Difference-in-Differences Designs: A Practitioner’s Guide

Andrew Baker∗ Brantly Callaway† Scott Cunningham‡

Andrew Goodman-Bacon§ Pedro H. C. Sant’Anna¶
arXiv:2503.13323v2 [econ.EM] 11 Jun 2025

June 12, 2025

Abstract

Difference-in-differences (DiD) is arguably the most popular quasi-experimental research

design. Its canonical form, with two groups and two periods, is well-understood. However,
empirical practices can be ad hoc when researchers go beyond that simple case. This article
provides an organizing framework for discussing different types of DiD designs and their
associated DiD estimators. It discusses covariates, weights, handling multiple periods, and
staggered treatments. The organizational framework, however, applies to other extensions of
DiD methods as well.

1 Introduction
Dating to the 1840s, difference-in-differences (DiD) is now the most common research design for
estimating causal effects in the social sciences.1 A basic DiD design requires two time periods,
one before and one after some treatment begins, and two groups, one that receives a treatment
and one that does not. The DiD estimate equals the change in outcomes for the treated group
minus the change in outcomes for the untreated group: the difference of two differences. If the
average change in the outcomes would have been the same in the two groups had the treatment
not occurred, which is referred to as a “parallel trends” assumption, this comparison estimates the
average treatment effect among treated units.

∗
University of California, Berkeley
†
University of Georgia
‡
Baylor University
§
Opportunity and Inclusive Growth Institute, Federal Reserve Bank of Minneapolis
¶
Emory University
1
Currie, Kleven and Zwiers (2020) find that almost 25% of all NBER empirical working papers and 17%
of empirical articles in five leading general-interest economics journals in 2018 mention DiD. The earliest DiD
applications we are aware of are from Ignaz Semmelweis from the 1840s (Semmelweis, 1983) and Snow (1855). For
a brief overview of the long history of DiD in economics, see Section 2 of Lechner (2011).

1
 In practice, however, researchers apply DiD methods to situations that are more complicated
than the classic two-period and two-group (2 ˆ 2) setup. Most datasets cover multiple periods, and
units may enter (or exit) treatment at different times. Treatment might also vary in its amount or
intensity. Other variables are often used to make treated and untreated units more comparable.
Today’s typical DiD study includes at least one of these deviations from the canonical 2 ˆ 2 setup.
For many years, the common practice in applied research was to estimate complex DiD designs
using linear regressions with unit and time fixed effects (two-way fixed effects, henceforth TWFE).
Their identifying assumptions and interpretation were informally traced to the fact that, in the
2 ˆ 2 case, a TWFE estimator gives the same estimate as a DiD estimator calculated directly
from sample means, and thus inherits a clear causal interpretation under a specific parallel trends
identification assumption. This appeared to justify the use of a single technique for any type
of design or specification. Recent research, however, has shown that simple regressions can fail
to estimate meaningful causal parameters when DiD designs are complex and treatment effects
vary, producing estimates that are not only misleading in their magnitudes but potentially of the
wrong sign. The significance of these findings is substantial; given the prevalence of DiD analysis
in modern applied econometrics work, common empirical practices have almost certainly yielded
misleading results in several concrete cases (Baker, Larcker and Wang, 2022).
So, what should applied researchers do instead? This paper proposes a unified framework
for discussing and conducting DiD studies that is rooted in the principles of causal inference in
the presence of treatment effect heterogeneity. The central conclusion of recent methodological
research is that even complex DiD studies can be understood as aggregations of 2 ˆ 2 comparisons
between one set of units for which treatment changes and another set for which it does not.
This fact links a wide variety of DiD designs used in practice and guides methodological choices
about estimating them. Viewing DiD studies through the lens of 2 ˆ 2 “building blocks” aids in
interpretability by clarifying that they yield causal quantities that aggregate the treatment effects
identified by each 2ˆ2 component. It also means that identification comes from the simple parallel
trends assumptions required for each 2ˆ2 building block. Practically, this framework suggests first
estimating each 2 ˆ 2 building block and then aggregating them. As long as the effective sample
size is large, this approach allows for asymptotically valid inference using standard techniques.
This framework is a “forward-engineering” approach to DiD that embraces treatment effect
heterogeneity and constructs estimators that recover well-motivated causal parameters under ex-
plicitly stated assumptions. By fixing the goals of the study (the target parameters) and deriving
analytical techniques, forward engineering provides clear benefits over “reverse-engineering” ap-
proaches that begin with a familiar regression specification and derive the assumptions under
which it has some causal interpretation. The methods we describe in this paper combine familiar
techniques with some newer ones, but expressly avoid the difficulties of interpretation inherent
in common regression estimators (Goodman-Bacon, 2021; de Chaisemartin and D’Haultfoeuille,
2020; Sun and Abraham, 2021; Borusyak, Jaravel and Spiess, 2024). Moreover, the interpretation

