NBER WORKING PAPER SERIES

CAUSAL MODELS FOR LONGITUDINAL AND PANEL DATA:

A SURVEY

Dmitry Arkhangelsky
Guido Imbens

Working Paper 31942

http://www.nber.org/papers/w31942

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
December 2023

This paper came out of the Sargan lecture presented online at the 2021 Royal Economic Society
Meetings. We are grateful for comments by Manuel Arellano, Apoorva Lal and Ganghua Mei.
This work was partially supported by the Office of Naval Research under grants
N00014-17-1-2131 and N00014-19-1-2468 and a gift from Amazon. This work was partially
supported by the Office of Naval Research under grants N00014-17-1-2131 and
N00014-19-1-2468 and a gift from Amazon. The views expressed herein are those of the authors
and do not necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.

© 2023 by Dmitry Arkhangelsky and Guido Imbens. All rights reserved. Short sections of text,
not to exceed two paragraphs, may be quoted without explicit permission provided that full
credit, including © notice, is given to the source.
Causal Models for Longitudinal and Panel Data: A Survey
Dmitry Arkhangelsky and Guido Imbens
NBER Working Paper No. 31942
December 2023
JEL No. C23

ABSTRACT

This survey discusses the recent causal panel data literature. This recent literature has focused on
credibly estimating causal effects of binary interventions in settings with longitudinal data, with
an emphasis on practical advice for empirical researchers. It pays particular attention to
heterogeneity in the causal effects, often in situations where few units are treated. The literature
has extended earlier work on difference-in-differences or two-way-fixed-effect estimators and
more generally incorporated factor models or interactive fixed effects. It has also developed novel
methods using synthetic control approaches.

Dmitry Arkhangelsky
CEMFI
5 Calle Casado del Alisal
Madrid 28014
Spain
[email protected]

Guido Imbens
Graduate School of Business
Stanford University
655 Knight Way
Stanford, CA 94305
and NBER
[email protected]
 1. Introduction

In recent years there has been a fast-growing and exciting body of research on new
methods for estimating causal effects in panel data settings. This literature has taken
some of the elements of the earlier panel data literature and brought in insights from
the causal inference literature. It has largely focused on the case with binary treatments,
although the insights obtained in this body of work extend beyond that setting. Much
of this work focuses on settings where traditionally Difference-In-Differences (DID)
methods have been popular among empirical researchers. In this survey, we review
some of the methodological research and make connections to various other parts of
the panel data literature. Part of the challenge in writing this survey is the organization.
We have chosen one that is different from that in the rest of the literature (with the
exception of [ABD+ 21a], but one we felt was helpful for organizing our own thinking. It
is clear that one could organize the material differently.
Although we intend to make this survey of interest to empirical researchers, it is
not a guide with recommendations for specific cases. Rather, we intend to lay out out
views on this literature and present the key features of various methods in order that
practitioners can decide which of the methods they wish to use in particular settings.
In line with most of the literature we see the models and assumptions used in this
literature not as either holding or not, but as approximations that may be accurate
and useful in particular settings. For example, the two-way-fixed-effect set up has been
critized as making assumptions that are too strong. At some level that is true almost
by definition: parallel trends are unlikely to hold over any extended period of time.
Similarly, assuming the absence of dynamic effects is unlikely to ever hold, and similarly
treatment effects are almost always heterogeneous. Nevertheless, in many cases fixed
effect models with constant treatment effects may be good working models that lead to
robust estimates and inferences. Understanding when they do so, and when relaxing
them is likely to improve the combination of bias-squared and variance, and in what
way to generalize them without sacrificing too much precision is close to what we intend
to do here.
After this introduction we first discuss in Section 2 some of the earlier econometric
panel data literature. This serves both to set the stage for the framing of the questions
of the current literature, as well as to clarify differences in emphasis between the
traditional and new literatures. We also point out conceptual issues that have been
raised in the earlier literature and that have received less attention, or even are in

1
danger of being entirely ignored, in the current literature.
Next, in Section 3 we discuss three ways of organizing the panel data literature, first
by types of data available, e.g., proper panel data, repeated cross-sections, or two-way
data structures, second by shape of the data frame, many units or many periods, and
third by the assignment for the causal variable of interest. We find this organization
useful because it matters for the relevance of various methods that have been proposed.
Although the earlier econometric panel data literature also stressed the importance of
the relative magnitude of the time and unit dimension as we do in our second classifica-
tion, the realization that the structure of the assignment matrix is important is a more
recent insight. Many of the recent papers focus on particular parts of the general space
of panel data inferential problems. For example, the vast unconfoundedness literature
has focused largely on the setting with panel data with large N and relatively small T,
and a subset of the units treated in the last period. In contrast, the Synthetic Control (SC)
literature has primarily focused on the setting where N and T are comparable in size,
and one or few units are treated from some period onwards. The recent DID literature
has paid particular attention to the setting with staggered adoption patterns in the as-
signment. The singular focus of some of these literatures has helped in advancing them
more rapidly, but occasionally insights from related settings have been overlooked.
In Section 4 we introduce some of the notation and estimands. We use, as in much
of the causal inference literature, the potential outcome notation that makes explicit
the causal nature of the questions.
In Section 5 we introduce the standard DID and Two-Way-Fixed-Effect (TWFE) setup
as a stepping stone to our following discussion of the recent developments in causal
panel data literature in Sections 6 through 10. We see four main threads to the new causal
panel literature. First, one strand of the literature has focused on a more general setup
where different groups adopt treatments at different points in time. In this staggered
adoption case recent research has highlighted some concerns with the standard TWFE.
In particular, in cases with general treatment effect heterogeneity, the implicit weights
on the building blocks of the TWFE estimator have been argued to be counter-intuitive,
and alternatives have been proposed. We discuss this literature in Section 6.
Second, the literature has generalized the popular two-way additive fixed effect
structure. One direction has allowed for factor models, whereas another strand of the
literature generalized the models to allow for nonlinear structures. An important part
of the first extension that allows for factor structures is the synthetic control approach
developed in a series of papers by Alberto Abadie and coauthors [ADH10]. Although

2
 TABLE 1: ACRONYMS

DID Difference In Differences

TWFE Two Way Fixed Effect
SC Synthetic Control
GRCS Grouped Repeated Cross Section
RCED Row Column Exchangeable Data
SDID Synthetic Difference in Differences
CIC Changes in Changes

this literature shares features with the TWFE and DID literature, it has partly developed
separately, ignoring some of the gains arising from combining the insights from each
of them. We discuss this literature in Section 7. The second generalization, allowing for
non linear models, is discussed in Section 8.
Fourth, the modern causal panel literature has sometimes taken a design-based
approach to inference where the focus is on uncertainty arising from the assignment
mechanism, rather than a model-based or sampling-based perspective that is common
in the earlier literature. We discuss this literature in Section 9.
Finally, in Section 10 we discuss open questions which we view as exciting avenues
for future research.
There are some excellent recent surveys of the new DID and causal panel data litera-
ture. They differ in their focus, and in the perspectives of the authors, and complement
ours in various ways. Some of these surveys [DCd23, RSBP23] focus primarily on the DID
setting with heterogenous treatment effects, with less attention paid to the connections
with the synthetic control methods and factor models that we view as an important
feature of the current panel data literature. In contrast, [Aba19] focuses primarily on
synthetic control methods. In the current survey we stress deeper linkages between
these ideas as well as potential benefits from combining them. Recent surveys in the
political science literature, [Xu23, LWX22], like the current survey, also discusses the
connections between synthetic control, unconfoundedness, and TWFE approaches.
In this discussion we use a number of acronyms. We list those in Table 1.

2. The Econometrics Panel Data Literature

Although the new panel literature focuses on different estimands and settings, and
emphasizes different concerns about internal and external validity, many of the methods

3
overlap with those discussed in the earlier econometric panel data literature. Here
we discuss at a high level some of the key insights from the earlier literature, in so
far they relate to the current literature, and some marked differences between the
two. We come back to some of the specific areas of overlap in later sections. We do
not attempt to review this entire literature, partly because that is a vast literature in
itself, but mainly because there are many excellent surveys and textbooks, including
[Are03, Bal08, Hsi22, Woo10, AB11b].
First of all, by the econometric panel data literature, we mean primarily the literature
from the 1980s to the early 2000s, as for example reviewed in the surveys and textbooks
including [Cha82, Cha84, HSCKW12, Are03, AH01, Hsi22, Bal08, Woo10, AB11b]. This
literature was initially motivated by the increased availability of various large public-
use longitudinal data sets starting in the 1960s. These data sets included the Panel
Study of Income Dynamics, the National Longitudinal Survey of Youth, which are
proper panels where individuals are followed over fairly long periods of time, and the
Current Population Survey, which is primarily a repeated cross-section data set with
some short term panel features. These data sets vary substantially in the length of the
time series component, motivating different methods that could account for such data
configurations.
The primary focus of the econometric literature has been on estimating invariant
or structural parameters in the sense of [Gol91]. Part of the literature analyzed fully
parametric models, but more often semiparametric settings were considered. The
parameters of interest could be causal in the modern sense, but the term would rarely
be used explicitly. A major concern in this literature has been the presence of time-
invariant unit-specific components. The literature distinguished between two types
of such components, fixed effects that were conditioned on in the analyses and were
modeled as unrestricted in their correlation with other variables, versus random effects
that were treated as stochastic and often uncorrelated with observed covariates (though
not always, see the correlated random effects discussion in [Cha84]). See for general
dsicussions [Hsi22, BJ15]. This distinction was often used as an organizing principle
for the panel data literature, in combination with the reliance on fixed T versus large T
asymptotic approximations. A substantial literature was devoted to identification and
inference results in settings with fixed effects leading to various forms of what Neyman
and Scott labeled the incidental parameter problem [NS48]. Especially when the fixed
effects entered in in non-additive and non-linear ways in short (fixed length) panels,
with limited dependent or discrete outcomes, this led to challenging identification

4
problems e.g., [Cha80, Mag04, Hon92]. In cases where identification in fixed T settings
was not feasible the literature introduced various methods for bias-correction (see
[AH07b] for a survey), or developed bounds analyses (e.g., [HT06]). More recently, these
bias-reduction ideas were extended to nonlinear two-way models [FVW16, FVW18].
The earlier econometic panel data literature paid close attention to the dynamics in
the outcome process, arising from substantive questions such as the estimation of struc-
tural models for production functions and dynamic labor supply. Motivated by these
questions this literature distinguished between state dependence and unobserved het-
erogeneity (e.g., [Hec81, Cha84]) and various dynamic forms of exogeneity (e.g., weak,
strong and strict exogeneity, and predeterminedness, see [EHR83, AB91]), issues that
have not received as much attention yet in the current literature. The earlier literature
also studied models that combined the presence of unit-fixed effects with lagged depen-
dent variables, leading to concerns about biases of least squares estimators in short pan-
els and the use of instrumental variable approaches [Nic81, AB91, BB98, HK02, AA03].
This literature had a huge impact on empirical work in social sciences.
In contrast, an important theme in the current literature is allowing for general het-
erogeneity in causal effects, both over time and across units, associated with observed
as well as unobserved characteristics. Much of the earlier literature often restricted
the heterogeneity in treatment effects to parametric forms or assumed its absence
entirely. Exceptions include [Cha92, AB11a, GP12, CFVHN13], which emphasized het-
erogeneity in a way that is more in line with the current literature. The recognition of
the importance of heterogeneity has led to findings that previously popular estimators
are sensitive to the presence of heterogeneity and the development of more robust
alternatives.

3. Set Up and Data Configurations

In this section, we consider different types of data as well as two classifications of panel
data configurations. One of these classifications, in terms of the relative size of the
cross-section and time-series dimensions, is familiar from the earlier literature, but the
second, in terms of the assignment mechanism, is original to the current literature. In
the earlier literature, there was an additional classification that made a distinction that
depended on the heterogeneity between cross-section units being modeled as fixed
effects or random effects, e.g., [Cha84]. This distinction plays less of a role in the current
literature.

5
 Both these classifications are helpful in understanding which specific methods may
be useful and what type of asymptotic approximations for inference are credible. In
addition, they allow us to place the individual papers in context.
To put the following discussions into context, it is also helpful to keep in mind that
in most cases we are interested in the average causal effect of some intervention on
the outcomes for the treated units, during the periods they were treated. Later we are
more precise about the exact estimands we focus on, and in particular, how some of the
assumptions such as the absence of dynamic effects affect the interpretation of those
estimands.

3.1. Data Types

Although we focus in this paper mostly on the proper panel data setting where we
observe outcomes for a number of units over a number of time periods, we also consider
grouped repeated cross-section settings. Here we want to clarify the distinction and be
precise about the notation.

3.1.1. Panel Data

In the proper panel data case we have observations on N units, indexed by i = 1, . . . , N,

over T periods, indexed by t = 1, . . . , T. The outcome of interest is denoted by Yit , and
the treatment is denoted by Wit , both doubly indexed by the unit and time indices.
These observations may themselves consist of averages over more basic units as in the
grouped repeated cross-section case from Section 3.1.2. We collect the outcomes and
treatment assignments into two N × T matrices:
   
Y Y12 Y13 ... Y1T W W12 W13 ... W1T
 11   11 
 Y
 21 Y22 Y23 . . . Y2T   W
  21 W22 W23 ... W2T 

   
 Y31 Y32 Y33
Y= . . . Y3T 
,  W31 W32 W33
W= . . . W3T 
,
 . .. .. .. ..   . .. .. .. .. 
 .. . . . .   .. . . . . 
   
YN1 YN2 YN3 . . . YNT WN1 WN2 WN3 . . . WNT

with the rows corresponding to units and the columns corresponding to time periods.
We may also observe other exogenous variables, denoted by Xit or Xi , depending on
whether they vary over time. Most of the time we focus on a balanced panel where for all
units i = 1, . . . , N we observe outcomes for all t = 1, . . . , T periods. In practice, concerns

6
can arise from the panel being unbalanced where we observe units for different lengths
of time, or from missing data. We ignore both complications in the current discussion.
Classic examples of this proper panel setting include [Ash78] with information on
earnings for over 90,000 individuals for 11 years, and [AC86] with information on wages
for 1,448 individuals also for 11 years. Another classic example is [CK94] with data for
two periods and 399 fast food restaurants.

3.1.2. Grouped Repeated Cross-section Data

In a Grouped Repeated Cross-section Data (GRCD) setting we have observations on N

units. Unit i is observed only once, in period Ti , which is indexed by i to account for the
fact that different units may be observed at different points in time. Typically Ti takes
on only a few values (the repeated cross-sections), e.g., just two or three, with many
units sharing the same value for Ti , although for some of the methods this is formally
not required. The outcome and treatment received for unit i are denoted by Yi and Wi
respectively, both indexed just by the unit index i.1 The set of units is partitioned into two
or more groups, with the group that unit i belongs to denoted by Gi ∈ G = {1, 2, . . . , G}.
Define the average outcome for each group/time-period pair:

N
, N
X X
Y gt ≡ 1Gi =g,Ti =t Yi 1Gi =g,Ti =t ,
i=1 i=1

and similar for W gt . If we view the G × T group averages Y gt , instead of the original
Yi , as the unit of observation, this grouped repeated cross-section setting is just like
a panel as in Section 3.1.1, immediately allowing for methods that require repeated
observations on the same unit as pointed out in [Dea85] and [Woo10]. Many methods in
this literature do not use the data beyond the group/time averages, and so the formal
distinction between the grouped repeated cross-section and proper panel case becomes
moot. However, in practice the settings with grouped repeated cross-section data have
typically many fewer groups than proper panel data have units, sometimes as few as
two or three, limiting the scope for high-dimensional parametric models and raising
concerns about the applicability of large-N asymptotics.
In a seminal application of DID estimation with repeated cross-section data, with two
1 Some empirical studies continue to use the panel notation that includes two indices for the outcomes
and treatments in the repeated cross-section case, but that is confusing because Yit and Yit ′ do not refer
to the same unit i in this repeated cross-section case.

7
groups and two periods, [MVD95], the units are individuals getting injured on the job,
and we observe individuals getting injured at most once. The time periods correspond to
the year the individuals are injured, with data available for two years. Similarly, in [EL96]
the units are different taxpayers in two different years, with the number of groups again
equal to two. The case with multiple groups is studied in [BDM04], and in countless other
studies, often with the groups corresponding to states, and the treatment regulations
implemented at the state level.

3.1.3. Row and Column Exchangeable Data

One data type that has not received as much attention as panel or repeated cross-
section data corresponds to what we refer to as row-column exchangeable data (RCED)
[Ald81, Lyn84]. Like proper panel data, these data are doubly indexed, with outcomes
denoted by Yi j , i = 1, . . . , N, j = 1, . . . , J. The difference with panel data is that there
is no ordering for the second index (time in the panel case). An example of such a
data type is supermarket shopping data, where we observe expenditures on item j
for shopper i, or data from a rideshare company where we observe outcomes for trips
involving customer i and driver j . Although this is not a particularly common data
configuration, it is useful to discuss it explicitly to see that panel data differ in two
aspects from cross-section data: the double indexing and the time ordering.
In this case it is more plausible to model the units i = 1, . . . , N as exchangeable, but
also model the units j = 1, . . . , J as exchangeable, whereas with proper panel data the
exchangeability of the time periods is typically implausible. Many, but not all, methods
ostensibly developed for use with panel data are also applicable in this RCED setting, and
this is an important reason to distinguish between the two types of data. For example,
TFWE methods, factor models, and many SC estimators, all discussed in more detail
below, can be used with such data. The fact that those methods can be used in the RCED
setting directly means that such estimators do not place any value on knowledge of the
time series ordering of the data. If ex ante one believes such information is valuable,
these methods should be used with caution in proper panel data settings.
A related, but even more general data type involves RCED with repeated observations.
An example of such a data frame is a panel of matched employer-employee data [AKM99,
CRY22]. See [Bon20] for a recent survey of the relevant methods.

