.1 Related Problems in Other Empirical Designs
5.1.1 Difference-in-differences with Group-specific Time Trend

Consider first the standard difference-in-difference setup where all units in the treatment group (Gi = 1) are
treated in same period E, whereas control group units (Gi = 0) are never treated. The treatment indicator is
Dit = P ostt · Gi , where P ostt = 1 {t ≥ E}. While the plain vanilla event study specification

∞
X
Yit = αi + βt + γk 1 {t − E = k} · Gi + noise
k=−∞

does not suffer from any identification problems, sometimes researchers are not confident in the quality of their
control group and would like to include a group-specific trend,
∞
X
Yit = αi + βt + µt · Gi + γk 1 {t − E = k} · Gi + noise.
k=−∞

Because E is the same for all units, the group-specific time trend is collinear with (t − E) Gi in presence of individual
fixed effects.
As a consequence, restricted specifications, such as

Yit = αi + βt + µt · Gi + γDit + noise, (7)

suffer from negative weighting of dynamic effects, regardless of the relative size of the control group. Figure 6
illustrates this point for the case when there are t = 1, . . . , 20 periods, and treatment happens in the middle,
E = 11.
One can replace individual fixed effects with group-specific intercepts in (7),

Yit = βt + (α + µt) · Gi + γDit + noise,

or include unit-specific trends instead,

Yit = αi + βt + µi · t + γDit + noise.

All of these specifications are affected by the same problem.24

As we mentioned in the previous sections, the issues equally arise in staggered designs with a control group if
either group- or unit-specific time trend is included. Indeed, the relative time in the treatment group, (t − Ei ) Gi ,
can be recovered from αi + µt · Gi .

5.1.2 Unit-Specific Trends without Control Group

Return now to the staggered design without a control group. When the empirical strategy is not perfectly convincing,
it is sometimes recommended to check robustness to including unit-specific (e.g., state-specific) time trends:

∞
X
Yit = αi + βt + µi · t + γk 1 {Kit = k} + noise.
k=−∞

24 Inall of these cases, semi-dynamic specifications, e.g. Yit = αi + βt + µt · Gi + ∞

P
k=0 γk 1 {t − E = k} · Gi + noise, are fine. In this
regression µ is estimated only using non-treated observations, while in restricted specifications treatment effects influence the estimate
of µ, biasing γ.

16
Without the control group, this creates an additional problem. Not only a linear term Kit can now be recovered
2
by fixed effects, but also a quadratic term (Kit ) . Indeed,

2 2
(Kit ) = (Ei ) + t2 − 2Ei t.

These three components are nested by αi , βt , and µi ·t, respectively. As a result, the fully dynamic path of treatment
effects is identified only up to a quadratic polynomial. Three restrictions instead of two must be imposed to regain
identification, and the F -test described above is only powerful against pre-trends which have more complicated
shape.
Correspondingly, weighting in the canonical-type regression

Yit = αi + βt + µi · t + γDit + noise

becomes even worse than before.25

Given our results in the previous subsection and here, we do not recommend including unit-specific time trends
in any difference-in-difference or event study specifications (except for the case discussed in footnote 24).

5.2 Treatment Effect Heterogeneity

Throughout the paper, we imposed a strong assumption that the fully dynamic specification characterizes the true
data generating process. Treatment effects are required to depend only on the time relative to treatment, Kit ,
but otherwise are homogenous across units and calendar time periods. Formally, we define treatment effects in
(k) (−1)
terms of potential outcomes as τitk = Yit − Yit , and the fully dynamic specification requires τitk ≡ γk . Such
homogeneity, particularly across time periods, is very restrictive and difficult to reconcile with economic models—we
provide an example related to the optimal choice of consumption and saving under the permanent income hypothesis
in Appendix E.
What is the estimand of the semi-dynamic regression when heterogeneity is allowed for? To understand this,
we again invoke the result on OLS as a weighting estimator, which weights each observation by the residual from
the linear propensity score regression (see Appendix C). The propensity score regression behind γk , k ≥ 0, in the
semi-dynamic specification is the following one:
∞
X
1 {Kit = k} = F Ei + F Et + ρkl 1 {Kit = l} + noise. (8)
l=0
l6=k