2
of common regression estimators changes across specifications, which makes it hard to understand
the difference between non-robustness and a shifting target parameter. In contrast, our proposed
framework naturally leads to estimation procedures that target the same parameter under differ-
ent transparent identification assumptions. Thus, two estimates can be distinguished easily by
their identifying assumptions. Finally, the principles of the forward-engineering approach provide
guidance to good econometric practices even in settings without well-established methodological
findings.
This paper is not designed to be a comprehensive literature review; its goal is to provide
guidelines for practitioners who want to better understand DiD and its various forms. Because
of the tremendous variations in design, data, and specification that practitioners encounter, we
opt to focus on three of the most common aspects of modern DiD studies: the use of weights,
covariates, and staggered treatment timing. We apply techniques to address these issues to a
specific example: the causal effect of recent public health insurance expansions in the US on county-
level mortality. In an appendix, we briefly discuss related DiD designs with different treatment
variables (ones that turn on and off or take many values), additional comparisons (i.e., triple-
difference designs), distributional target parameters, or different data structures (repeated cross-
sections or unbalanced panels). Several recent reviews follow the logic laid out here and cover
additional DiD-related topics and technical details:Roth, Sant’Anna, Bilinski and Poe (2023);
de Chaisemartin and D’Haultfoeuille (2023b); Callaway (2023).
The rest of the paper is structured as follows. Section 2 introduces the Medicaid example.
Section 3 discusses the canonical 2ˆ2 DiD setups with and without weights, and Section 4 discusses
threats to the identification assumptions, how to assess them, and how to incorporate covariates.
Section 5 extends the 2ˆ2 setup to multiple periods with potentially staggered treatment adoption.
Section 6 concludes and briefly discusses some extensions that involve more complex DiD designs.

2 Medicaid and mortality: The running example

To make our methodological discussion concrete, we revisit a timely and important causal question:
How did the expansion of public health insurance (Medicaid) under the Affordable Care Act (ACA)
affect mortality?
Medicaid expansion is a great example of a staggered treatment adoption. The ACA originally
mandated that in 2014 all states expand Medicaid eligibility to adults with incomes up to 138%
of the federal poverty threshold. In upholding the law’s constitutionality in a 2012 decision,
however, the Supreme Court made Medicaid expansion optional. As a result, many states expanded
Medicaid after 2014, but several have not done so as of 2024.
Columns 1 and 2 of Table 1 illustrate the variation in Medicaid expansion dates.

3
 Table 1: Medicaid Expansion under the Affordable Care Act

Expansion Year States Share of States Share of Counties Share of Adults (2013)

Pre-2014 DE, MA, NY, VT 0.08 0.03 0.09

2014 AR, AZ, CA, CO, CT, HI, IA, IL, KY, MD, 0.44 0.36 0.45
MI, MN, ND, NH, NJ, NM, NV, OH, OR,
RI, WA, WV
2015 AK, IN, PA 0.06 0.06 0.06
2016 LA, MT 0.04 0.04 0.02
2019 ME, VA 0.04 0.05 0.03
2020 ID, NE, UT 0.06 0.04 0.02
2021 MO, OK 0.04 0.06 0.03
2023 NC, SD 0.04 0.05 0.03
Non-Expansion AL, FL, GA, KS, MS, SC, TN, TX, WI, WY 0.20 0.31 0.26

The table shows which states adopted the ACA’s Medicaid expansion in each year as well as the share of all states, counties, and adults in each
expansion year.