8
3.2. Shapes of Data Frames

One classification the panel data literature can be organized by the shape of the data
frame. This is not an exact classification, and which category a particular data set fits,
and which methods are appropriate, in part depends on the magnitude of the cross-
section and time-series correlations. Nevertheless, it is useful to reflect on the relative
magnitude of the cross-section and time-series dimensions as it has implications for
the relative merits of statistical methods for the analysis of such data. In particular it
often motivates the choice of asymptotic approximations based on large N and fixed T,
or large N and large T. See [AA03] for a discussion of the properties of various panel
data estimators under different asymptotic sequences.

3.2.1. Many Units, Few Time Periods: N >> T

Much of the traditional panel data case considers the setting where the number of
cross-section units is large relative to the number of time periods:
 
Y11 Y12 Y13
 

 Y23 Y23 Y23 

 

 Y33 Y33 Y33 

Y thin 
=  Y43 Y43 Y43 

 
(N >> T) 
 Y53 Y53 Y53 

 

 Y63 Y63 Y63 

 .. .. .. 

 . . . 

YN3 YN3 YN3

This is a common setting when the units are individuals and it is challenging or ex-
pensive to get repeated observations for many periods for the same individual. The
PSID and NLS panel data fit this setting, with often thousands of units. In this case
inferential methods often rely on asymptotic approximations based on large N for fixed
T. Incidental parameter problems of the type considered by [NS48] are relevant (see
[Lan00] for a modern discussion). Specifically, if there are unit-specific parameters, e.g.,
fixed effects, it is not possible to estimate those parameters consistently. This does not
necessarily imply that one cannot estimate the target parameters consistently, and the
traditional literature developed many procedures that allowed for the elimination of
these fixed effects, even if they enter nonlinearly, e.g., [Hon92, Cha10, Bon12]. However,

9
the fact that the time series dimension is small or modest does mean that random
effect assumptions are very powerful because they place a stochastic structure on the
individual components so that these individual components can be integrated out .

3.2.2. Few Units, Many Time Periods: N << T

The second setting is one where the number of time periods is large relative to the
number of cross-section units:
 
Y11 Y12 Y13 Y14 Y15 Y16 Y17 Y18 . . . Y1T
 
Yfat =  Y21 Y22 Y23 Y24 Y25 Y26 Y27 Y28 . . . Y2T 

 
(N << T)   Y31 Y 32 Y33 Y 34 Y 35 Y36 Y37 Y38 . . . Y3T


Y41 Y42 Y43 Y44 Y45 Y46 Y47 Y48 . . . Y4T

This setting is more common when the cross-section units are aggregates, e.g., states or
countries, for which we have observations over many time periods, say output measures
for quarters, or unemployment rates per month.
This setting is closely related to the traditional time series literature, and the insights
from that literature have not always been fully appreciated in the modern causal panel
literature. There are some exceptions that take more of a time-series approach to this
type of panel data, e.g., [BGK+ 15, BMAF+ 23]. The work on inference using conformal
methods is also in this spirit, [CWZ21].

3.2.3. Comparable Number of Units and Time Periods: N ≈ T

In the third case the number of time periods and cross-section units is roughly compa-
rable.  
Y Y12 Y13 . . . Y1T
 11 
 Y Y22 Y23 . . . Y2T 
21
Ysquare = 
 

 Y31 Y32 Y33 . . . Y3T 
(N ≈ T)  .
 
 .. .. .. .. .. 
 . . . . 

YN1 YN2 YN3 . . . YNT
A common example is that were the units are states and the time periods are years or
quarters. We may have observations on 50 states, for 30 years, or for 80 quarters. This is
a particularly challenging case and at the same time increasingly common in practice.
Many empirical studies using DID, SC, or related estimators fit into this setting.

10
 Whether in this case asymptotic approximations based on large N and fixed T, or
large N and large T, or neither, are more appropriate is not always obvious. Simply
looking at the magnitudes of the time series and cross-section dimension itself is not suf-
ficient to make that determination because the appropriate approximations depend also
on the magnitude of cross-section and time-series correlations. There is an important
lesson in this regard in the weak instrument literature. In the influential Angrist-Krueger
analysis of the returns to schooling [AK91] the authors report results based on over
300,000 units and 180 instruments. Because of the relative magnitude of the number of
units and instruments one might have expected that asymptotic approximations based
on a fixed number of instruments and an increasing number of units would be appropri-
ate. Nevertheless it turned out that the Bekker asymptotic approximation [Bek94] based
on letting the number of instruments increase proportionally to the number of units is
substantially more accurate because of the weak correlation between the instruments
and the endogenous regressor (years of education in the Angrist-Krueger study).

3.3. Assignment Mechanisms

The second classification for panel data methods we consider is based on features
of the assignment process for the treatment. As in the classification based on the
relative magnitudes of the components of the data frame, features of the assignment
process are important for determining which statistical methods and which asymptotic
approximations are reasonable. This classification is not present in the earlier panel
data literature, but features prominently in the current literature. This reflects the more
explicit focus on causal effects in general in the econometric literature of the last three
decades.

11
3.3.1. The General Case

In the most general case the treatment may vary both across units and over time, with
units switching in and out of the treatment group:
 
1 1 0 0 ... 1
 

 0 0 1 0 ... 0 
 
Wgen =  1 0 1 1 ... 0  
(1)  
(general) 
 1 0 0 1 ... 0  
 .. .. .. .. . . . .. 

 . . . . . 

1 0 1 0 ... 0

With this type of data we can use variation of the treatment within units and variation
of the treatment within time periods to identify causal effects. Especially in settings
without dynamic effects the presence of both types of variation may be important
for credibly estimating causal effects. This setting is particularly relevant for the row-
column exchangeable data configuration. It is not very common in proper panel data
settings. Some examples include marketing settings with the units corresponding to
products and the treatment corresponding to promotions or discounts.
In this setting assumptions about the absence of dynamic treatment effects are
crtical, and in settings where such assumptions are not plausible many of the commonly
used methods may lead to results that are difficult to interpret.

3.3.2. Single Treated Period

One important special case arises when a substantial number of units is treated, but
these units are only treated in the last period.
 
0 0 0 0 ... 0
 

 0 0 0 0 ... 0 

 
Wlast = 
 0 0 0 0 ... 1 

 
(last period) 
 0 0 0 0 ... 1 

 .. .. .. .. . . .. 
 . . . . . . 
 
0 0 0 0 ... 1

12
In settings where the number of time periods is relatively small, this case is often
analyzed as a cross-section problem. The lagged outcomes are simply used as exogenous
covariates or pre-treatment variables that should be adjusted for in treatment-control
comparisons based on an unconfoundedness assumption [RR83]. A classic example
in the economics literature is the Lalonde-Dehejia-Wabha data originally collected
in [LaL86] with the data set now commonly used first analyzed in [DW99]. In that
case there are three periods of outcome data (earnings), but only in the last period
are some units treated. The original study [LaL86] reported results for a variety of
models, including some two-way-fixed-effect regressions. Much of the subsequent
literature since [DW99, DW02], however, has focused more narrowly on methods relying
on unconfoundedness, sometimes in combination with functional form assumptions.
Asymptotics are typically based on large N and fixed T.
With outcomes given the treatment observed only in a single period the presence
of dynamic effects is of course not testable, and dynamics do not really matter in
the sense their presence only leads to a minor change in the interpretation of the
estimand, typically the average effect for the treated units and time period. Because of
the shortness of the panel these are obviously short-term effects, with little evidence
regarding long term impacts of the interventions.

3.3.3. Single Treated Unit

Another key setting is that with a single treated unit, treated in multiple periods.
 
0 0 0 0 ... 0
 

 0 0 0 0 ... 0 

 
Wsingle =  0 0 0 0 ... 0 
 
(single unit) 
 0 0 0 0 ... 0 
 .. .. .. .. . . .. 
 . . . . . . 
 
0 0 1 1 ... 1

This setting is prominent in the original applications of the synthetic control literature
[AG03, ADH10, Aba19]. This literature has exploded in terms of applications and theo-
retical work in the last twenty years. Here the number of time periods is typically too
small to rely credibly on large T asymptotics, creating challenges for inference that are
not entirely resolved.

13
3.3.4. Single Treated Unit and Single Treated Period

An extreme case is that with only a single unit ever treated, and this unit only treated in
a single period, typically the last period:
 
0 0 0 0 ... 0
 

 0 0 0 0 ... 0 

 
Wone =  0 0 0 0 ... 0 
 
(single unit/period) 
 0 0 0 0 ... 0 
 .. .. .. .. . . .. 
 . . . . . . 
 
0 0 0 0 ... 1

This is a challenging setting for inference: we cannot rely on large sample approxi-
mations for the average outcomes for the treated unit/periods because there is only
a single treated unit/period pair. Instead of focusing on population parameters it is
natural here to focus on the effect for the single treated/time-period pair and construct
prediction intervals. Because it is a special case of both the single treated period and the
single treated unit case it is conceptually important for comparing estimation methods
popular for those settings.

3.3.5. Block Assignment

Another important case in practice is that with block assignment, where a subset of
units is treated after a common starting date:
 
0 0 0 0 ... 0 0
 

 0 0 0 0 ... 0 0 
 
W block =  0 0 0 0 ... 0 0 
 
 
(block)  0 0 0 1 ... 1 1  
 .. .. .. .. . . .. .. 
 . . . . . . . 
 
0 0 0 1 ... 1 1

This assignment matrix is the basis of the simulations reported in [BDM04, AAH+ 21].
In this case there is often a sufficient number of treated unit/time-period pairs to allow
for reasonable approximations to be based on that number being large.
Here the presence of dynamic effects changes the interpretation of the average effect

14
for the treated. The average effect for the treated now becomes an average over short
and medium term effects during different periods. There is limited ability to separate
out heterogeneity across calendar time and dynamic effects because in any given time
period there are only treated units with an equal number of treated periods in their
past.

3.3.6. Staggered Adoption

The recent difference-in-differences literature has focused on the staggered adoption

case where units remain in the treatment group once they adopt the treatment, but they
vary in the time at which they adopt the treatment, with some early adopters, and some
later adopters:
 
0 0 0 0 ... 0 0
 

 0 0 0 0 ... 0 1 

 
Wstag = 
 0 0 0 0 ... 1 1 

 
(staggered adoption) 
 0 0 0 1 ... 1 1 

 .. .. .. .. . . .. .. 
 . . . . . . . 
 
0 1 1 1 ... 1 1

This case is also referred to as the absorbing treatment setting. Clearly this setting leads
to much richer information about the possible presence of dynamic effects, with the
ability, under some assumptions, to separate dynamic effects from heterogeneity across
calendar time.

15
3.3.7. Event Study Designs

A closely related design is the event-study design where units are exposed to the treat-
ment in at most one period.
 
0 0 0 0 ... 0 0
 

 0 0 0 0 ... 0 1  
 
 1 0 0 0 ... 0 0 
Wevent =  
... 0 0 
 
 0 0 1 0
(event study)  

 0 0 1 0 ... 0 0  
 .. .. .. .. . . . .. .. 

 . . . . . . 
0 0 0 1 ... 0 0

In this setting there are often dynamic effects of the treatment past the time of initial
treatment. If these effects are identical to the initial effect the analysis can end up being
very similar to that of staggered adoption designs. Canonical applications include some
in finance, e.g., [FFJR69]

3.3.8. Clustered Assignment

Finally, in many applications units are grouped together in clusters, with units within
the same the clusters always assigned to the same treatment. The example below has C
clusters, with a subset of the clusters assigned to the treatment from a common period

16
onwards, in a block assignment structure.
 
cluster
 

 ↓ 

 

 0 0 0 0 ... 0 0 1 

0 0 0 0 ...
 
 0 0 1 
 
 0 0 0 0 ... 0 0 1 
W cluster 
= 



 0 0 0 0 ... 0 0 2 

(cluster/block)  

 0 0 0 0 ... 0 0 2 

1 ... 1 1
 
 0 0 0 3 
 .. .. .. .. . . .. .. .. 

 . . . . . . . .


 

 0 0 0 1 ... 1 1 C 

0 0 0 1 ... 1 1 C

The clustering creates particular complications for inference, whether it is in the block
assignment case, or other settings, in particular because often there are relatively few
clusters. It also creates challenges for estimation if there are cluster components to the
outcomes.

4. Potential Outcomes, General Assumptions, and Estimands

In this section we collect in a single section the notation that allows us to cover various
parts of the literature. We focus on the panel data case with N units and T periods. We
use the potential outcome notation [Rub74, IR15]. We also discuss basic estimands that
have been the focus of this literature and some of the maintained assumptions.
Let w denote the full T-component vector of treatment assignments,

w ≡ (w1 , . . . , wT )⊤ ,

and wi the vector of treatment values for unit i. Let wt the t-component vector of
treatment assignments up to time t:

wt ≡ (w1 , . . . , wt )⊤ ,

so that wT = w, and similar for wti . In general we can index the potential outcomes for

17
unit i in period t by the full T-component vector of assignments w:

Yit (w).

Even this notation already makes a key assumption, the Stable Unit Treatment Value
Assumption, or SUTVA [Rub78, IR15]. SUTVA requires that there is no interference or
spillovers between units. This is a strong assumption and in many applications it may be
violated. There has been little attention paid to models allowing for such interference
in the recent causal panel data literature to date, although there is extensive literature
on interference in cross-section settings [HH08, AS17].
Without further restrictions, this setup describes for each unit T potential outcomes
as a function of multi-valued treatment w. As a result we can define for every period t
unit-level treatment effects for every pair of assignment vectors w and w′ :

w,w′
(2) τit ≡ Yit (w′ ) – Yit (w),

with the corresponding average effect (averaged over all units) defined as

N
w,w′ 1 X 
τt ≡ Yit (w′ ) – Yit (w) .
N
i=1

These unit-level and average causal effects are the basic building blocks of many of the
estimands considered in the literature.
w,w′
If all we are interested in is average causal effects of the form τt , we have in
essence a problem similar to the cross-sectional version of the problem of estimating
average causal effects. One approach would be to analyze such problems using standard
methods for multi-valued treatments under unconfoundedness, e.g., [Imb02]. Here this
implies we could compare outcomes in period t for units with treatment vectors w and
w′ .
w,w′
If we have completely random assignment, all τt are identified given sufficient
variation in the treatment paths. That also means we can identify in this setting dynamic
treatment effects. For example, in the two-period case

(1,1),(0,1)
τ2 ,

is the average effect in the second period for of being exposed to the sequence (1, 1)
rather than the sequence (0, 1), so that measures the dynamic effect of being exposed to

18
the treatment in period one on period 2 outcomes for units exposed to the treatment in
the second period.
A key challenge is that there are many, 2T–1 × (2T – 1) to be precise, distinct average
w,w′
effects of the form τt . Even with T = 2 this means there are six different average
causal effects, and with T larger this number quickly increases. This means that in
practice we need to limit or focus on summary measures of all these causal effects, e.g.,
averages over effects at different times. Typically we also put additional structure on
these causal effects in the form of cross-temporal restrictions on the potential outcomes
Yit (w). That enables us to give comparisons of outcomes from different time periods a
causal interpretation. See [Cha84] for a discussion of this in the case of linear models.
w,w′
Note that without additional restrictions all the average treatment effects τt are
just-identified, so any additional assumptions will typically imply testable restrictions.
The first restriction that we consider is the commonly made no-anticipation assump-
tion [AI21, CS20, SA20].

ASSUMPTION 1. (NO ANTICIPATION) The potential outcomes satisfy

Yit (w) = Yit (w′ ),

for all i, and for all combinations of t, w and w′ such that wt = w′ .

With experimental data, and sufficient variation in treatment paths, this assumption
is testable. We can compare outcomes in period t for units with the same treatment
path up to and including t, but whose treatments paths diverge in the future, that is,
after period t.
This substantive assumption is appealing in situations where units are not active
decision-makers but rather passive recipients of the treatment. In such cases, the no-
anticipation assumption can in principle be guaranteed by design. If random units
are assigned treatment each period, or, in the staggered adoption case, if the adoption
date is randomly assigned, potential outcomes cannot vary with the future random
assignment. In observational studies of course the assumption need not hold. In many
applications treatments are state-level regulations that are anticipated some time prior
to the time they formally take effect. One remedy for this problem is to impose limited
anticipation, assuming the treatment can be anticipated for a fixed number of periods,
as in [CS20]. Algorithmically, this amounts to redefining the variable w by shifting it by
the fixed number of periods.

19
 At the same time, there are numerous economic applications where units are in-
volved into active decision-making process. This process can be about w itself, e.g.,
firms deciding on the adoption of a new technology, or students deciding on exiting
schools. Alternatively, agents can decide on other variables for which w is an important
input, e.g., current and future tax changes can be important determinants of the cur-
rent consumption. In such cases, no anticipation assumption cannot be guaranteed by
design, because units can change their behavior in response to different experimental
designs [LJ76]. See [AVdB03] for a related discussion of the no anticipation assumption,
and [HN07] for a distinction between this assumption for structural and reduced-form
models in the context of staggered adoption designs.
This structure still allows us to distinguish between static treatment effects, i.e.,
(wt–1 ,0),(wt–1 ,1)
τit , which measures the response of current outcome to the current treat-
(wt–1 ,wt ),(w′ t–1 ,wt )
ment, holding the past one fixed, and dynamic ones, i.e., τit , which does
the opposite. In the earlier panel data literature, the dynamic effects were explicitly
modeled by putting a particular structure on them, but in principle, one can estimate
them without imposing additional restrictions on the potential outcomes given assump-
tions on the assignment mechanism, such as random assignment [BRS21]. There is also
a large literature in biostatistics on dynamic models that is relevant for these problems,
e.g., [RHB00].
The no anticipation assumption reduces the total number of potential treatment
effects from 2T–1 × (2T – 1) to ( Tt=1 2t–1 )( Tt=1 2t – 1). The basic building blocks, unit-
P P

period specific treatment effects, are now of the type

′
wt ,wt ′
(3) τit ≡ Yit (wt ) – Yit (wt ),

with the potential outcomes for period t indexed by treatments up to period t only.
A stronger assumption is that the potential outcomes only depend on the contempo-
raneous assignment, ruling out dynamic effects of any type.