An observation is weighted negatively in two cases: if Kit = k and the fitted value in (8) is above one, or if Kit 6= 0
and the fitted value is any positive number. While we did not observe the former situation in simulations, the latter
was quite prevalent among treated observations. The semi-dynamic estimand γk is a positively-weighted average
of observations with Kit = k, minus a weighted average of control observations„ plus additional terms for treated
observations with Kit = l 6= k which have both positive and negative weights summing up to zero.
The intuition for having these additional terms is that the semi-dynamic specification imposes τitk ≡ γk , allowing
for substantial extrapolation. For instance, γ1 can be estimated by comparing units treated at Ei = 1 and 2, both
observed at periods t = 1 and 3. Alternatively, it can be estimated from units treated at Ei = 5 and 6 observed
at t = 5 and 7. The difference between the resulting estimates is consistent for zero when the semi-dynamic
specification is true, so an estimator for γ0 can add or subtract this difference multiplied by any constant. Such
extrapolation can improve efficiency, but also makes OLS non-robust to heterogeneity.
25 Illustrative simulations are available from the authors upon request.

17
 Severe extrapolation is inevitable in some cases. Suppose for example that treatment happens to all units
between periods tF and tF + 3, but we want to estimate γ7 —the effect of being treated seven periods ago compared
to not having been treated yet. There is no direct difference-in-differences quasi-experiment that would help for
this task, because in calendar periods where some unit has been treated for 7 periods, all units have already been
treated.
However, for k smaller than the range of treatment periods, γ̃k can be estimated without extrapolation. As
long as individual heterogeneity is captured by unit fixed effects, as we have always assumed, one can find periods
t0 < t, as well as units i and j treated at Ei < Ej and observed in these periods, which satisfy t0 < Ei < t < Ej
and t = Ei + k. That is, unit i has been treated at t for k periods, but not treated yet at t0 , whereas unit j has not
been treated in either period. When there are no pre-trends, j provides a valid counterfactual to i.
The simplest way to estimate treatment effects without extrapolation is to run the semi-dynamic regression

Yit = αi + βt + γk Dit + noise (9)

for each k ≥ 0 separately on the subsample of the data which includes observations with Kit = k, as well as the
entire control group Kit < 0. This regression cannot suffer from extrapolation of treated observations because the
only type of treatment effects the is present in the estimation sample here is γk . We will use this strategy in the
application in Section ??.
Another approach is matching. For each treated observation (i, t) there are multiple t0 and potentially many
units j that i can be matched with, subject to the rules discussed above. For efficiency reasons, all of them should
be used with some weights wit,jt0 , and the resulting matching estimator can be written as


Yit − YitCF 1 {Kit = k}
P
i,t
γ̂k = P
i,t 1 {Kit = k}

where YitCF =
P P
j,t0 wit,jt0 · (Yit0 − Yjt0 + Yjt ) is the counterfactual outcome and j,t0 wit,jt0 = 1.
Importantly, γ̂k constructed in this way weights treated observations positively, in fact equally, to obtain the
average treatment effect on the treated. This matching estimator can be rewritten as a weighting estimator, which
extends Hirano et al. (2003) (on regressions with controls) and Abadie (2005) (on difference-in-differences with a
single pre-period) to our event study design.
The only paper known to us which uses a similar estimator is Fadlon and Nielsen (2015), except that they
require Ej = Ei + 5 and t0 = Ei − 2. That reduces efficiency without relaxing any assumptions.26

6 Empirical Relevance
This section is available from the authors upon request.

7 Conclusion
Difference-in-difference estimators are some of the most prevalent tools in economics for estimating causal effects of
treatments. While the classical design compares two groups of units before and after a single date when one of them
gets treated, researchers also frequently work with panels where different units are treated at different moments.
For instance, in broadly used event studies all units in a panel eventually get treated, but some earlier than others.
26 Fadlon and Nielsen (2015) test the identification assumptions by comparing 5-year pre-trends of treated observations and their

counterfactuals. In our notation this is a test for γ̃−1 − γ̃−6 = · · · = γ̃−5 − γ̃−10 , a hypothesis weaker than the linearity of pre-trends
discussed in Section 3.2.2.