ASSUMPTION 2. (NO DYNAMIC / CARRY-OVER EFFECTS ) The potential outcomes satisfy

Yit (w) = Yit (w′ ),

for all i and for all combinations of t, w and w′ such that wit = wit
′ .

Like the no-anticipation assumption this is not a design assumption that can be
guaranteed by randomization. It restricts the treatment effects, and thus the potential

20
outcomes for the post-treatment periods. Again like the no-anticipation assumption it
has testable restrictions given random assignment of the treatment, and given sufficient
variation in the treatment paths. Note that it does not restrict the time-path of the
potential outcomes in the absence of any treatment, Yit (0), where 0 is the vector with
all elements equal to zero, which can exhibit arbitrary correlations.
If we are willing to make the no-dynamics assumption, we can write the potential
outcomes, with some abuse of notation, as Yit (0) and Yit (1) with a scalar argument. In
this case, the total number of treatment effects for each unit is greatly reduced to T (one
per period) and we can simplify them to

(4) τit ≡ Yit (1) – Yit (0),

where τit has no superscripts because there are only two possible arguments of the
potential outcomes, w ∈ {0, 1}.
So far we have discussed assumptions on potential outcomes themselves. A concep-
tually different assumption is that of absorbing treatments, that is where the assignment
mechanism corresponds to staggered adoption.

ASSUMPTION 3. (STAGGERED ADOPTION)

Wit ≥ Wit–1 ∀t = 2, . . . , T.

Defining the adoption date Ai as the date of first treatment, Ai ≡ T + 1 – Tt=1 Wit .
P

In the staggered adoption case, we can write the potential outcomes, again with some
abuse of notation, in terms of the adoption date, Yit (a), for a = 1, . . . , T, and the realized
outcome as Yit = Yit (Ai ).
There are two broad classes of settings that are sometimes viewed as staggered
adoption designs but that are substantially different. First, there may be interventions
that are adopted and remain in place. States or other administrative units adopt new
regulations at different times. For example, states adopted speed limits or minimum
drinking ages at different times [AG04], and counties adopted enhanced 911 policies
at different times [AS02]. Second, there may be one-time interventions that have a
long-term or even permanent impact. We refer to such settings as event studies. In that
case the post intervention period effects would be dynamic effects.
Given staggered adoption, but without the no-anticipation and no-dynamics assump-

21
tions, we can write the building blocks as

a,a′
(5) τit ≡ Yit (a′ ) – Yit (a),

with the corresponding average

N 
a,a′ 1 
Yit (a′ ) – Yit (a) .
X
τt ≡
N
i=1

We also introduce notation for the average for subpopulations defined by the adoption
date:
a,a′ 1 X  
τt|a′′ ≡ ′
Yit (a ) – Yit (a) ,
Na′′ ′′
i:Ai =a
PN
where Na ≡ i=1 1Ai =a is the number of units with adoption date a. Compared to previ-
ously defined estimands, this one explicitly depends on the details of the assignment
process, which determines which units adopt the treatment and when they do so.
In the two period case where all units are exposed to the control treatment the
0,1
estimand τt|1 , the average effect of the treatment in the second period for those who
adopt in the second period is very much like the average effect for the treated, In settings
with more variation in the adoption date there are many such average effects, depending
on when the units adopted, and which period we are measuring the effect in.

5. Two-Way-Fixed-Effect Estimators and Difference-In-Differences

In this section we give a brief introduction to conventional Difference-In-Differences

(DID) estimation. The discussion is framed in terms of the potential outcomes frame-
work from the modern causal inference literature, but otherwise it is largely standard
and following textbook discussions. For other recent surveys of this literature see
[CLLX23, DCd23, RSBP23].

5.1. The Two-Way-Fixed-Effect Characterization

We start with the Two-Way-Fixed-Effect (TWFE) specification in a proper panel setting

with no anticipation and no dynamics and constant treatment effects. Traditionally this
specification often motivates the DID estimator.

22
ASSUMPTION 4. (THE TWO-WAY-FIXED-EFFECT MODEL) The control outcome Yit (0) satis-
fies

(6) Yit (0) = αi + βt + εit .

The unobserved component εit is mean-independent of the treatment assignment Wit .

ASSUMPTION 5. (CONSTANT STATIC TREATMENT EFFECTS) The potential outcomes satisfy

Yit (1) = Yit (0) + τ ∀(i, t).

The combination of these two assumptions leads to a model for the realized outcome,
Yit = Wit Yit (1) + (1 – Wit )Yit (0),

(7) Yit = αi + βt + τWit + εit .

We can estimate the parameters of this model by least squares:

N X
T
X
2
(8) (τ̂TWFE , α̂, β̂) = arg min Yit – αi – βt – τWit .
τ,α,β
i=1 t=1

Here we need to impose one restriction on the αi or βt to avoid perfect collinearity,

but this normalization does not affect the value for the estimator of the parameter of
interest, τ.
Under a block assignment structure we have Wit = 1 only for a subset of the units (the
“treatment group” with i ∈ I, where the cardinality for the set I is N tr , and N co = N – N tr ),
and for those units only for t > T0 (“post-treatment”). Defining the averages in the four
groups as P P P P
tr,post i∈I t>T0 Yit tr,pre i∈I t≤T0 Yit
Y = , Y = ,
N tr (T – T0 ) N tr T0
P P P P
co,post i∈I
/ t>T0 Yit co,pre i∈I
/ t≤T0 Yit
Y = , and Y = ,
(N co (T – T0 ) N co T0
we can write the estimator for the treatment effect as
   
tr,post tr,pre co,post co,pre
τ̂TWFE = Y –Y – Y –Y ,

in the familiar double difference form that motivated the DID terminology.
It is convenient to use the TWFE characterization based on least squares estimation

23
of the regression function in (7) because this characterization also applies in settings
where the estimator does not have the double difference form.

5.2. The Difference-In-Differences Estimator in the Grouped Repeated Cross-Section

Setting

Here we study the grouped repeated cross-section case where we observe each physical
unit only once, obviously implying that we observe different units at different points
in time. We continue to focus on the case with blocked assignment. To reflect this our
notation now only has a single index for the unit, i = 1, . . . , N. Let Gi ∈ G = {1, . . . , G}
denote the cluster or group unit i belongs to, and Ti ∈ {1, . . . , T} the time period unit i
is observed in. The set of clusters G is partitioned into two groups, GC and a treatment
group GT , with cardinality GC and GT .
Units belonging to a group Gi with Gi ∈ GC are not exposed to the treatment, irrespec-
tive of the time the units are observed. Units with Gi ∈ GT are exposed to the treatment
if and only if they are observed after the treatment date T0 , so that Wi = 1Gi ∈GT ,Ti >T0 .
Let Di = 1Gi ∈GT be an indicator that unit i is in one of the treated groups, irrespective of
whether the unit is observed in the post-treatment period.
To define the DID estimator we first average outcomes and treatments for all units
within a cluster/time period and construct Y gt and W gt . By assumption that the treat-
ment within group and time period pairs is constant, the cluster/time-period average
treatment W gt is binary if the original treatment is. The DID estimator is then the double
difference
1 X 1 X
τ̂DID = Y gt – Y gt
GT (T – T0 ) GC (T – T0 )
g∈GT ,t>T0 g∈GC ,t>T0

1 X 1 X
– Y gt + Y gt
GT T0 GC T0
g∈GT ,t≤T0 g∈GC ,t≤T0

Alternatively we can use the TWFE specification at the group level where we do have a
proper panel set up:

(9) Y gt (0) = αg + βt + εgt , Y gt (1) = Y gt (0) + τ,

similar to that in (6). The potential outcomes Y gt (0) and Y gt (1) should here be interpreted
as the average of the potential outcomes if all units in a group/time-period pair are
exposed to the control (active) treatment.

24
5.3. Inference

To conduct inference about τ̂DID or τ̂TWFE we need to be explicit about the sampling
scheme. In the proper panel setting one often assumes that the units are randomly
sampled from a large population, and thus independent across each other. In this
case, inference about τ̂TWFE reduces to joint inference about four means with i.i.d.
observations. This approach was used by [Car94] to quantify the uncertainty about the
estimated effect of the minimum wage.
With GRCS data the situation can be different, and instead of dealing with uncertainty
at the unit-level, one can allow for non-vanishing errors at the group level, as in the
model in Equation (9). This clustering approach was advocated in [LZ86, Are87, BDM04],
and is routinely used in situations where the number of groups is large. See [AAIW23]
for a recent discussion in a design setting. Note that clustering adjustments to standard
errors is infeasible in the two group case, as in [Car94].
Regardless of the level of aggregation, since [BDM04] the inference for TWFE/DID
estimators has typically taken into account correlation in outcomes over time within
units. This implies that if one estimates the average treatment effect by least squares
regression for (8), it is not appropriate to use the robust, Eicker-Huber-White standard
errors. Instead one can use clustered standard errors [LZ86, Are87], based on clustering
observations by units. The appropriate standard errors can also be approximated by
bootstrapping all observations for each unit. [Han07] discusses a more general hierar-
chical set up.

5.4. The Parallel Trend Assumption

The fundamental motivation for the DID estimator, in one form or another, is based
on a parallel trend assumption that, in one form or another, the units who are treated
would have followed, in the absence of the treatment, a path that is parallel to the path
followed by the control units, in an average sense. The substantive content and the exact
form of the assumption depend on the specific set up, the proper panel case versus the
grouped repeated cross-section case, whether one takes a model-based or design-based
perspective, the number of groups, and the averaging that is performed.
Let us first consider the proper panel case, with block assignment and Di the indi-
cator for the event that unit i will be exposed to the treatment in the post-treatment
period (after T0 ). Then assumption is that the expected difference in control outcomes
in any period for units who later are exposed to the treatment, and units who are always

25
in the control group is constant:

ASSUMPTION 6. For all t, t ′ ,

(10) E[Yit (0)|Di = 1] – E[Yit (0)|Di = 0] = E[Yit ′ (0)|Di = 1] – E[Yit ′ (0)|Di = 0].

We can also formulate this assumption in terms of changes over time. In that formu-
lation the assumption is that the expected change in control outcomes is the same for
those who will eventually be exposed to the treatment and those who will not:

E[Yit (0) – Yit ′ (0)|Di = 1] = E[Yit (0) – Yit ′ (0)|Di = 0] ∀t, t ′ .

To motivate this assumption for the panel case an alternative is to postulate a TWFE
model for the control outcomes, as in (7), with the additional assumption that the
treatment assignment Di is independent of the vector of residuals εit , t = 1, . . . , T
conditional on fixed effects:

Di ⊥
⊥ (εi1 , . . . , εiT )|αi ,

as in, for example, [Are03]. From the point of view of the causal inference literature, the
parallel trend assumption is somewhat non-standard because it combines restrictions
on the potential outcomes through the TWFE specification with restrictions on the
assignment mechanism.
Consider next the grouped repeated cross-section (GRCS) case. Suppose in the popu-
lation all groups are large (infinitely large) in each period, and we have random samples
from these populations for each period. Then the expectations are well defined as
population averages. In that case the parallel trends assumption can be formulated as
requiring that the difference in expected control outcomes between two groups remains
constant over time:

ASSUMPTION 7. For all pairs of groups g, g ′ and for all pairs of time-periods t, t ′ , the average
difference between the groups remains the same over time, irrespective of their treatment
status:
h i h i h i h i
(11) E Ygt (0) Di = 1 – E Yg ′ t (0) Di = 0 = E Ygt ′ (0) Di = 1 – E Yg ′ t ′ (0) Di = 0 .

Alternatively we could formulate this as the assumption that the expected change

26
between periods t ′ and t is the same for all groups:
h i h i h i h i
E Ygt (0) Di = 1 – E Ygt ′ (0) Di = 1 = E Yg ′ t (0) Di = 0 – E Yg ′ t ′ (0) Di = 0 ,

for all g, g ′ , t, t ′ . If we were to observe Ygt (0) for all groups and time periods, then the
presence of two groups and two time periods would be sufficient for this assumption to
have testable implications. However, with at least one of the four cells defined by group
and time period exposed to the treatment, there are not testable restrictions implied by
this assumption in the two group / two period case.
Because we can view the panel case as a two-group setting, there are only testable
restrictions from this assumption when we have more than two periods. With more than
two groups, just as in the case with more than two periods, there are testable restrictions
implied by the parallel trend assumption. In an early paper [AC85] argued against using
the TWFE model for evaluation of training programs based on the observed failure of
parallel trends. See [RSBP23, RR23, Jak21] for a discussion and bounds based on limits
on the deviations from parallel trends. Bridging some of the gap between design and
sampling based approaches [RS23, GSW22] show how parallel trends can be implied by
random assignment of treatment. They also discuss the sensitivity to transformations
of the parallel trend assumption.

5.5. Pre-treatment Variables

Often researchers observe time-invariant characteristics of the units in addition to the

time path of the outcome. Such characteristics cannot be incorporated simply by adding
them to the TWFE specification in (7) because they would be perfectly collinear with
the individual components αi [PT07]. Nevertheless, they can be important by relaxing
some of the key assumptions. Specifically one can assume that the parallel trend and
constant treatment effect assumptions hold only within subpopulations defined by
these characteristics. Two specific proposals have been made.

5.5.1. Abadie (2005)

In an early paper [Aba05] proposes flexible ways of adjusting for time-invariant covari-
ates, while continuing with the conditional parallel trends assumption. His solution
was based on re-weighting the differences in outcomes by the propensity score.

27
5.5.2. Sant’Anna and Zhao (2020)

[SZ20] use recent advances in the cross-section causal literature to adjust for time-
invariant covariates in a more robust way, by combining inverse-propensity score
weighting with outcome modeling. In cross-sectional settings, such doubly robust
methods have been found to be more attractive than either outcome modeling or
inverse-propensity-score weighting on their own. They do maintain the parallel trends
assumption conditional on covariates.

5.6. Unconfoundedness

One key distinction between the repeated cross-section and proper panel case is that
in the case with proper panel data there is a natural alternative to the TWFE model.
This is most easily illustrated in the case with blocked assignment, where the treatment
group is only exposed to the treatment in the last period. Viewing the pre-treatment
outcomes as covariates, one could assume unconfoundedness:
 
(12) Di ⊥⊥ YiT (0), YiT (1) Yi1 , . . . , YiT–1 .

If one is willing to make this assumption the larger lierature on estimation of treatment
effects under unconfoundedness applies. See [Imb04] for a survey. Modern methods
include doubly robust methods that combine modelling outcomes with propensity score
weighting. See, for example, [BR05, CCD+ 17, AIW18].
Consider the case with two periods, T = 2. Because unconfoundedness is equivalent
to assuming  
Di ⊥⊥ Yi2 (0) – Yi1 , Yi2 (1) – Yi1 Yi1 ,

it follows that the issue in the choice between DID/TWFE and unconfoundedness is
really whether one should adjust for differences between treated and control units in
the lagged outcome, Yi1 . The DID/TWFE approach implies one should not, and that
doing so would introduce biases that are otherwise absent, and the unconfoundedness
approach implies one should adjust for these differences.
The unconfoundedness assumption and TWFE model validate different non-nested
comparisons, and applied researchers often do not motivate their choices clearly in
this regard. The key difference between the two models is the underlying selection
mechanism. The TWFE model assumes that the treated units differ from the control
ones in unobserved characteristics that are potentially correlated with a persistent

28
component of the outcomes – the fixed effect αi . The unconfoundedness assumption, on
the other hand, is satisfied when the selection is based on past outcomes and potentially
other observed pre-treatment variables.
Methodological literature does not provide a lot of guidance on the choice between
these two strategies, with exceptions in [AP08, Xu23], and is somewhat segmented. For
example, there is a large literature re-analyzing the data originally studied in [LaL86]
(see also [DW99]) where the researcher has observations on two lagged outcomes. Al-
though Laonde reports estimates from various TWFE models in addition to estimates
that adjust for initial period outcomes, in the subsequent literature the focus is almost
entirely on methods assuming unconfoundedness. In contrast, most of the literature
reanalyzing the data originally studied in [CK94] where the researcher observes out-
comes for a single pre-treatment period has focused on TWFE and related methods
with relatively little attention paid to unconfoundedness approaches. It is not clear
what motivates the differences in emphasis in these two applications. In an early study
[AC85] point out the limitations of the TWFE model and in particular its inability to
capture temporary declines in earnings prior to enrollment in labor market programs,
the so-called Ashenfelter dip.
There are two important cases where the unconfoundedness and TWFE approaches
lead to similar results. Again this is most easily seen in the two period case. The results
from the two approaches are similar if the averages of the initial period outcomes are
similar for the two groups, or if the average in the control group did not change much
over time. One way to think about this case is to view it as one where there are multiple
potential control groups. One can use the contemporaneous control group, or one can
use the treatment group in the first period. If either the control group does not change
over time, or if the treatment group and the control group do not differ in the first
period, than the two potential control groups deliver the same results. See for more on
this perspective [Ros02].
When the control group changes over time, and the control group and treatment
group differ in the initial period, then the TWFE and unconfoundedness approach give
different results. However, the differences can be bounded [AP08, DL19]. Suppose the
true average effect is positive, and suppose unconfoundedness holds. Then the TWFE
estimator will over-estimate the true effect. On the other hand, if the effect is positive,
and the TWFE model holds, then assuming unconfoundedness and adjusting for the
lagged outcome will under-estimate the true effect.
[IKW21] takes a middle ground between unconfoundedness assumptions and TWFE/DID

29
methods by conditioning on lagged outcomes other than the most recent one, which is
differenced out in a TWFE approach.

6. The Staggered Adoption Case

Although much of the DID literature starts with the TWFE model for control outcomes in
(Assumption 4) with constant treatments effects (Assumption 5), the constant treatment
effect assumption is not important in the setting with block assignment. Maintaining the
TWFE for control outcomes, but allowing unrestricted heterogeneity in the treatment
effects Yit (1) – Yit (0), the TWFE/DID estimator continues to estimate a well-defined
average causal effect, namely for the average treatment effect for the treated,

N X T
1 X  
PN PT Wit Yit (1) – Yit (0) .
i=1 t=1 Wit i=1 t=1

This robustness to treatment effect heterogeneity does not extend to settings outside of
block assignment. Part of the new DID literature builds on traditional TWFE methods
in the staggered adoption setting, allowing for general heterogeneity in the treatment
effects. The twin goals are understanding what is estimated by TFWE estimators in
this setting that is common in empirical work (see [DCd23]), and proposing modifica-
tions that ensure that a meaningful average causal effect is estimated. We review that
literature here. Recall the staggered adoption setting,
 
0 0 0 0 ... 0 0 (never adopter)
 

 0 0 0 0 ... 0 1 
 (very late adopter)
 
Wstag = 
 0 0 0 0 ... 1 1 
 (late adopter)
 
(staggered adoption) 
 0 0 0 1 ... 1 1 
 (medium adopter)
 .. .. .. .. . . .. ..  ..
 . . . . . . .  .
 
0 1 1 1 ... 1 1 (early adopter)

Let Ai = arg min{t = 1, . . . , T|Wit = 1} = T + 1 – Tt=1 Wit be the adoption date (the first
P

time unit i is treated), with Ai = ∞ for units who never adopt the treatment, and let Na
be the number of units with adoption date Ai = a. Define also the average treatment

30
effect by time and adoption date,
 
τt|a = E Yit (1) – Yit (0)| Ai = a .

Key is that these average treatment effects can vary both by time and by adoption
date. Such heterogeneity was rarely allowed for in the earlier literature, with an early
exception in [Cha92] and more recently in [AB11a, GP12, CFVHN13], which allow for
heterogeneity over units, but not over time.

6.1. Decompositions of the TWFE Estimator

Here we discuss the interpretation of the TWFE estimator based on the least squares
regression
N XT
X
2
min Yit – αi – βt – τWit ,
α,β,τ
i=1 t=1
in the staggered adoption case, following the discussion in [GB21]. We maintain the
TWFE structure for the control outcomes in (6), as well as the mean-independence of
the residuals and the treatment group indicator.
Define for all time-periods t and all adoption dates a the average outcome in period
t for units with adoption date a:

1 X
Y t|a ≡ Yi,t .
Na
i:Ai =a

Then, for all pairs of time periods t, t ′ and pairs of adoption dates a, a′ such that t ′ < a ≤ t
and either a′ ≤ t ′ or t < a′ , define the building block for the TWFE estimator:

a,a′
   
(13) τ̂t,t ′ ≡ Y t|a – Y t ′ |a – Y t|a′ – Y t ′ |a′

The interpretation of this building block plays a key role. The group with adoption date
a changes treatment status between periods t ′ and t, so the difference Y t,a – Y t ′ ,a reflects
a change in treatment but this treatment effect is contaminated by the time trend in the
control outcome: h i
E Y t|a – Y t ′ |a′ = βt – βt ′ + τt|a .

For the group with an adoption date a′ , the difference Y t,a′ – Y t ′ ,a′ does not capture
a change in treatment status. If t < a′ , it is a difference in average control outcomes,

31
 a,a′
and τ̂t,t ′ is a standard DID estimand, which under the TWFE model for the control
outcomes has an interpretation as an average treatment effect. [RSBP23] refer to this as
a “clean” comparison.
However, if a′ < t ′ , the difference Y t|a′ – Y t ′ |a′ is a difference in average outcomes
under the treatment. In the presence of treatment effect heterogeneity, and in the
absence of a TWFE model for the outcomes under treatment, its expectation can be
written as h i  
E Y t|a′ – Y t ′ |a′ = βt – βt ′ + τt|a′ – τt ′ |a′ .

Hence, in the case with a′ < t ′ , the basic building block in (13) has expectation

a,a′
h i
E τ̂t,t ′ = τt|a – τt|a′ + τt ′ |a′ ,

a weighted average of treatment effects, with the weights adding up to one, but with
some of the weights negative. This is sometimes referred to as a “forbidden” comparison.
If the treatment effects are all identical, this does not in fact create a concern. However,
if there is reason to believe there is substantial heterogeneity, researchers may be
reluctant to report weighted averages with negative weights. Note that the concern with
a,a′
the comparisons τ̂t,t ′ when a′ < t ′ but not when a′ > t fundamentally treats the treated
state and the control state as different: the parallel trends assumption is assumed for
the control outcomes, but not for the treated outcomes.
The TWFE estimator τ̂TWFE can be characterized as a linear combination of the
a,a′
building blocks τ̂t,t ′ , including those where the non-changing group has an early adop-
tion date a′ < t ′ . The coefficients in that linear combination depend various aspects
of the data, including the number of units Na in each of the corresponding adoption
groups, as discussed in detail in [GB21, BLW21] for the staggered adoption case and in
[IK21] for the general case. As a result, the TWFE estimator has two distinct problems.
First, without further assumptions, the estimator does not have an interpretation as an
estimate of an average treatment effect with nonnegative weights in general. Second, the
combination of weights on the building blocks chosen by the TWFE regression depends
on the data, in particular on the distribution of units across the adoption groups. As
a result, given two identical populations in terms of potential outcome distributions
that, for some reason, have different adoption patterns, one would estimate different
quantities.
We emphasize that the expectations above are computed with respect to the errors
εit holding the adoption dates fixed. This is in line with the fixed-effects tradition in

32
the panel data literature, which does not restrict the conditional distribution of unit-
specific parameters, such as τit , given the covariates of interest, which in our case
corresponds to Ai . In some situations, e.g., in randomized experiments, the adoption
date is unrelated to the τit and thus the conditional distribution of the τit is equal to its
marginal distribution, and the negative weights issue does not necessarily arise. We can
also immediately see that this problem does not arise if τit ≡ τi , which was assumed in
[Cha92] and the follow-up literature.
We also note that in other settings, including linear regression, researchers of-
ten report estimates that in the presence of treatment effect heterogeneity represent
weighted averages of treatment effects with some of the weights negative. While that is
not necessarily ideal, there are in the current set up tradeoffs with other assumptions,
including the parallel trend assumptions, that may force the researcher to make some
assumption that are at best approximations. Similar tradeoffs motivate the use of higher
order kenels which also lead to estimators with negative weights. We therefore do not
view the negative weights of some estimators as necessarily disqualifying and find the
terminology “clean” and “forbidden” not doing justice to the potential benefits from
such methods.

6.2. Alternative DID-type Estimators for the Staggered Adoption Setting

To deal with the negative weights researchers have recently, more or less contempora-
neously, proposed a number of different modifications to the TWFE estimator. Here we
discuss four of these modifications that have attracted widespread attention. It should
be noted that all maintain the TWFE assumption for the control outcomes, and all four
avoid additional assumption on treatment effect heterogeneity.

6.2.1. Callaway and Sant’Anna (2020)

[CS20] propose two ways of dealing with the negative weights. The first compares out-
comes for a group that adopts the treatment in any period after the adoption (Y t,a for
t ≥ a) to average outcomes for the same group immediately prior to the adoption (Y a–1,a ,
and subtracts the difference in outcomes for the same two time periods for the group
that never adopts the treatment. Formally, consider, for t ≥ a, the double difference

a,∞
   
τ̂t,a–1 = Y t,a – Y a–1,a – Y t,∞ – Y a–1,∞ .

33
The particular control group, those who never adopt the treatment, may not be partic-
ularly attractive. In some cases one would be concerned that the fact that this group
never adopts the treatment is an indication that they are fundamentally different from
the other groups, and thus less suitable as a comparison for the trends in the absence
of the treatment. In addition there may be very few of these never-adopters, especially
in long panels, so that the precision of the natural estimators for such estimands may
make them unattractive.
Recognizing this concern [CS20] suggest using as an alternative control group the
average of the groups that do adopt the treatment, but restricting this to those who
adopt after period t:

T
a,>t
  1 X 
τ̂t,a–1 = Y t,a – Y a––1,a – Y t,a′ – Y a––1,a′
T–t ′
a =t+1

Given these two estimators, [CS20] suggest reporting averages using a variety of
possible weight functions ω(a, t) that depend on the adoption date and the time period.
One of their preferred weight functions is

ωe (a, t) = 1a+e=t · pr(Ai = a|Ai ≤ T – e),

which leads to an average of treatment effects e periods after adoption, for their two
control groups,

T–e T–e
a,∞ a,>t
X X
τ̂CS,I (e) = ωe (a, t) · τ̂t,a–1 , or τ̂CS,II (e) = ωe (a, t) · τ̂t,a–1 .
a=2 a=2

We should note that [CS20] also allow for the possibility that the treatment is antic-
ipated, and so that up to some known number of periods prior to the treatment the
outcome may already be affected by this. See [CS20] for details.

6.2.2. Sun and Abraham (2020)

[SA20] propose a different way of dealing with the negative weights by comparing the
change for a group that adopted in period a between the period just before the adoption
a – 1 and an arbitrary later period t, Y t,a – Y a–1,a , relative to the change for the never-

34
adopters over the same period:

a,∞
   
τ̂t,a–1 = Y t,a – Y a–1,a – Y t,∞ – Y a–1,∞ .

Given double differences of this type they suggest reporting the average of this:

T X
t
a,∞ pr(Ai = a|2 ≤ Ai ≤ t) · 12≤a≤t≤T
X
τ̂SA = τ̂t,a–1 · .
T–1
t=2 a=2

This is a simple unweighted average over the periods t after the first period, with the
weights within a period equal to the fraction of units with an adoption date prior to that,
excluding first period adopters.
An important additional issue emphasized by [SA20] is related to the validation of
the two-way model. In applications, this validation is done by testing for parallel trends
using pre-treatment data. [SA20] show that common implementation of such tests using
two-way specifications with leads of treatments also include comparisons with negative
weights. As a result, they caution against such procedures.

6.2.3. de Chaisemartin and d’Haultfœuille, 2020

[DCd20] deal with the negative weights by focusing on one-period ahead double differ-
ences, with control groups that adopt later (a > t):

t,a
   
τ̂t,t–1 = Y t,t – Y t–1,t – Y t,a – Y t–1,a .

They aggregate these by averaging over all groups that adopt later:

1 X t,a
τ̂+,t = τ̂t–1,t .
T – (a – 1)
a>t

Then they average over the time periods, weighted by the fraction of adopters in each
period:
T
X
τ̂CH = τ̂+,t · pr(Ai = a|Ai ≥ 2).
t=2

35
6.3. Borusyak, Jaravel, and Spiess, 2021

[BJS21] focus on a model for the baseline outcomes that is richer than the TWFE model:

Yit (0) = A⊤ ⊤
it λi + Xit δ + ϵit

where Ait and Xit are observed covariates, leading to a factor-type structure. This setup
reduced to the TWFE for Ait ≡ 1 and Xit ≡ ({t = 1}, . . . , {t = T}). They propose estimat-
ing λi and δ by least squares using only observations for control units only, and later
construct unit-time specific imputations for treated units:

τ̂it = Yit – A⊤ ⊤
it λ̂i + Xit δ̂.

These unit-specific estimators can then be aggregated into an estimator for the target
of interest, let us call the estimator τ̂BJS . Notably, despite each unit-time specific treat-
ment effect estimator τ̂it being inconsist, after these objects are averaged the estimator
behaves well. Moreover, [BJS21] show that the resulting estimator is efficient as long as
ϵit are i.i.d. over i and t, which relies on a version of Gauss-Markov theorem for their
setup.

6.3.1. Discussion

If one is concerned with the negative weights in the TWFE estimator in a setting with
staggered adoption, how should one choose between these four alternatives, τ̂CS,I (or
τ̂CS,II ), τ̂SA , τ̂CH , and τ̂BJS ? There are a couple of substantive arguments that matter for
this choice: (i) the never-adopter group may well be substantively different from groups
that eventually adopt, (ii) for long differences (where we compare outcomes for time
periods far apart) the assumption that differences between units are additive and stable
over time becomes increasingly less plausible, (iii) one-period differences may be quite
different from differences based on comparisons multiple periods apart if there are
dynamic effects, and (iv) efficiency considerations. These concerns do not lead to one
of the proposals clearly dominating the others, and in practice, looking for a single
estimator may be the wrong goal.
What should one to instead? One option is to report all of the proposed estimators,
as for example, [BLM22b] who reports estimates based on all four approaches in ad-
dition to the standard TWFE estimator. However, that does not do justice to the fact
that the estimators rely on different assumptions, in particular about treatment effect

36
heterogeneity, and focus on different estimands. Moroever, some of these comparisons
may have little power in terms of uncovering heterogeneity of particular forms. Instead
of reporting all estimators, one may therefore wish to explore directly the presence
a,a′
of systematic variation in the τ̂t,s , by adoption date, a, by the length of the period
between before and after, t – s, and the time since adoption, t – a.

7. Relaxing the Two-Way Fixed Effect Structure

A key strand of the recent causal panel literature starts with the introduction of the
Synthetic Control (SC) method by Alberto Abadie and coauthors, ininitially in [AG03],
with more detailed methodological discussions in [ADH10] and [ADH15]. This brought a
substantially different perspective to the questions studied in the DID/TWFE literature.
Initially the SC literature remained very separate from the earlier DID/TWFE discussions.
The SC literature focused on imputing missing potential outcomes by creating synthetic
versions of the treated units, constructed as convex combinations of control units.
This more algorithmic, as opposed to model-based, approach has inspired much new
research, ranging from factor-model approaches that motivate synthetic-control type
algorithms, to hybrid approaches that link synthetic control methods to the earlier
DID/TWFE methods and highlight their connections.
In this section we first discuss the basic synthetic control method, in Section 7.1. Next,
in Section 7.2, we discuss direct estimation of factor models. In Section 7.3 we discuss
some hybrid methods that combine synthetic control and DID/TWFE components.

7.1. Synthetic Control Methods

In the original paper, [AG03], Abadie and Gardeazabal were interested in estimating
the causal effect on terrorism on the Basque region economy. They constructed a
comparison for the Basque region based on a convex combination of other regions in
Spain, the synthetic Basque region, with the weights chosen to ensure that this synthetic
Basque region matched the actual Basque region closely in the years pre-treatment
(prior to the terrorism) years.
In a short period of time this synthetic control method has become a very widely
used approach, popular in empirical work in social sciences, as well as in the popular
press (including The Economist and The Guardian), with many theoretical advances in
econometrics, statistics, and computer science. The key papers by Abadie, Diamond

37
and Hainmueller that discuss the details of the original synthetic control proposals are
[ADH10, ADH15]. For a recent review see [Aba19, SSP+ 19].

7.1.1. Estimation

Here we use a slightly different characterization of the SC method as a least squares

estimator, as discussed in [DI16], focusing on the case without covariates. Suppose unit
N is the sole treated unit, and is treated in period T only. Define the weights ω̂ as
 2
T–1
X N–1
X
(14) ω̂ ≡ arg min
P YNt – ω j Y j t ,
ω|ω≥0, j ω j =1
t=1 j =1

and then impute the missing potential outcome as

N–1
X
ŶNT (0) = ω̂ j Y j T .
j =1

The nonnegative weights ω j define the “synthetic” control that gave the methods its
name. One remarkable finding in the initial papers by Abadie and coauthors is that this
solution is typically sparse, with positive weights ω j > 0 for only a small subset of the
control units. Although this is not always important substantively, it greatly facilitates
the interpretation of the results. For example, in the German re-unification application
in [ADH15] where the full set of potential controls consists of sixteen OECD countries,
only five countries, Austria, Japan, The Netherlands, Switzerland and the US, have
positive weights.
The characterization of the SC estimator in (14) allows for an interesting comparison
with methods based on unconfoundedness assumption discussed in Section 5.6. With a
linear model specification unconfoundedness would suggest an estimator

N–1 T–1
!2
X X
(15) β̂ = arg min YiT – β0 – βs Yis ,
β
i=1 s=1

and then impute the missing potential outcome as

T–1
X
ŶNT (0) = β̂0 + β̂s YNs .
s=1

38
The difference is that in, using the terminology of [ABD+ 21b], SC in the regression
in (14) relies on vertical regression with T – 1 observations and N – 1 predictors, with
some restrictions on the parameters, and (15) relies on horizontal regression with N – 1
observations and T regressors (by including an intercept). If the minimizers in these least
squares regressions are not unique, we take the solution to be the one that minimizes
the L2 norm [SIV23]. See [SDSY22] for more insights into the comparison between the
horizontal and vertical regressions in this setting.
One interesting aspect of the synthetic control approach is that it is more algorithmic
that many other methods used in these settings. Consider the estimator based on
unconfoundedness in (15). Such an approach is typically motivated by a linear model

T–1
X
YiT = γ0 + γs Yis + εi ,
s=1

with assumptions on the εi given the lagged outcomes. The corresponding model for
the SC estimator would be
N–1
X
YNt = ω j Y j t + ηt ,
j =1

with assumptions on ηt given the contemparenous outcomes for other units. However,
such a model is never postulated, and for good reason. It would postulate a relationship
between the cross-section units, e.g., states, that is oddly asymmetric. If, as in the
application in [ADH10], California is the treated state, this model would postulate a
relationship between California and the other states of a form that cannot also hold
for the other states. Attempts to specify models that motivate the synthetic control
estimator have met with limited success. [ADH10] discusses factor models as data
generating processses, but that begs the question why one would not directly estimate
the factor model. Researchers have done so, as discussed in Secction 7.2 below, but such
attempts have not always outperformed synthetic control methods. See the review in
[Aba19] for more discussion on conditions under which synthetic control methods are
appropriate.