18
This paper showed that because of underidentification issues and negative weighting in event study designs, results
from common specifications are likely to seem non-robust. These problems can be alleviated in a principled way by
using parametric and semi-parametric estimators and tests.
Two approaches were introduced to address the underidentification issue. The first is to restrict the pre-trends
in the fully dynamic specification, while keeping unit fixed effects. The second consists in replacing unit fixed effects
with unit random effects. Statistical and graphical tests were developed to prove the validity of these approaches.
The paper also showed that the canonical estimator does not provide a reasonable average of treatment effects along
an important dimension---the time since first treatment---although other methods can. Specifically, the canonical
estimator has a severe short-term bias and in many cases weights long-term effects negatively. To address the
problem, alternative parametric and semi-parametric estimation techniques were introduced, which always average
dynamic effects in a convex way. The parametric approach amounts to estimating the dynamic regression that
includes dummies for all lags of treatment and then manually averaging its coefficients, whereas the semi-parametric
one is a new weighting estimator in the spirit of Hirano et al. (2003).
The practical relevance of this problem was established in light of the existing literature and in a specific
application about the estimation of the marginal propensity to consume out of tax rebates.

References
Abadie, Alberto, “Semiparametric Difference-in-Difference Estimators,” Review of Economic Studies, 2005, 72,
1–19.

, Alexis Diamond, and Jens Hainmueller, “Comparative Politics and the Synthetic Control Method,” Amer-
ican Journal of Political Science, 2015, 59 (2), 495–510.

Allegretto, Sylvia, Arindrajit Dube, Michael Reich, and Ben Zipperer, “Credible Research Designs for
Minimum Wage Studies,” IZA Discussion Papers, 2013, 7638 (7638).

Angrist, Joshua, “Estimating the labor market impact of voluntary military service using social security data on
military applicants,” Econometrica, 1998, 66 (2), 249–288.

Angrist, Joshua D. and JS Pischke, Mostly harmless econometrics: An empiricist’s companion, Princeton
University Press, 2008.

Broda, Christian and Jonathan A. Parker, “The economic stimulus payments of 2008 and the aggregate
demand for consumption,” Journal of Monetary Economics, 2014, 68 (S), S20–S36.

Dobkin, Carlos, Amy Finkelstein, Raymond Kluender, and Matthew J Notowidigdo, “The Economic
Consequences of Hospital Admissions,” 2015.

Fadlon, Itzik and Torben Heien Nielsen, “Household Responses to Severe Health Shocks and the Design of
Social Insurance,” 2015.

Hirano, Keisuke, Guido W. Imbens, and Geert Ridder, “Efficient Estimation of Average Treatment Effects
Using the Estimated Propensity Score,” Econometrica, 2003, 71 (4), 1161–1189.

Imbens, Guido W., “Matching methods in practice: Three examples,” Journal of Human Resources, 2015, 50
(2), 373–419.

Meer, Jonathan and Jeremy West, “Effects of the Minimum Wage on Employment Dynamics,” Journal of
Human Resources, 2015.

19
Parker, Jonathan A., Nicholas S. Souleles, David S. Johnson, and Robert McClelland, “Consumer
Spending and the Economic Stimulus Payments of 2008,” American Economic Review, 2013, 103 (6), 2530–2553.

Sloczynski, Tymon, “A General Weighted Average Representation of the Ordinary and Two-Stage Least Squares
Estimands,” 2016.

Wolfers, Justin, “Did Unilateral Divorce Laws Raise Divorce Rates? A Reconciliation and New Results,” American
Economic Review, 2006, pp. 1802–1820.

A Additional Figures and Tables

Figure 4: Biases in the Capped Regression

Figure 5: Canonical Weights with Control Group

20
 Figure 6: Weights in Diff-in-Diff Specifications

Notes: “Classical” refers to specification Yit = αi + βt + γDit . The specification with group-specific trends is given
by (7).