7.1.2. Modifications

A number of modifications have been suggested to the basic version of the SC estimator.
[DI16] and [FP21] suggest making the estimator more flexible by allowing for an intercept
in the regression (or, equivalently, applying the method to outcomes in deviations from

39
time-averages).
[DI16] and [AL21] are concerned with settings where the number of potential control
units, N – 1 is large relative to the number of time periods that is used to estimate the
weights. [DI16] suggest regularizing the weights by imposing an elastic net penalty on
the weights ωi , with the penalty chosen by cross-validation. [SIV23] avoid the choice
of a penalty term by choosing the minimum L2 norm value for the weights within the
set of weight combinations that lead to the optimal within sample fit, in the spirit of
the recent double descent literature [BHMM19]. [AL21] recognizes that weights that a
convex combination of control units that are all far away from the treated unit is not
as attractive as a convex combination of control units that are all close to the target
treated unit. They suggest choosing the weights by minimizing the sum of the original
synthetic control criterium and a term that penalizes the distance between any of the
control units and the target unit
 2
T–1
X N–1
X N–1
X T–1
X
ω̂ = arg min
P YNt – ω j Y j t + λ ωj (YNt – Y j t )2 ,
ω|ω≥0, j ω j =1
t=1 j =1 j =1 t=1

with the tuning parameter λ chosen through cross-validation, for example on the control
units.
Typically in synthetic control methods only the control units are weighted. In prin-
ciple, however, one could also weight the treated units to make it easier to find a set
of (weighted) control units that are similar to these weighted treated units during the
pre-treatment period, as suggested in [KZEM21].
[KMPT21] suggest combining matching and synthetic control methods. Whereas
synthetic control methods avoid extrapolation at any cost, combining it with matching
allows researchers to lower the bias from either method.

7.1.3. Inference

Inference has been a major challenge in synthetic control settings, and there is as of
yet not consensus regarding the best way to estimate variances or construct confidence
intervals. One particular challenge is that the methods are often used in settings with
just a single treated unit/period, or relatively few treated unit/period pairs, making it
difficult to rely on central limit theorems for the distribution of estimators.
One approach has been to use placebo methods to test sharp null hypotheses, typi-
cally for the null hypothesis of no effect of the intervention. [ADH10] proposes such a

40
method. Suppose there is a single treated unit, say unit N. [ADH10] construct a distri-
bution of estimates based on each unit in term being analyzed as the treated unit, and
then calculating the p-value for unit N as the quantile in that distribution of placebo
estimates.
[DI16] suggests that the same placebo approach can be based on changing the time
period that was treated. Essentially here the idea is to think of the time of the treat-
ment as random, generating a randomization distribution of estimates. Closely related
[CWZ21] builds on [LC20] to develop conformal inference procedures that rely on ex-
changeability of the residuals from some model over time. [CFT21] does not rely on
large samples of treated units by proposing the construction of prediction intervals for
the counteractual outcome.

7.2. Matrix Completion Methods and Factor Models

A second set of methods that relaxes the TWFE assumptions focuses directly on factor
models, where the outcome is assumed to have the form

R
X
Yit (0) = αir βtr + εit .
r=1

If we fix the rank at R = 2, and set αi2 = 1 for all i and βt1 = 1 for all t, this is the TWFE
specification, but it obviously allows for more general dependence structures in the
data. Although these factor models have a long tradition in panel data, e.g., [And84,
CR83, SW98, BN02, Bai09], the recent causal literature has used them in different ways.

7.2.1. Matrix Completion with Nuclear Norm Regularization

[ABD+ 21b] take an approach that models the entire matrix of potential control outcomes
as
Yit (0) = Lit + αi + βt + εit ,

where the εit is random noise, uncorrelated with the other components. The matrix L
with typical element Lit is a low-rank matrix. As mentioned above the unit and time
components αi and βt could be subsumed in the low-rank component as they on
their own form a rank-two matrix, but in practice it improves the performance of the
estimator to keep these fixed effect components in the specification separately from
the low-rank component L. The reason is that we regularize the low rank component L,

41
but not the individual and time components. The parameters L, α and β are estimated
by minimizing
N XT
X
2
(1 – Wit ) Yit – Lit – αi – βt + λ∥L∥∗ .
i=1 t=1
Here the nuclear norm ∥L∥∗ is the sum of the singular values σl (L) of the matrix L,
based on the singular value decomposition L = SΣR, where S is N × N, Σ is the N × T
diagonal matrix with the singular values and R is T × T. The penalty parameter λ is
chosen through out-of-sample crossvalidation. The nuclear norm regularization shrinks
towards a low rank estimator for L, similar to the way LASSO shrinks towards a sparse
solution in linear regression.

7.2.2. Robust Synthetic Control

[ASS18] focusing on the case with a single treated unit. They start with a factor model
Y = L + ε. They would like to use a synthetic control estimator with denoised matrix L as
the control outcomes, rather than the actual outcomes Y. They implement this through
a two step procedure. In the first step the matrix L is estimated by taking the singular
value decomposition, and setting all singular values below a threshold µ equal to zero.
This leads to a low-rank estimate L̂, which is then scaled by one over p, where p is the
maximum of the fraction of observed outcomes and 1/((N – 1)T).
In the second step [ASS18] use the part of this rescaled matrix corresponding to the
control units, in combination with the pre-treatment-period values of for treated unit,
in a standard synthetic control approach. The idea is that using de-noised outcomes L̂
instead of the actual outcomes Y leads to better predictors by removing an estimate of
the noise component ε. In this second synthetic control step [ASS18] do not impose the
convexity restrictions on the weights, but do add a regularization penalty.

7.2.3. Interactive Fixed Effect or Factor Models

Building on the factor model literature in econometrics [CR83, Pes06, Bai09, MW15,
MW18, Fre18] researchers have studied direct estimation of factor models as an alterna-
tive to synthetic control methods [Xu17]. The basic set up models the control potential
outcome as
R
X
Yit (0) = αir βtr + εit ,
r=1

42
where the εit are independent of the factors and loadings. Based on this model on can
impute the missing potential outcomes for the treated unit/time-period pairs and use
that to estimate the average effect for the treated. See [GM16] for an application.

7.2.4. Grouped Panel Data

[BM15, BLM22a] consider a factor model but impose a group structure. In our causal
setting, their set up would correspond to

Yit (0) = θGi ,t + εit ,

with the group membership unknown. They focus on the case with the number of
groups G known. In that case one can write the model as a factor model with G factors
λrt and the loadings equal to indicators, αir = 1Gi =r , so that

G
X
Yit (0) = θGi ,t + εit = αir λrt + εit .
r=1

Computationally this grouped structure creates substantial challenges. [Mug22, CM22]

suggest alternative estimation methods that are computationally more attractive.

7.2.5. Tuning

One disadvantage the methods discussed in this section share is the need to specify the
tuning parameters. This sets them apart from the conventional two-way methods we
discussed before and makes them harder to adopt in practice. In the case of the matrix
completion, this tuning parameter is the regularization parameter λ that quantifies
the importance of the nuclear norm penalty. In the context of the standard interactive
fixed effects estimators, one needs to specify the rank of the underlying factor model.
The same applies to the estimator based on finite groups. In principle, one can use
traditional techniques from the machine learning literature, such as cross-validation,
to find appropriate values of these parameters. The panel dimension, however, creates
an additional challenge on how exactly to implement the cross-validation. It is thus
attractive to have methods that generalize the two-way methodology and do not require
explicit tuning. One such proposal is [MW18], where the authors analyze the limiting
version of the estimator from [ABD+ 21b] with λ approaching zero. They show that
the resulting estimator is consistent under relatively weak assumptions, albeit it can

43
converge at a slower rate.

7.3. Hybrid Methods

Two recent methods combine some of the benefits from the synthetic control approach
with either TWFE/DID ideas or with unconfoundedness methods.

7.3.1. Synthetic Difference In Differences

For expositional reasons let us consider the case with a single treated unit and time
period, say unit N in period T, although the insights readily extend to the block assign-
ment case. Once one has calculated the SC weights, the SC estimator for the treatment
effect can be characterized as a weighted least squares regression,

N X
T
X
2
(16) min ωi Yit – βt – τWit .
β,τ
i=1 t=1

It is useful to contrast this with the TWFE estimator, which is based on a slightly different
least squares regression:

N X
T
X
2
(17) min Yit – αi – βt – τWit .
β,α,τ
i=1 t=1

The two differences are that (i), the SC regression in (16) uses weights ωi , and (ii) the
TWFE regression in (17) has unit-specific fixed effects αi .
In the light of of this comparison the omission of the unit fixed effects from the
synthetic control regression may seem surprising. [AAH+ 21] exploit this by proposing
what they call the Synthetic Difference In Difference (SDID) estimator that includes
both the unit fixed effects αi and the SC weights ωi , as well as analogous time weights
λt , leading to
N X
T
X
2
min ωi λt Yit – αi – βt – τWit .
β,α,τ
i=1 t=1
The time weights λt are calculated in a way similar to the unit weights,

N–1 T–1
!2
X X
min YiT – λs Yis ,
λ
i=1 s=1

44
 PT–1
subjet to the restriction that λs ≥ 0, and s=1 λs = 1.

7.3.2. Augmented Synthetic Control

[BMFR21] augment the SC estimator using a linear regression of the treated period
outcomes on the lagged outcomes for the control units. Following [BMFR21] we use
ridge regression for this first step, again in the setting with unit N and period T the only
treated unit/time-period pair:

N–1 T–1
!2 T–1
X X X
η̂ = arg min YiT – η0 – ηs Yis +λ η2s ,
η
i=1 s=1 s=1

with ridge parameter λ chosen through cross-validation. A standard unconfoundedness

approach would predict the potential control outcome for the treated unit/time period
pair as
T–1
X
ŶNT = η̂0 + η̂s YNs .
s=1
The augmented SC estimator modifies this by combining it with SC weights in a way
that can be seen either as a bias-adjustment to the unconfoundedness estimator, or a
bias-adjustment to the SC estimator:

T–1 N–1 T–1

!
X X X
ŶNT = η̂0 + η̂s YNs + ωi YiT – η̂0 – η̂s Yis
s=1 i=1 s=1
 
N–1
X T–1
X N–1
X
= ωi YiT + η̂s YNs – ω j Y j s .
i=1 s=1 j =1

[BMFR22] extend this approach to the case with staggered adoption.

7.3.3. The Connection Between Unconfoundedness, Difference-In-Differences, Syn-

thetic Control and Matrix Completion

Although methods based on unconfoundedness, and synthetic control estimators,

difference-in-differences, and matrix completion estimators appear to be quite dif-
ferent, they are in fact closely related. We want to highlight two insights regarding these
connections.

45
 We focus on the case with a single treated unit/period pair, say unit N in period T,
The observed control outcomes are Y, an N × T matrix with the (N, T) entry missing.
We partition this matrix as
!
Y 0 y1
Y= ,
y⊤
2 ?
where Y0 is a (N – 1) × (T – 1) matrix, and y1 and y2 are (N – 1) and (T – 1) component
vectors, respectively.
First, [SDSY22] discuss an interesting connection between SC estimators and esti-
mators based on unconfoundedness in combination with linearity. In that case we first
estimate a linear regression

T–1
X
YiT = γ0 + γs Yis + εi ,
s=1

and then impute the missing outcome as ŶNT = γ̂0 + T–1

P
s=1 γ̂s YNs . [SDSY22] show that
if we drop the intercept from this regression, γ0 = 0, and we impose the restrictions
that the γs are nonnegative and sum to one, then the unconfoundedness imputation
is identical to the SC imputation (where we also restrict the weights to be nonnegative
and sum to one).
In an alternative connection, [ABD+ 21b] show that in some cases linear versions
of all four estimators can all be characterized as solutions to the same optimization
problem, subject to different restrictions on parameters of that optimization problem.
Now define for a given positive integer R, an N × R matrix U, an T × R matrix V, an
N-vector α and a T-vector β, and a scalar λ the objective function
!2  
N X
X T R
X N X
X R T X
X R
(18) (1 – Wit ) Yit – Uir Vtr – αi – βt + λ 2 +
Uir 2.
Vtr
i=1 t=1 r=1 i=1 r=1 t=1 r=1

When R = 0, we take the product UV⊤ to be the N × T matrix with all elements equal to
zero. Given U, V, α and β the imputed value for YNT is ŶNT = R
P
r=1 UNr VTr – αi – βt .
First note that minimizing the objective function (18) over the rank R, the matrices
U, V and the vectors α and β given λ = 0, does not lead to a unique solution. By choosing
the rank R to the minimum of N and T, we can find for any pair α and β a solution
for U and V such that (1 – Wit )(Yit – R
P
r=1 Air Btr – γi – δt ) = 0 for all (i, t), with different
imputed values for YNT . The implication is that we need to add some structure to the

46
optimization problem. The next result shows that unconfoundedness regression, the SC
estimator, the DID estimator, and the MC estimator can all be expressed as minimizing
the objective function under different restrictions on, or with different approaches to
regularization of, (R, U, V, α, β).
NUCLEAR NORM MATRIX COMPLETETION The nuclear norm matrix completion estima-
tor chooses λ through cross-validation.
UNCONFOUNDEDNESS The unconfoundedness regression is based on regressing Y1 on
y⊤
2 and an intercept. It can also be characterized as the solution to minimizing (18) with
the restrictions
!
Y0
R = T – 1, U= , α = 0, β1 = β2 = . . . = βT–1 = 0, λ = 0.
y⊤
2

SYNTHETIC CONTROL The SC estimator imposes the restrictions subject to

N–1
!
Y⊤
0
X
R = N – 1, V= , α = 0, β = 0, ∀ i, UiT ≥ 0, AiT = 1, λ = 0.
y⊤
1 i=1

DIFFERENCE IN DIFFERENCES The DID estimator fixes R = 0.

7.3.4. The Selection Mechanism

Another set of insights concerning the differences between the various estimators
emerges from a focus on the selection mechanism. See [AH23, IV23] for two recent
discussions. We can see this in the setting with the block design, where there are T0 + 1
periods, and some units are treated in the last period, with Di being the treatment
indicator. Suppose the underlying potential outcomes follow a static two-way model of
Section 5:
Yit (0) = αi + βt + εit , εit ⊥
⊥ αi , τ = Yit (1) – Yit (0).

The key feature that determines the performance of different algorithms in this en-
vironment is the relationship between Di and the vector of errors (εi1 , . . . , εiT0 +1 ). As
long as Di is mean-independent from (εi1 , . . . , εiT0 +1 ), then the discussed estimators will
have good statistical properties as long as T0 goes to infinity fast enough. This should
not be surprising for the DID (which does not rely on large T0 ) or matrix completion
estimator because their statistical properties are established under this assumption.
The fact that the SC estimator would work well in this situation follows from the results

47
in [AAH+ 21, AH23].
This conclusion changes dramatically if we allow (εi1 , . . . , εiT0 +1 ) to be correlated
with Di . If this correlation is completely unrestricted, then any observed differences
in outcomes in the two groups can be attributed to differences in errors, and it is
impossible to identify the effect using any method. Suppose, however, that we make a
natural selection assumption

Di ⊥⊥ YiT0 +1 (0)|αi , Yi1 , . . . , YiT0 +1 (0),

which restricts the correlation of Di with εi,T0 +1 . Note that this restriction combines
both the selection on fixed effect assumption discussed in Section 5.4 and the uncon-
foundedness assumption discussed in Section 5.6.
As long as εit are autocorrelated, the DID estimator is inconsistent, even when T0
goes to infinity. The reason for this failure is that εiT0 +1 – T1
P
0 t≤T0 εit remains correlated
with Di which introduces bias. The performance of the SC estimator is different, and the
results in [AH23] show that the SC estimator is consistent and asymptotically unbiased
as long as T0 goes to infinity fast enough. The consistency properties of the matrix
completion estimator and the unconfoundedness regression are not established for
this setting.
[IV23] focuses on a factor model with block assignment where Di can be correlated
with the factor loadings and the time of initial exposure can be correlated with the
factors. They present conditions under which the SC estimator is consistent.
This discussion illustrates that to analyze the behavior of algorithmically related es-
timators, one needs to take a stand on the underlying selection mechanism. Most of the
recent results in the causal panel data literature are established under strict exogeneity,
which does not allow Di to be correlated with εit . Understanding the performance of
different estimators in environments where such correlation is present is an attractive
area of future research that can benefit from the econometric panel data literature.

8. Nonlinear Models

In this section we discuss some nonlinear panel data models. By nonlinear models, we
mean here models that do not specify a model for the conditional mean that is linear in
parameters. Part of this literature is motivated by the concern that the standard fixed
effect models maintain additivity and linearity in a way that does not do justice to the
type of data that are often analyzed. With binary outcomes, it is difficult to justify the

48
standard TWFE model. At the same time, estimating the unit and time fixed effects
inside a logistic or probit model does not lead to consistent estimators for the effects of
interest in typical settings.

8.1. Changes-In-Changes

[AI06] focus on the repeated cross-section case with two periods and two groups, one
treated in the second period, and one always in the control group. They are concerned
with the functional-form dependence of the standard TWFE specification in levels. If
the model
Yi (0) = µ + α1Ci =1 + β1Ti =1 + εi ,

holds in levels, then it cannot hold in logarithms. In fact, in some cases on can test
that the model cannot hold in levels. Suppose the outcome is binary, and suppose that
the potential control outcome averages by group and time period are Y 11 (0) = 0.2 (for
the first period control group), Y 12 = 0.8 (for the second period control group), and
Y 21 (0) = 0.7 (for the first period treatment group). Then the additive TWFE model implies
that the second period treatment group, in the absence of the treatment, would have had
average outcome 0.7+(0.8-0.2)=1.3, which of course is not feasible with binary outcomes.
To address this concern [AI06] propose a scale-free changes-in-changes (CIC) model
for the potential control outcomes,

Yi = g(Ui , Ti ),

where the Ui is an unobserved component that has a different distribution in the treat-
ment group and the control group, but a distribution that does not change over time.
The standard TWFE model can be viewed as the special case where g(u, t) is additively
separable in u and t:
g(u, t) = β0 + u + β1 t,

implying that the expected control outcomes can be written in the TWFE form as

E[Yi (0)|Ti = t, Gi = g) = β0 + β1 t + αE[Ui |Gi = g] – E[Ui |Gi = 1].

[AI06] show that the if Ui is a scalar, and g(u, t) is strictly monotone in u, one can infer
the second period distribution of the control potential outcome in the treatment group

49
as   
–1
FYi (0)|Ti =2,Gi =2 ( y) = FYi (0)|Ti =1,Gi =2 FY (0)|T =1,G =1 FYi (0)|Ti =2,Gi =1 ( y) .
i i i

This in turn can be used to estimate the average effect of the intervention on the second
period outcomes for the treatment group.
The expression for the counterfactual distribution of the control outcome for the
second-period treatment group has an analogue in the literature on wage decompo-
sitions, see [AB99]. [Ark19] discusses a similar approach to the changes-in-changes
estimator in [AI06], where the role of the groups and time periods are reversed and also
considers an extension for multiple outcomes. [Woo22] also studies nonlinear versions
of DID and TWFE approaches. In the two-period-two-group setting this starting point
assumes there is a known function g : R 7→ R such that E[Yit (0)|Di ] = g(µ + αD + γt ), so
that there is a parallel trend inside the known transformation g(·). The transformation
g(·) could be the exponential function, g(a) = exp(a) in case of non-negative outcomes,
or the logistic function g(a) = exp(a)/(1 + exp(a)) in case of binary outcomes.

8.2. Distributional Synthetic Controls

[Gun23] develops a model that has similarities to both the CIC and SC control approaches.
He focuses on a setting with repeated cross-sections, where we have a relatively large
number of units observed in a modest number of groups, with a modest number of
time periods. As in the canonical synthetic control case there is a single treated group.
Whereas the synthetic control method chooses weights on the control units so that the
weighted controls match the treated outcomes in the pre-treatment periods, the [Gun23]
approach chooses weights on the control groups so that the marginal distribution for
the weighted controls matches that for the treated group. The metric is based on the
quantile function FY–1gt (v), for group g and period t. First weights ω̂tg are calculated
separately for each pre-treatment period t based on the following objective:
 2
M G–1
1 X X
ω̂tg = arg min  ωgt F̂Y–1gt (Vm ) – F̂Y–1Gt (Vm ) ,
ω:ω≥0,
PG–1
ωgt =1 M
g=1 m=1 g=1

where the quantile functions are evaluated at M randomly choosing values v1 , . . . , vM .

50
 In the next step the weights are averaged over time,

T–1
1 X
ω̂g = ω̂gt .
T–1
t=1

Finally the quantile function for the treated group in the absence of the treatment is
estimated as the synthetic control average of the control quantile functions:

G–1
X
–1 (v) =
F̂GT –1 (v).
ω̂g F̂gT
g=1

Note that in the case with G = 2, so there is just a single control group, the quantile
function for the treated group in the absence of the treatment is identical to the quantile
function for the control group, and the pre-treatment distributions are immaterial.

8.3. Balancing Statistics to Control for Unit Differences

[AI22] focus on settings where the treatment can switch on and off, as in the assignment
matrix in Equation (1), unlike the staggered adoption case where the treatment can only
switch on. They also assume there are no dynamic effects. Their focus is on flexibly
adjusting for differences between beyond additive effects. Allowing for completely
unrestricted differences between units would require relying solely on within-unit com-
parisons. Often the number of time periods is not sufficient to rely on such comparisons
and still obtain precise estimates. [AI22] balance these two concerns, the restrictiveness
of the TWFE model and the lack of precision when focusing purely on within-unit com-
parisons, by making assumptions that allow the between-unit differences to be captured
by a low-dimensional vector, which then can be adjusted for in a flexible, nonlinear
way using some of the insights from the cross-section causal inference literature.
To see the insights most clearly it is useful to start with a simpler setting. Specifically,
let us first consider a clustered sampling setting with cross-section data studied in
[AI18a]. In that case a common approach is based on a fixed effect specification

Yi = αCi + τWi + βXi + εi ,

where Ci is the cluster indicator for unit i. Estimating τ by least squares is the same as

51
estimating the following regression function by least squares,

Yi = µ + τWi + γW Ci + βXi + δX Ci + ηi ,

W c is the cluster average of the treatment for cluster c, and similar for X c . This equiva-
lence has been known since [Mun78].
[AI18a] build on the Mundlak insight, still in the clustered setting, by making the
unconfoundedness assumption that
 
Wi ⊥⊥ Yi (0), Yi (1) Xi , X Ci , W Ci .

Implicitly this uses the two averages X Ci and W Ci as proxies for the differences between
the clusters. This idea is directly related to [AM05], who also use exchangeability to
control for unobserved heterogeneity. Given this uconfoundedness assumption, one can
then adjust for differences in (Xi , X Ci , W Ci ) in a flexible way, through non-parametric
adjustment methods, possibly in combination with inverse propensity score weighting.
[AI18a] then generalize this by assuming that
 
Wi ⊥⊥ Yi (0), Yi (1) Xi , SCi ,

where the sufficient statistic Sc captures the relevant features of the cluster, possibly
including distributional features such as the average of Wi in the cluster, but also other
averages such as the average of the product of Xi and Wi in the cluster.
[AI22] extend these ideas from the clustered cross-sction case to the panel data case.
They focus on the no-dynamics case where the potential outcomes are indexed only
by the binary contemporaneous treatment. In panel data settings, an alternative to
two-way fixed effect regressions is the least squares regression

Yit = τẄit + εit ,

where Ẅit is the double difference

Ẅit = Wit – W i· – W ·t + W ,

with
T N N T
1X 1 X 1 XX
W i· = Wit , W ·t = Wit , and W = Wit .
T N NT
t=1 i=1 i=1 t=1

52
See, for example, [Vog12]. [Woo21] shows the same estimator can be obtained by through
what he calls the Mundlak regression

Yit = τWit + γW i· + δW ·t + εit .

[AI22] postulate the existence of a known function Si (Wi1 , . . . , WiT ) that captures all
the relevant components of the assignment vector W i = (Wi1 , . . . , WiT ) (and possibly
other covariates, time-varying or time-invariant). Given this balancing statistic, they
assume that the potential outcomes are independent of the treatment assignment vector
given this balancing statistic:

(19) Wi ⊥
⊥ Yit (w) Si .

Consider the case where the balancing statistic is fraction treated periods, Si = W i . The
unconfoundedness assumption in (19) implies that one can compare treated and control
units in the same period, as long as they have the same fraction of treated periods over
the entire sample. More generally Si could capture both the fraction of treated periods,
as well as the number of transitions between treatment and control groups.
The estimator proposed by [AI22] has a built-in robustness property: it remains
consistent if the two-way model is correctly specified, or the unconfoundedness given
Si holds. As a result, it does not require researchers to commit to a single identification
strategy.

8.4. Negative Controls, Proxies, and Deconvolution

The results in [AI22] show how to use panel data to construct a variable that elimi-
nates the unobserved confounding. A related but different strategy is to use a panel to
construct a set of proxy measures for the unobservables. If these proxy measures do
not directly affect either outcomes or treatments, then this restriction can be used for
identification. In biostatistics, such proxy variables are called negative control variables.
To emphasize the connections between this literature and economic applications, we
use these two terms interchangeably. In biostatistics a recent literature focuses on non-
parametric identification results for average treatment effects that are based on negative
controls [SRC+ 16, SMNT20]. See [YMST21] for an introductory article. This literature
is tightly connected to econometric literature on non-parametric identification with
measurement error [HS08] and the changes-in-changes model [AI06]. In a DID setting

53
one can view the pre-treatment outcomes as proxies or negative controls in the sense
of this literature. Recently, these arguments have been extended to prove identification
results for a class of panel data models [Dea21b].
Proxy variables have a long history in economics. In early applications [Gri77, Cha77]
use data on several test scores to estimate returns to schooling accounting for unob-
served ability (see also [Dea21a]). Using modern terminology, these test scores serve as
negative controls. Versions of these strategies have also been successfully used in the
traditional panel data literature. For example, [HENR88] use data on past outcomes to
estimate a dynamic linear panel data model with interactive fixed effects with a finite
number of periods. They achieve this by eliminating the interactive fixed effects via
a quasi-differencing scheme, which is called a bridge function in the negative control
literature (see [IKM21] for a discussion).
A similar idea is used in [FHS19], where the authors consider a setting with an
unobserved confounder that can vary arbitrarily over i and t. To eliminate this con-
founder, the authors assume a presence of a proxy variable that is affected by the same
confounder but is not related to the treatment. As a result, one can eliminate the un-
observables by subtracting a scaled proxy variable from the outcome of interest. The
appropriate scaling is estimated using the pre-treatment data. In essence, this strategy
is analogous to quasi-differencing and is another example of using bridge functions.
An important aspect of the negative control literature, which it shares with most of
the methods discussed in this survey, is that it aims to isolate and eliminate the unob-
served confounders rather than identify causal effects conditional on unobservables.
Alternatively, one can obtain identification under different distributional assumptions
that connect the unobservables to outcomes and treatments using general deconvo-
lution techniques. This approach has been successfully employed to answer causal
questions in linear panel data models [BS11, AB11a] and nonlinear quantile panel data
models [AB16], but so far has not been widely adopted by a broader causal community.

8.5. Combining Experimental and Observational Data

Another direction this literature has explored is the combination of experimental and
observational data. [ACI20] study the case with an experimental data set that has obser-
vations on short-term outcomes, and an observational sample that has information on
the short-term outcome and the primary outcome. A key assumption is that the obser-
vational sample has an unobserved confounder that leads to biases in the comparison
of the short-term outcome by treatment group. The experimental data allows one to

54
remove the bias and isolate the unobserved confounder, which then can be used to
eliminate biases in the primary outcome comparisons essentially as a proxy variable as
discussed in the previous section. See also [GYR+ 22, IKM21, KM20].

9. Design-based Approaches to Estimation and Inference

An issue that features prominently in the recent panel data literature, but is largely
absent in the earlier one, is the interpretation of the uncertainty in the estimates. In
most empirical analyses in economics, and in most of the methodological literature in
econometrics, uncertainty is assumed to be arising from sampling variation. This is a
natural perspective if, say, we have data on individuals that can be at least approximately
viewed as a random sample from a well-defined population. Had we sampled a different
set of individuals our estimates would have been different, and the standard errors
reflect the variation that would be see if we repeatedly obtained different random
samples from that population. This sampling-based perspective is still a natural one in
panel data settings when the units can be viewed as a sample from a larger population,
e.g., individuals in the the Panel Study of Income Dynamics or the National Longitudinal
Survey of Youth.
The sampling-based perspective is less natural in cases where the sample is the
same as the population of interest. This is quite common in panel data settings, for
example when we analyze state-level data from the United States, or country level data
from regions of the world, or all firms in a particular class. It is not clear why viewing
such a sample as a random sample from a population is appropriate. Researchers have
struggled with interpreting the uncertainty of their estimates in that case. Manski and
Pepper write in their analysis of the impact of gun regulations with data from the fifty
US states: “measurement of statistical precision requires specification of a sampling
process that generates the data. Yet we are unsure what type of sampling process would
be reasonable to assume in this application. One would have to view the existing United
States as the sampling realization of a random process defined on a superpopulation of
alternative nations.” ([MP18], p. 234).
An alternative approach to formalizing uncertainty focuses on the random assign-
ment of causes, taking the potential outcomes as fixed. This approach has a long history
in the analysis of randomized experiments [Fis37, Ney90] where the justification for
viewing the causes as random is immediate. For modern discussions see [IR15, Ros23].
Recently these ideas have been used to capture uncertainty in observational studies, see

55
[AAIW20, AAIW23]. The justification in panel data settings is not always quite as clear.
Consider one of the canonical applications of synthetic control methods, to estimate
the causal effect of German re-unification in 1989 on West German Gross Domestic
Product. A design based approach would require the researcher to contemplate an
alternative world where either other countries would have joint with East Germany, or
an alternative world where the re-unification with West Germany would have happened
in a different year. Both are difficult to consider. On the other hand, a sampling-based
approach would require the researcher to consider a world with additional countries
that could experience a unification event, again not an easy task.

9.1. The TWFE Estimator in the Staggered Adoption Case with Random Adoption
Dates

[AI18b] analyze the properties of the TWFE estimator under assumptions on the assign-
ment process in the staggered adoption setting, keeping the potential outcomes fixed.
In that case the assignment process is fully determined by the distribution of the adop-
tion date. [AI18b] derive the randomization-based distribution of the TWFE estimator
under the random assignment assumption alone and present an interpretation for the
estimand corresponding to that estimator. They show that as long as the adoption date
is randomly assigned, the estimand can be writte as a linear combination of average
causal effects on the outcome in period t if assigned adoption date a′ relative to being
assigned adoption date a:

N 
a,a′ 1 
Yit (a′ ) – Yit (a) ,
X
(20) τt =
N
i=1

with the weights summing to one, but generally including negative weights.
[AI18b] show the implications for the estimand of the assumption that there is no
anticipation of the treatment (so that the potential outcomes are invariant to the future
date of adoption). They also show how the interpretation of the estimand changes
further under the additional assumption that there are no dynamic effects, so that the
potential outcomes only depend on whether the adoption has taken place or not, but
not on the actual adoption date. [RR20] discuss the implications of variation in the
assignment probabilities and the biases this can create.

56
9.2. Switchback Designs

One design that has recently received considerable attention, after a long history, is
what [Coc39] called the rotation experiment, and what more recently has been referred
as a switchback experiment [BSLZ20] or crossover experiment [BJ80]. In such experiments
units are assigned to treatment or control in each of a number of periods, with individual
units potentially switching between treatment and control groups. Such experiments
were originally used in agricultural settings, where, for example, cattled were assigned
to different types of feed for some period of time. Using each unit as its own control
can substantially improve the precision of estimators compared to assigning each
unit to the treatment or control group for the entire study period. Such design have
become popular in tech company settings to deal with spillovers. For example, Lyft and
Uber often randomize markets to treatment and control groups, with the assignment
changing over time.

9.3. Experimental Design with Staggered Adoption

This subsection focuses on the design of experiments where the adoption date, rather
than the treatment in each period, is randomly assigned. Early studies, including
[HHC+ 15, HH07, BMDC16] focused on simple designs, such as those where a constant
fraction of units adopted the treatment in each period after the initial period. Sometimes
these designs suggested analyses that allowed for spillovers so outcomes for one or two
periods after the adoption would be discarded from the analyses if the focus was on the
average treatment effect.
[XABI19] focused on the question of optimally choosing the fraction adopters in each
period and showed that instead of it being constant, it was initially small and then larger
for some periods, after which it declined again. [BBI+ 23] discuss randomization-based
inference for some of these settings and present exact variances for some estimators.

9.4. Design with Dynamic Effects

[BRS21] propose unbiased estimators and derive their properties under the random-
ization distribution. They allow for dynamics in the treatment effects and essentially
unrestricted heterogeneity. They also discuss the biases of the conventional TWFE spec-
ifications in their setting. [BSLZ22] discuss optimal design from a minimax perspective,
allowing for carryover effects where the treatment status in recent periods may affect
current outcomes.

57
9.5. Robust Methods

The design-based approach to estimation and inference is natural in the context of

randomized experiments. However, in practice, applied researchers continue using
conventional panel data methods, such as TWFE, even with experimental data (e.g.,
[BP14, CP22]). There are multiple practical reasons for this: one can believe that the
experiment’s description is inconsistent with how it was implemented or think that
the TWFE estimator is more precise. These concerns are even more salient in quasi-
experimental environments where the data does not come from an experiment, but it
is appealing to treat it as such (e.g., [BH22]).
To address these issues [AILL21] propose a version of the TWFE estimator that
incorporates the design information. The key property of this method is that it delivers
a consistent estimator even if the design assumptions do not hold as long as the TWFE
model is correctly specified. Algorithmically, it amounts to estimating a weighted version
of the standard TWFE model:
N X
T
(τ̂rob , α̂, β̂) = arg min
X
2
(21) Yit – αi – βt – τWit ωi ,
τ,α,β
i=1 t=1

where the weights {ωi }ni=1 are constructed using the information about the design.

10. Open Questions for Further Research

Here we discuss some open questions in the current causal panel data literature.

10.1. Modeling Dynamics in Potential Outcomes

The recent panel data literature has only paid limited attention to dynamic treatment
effects (e.g., [Han20]). For example, a curious feature of many of the current methods,
including factor models and synthetic control methods, is that they pay essentially no
attention to the time-ordering of the observations. If the time labels were switched, the
estimated causal effects would not change. This seems implausible. Suppose one has
data available for T0 pre-treatment periods. For many of the methods, the researcher
would be indifferent between having available the first T0 /2 pre-treatment period versus
the second T0 /2 pre-treatment observations, whereas in practice one would think that
the more recent data would be more valuable.

58
 It seems likely the current literature will take the dynamics more seriously. One
direction may be the follow [Rob86] and a number of follow-up studies that developed
a sequential unconfoundedness approach. See also [BGK+ 15, BMAF+ 23]. Another di-
rection is the earlier panel data literature (e.g., [AH81]), with [AH07a] providing an
overview.

10.2. Validation

[LaL86] has become a very influential paper in the causal inference literature because
it provided an experimental data set that could be used to validate new methods for
estimating average causal effects under unconfoundedness. There are no data sets that
can deliver the same in many other panel data settings. However, there are methods
that can be used to assess the performance of proposed estimators. An early paper with
suggested tests is [HH89]. Currently many approaches in panel data rely on placebo
tests where the researcher pretends the treatment ocurred some periods prior to when
it actually did, The researcher then estimates the treatment effect for these periods
where, in the absence of anticipation effects, the treatment effect is known to be zero.
Finding estimates close to zero, both substantially and statistically, is then taken as
evidence in favor of the proposed methods. See for examples [IRS01] and [ADH15].