B Additional Results and Proofs

B.1 Proofs
Proof of Proposition 1. [TBA]

Proof of Corollary 1. It is straightforward to verify that weights from Proposition 1 satisfy

(k + 1)ωk = −(l + 1)ωl

whenever k + l = T − 1. Using this fact and Proposition 1, we can write

−1 −1 −1
T T T
!
X 1 X X
γ= (k + 1) ωk = (k + 1) ωk + (l + 1) ωl = 0.
2
k=0 k=0 l=0

B.2 Weights in a Simplified Specification

Recall that two-way fixed effects can reproduce a linear trend in the relative time Kit , which is at the core of the
underidentification problem. Consider the following regression that is nested within the canonical one:27

Yit = µ + hKit + γ R Dit + εit . (10)

As in Lemma 1, γ R = ωkR γ̃k with weights that can be estimated from a series of regressions
P
k≥0

1 {Kit = k} = µ + hR R
k Kit + ωk 1 {Kit ≥ 0} + noise, k ≥ 0. (11)

This regression (11) is defined solely in terms of a single variable Kit . Figure 7 illustrates it in the context of
Proposition 1 for two values of k: small (k = 3  T = 20) and large (k = 12). Specification (11) fits the left-hand
side variable with two parallel lines of any location and slope, and ωkR is the shift between them. When k is large,
27 To get it, plug in αi = −hEi and βt = µ + ht.

21
this shift can be negative (see panel B). Moreover, short-run effects are always overweighted, as formalized by the
following proposition:28

Proposition 2. The weight schedule implied by regression (10) is always short-run biased. That is, ωkRestricted /sk is
P∞
decreasing in k. It is greater than 1 if and only if k is smaller than the average Kit post-treatment, i.e. k < l=0 lsl .

Figure 7: Estimation of Specification (11)

A: For Small k = 3 B: For Large k = 12

Proof. [TBA]

C Negative Weighting
The derivation below illustrates why in regressions with non-saturated controls OLS may assign “negative weights”
to treated observations, although one wishes to have positive weights for treated observations and negative weights
for control observations. We start from a simple setup with treatment indicator D and control variables X.

Y = βD + αX + ε

By the Frisch-Waugh-Lovell theorem,

Y − f 0 X = β (D − π 0 X) + ε∗

where f 0 X and π 0 X are linear projections of Y and D on X, respectively. The latter is the propensity score
estimated using the linear probability model. Then,

Cov(Y − f 0 X, D − π 0 X)
β=
V ar(D − π 0 X)
We can write this as a sum of Yi s multiplied by weights, or using the population notation,

1 D − π0 X
β = E [(Y − f 0 X)(D − π 0 X)] = E [Y · ω(D, X)] , for ω(D, X) = ,
V ar(D − π 0 X) V ar(D − π 0 X)
28 It should be primarily viewed as providing intuition because ωk 6= ωkR in general.

22
where the second equality uses the fact that f 0 X is a linear function of X, hence must be uncorrelated with
(D − π 0 X). It is clear that the OLS-implied weights are proportionate to the residual in the propensity score
regression.
It is straightforward to show that E [ω · D] = 1 and E [ω · (1 − D)] = −1. That is, β is an average of the outcomes
in the treatment group minus an average of outcomes in the control group, with weights adding up to one in both
cases. In the standard Rubin causal model, one can write Y = Y0 + τ D, where Y0 is the no-treatment potential
outcome, and τ is the (heterogenous) treatment effect. Therefore, OLS estimates

β = E [Y0 · ω] + E [τ · (ω · D)] .

The first term (with the weights adding up to zero) represents selection bias, and the second one—the average of
treatment effects. Even if selection is not a problem, the second term is problematic when ω < 0, i.e. π 0 X > D = 1.
This can never happen in a saturated regression, but is very likely for at least some observations in regressions with
continuous controls or multiple sets of fixed effects.
As simple as the result is, we are unaware of any other paper showing it. Abadie et al. (2015) note for the
setting with one treated and many untreated observations that the untreated ones can be weighted negatively.
However, they do not connect this to the propensity score regression or, since they do not allow for multiple treated
observations, to the averaging of treatment effects.

In the setting discussed in Section 4, X is a set of individual and time dummies. Then, people who are
treated earlier in the sample have more observations with treated status D = 1. Also, there are also more treated
observations in the later periods because treatment status doesn’t revert to 0. Therefore, the largest treated values
are for the long-run treatment effects (high Kit ) and we can get negative weights. These weights just depend on
the “grid”—the distribution of calendar time t and initial treatment periods Ei in the sample—because there is no
other variable in the propensity score regression.
This result is particularly worrisome if the treatment effect is dynamic because large treatment effects in the
long run get assigned a weight of the wrong sign. Because of this, the γ estimated in the canonical regression could
be outside of the convex hull of the true treatment effects γ̃k . For instance if the treatment effect is positive and
growing over time, the estimated γ could be negative although we were hoping it would be a weighted average of
γ̃k (we show in Section 6 that this in fact happens in several important empirical applications).

D Unbalanced vs. Balanced Panels and Individual Fixed Effects

As discussed in Section 3.2.3, one approach to regain identification is to do away with unit fixed effects and hope
that this does not pose a threat to identification. In this section, we clarify the nature of the potential threats
to identification when excluding unit fixed effects for both unbalanced and balanced panels. For this section,
assume throughout that there are no year effects to simplify the analysis (also note that since we are considering
specifications without individual fixed effects, the other concerns with the canonical regression discussed in 4 do not
apply). We believe there is a conventional wisdom in applied work that omitting unit fixed effects when working
with unbalanced panels is a big assumption because of selection into treatment, while omitting unit fixed effects in
the case of balanced panels is much less problematic. We discuss below why the two settings in fact pose similar
issues (as in the rest of this note, the discussion considers setting without a control group of units that never
experience treatment).
Unbalanced panels. It is well understood that when panels are unbalanced, if there is a correlation between
the time of treatment and the unit fixed effects (in our notation, Cov(αi , Ei ) 6= 0), then including individual

23
fixed effects is key. Such a correlation could result from intuitive patterns of endogenous selection into treatment.
Consider for instance a setting where i) treatment has a positive and constant effect, ii) unit fixed effects αi reflect
the bargaining power of the unit, which allows for better outcomes Yit in general and also for earlier selection into
treatment (e.g. Cov(αi , Ei ) < 0). In this setting, running a regression of the form Yit = βt + γTit + uit (without
unit fixed effects) yields an upwardly biased estimate of the true constant treatment effect. Intuitively, Tit conveys
information about the “type” of the unit: units with higher individual fixed effects are treated earlier in the sample,
i.e. for a longer period of time in the sample, and the estimated treatment effect coefficient partly captures these
higher fixed effects. In simulated data, this can be checked by running a regression of the form Tit = λαi + uit ,which
yields λ < 0. Balanced panel may at first glance appear to be impervious to this issue.
Balanced panels. By construction, in a balanced sample each unit gets treated for the same number of
periods of the observed sample. Therefore, in simulated data with the same data generating process as discussed
for unbalanced panels, running a regression of the form Tit = λαi + uit yields λ ≈ 0. Is it sufficient to restrict the
sample to a balanced panel to address the concerns resulting from endogenous selection into treatment discussed
in the case of unbalanced panels? And if not, why not and what is the link with unbalanced panels? We have
verified in simulations that balancing the sample does not solve the problem, and here we provide intuition for why
selection into treatment correlated with individual fixed effects in fact poses exactly the same problem in balanced
and unbalanced panels. The intuition can be best seen based on the following figure:

Figure 8: Share of Treated Units and Unit Fixed Effects over Time in Balanced Sample
1.5
.2
Share of Units Experiencing Treatment

.1

1
(in Balanced Sample)

(in Balanced Sample)

Unit Fixed Effects
0

.5
-.1

0
-.2
-.3

-.5
0 5 10 15 20 0 5 10 15 20
Year Year

Figure 8 is based on a simulation with Cov(αi , Ei ) < 0 (“better” units get treated earlier in calendar time),
where the sample used for regressions is restricted such that each unit is observed for k years before the first year of
treatment, as well as during the year of treatment and for k − 1 years after. In other words, each unit is observed for
k years under treated status and for k years under untreated status (in our simulation, k = 7). Figure 8 illustrates
two points. First, in the first k years of the balanced sample none of the units are treated, and in the last k years
of the sample all units are treated. This means that the year fixed effects for years that are early and late in the
sample will apply to a homogeneous group of units, which are all either treated or untreated - in other words, they
will absorb the treatment effect in those years. This means that identification of the treatment effect coefficient will
come entirely from observations in the interval of time between those cutoff dates (i.e. with t such that k < t < T̄ −k,
which in our simulation amounts to 6 years, with a total number of years given by T̄ = 20). Second, there are
composition effects over the years in terms of the unit fixed effects: units that show up earlier in the sample tend
to have higher fixed effects (because Cov(αi , Ei ) < 0). The composition of unit fixed effects is stable in a balanced