10.3. Dynamic Treatment Effects

The modern panel literature has paid relatively little attention to dynamic effects, com-
pared to the earlier literature [HN07], as well as compared to its importance in practice.
An interesting approach, somewhat related to synthetic control methods and factor
models, and taking a Bayesian perspectcive, is developed in [BGK+ 15]. Other recent
work includes [Han20, BB23].

10.4. Bridging Unconfoundedness and the TWFE Approach

Much of the discussion on unconfoundedness and the TWFE model has been framed
in terms of a choice. It is difficult to imagine that a clear consensus will emerge, and
finding practical methods that build on both approaches would be useful.

59
10.5. Continuous Treatments

Much of the recent literature has emphasized the binary treatment case. This has led
to valuable new insights, but it is clear that many applications go beyond the binary
treatment case. There is a small literature studying these cases, including [CGBS21]
and [DCd23], but more work is needed. Note that the earlier econometric panel data
literature did not distinguish between settings where the variables of interest were
binary or continuous.

11. Conclusion

The recent literature has expanded the set of methods available to empirical researchers
in social sciences in settings that are important in practice. This survey is an attempt
to put these methods in context, and show the close relationship between various
approaches, including two-way-fixed-effect and synthetic control methods, to provide
practitioners with additional guidance on when to use the various methods.

References
[AA03] Javier Alvarez and Manuel Arellano. The time series and cross-section asymptotics
of dynamic panel data estimators. Econometrica, 71(4):1121–1159, 2003.
+
[AAH 21] Dmitry Arkhangelsky, Susan Athey, David A Hirshberg, Guido W Imbens, and
Stefan Wager. Synthetic difference-in-differences. American Economic Review,
111(12):4088–4118, 2021.
[AAIW20] Alberto Abadie, Susan Athey, Guido W Imbens, and Jeffrey M Wooldridge. Sampling-
based versus design-based uncertainty in regression analysis. Econometrica,
88(1):265–296, 2020.
[AAIW23] Alberto Abadie, Susan Athey, Guido W Imbens, and Jeffrey M Wooldridge. When
should you adjust standard errors for clustering? The Quarterly Journal of Economics,
138(1):1–35, 2023.
[AB91] Manuel Arellano and Stephen Bond. Some tests of specification for panel data:
Monte carlo evidence and an application to employment equations. The review of
economic studies, 58(2):277–297, 1991.
[AB99] Joseph G Altonji and Rebecca M Blank. Race and gender in the labor market.
Handbook of labor economics, 3:3143–3259, 1999.
[AB11a] Manuel Arellano and Stéphane Bonhomme. Identifying distributional character-
istics in random coefficients panel data models. The Review of Economic Studies,
79(3):987–1020, 2011.
[AB11b] Manuel Arellano and Stéphane Bonhomme. Nonlinear panel data analysis. 2011.

60
 [AB16] Manuel Arellano and Stéphane Bonhomme. Nonlinear panel data estimation via
quantile regressions, 2016.
[Aba05] Alberto Abadie. Semiparametric difference-in-differences estimators. The Review
of Economic Studies, 72(1):1–19, 2005.
[Aba19] Alberto Abadie. Using synthetic controls: Feasibility, data requirements, and
methodological aspects. Journal of Economic Literature, 2019.
+
[ABD 21a] Susan Athey, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens, and Khashayar
Khosravi. Matrix completion methods for causal panel data models. Journal of the
American Statistical Association, 2021.
+
[ABD 21b] Susan Athey, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens, and Khashayar
Khosravi. Matrix completion methods for causal panel data models. Journal of the
American Statistical Association, 116(536):1716–1730, 2021.
[AC85] Orley Ashenfelter and David Card. Using the longitudinal structure of earnings
to estimate the effect of training programs. The Review of Economics and Statistics,
67(4):648–660, 1985.
[AC86] John M Abowd and David Card. On the covariance structure of earnings and hours
changes, 1986.
[ACI20] Susan Athey, Raj Chetty, and Guido Imbens. Combining experimental and obser-
vational data to estimate treatment effects on long term outcomes. arXiv preprint
arXiv:2006.09676, 2020.
[ADH10] Alberto Abadie, Alexis Diamond, and Jens Hainmueller. Synthetic control methods
for comparative case studies: Estimating the effect of california’s tobacco control
program. Journal of the American statistical Association, 105(490):493–505, 2010.
[ADH15] Alberto Abadie, Alexis Diamond, and Jens Hainmueller. Comparative politics and
the synthetic control method. American Journal of Political Science, pages 495–510,
2015.
[AG03] Alberto Abadie and Javier Gardeazabal. The economic costs of conflict: A case
study of the basque country. American Economic Review, 93(-):113–132, 2003.
[AG04] Orley Ashenfelter and Michael Greenstone. Using mandated speed limits to mea-
sure the value of a statistical life. Journal of political Economy, 112(S1):S226–S267,
2004.
[AH81] Theodore Wilbur Anderson and Cheng Hsiao. Estimation of dynamic models with
error components. Journal of the American statistical Association, 76(375):598–606,
1981.
[AH01] Manuel Arellano and Bo Honoré. Panel data models: some recent developments.
Handbook of econometrics, 5:3229–3296, 2001.
[AH07a] Jaap H Abbring and James J Heckman. Econometric evaluation of social programs,
part iii: Distributional treatment effects, dynamic treatment effects, dynamic dis-
crete choice, and general equilibrium policy evaluation. Handbook of econometrics,
6:5145–5303, 2007.
[AH07b] Manuel Arellano and Jinyong Hahn. Understanding bias in nonlinear panel models:
Some recent developments. Econometric Society Monographs, 43:381, 2007.

61
 [AH23] Dmitry Arkhangelsky and David Hirshberg. Large-sample properties of the
synthetic control method under selection on unobservables. arXiv preprint
arXiv:2311.13575, 2023.
[AI06] Susan Athey and Guido W Imbens. Identification and inference in nonlinear
difference-in-differences models. Econometrica, 74(2):431–497, 2006.
[AI18a] Dmitry Arkhangelsky and Guido Imbens. The role of the propensity score in fixed
effect models. Technical report, National Bureau of Economic Research, 2018.
[AI18b] Susan Athey and Guido Imbens. Design-based analysis in difference-in-differences
settings with staggered adoption. 2018.
[AI21] Susan Athey and Guido W Imbens. Design-based analysis in difference-in-
differences settings with staggered adoption. Journal of Econometrics, 2021.
[AI22] Dmitry Arkhangelsky and Guido W Imbens. Doubly robust identification for causal
panel data models. The Econometrics Journal, 25(3):649–674, 2022.
[AILL21] Dmitry Arkhangelsky, Guido W Imbens, Lihua Lei, and Xiaoman Luo. Double robust
two-way fixed effect regression for panel data. arXiv preprint arXiv:1909.09412, 2021.
[AIW18] Susan Athey, Guido W Imbens, and Stefan Wager. Approximate residual balancing:
debiased inference of average treatment effects in high dimensions. Journal of the
Royal Statistical Society: Series B (Statistical Methodology), 80(4):597–623, 2018.
[AK91] Joshua D Angrist and Alan Krueger. Does compulsory schooling affect schooling
and earnings. Quarterly Journal of Economics, CVI(4):979–1014, 1991.
[AKM99] John M Abowd, Francis Kramarz, and David N Margolis. High wage workers and
high wage firms. Econometrica, 67(2):251–333, 1999.
[AL21] Alberto Abadie and Jérémy L’hour. A penalized synthetic control estimator for
disaggregated data. Journal of the American Statistical Association, 116(536):1817–1834,
2021.
[Ald81] David J Aldous. Representations for partially exchangeable arrays of random
variables. Journal of Multivariate Analysis, 11(4):581–598, 1981.
[AM05] Joseph G Altonji and Rosa L Matzkin. Cross section and panel data estimators for
nonseparable models with endogenous regressors. Econometrica, 73(4):1053–1102,
2005.
[And84] Theodore Wilbur Anderson. Estimating linear statistical relationships. The Annals
of Statistics, 12(1):1–45, 1984.
[AP08] Joshua Angrist and Steve Pischke. Mostly Harmless Econometrics: An Empiricists’
Companion. Princeton University Press, 2008.
[Are87] Manuel Arellano. Computing robust standard errors for within group estimators.
Oxford bulletin of Economics and Statistics, 49(4):431–434, 1987.
[Are03] Manuel Arellano. Panel data econometrics. OUP Oxford, 2003.
[Ark19] Dmitry Arkhangelsky. Dealing with a technological bias: The difference-in-difference
approach. Centro de estudios monetarios y financieros, 2019.
[AS02] Susan Athey and Scott Stern. The impact of information technology on emergency
health care outcomes. The RAND Journal of Economics, 33(3):399–432, 2002.
[AS17] Peter M Aronow and Cyrus Samii. Estimating average causal effects under general

62
 interference, with application to a social network experiment. Annals of Applied
Statistics, 11(4):1912–1947, 2017.
[Ash78] Orley Ashenfelter. Estimating the effect of training programs on earnings. The
Review of Economics and Statistics, pages 47–57, 1978.
[ASS18] Muhammad Amjad, Devavrat Shah, and Dennis Shen. Robust synthetic control.
The Journal of Machine Learning Research, 19(1):802–852, 2018.
[AVdB03] Jaap H Abbring and Gerard J Van den Berg. The nonparametric identification of
treatment effects in duration models. Econometrica, 71(5):1491–1517, 2003.
[Bai09] Jushan Bai. Panel data models with interactive fixed effects. Econometrica,
77(4):1229–1279, 2009.
[Bal08] Badi Baltagi. Econometric analysis of panel data. John Wiley & Sons, 2008.
[BB98] Richard Blundell and Stephen Bond. Initial conditions and moment restrictions in
dynamic panel data models. Journal of econometrics, 87(1):115–143, 1998.
[BB23] Nicholas Brown and Kyle Butts. Dynamic treatment effect estimation with interac-
tive fixed effects and short panels. 2023.
[BBI+ 23] Patrick Bajari, Brian Burdick, Guido W Imbens, Lorenzo Masoero, James McQueen,
Thomas S Richardson, and Ido M Rosen. Experimental design in marketplaces.
Statistical Science, 1(1):1–19, 2023.
[BDM04] Marianne Bertrand, Esther Duflo, and Sendhil Mullainathan. How much should
we trust differences-in-differences estimates? The Quarterly Journal of Economics,
119(1):249–275, 2004.
[Bek94] Paul A Bekker. Alternative approximations to the distributions of instrumental
variable estimators. Econometrica: Journal of the Econometric Society, pages 657–681,
1994.
[BGK+ 15] Kay H Brodersen, Fabian Gallusser, Jim Koehler, Nicolas Remy, and Steven L Scott.
Inferring causal impact using bayesian structural time-series models. The Annals
of Applied Statistics, pages 247–274, 2015.
[BH22] Kirill Borusyak and Peter Hull. Non-random exposure to exogenous shocks. NBER
Working Paper, (27845), 2022.
[BHMM19] Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. Reconciling modern
machine-learning practice and the classical bias–variance trade-off. Proceedings of
the National Academy of Sciences, 116(32):15849–15854, 2019.
[BJ80] Byron Wm Brown Jr. The crossover experiment for clinical trials. Biometrics, pages
69–79, 1980.
[BJ15] Andrew Bell and Kelvyn Jones. Explaining fixed effects: Random effects modeling
of time-series cross-sectional and panel data. Political Science Research and Methods,
3(1):133–153, 2015.
[BJS21] Kirill Borusyak, Xavier Jaravel, and Jann Spiess. Revisiting event study designs:
Robust and efficient estimation. arXiv preprint arXiv:2108.12419, 2021.
[BLM22a] Stéphane Bonhomme, Thibaut Lamadon, and Elena Manresa. Discretizing unob-
served heterogeneity. Econometrica, 90(2):625–643, 2022.
[BLM22b] Luca Braghieri, Ro’ee Levy, and Alexey Makarin. Social media and mental health.

63
 American Economic Review, 112(11):3660–3693, 2022.
[BLW21] Andrew Baker, David F Larcker, and Charles CY Wang. How much should we trust
staggered difference-in-differences estimates? Available at SSRN 3794018, 2021.
[BM15] Stéphane Bonhomme and Elena Manresa. Grouped patterns of heterogeneity in
panel data. Econometrica, 83(3):1147–1184, 2015.
[BMAF+ 23] Eli Ben-Michael, David Arbour, Avi Feller, Alexander Franks, and Steven Raphael.
Estimating the effects of a california gun control program with multitask gaussian
processes. The Annals of Applied Statistics, 17(2):985–1016, 2023.
[BMDC16] Daniel Barker, Patrick McElduff, Catherine D’Este, and MJ Campbell. Stepped
wedge cluster randomised trials: a review of the statistical methodology used and
available. BMC medical research methodology, 16(1):1–19, 2016.
[BMFR21] Eli Ben-Michael, Avi Feller, and Jesse Rothstein. The augmented synthetic control
method. Journal of the American Statistical Association, 116(536):1789–1803, 2021.
[BMFR22] Eli Ben-Michael, Avi Feller, and Jesse Rothstein. Synthetic controls with staggered
adoption. Journal of the Royal Statistical Society Series B: Statistical Methodology,
84(2):351–381, 2022.
[BN02] Jushan Bai and Serena Ng. Determining the number of factors in approximate
factor models. Econometrica, 70(1):191–221, 2002.
[Bon12] Stéphane Bonhomme. Functional differencing. Econometrica, 80(4):1337–1385, 2012.
[Bon20] Stéphane Bonhomme. Econometric analysis of bipartite networks. In The econo-
metric analysis of network data, pages 83–121. Elsevier, 2020.
[BP14] Christian Broda and Jonathan A Parker. The economic stimulus payments of
2008 and the aggregate demand for consumption. Journal of Monetary Economics,
68:S20–S36, 2014.
[BR05] Heejung Bang and James M Robins. Doubly robust estimation in missing data and
causal inference models. Biometrics, 61(4):962–973, 2005.
[BRS21] Iavor Bojinov, Ashesh Rambachan, and Neil Shephard. Panel experiments and
dynamic causal effects: A finite population perspective. Quantitative Economics,
12(4):1171–1196, 2021.
[BS11] Stéphane Bonhomme and Ulrich Sauder. Recovering distributions in difference-
in-differences models: A comparison of selective and comprehensive schooling.
Review of Economics and Statistics, 93(2):479–494, 2011.
[BSLZ20] Iavor Bojinov, David Simchi-Levi, and Jinglong Zhao. Design and analysis of switch-
back experiments. Available at SSRN 3684168, 2020.
[BSLZ22] Iavor Bojinov, David Simchi-Levi, and Jinglong Zhao. Design and analysis of switch-
back experiments. Management Science, 2022.
[Car94] David Card. Intertemporal labour supply: an assessment, volume 2 of Econometric
Society Monographs, page 49–78. Cambridge University Press, 1994.
[CCD+ 17] Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian
Hansen, and Whitney Newey. Double/debiased/neyman machine learning of treat-
ment effects. American Economic Review, 107(5):261–65, 2017.
[CFT21] Matias D Cattaneo, Yingjie Feng, and Rocio Titiunik. Prediction intervals for syn-

64
 thetic control methods. Journal of the American Statistical Association, 116(536):1865–
1880, 2021.
[CFVHN13] Victor Chernozhukov, Iván Fernández-Val, Jinyong Hahn, and Whitney Newey.
Average and quantile effects in nonseparable panel models. Econometrica, 81(2):535–
580, 2013.
[CGBS21] Brantly Callaway, Andrew Goodman-Bacon, and Pedro HC Sant’Anna. Difference-
in-differences with a continuous treatment. arXiv preprint arXiv:2107.02637, 2021.
[Cha77] Gary Chamberlain. Education, income, and ability revisited. Journal of Econometrics,
5(2):241–257, 1977.
[Cha80] Gary Chamberlain. Analysis of covariance with qualitative data. The review of
economic studies, 47(1):225–238, 1980.
[Cha82] Gary Chamberlain. Multivariate regression models for panel data. Journal of
econometrics, 18(1):5–46, 1982.
[Cha84] Gary Chamberlain. Panel data. Handbook of econometrics, 2:1247–1318, 1984.
[Cha92] Gary Chamberlain. Efficiency bounds for semiparametric regression. Econometrica:
Journal of the Econometric Society, pages 567–596, 1992.
[Cha10] Gary Chamberlain. Binary response models for panel data: Identification and
information. Econometrica, 78(1):159–168, 2010.
[CK94] David Card and Alan Krueger. Minimum wages and employment: A case study
of the fast-food industry in new jersey and pennsylvania. The American Economic
Review, 84(4):772–793, 1994.
[CLLX23] Albert Chiu, Xingchen Lan, Ziyi Liu, and Yiqing Xu. What to do (and not to do) with
causal panel analysis under parallel trends: Lessons from a large reanalysis study.
Available at SSRN 4490035, 2023.
[CM22] Denis Chetverikov and Elena Manresa. Spectral and post-spectral estimators for
grouped panel data models. arXiv preprint arXiv:2212.13324, 2022.
[Coc39] WG Cochran. Long-term agricultural experiments. Supplement to the Journal of the
Royal Statistical Society, 6(2):104–148, 1939.
[CP22] Emanuele Colonnelli and Mounu Prem. Corruption and firms. The Review of
Economic Studies, 89(2):695–732, 2022.
[CR83] Gary Chamberlain and Michael Rothschild. Arbitrage, factor structure, and mean-
variance analysis on large asset markets. Econometrica: Journal of the Econometric
Society, pages 1281–1304, 1983.
[CRY22] David Card, Jesse Rothstein, and Moises Yi. Industry wage differentials: A firm-
based approach. Unpublished draft, University of California, Berkeley, 2022.
[CS20] Brantly Callaway and Pedro HC Sant’Anna. Difference-in-differences with multiple
time periods. Journal of Econometrics, 2020.
[CWZ21] Victor Chernozhukov, Kaspar Wüthrich, and Yinchu Zhu. An exact and robust
conformal inference method for counterfactual and synthetic controls. Journal of
the American Statistical Association, 116(536):1849–1864, 2021.
[DCd20] Clément De Chaisemartin and Xavier d’Haultfœuille. Two-way fixed effects estima-
tors with heterogeneous treatment effects. American Economic Review, 110(9):2964–