24
sample for years t such that k ≤ t ≤ T̄ − k, which in our simulation amount to 8 years. Intuitively, a balanced
sample appears to solve the “selection issue” discussed in the setting of unbalanced samples because each unit gets
treated for the same number of years during the sample: however, in practice year fixed effects absorb all of the
variation for years t that are not such that k < t < T̄ − k (i.e. that are at the beginning or end of the sample). For
this reason, the point estimate we obtain for the treatment effect by running Yit = βt + γTit + uit in the balanced
sample is exactly the same as the one we obtain by running the same specification in the balanced sample further
restricted to t such that k < t < T̄ − k (the middle of the sample).29 In this restricted sample (which is unbalanced !)
we have Cov(αi , Ei ) < 0.30 Intuitively, once we eliminate the observations that are at the very beginning and the
very end of the sample, we are left with a sample where units with higher fixed effects spend a bigger share of the
sample under treated status.31 This shows that endogenous selection effects in the data generating process affects
the consistency of the point estimate in the same way under balanced and unbalanced panels.
Another way to summarize the intuition is as follows: the idea of irrelevance of unit fixed effects in balanced
panel is coming from the observation that Tit is orthogonal to the unit dummies in the balanced sample. This holds
unconditionally but fails conditionally on the time dummies, which are always included in these regressions. This
failure is obvious: if ones fixes t, the dependence between Tit (equivalently, Ei ) and the true unit fixed effects is
precisely the problem we were hoping to address - and it exists even restricting the sample to individuals observed
at t in the balanced sample.

E An Economic Model: the Permanent Income Hypothesis

To see how restrictive the assumption of homogenous treatment effects across cohorts and time is, we consider the
simplest setting where event timing is fully random and unpredictable. All units (for simplicity, let us think of them
as people for now) have the same prior belief about when the event may happen, and they update it over time using
the Bayes rule. If a person does not receive treatment at the beginning of period t, this may be a minor shock if
she thinks she is likely to get it next period, or a large shock if the expected date of the event moves far into the
future, and behavior will respond accordingly. Only under very specific prior distributions can the treatment effect
be independent of t.
For a specific example, consider an agent who lives for t = 1, . . . , T periods and is expecting to get a permanent
raise of R from her baseline wage normalized to zero. The raise will happen at the beginning of a random period
drawn from some distribution, and she does not get any information until the date of the raise. Her consumption
follows the permanent income hypothesis, and there is no discounting. What is the effect of getting a raise on
consumption at impact, τit0 ? If she gets the raise at time t, she is certain that the permanent income per period is
R. If the event does not happen at t, it equals R · E [T − Ei + 1 | Ei > t] / (T − t + 1). The treatment effect equals
the difference between the two:
E [Ei − t | Ei > t]
τito = R · .
T −t+1
It is independent of t only if Ei is uniformly distributed across periods.

29 We have verified in simulation that the point estimate and standard errors for the treatment effect are indeed exactly the same in

the balanced sample and the balanced sample with the further restriction that k < t < T̄ − k.
30 We have verified this in the simulation by running T = λα + u in the balanced sample restricted to t such that k < t < T − k,
it i it
which indeed yields λ < 0
31 For instance, the units that were treated at the earliest possible time, i.e. with T = k + 1, are now treated in 100% of observations.
i
Before further restricting the balanced sample, they were (by definition!) treated in 50% of observations. Conversely, the units that
1
were treated at the latest possible time, i.e. with Ti = T̄ − k − 1, appear as treated in T̄ −2k % of observation, or 16.66% of observations
in our simulation.

25