65
 2996, 2020.
[DCd23] Clément De Chaisemartin and Xavier d’Haultfoeuille. Two-way fixed effects and
differences-in-differences with heterogeneous treatment effects: A survey. The
Econometrics Journal, 26(3):C1–C30, 2023.
[Dea85] Angus Deaton. Panel data from time series of cross-sections. Journal of econometrics,
30(1-2):109–126, 1985.
[Dea21a] Ben Deaner. Many proxy controls. arXiv preprint arXiv:2110.03973, 2021.
[Dea21b] Ben Deaner. Proxy controls and panel data. arXiv preprint arXiv:1810.00283, 2021.
[DI16] Nikolay Doudchenko and Guido W Imbens. Balancing, regression, difference-in-
differences and synthetic control methods: A synthesis. Technical report, National
Bureau of Economic Research, 2016.
[DL19] Peng Ding and Fan Li. A bracketing relationship between difference-in-differences
and lagged-dependent-variable adjustment. Political Analysis, 27(4):605–615, 2019.
[DW99] Rajeev H Dehejia and Sadek Wahba. Causal effects in nonexperimental studies:
Reevaluating the evaluation of training programs. Journal of the American statistical
Association, 94(448):1053–1062, 1999.
[DW02] Rajeev H Dehejia and Sadek Wahba. Propensity score-matching methods for
nonexperimental causal studies. Review of Economics and statistics, 84(1):151–161,
2002.
[EHR83] Robert F Engle, David F Hendry, and Jean-Francois Richard. exogeneity. Economet-
rica: Journal of the Econometric Society, pages 277–304, 1983.
[EL96] Nada Eissa and Jeffrey B Liebman. Labor supply response to the earned income
tax credit. The quarterly journal of economics, 111(2):605–637, 1996.
[FFJR69] Eugene F Fama, Lawrence Fisher, Michael C Jensen, and Richard Roll. The adjust-
ment of stock prices to new information. International economic review, 10(1):1–21,
1969.
[FHS19] Simon Freyaldenhoven, Christian Hansen, and Jesse M Shapiro. Pre-event trends
in the panel event-study design. American Economic Review, 109(9):3307–38, 2019.
[Fis37] Ronald Aylmer Fisher. The design of experiments. Oliver And Boyd; Edinburgh;
London, 1937.
[FP21] Bruno Ferman and Cristine Pinto. Synthetic controls with imperfect pretreatment
fit. Quantitative Economics, 12(4):1197–1221, 2021.
[Fre18] Joachim Freyberger. Non-parametric panel data models with interactive fixed
effects. The Review of Economic Studies, 85(3):1824–1851, 2018.
[FVW16] Iván Fernández-Val and Martin Weidner. Individual and time effects in nonlinear
panel models with large n, t. Journal of Econometrics, 192(1):291–312, 2016.
[FVW18] Iván Fernández-Val and Martin Weidner. Fixed effects estimation of large-t panel
data models. Annual Review of Economics, 10:109–138, 2018.
[GB21] Andrew Goodman-Bacon. Difference-in-differences with variation in treatment
timing. Journal of Econometrics, 225(2):254–277, 2021.
[GM16] Laurent Gobillon and Thierry Magnac. Regional policy evaluation: Interactive
fixed effects and synthetic controls. Review of Economics and Statistics, 98(3):535–551,

66
 2016.
[Gol91] Arthur Stanley Goldberger. A course in econometrics. Harvard University Press, 1991.
[GP12] Bryan S Graham and James L Powell. Identification and estimation of average
partial effects in “irregular”’ correlated random coefficient panel data models.
Econometrica, 80(5):2105–2152, 2012.
[Gri77] Zvi Griliches. Estimating the returns to schooling: Some econometric problems.
Econometrica: Journal of the Econometric Society, pages 1–22, 1977.
[GSW22] Dalia Ghanem, Pedro HC Sant’Anna, and Kaspar Wüthrich. Selection and parallel
trends. arXiv preprint arXiv:2203.09001, 2022.
[Gun23] Florian F Gunsilius. Distributional synthetic controls. Econometrica, 91(3):1105–1117,
2023.
[GYR+ 22] AmirEmad Ghassami, Alan Yang, David Richardson, Ilya Shpitser, and Eric Tchet-
gen Tchetgen. Combining experimental and observational data for identification
and estimation of long-term causal effects. arXiv preprint arXiv:2201.10743, 2022.
[Han07] Christian B Hansen. Generalized least squares inference in panel and multilevel
models with serial correlation and fixed effects. Journal of Econometrics, 140(2):670–
694, 2007.
[Han20] Sukjin Han. Identification in nonparametric models for dynamic treatment effects.
Journal of Econometrics, 2020.
[Hec81] James J Heckman. Statistical models for discrete panel data. Structural analysis of
discrete data with econometric applications, 114:178, 1981.
[HENR88] Douglas Holtz-Eakin, Whitney Newey, and Harvey S Rosen. Estimating vector
autoregressions with panel data. Econometrica: Journal of the econometric society,
pages 1371–1395, 1988.
[HH89] James J Heckman and V Joseph Hotz. Choosing among alternative nonexperimental
methods for estimating the impact of social programs: The case of manpower
training. Journal of the American statistical Association, 84(408):862–874, 1989.
[HH07] Michael A Hussey and James P Hughes. Design and analysis of stepped wedge
cluster randomized trials. Contemporary clinical trials, 28(2):182–191, 2007.
[HH08] Michael Hudgens and Elizabeth Halloran. Toward causal inference with interfer-
ence. Journal of the American Statistical Association, pages 832–842, 2008.
[HHC+ 15] Karla Hemming, Terry P Haines, Peter J Chilton, Alan J Girling, and Richard J
Lilford. The stepped wedge cluster randomised trial: rationale, design, analysis,
and reporting. Bmj, 350, 2015.
[HK02] Jinyong Hahn and Guido Kuersteiner. Asymptotically unbiased inference for a
dynamic panel model with fixed effects when both n and t are large. Econometrica,
70(4):1639–1657, 2002.
[HN07] James J Heckman and Salvador Navarro. Dynamic discrete choice and dynamic
treatment effects. Journal of Econometrics, 136(2):341–396, 2007.
[Hon92] Bo E Honoré. Trimmed lad and least squares estimation of truncated and censored
regression models with fixed effects. Econometrica: journal of the Econometric Society,
pages 533–565, 1992.

67
 [HS08] Yingyao Hu and Susanne M Schennach. Instrumental variable treatment of non-
classical measurement error models. Econometrica, 76(1):195–216, 2008.
[HSCKW12] Cheng Hsiao, H Steve Ching, and Shui Ki Wan. A panel data approach for program
evaluation: measuring the benefits of political and economic integration of hong
kong with mainland china. Journal of Applied Econometrics, 27(5):705–740, 2012.
[Hsi22] Cheng Hsiao. Analysis of panel data. Number 64. Cambridge university press, 2022.
[HT06] Bo E Honoré and Elie Tamer. Bounds on parameters in panel dynamic discrete
choice models. Econometrica, 74(3):611–629, 2006.
[IK21] Kosuke Imai and In Song Kim. On the use of two-way fixed effects regression
models for causal inference with panel data. Political Analysis, 29(3):405–415, 2021.
[IKM21] Guido Imbens, Nathan Kallus, and Xiaojie Mao. Controlling for unmeasured con-
founding in panel data using minimal bridge functions: From two-way fixed effects
to factor models. arXiv preprint arXiv:2108.03849, 2021.
[IKW21] Kosuke Imai, In Song Kim, and Erik H Wang. Matching methods for causal infer-
ence with time-series cross-sectional data. American Journal of Political Science,
2021.
[Imb02] Guido W Imbens. Generalized method of moments and empirical likelihood.
Journal of Business & Economic Statistics, 20(4):493–506, 2002.
[Imb04] Guido Imbens. Nonparametric estimation of average treatment effects under
exogeneity: A review. Review of Economics and Statistics, pages 1–29, 2004.
[IR15] Guido W Imbens and Donald B Rubin. Causal Inference in Statistics, Social, and
Biomedical Sciences. Cambridge University Press, 2015.
[IRS01] Guido W Imbens, Donald B Rubin, and Bruce I Sacerdote. Estimating the effect of
unearned income on labor earnings, savings, and consumption: Evidence from a
survey of lottery players. American Economic Review, pages 778–794, 2001.
[IV23] Guido Imbens and Davide Viviano. Identification and inference for synthetic
controls with confounding. 2023.
[Jak21] Pamela Jakiela. Simple diagnostics for two-way fixed effects. arXiv preprint
arXiv:2103.13229, 2021.
[KM20] Nathan Kallus and Xiaojie Mao. On the role of surrogates in the efficient estimation
of treatment effects with limited outcome data. arXiv preprint arXiv:2003.12408,
2020.
[KMPT21] Maxwell Kellogg, Magne Mogstad, Guillaume A Pouliot, and Alexander Torgovitsky.
Combining matching and synthetic control to tradeoff biases from extrapolation
and interpolation. Journal of the American statistical association, 116(536):1804–1816,
2021.
[KZEM21] Timo Kuosmanen, Xun Zhou, Juha Eskelinen, and Pekka Malo. Design flaw of the
synthetic control method. 2021.
[LaL86] Robert J LaLonde. Evaluating the econometric evaluations of training programs
with experimental data. The American economic review, pages 604–620, 1986.
[Lan00] Tony Lancaster. The incidental parameter problem since 1948. Journal of economet-
rics, 95(2):391–413, 2000.

68
 [LC20] Lihua Lei and Emmanuel J Candès. Conformal inference of counterfactuals and
individual treatment effects. arXiv preprint arXiv:2006.06138, 2020.
[LJ76] Robert E Lucas Jr. Econometric policy evaluation: A critique. In Carnegie-Rochester
conference series on public policy, volume 1, pages 19–46. North-Holland, 1976.
[LWX22] Licheng Liu, Ye Wang, and Yiqing Xu. A practical guide to counterfactual estimators
for causal inference with time-series cross-sectional data. American Journal of
Political Science, 2022.
[Lyn84] James Lynch. Canonical row-column-exchangeable arrays. Journal of multivariate
analysis, 15(1):135–140, 1984.
[LZ86] Kung-Yee Liang and Scott L Zeger. Longitudinal data analysis using generalized
linear models. Biometrika, 73(1):13–22, 1986.
[Mag04] Thierry Magnac. Panel binary variables and sufficiency: generalizing conditional
logit. Econometrica, 72(6):1859–1876, 2004.
[MP18] Charles F Manski and John V Pepper. How do right-to-carry laws affect crime rates?
coping with ambiguity using bounded-variation assumptions. Review of Economics
and Statistics, 100(2):232–244, 2018.
[Mug22] Martin Mugnier. Make the difference! computationally trivial estimators for
grouped fixed effects models. arXiv preprint arXiv:2203.08879, 2022.
[Mun78] Yair Mundlak. On the pooling of time series and cross section data. Econometrica,
pages 69–85, 1978.
[MVD95] Bruce D Meyer, W Kip Viscusi, and David L Durbin. Workers’ compensation and
injury duration: Evidence from a natural experiment. The American Economic
Review, 85(3):322, 1995.
[MW15] Hyungsik Roger Moon and Martin Weidner. Linear regression for panel with
unknown number of factors as interactive fixed effects. Econometrica, 83(4):1543–
1579, 2015.
[MW18] Hyungsik Roger Moon and Martin Weidner. Nuclear norm regularized estimation
of panel regression models. arXiv preprint arXiv:1810.10987, 2018.
[Ney90] Jerzey Neyman. On the application of probability theory to agricultural experi-
ments. essay on principles. section 9. Statistical Science, 5(4):465–472, 1923/1990.
[Nic81] Stephen Nickell. Biases in dynamic models with fixed effects. Econometrica: Journal
of the econometric society, pages 1417–1426, 1981.
[NS48] Jerzy Neyman and Elizabeth L Scott. Consistent estimates based on partially
consistent observations. Econometrica: Journal of the Econometric Society, pages
1–32, 1948.
[Pes06] M Hashem Pesaran. Estimation and inference in large heterogeneous panels with
a multifactor error structure. Econometrica, 74(4):967–1012, 2006.
[PT07] Thomas Plümper and Vera E Troeger. Efficient estimation of time-invariant and
rarely changing variables in finite sample panel analyses with unit fixed effects.
Political analysis, 15(2):124–139, 2007.
[RHB00] James M Robins, Miguel Angel Hernan, and Babette Brumback. Marginal structural
models and causal inference in epidemiology. Epidemiology, 11:550–560, 2000.

69
 [Rob86] James Robins. A new approach to causal inference in mortality studies with a
sustained exposure period application to control of the healthy worker survivor
effect. Mathematical modelling, 7(9-12):1393–1512, 1986.
[Ros02] Paul R Rosenbaum. Multiple control groups. Observational Studies, pages 253–275,
2002.
[Ros23] Paul R Rosenbaum. Causal inference. MIT Press, 2023.
[RR83] Paul R Rosenbaum and Donald B Rubin. The central role of the propensity score
in observational studies for causal effects. Biometrika, 70(1):41–55, 1983.
[RR20] Ashesh Rambachan and Jonathan Roth. Design-based uncertainty for quasi-
experiments. arXiv preprint arXiv:2008.00602, 2020.
[RR23] Ashesh Rambachan and Jonathan Roth. A more credible approach to parallel
trends. Review of Economic Studies, page rdad018, 2023.
[RS23] Jonathan Roth and Pedro HC Sant’Anna. When is parallel trends sensitive to func-
tional form? Econometrica, 91(2):737–747, 2023.
[RSBP23] Jonathan Roth, Pedro HC Sant’Anna, Alyssa Bilinski, and John Poe. What’s trending
in difference-in-differences? a synthesis of the recent econometrics literature.
Journal of Econometrics, 2023.
[Rub74] Donald B Rubin. Estimating causal effects of treatments in randomized and non-
randomized studies. Journal of educational Psychology, 66(5):688, 1974.
[Rub78] Donald B Rubin. Bayesian inference for causal effects: The role of randomization.
The Annals of statistics, pages 34–58, 1978.
[SA20] Liyang Sun and Sarah Abraham. Estimating dynamic treatment effects in event
studies with heterogeneous treatment effects. Journal of Econometrics, 2020.
[SDSY22] Dennis Shen, Peng Ding, Jasjeet Sekhon, and Bin Yu. A tale of two panel data
regressions. arXiv preprint arXiv:2207.14481, 2022.
[SIV23] Jann Spiess, Guido Imbens, and Amar Venugopal. Double and single descent in
causal inference with an application to high-dimensional synthetic control. arXiv
preprint arXiv:2305.00700, 2023.
[SMNT20] Xu Shi, Wang Miao, Jennifer C. Nelson, and Eric J. Tchetgen Tchetgen. Multiply
robust causal inference with double-negative control adjustment for categorical
unmeasured confounding. Journal of The Royal Statistical Society Series B-statistical
Methodology, 82(2):521–540, 2020.
+
[SRC 16] Tamar Sofer, David B Richardson, Elena Colicino, Joel Schwartz, and Eric J Tch-
etgen Tchetgen. On negative outcome control of unobserved confounding as a
generalization of difference-in-differences. Statistical science: a review journal of the
Institute of Mathematical Statistics, 31(3):348, 2016.
[SSP+ 19] Pantelis Samartsidis, Shaun R Seaman, Anne M Presanis, Matthew Hickman, and
Daniela De Angelis. Assessing the causal effect of binary interventions from ob-
servational panel data with few treated units. Statistical Science, 34(3):486–503,
2019.
[SW98] James H Stock and Mark W Watson. Diffusion indexes, 1998.
[SZ20] Pedro HC Sant’Anna and Jun Zhao. Doubly robust difference-in-differences estima-

70
 tors. Journal of Econometrics, 219(1):101–122, 2020.
[Vog12] Timothy J Vogelsang. Heteroskedasticity, autocorrelation, and spatial correlation
robust inference in linear panel models with fixed-effects. Journal of Econometrics,
166(2):303–319, 2012.
[Woo10] J.M. Wooldridge. Econometric Analysis of Cross Section and Panel Data. The MIT
Press. MIT Press, 2010.
[Woo21] Jeffrey M Wooldridge. Two-way fixed effects, the two-way mundlak regression, and
difference-in-differences estimators. Available at SSRN 3906345, 2021.
[Woo22] Jeffrey M Wooldridge. Simple approaches to nonlinear difference-in-differences
with panel data. Available at SSRN 4183726, 2022.
[XABI19] Ruoxuan Xiong, Susan Athey, Mohsen Bayati, and Guido Imbens. Optimal experi-
mental design for staggered rollouts. arXiv preprint arXiv:1911.03764, 2019.
[Xu17] Yiqing Xu. Generalized synthetic control method: Causal inference with interactive
fixed effects models. Political Analysis, 25(1):57–76, 2017.
[Xu23] Yiqing Xu. Causal inference with time-series cross-sectional data: a reflection.
Available at SSRN 3979613, 2023.
[YMST21] Andrew Ying, Wang Miao, Xu Shi, and Eric J Tchetgen Tchetgen. Proximal causal
inference for complex longitudinal studies. arXiv preprint arXiv:2109.07030, 2021.

71