States expanded Medicaid largely because of economic and political considerations (Sommers
and Epstein, 2013), which created observable differences between expansion and non-expansion
states. For instance, just four out of the 22 states that expanded Medicaid in 2014 are in the
southern Census region; conversely, seven out of 10 non-expansion states are in the South. This
suggests a potential role for covariates when analyzing Medicaid expansion.
Finally, mortality is measured in jurisdictions like states and counties, which are of very different
sizes. Choices about (population) weights determine not only how different estimation approaches
average the units within a given expansion group but also how a given estimation technique
averages estimated effects across those groups. California, for example, represented 4.5% of the
states that expanded Medicaid in 2014, 5.4% of the counties, but 27.7% of the adults ages 20-
64; its contribution to “the” average outcome for the 2014 expansion group is very different with
weights than without. The final three columns of Table 1 show that, in our data the entire 2014
expansion group contains 44% of the states, 36% of the counties, but 45% of all adults. Weighting
will therefore change how important the estimated treatment effects are for the 2014 group.
Several recent papers study the effect of ACA Medicaid expansion on mortality rates for lower-
income adults, who are most likely to gain insurance through Medicaid. Miller, Johnson and
Wherry (2021) and Wyse and Meyer (2024) use simple DiD methods to provide evidence that
Medicaid reduced adult mortality rates for targeted sub-populations. Unfortunately, their analyses
require restricted links between income and mortality data, which are important for overcoming
the low statistical power in studies using aggregate mortality data (Black, Hollingsworth, Nunes
and Simon, 2022). Our goal is to pursue a replicable and shareable example based on a related
analysis by Borgschulte and Vogler (2020). They use a sophisticated strategy to select and use
covariates in a weighted TWFE regression using restricted access data, and find that Medicaid
expansion reduced aggregate county-level mortality rates. We use publicly available data, which

4
we include in a fully-reproducible replication package, and consider only a handful of intuitive
demographic and economic covariates sufficient to illustrate several practical challenges that can
arise with DiD. This empirical exercise is meant solely to illustrate how to tackle several common
features of DiD designs. The results are pedagogical in spirit and do not represent the best possible
estimates of Medicaid’s effect on adult mortality.
Our outcome variable is the crude adult mortality rate, Yi,t , for people ages 20-64 (measured
per 100,000) by county (i) from 2009 to 2019 released by the Centers for Disease Control and
Prevention (2024).2 We denote county i’s adult population in 2013 by Wi and its socioeconomic
covariates in year t (discussed below) by Xi,t . The information in Table 1 defines the treatment
group variable Gi that equals the year in which county i’s state expanded Medicaid; Gi “ 8 for
the non-expansion states. Our final sample contains 2,604 counties in states with complete data
on mortality rates from 2009 to 2019 and covariates for 2013 and 2014.
Faced with a setup such as this, researchers need to make a range of tightly related choices.
Which treatment groups in Table 1 should be compared with each other and over what time
horizons? What must be true for those comparisons to identify causal effects, and how should one
empirically evaluate their plausibility? How can other information, such as covariates or pre-period
outcomes, be used to improve the credibility of the design? How do these methodological choices
affect the causal interpretation of a given analysis? The aim of this review is to demonstrate to
practitioners using DiD in realistic scenarios why and how to use state-of-the-art econometric tools
to answer these questions.

3 2ˆ2 DiD designs

We begin our discussion by focusing on the canonical 2ˆ2 DiD setup, which has two time periods—
one before and one after treatment—and two groups—one that remains untreated in both periods
and one that becomes treated in the second period. In our Medicaid example, we focus on com-
parisons in 2014 and 2013 between the 2014 expansion group (978 counties) and the group that
had not expanded by 2019 (1,222 counties). When we consider more complex designs, this kind of
comparison will still play a role: it will be one 2 ˆ 2 “building block” among many.
Using these basic ingredients, we can now define a 2ˆ2 DiD design, composed of a causal target
parameter, a treatment variable, an assumption under which it is identified, and an estimation
approach, which will be the classic difference of two differences. This may be familiar territory
in the simple case, but it is a crucial framework for building up appropriate techniques in more
complex cases.

2
It is common to adjust mortality rates by the county age distribution. Unfortunately, the CDC measurements
of age-specific deaths are restricted for counties with fewer than 10 annual deaths. We aim to use publicly available
and shareable data for pedagogical purposes; we follow Borgschulte and Vogler (2020) and use the crude mortality
rate.

5
 We first define a treatment group dummy Di that equals one for the treated units (expansion
states, Gi “ 2014) and zero for the untreated units (states that had not expanded by 2019,
Gi ą 2019). The treatment status dummy, Di,t “ Di ˆ 1tt ě 2014u, then equals one for counties in
2014 expansion states during post-expansion years. To highlight how weights enter different kinds
of DiD analyses, we use the following notation for expected values. For generic random variables
L
A and C, for a given set of non-negative weights ω, define Eω rA|Cs “ ErωA|Cs Erω|Cs as the
ω-weighted population expectation of A given C. When ω “ 1 for all units in the population, we
simply write Eω rA|Cs “ ErA|Cs. Henceforth, unless otherwise noted, we assume that we have a
balanced panel data random sample of pYt“1 , . . . , Yt“T , G, Xq.

3.1 Causal effects and target parameters: The ATT

The first step of any causal analysis is to define the causal quantity of interest, also called the
target parameter. We use the potential outcomes framework of Rubin (1974) and Robins (1986)
to do so. Let Yi,t p0, 0q denote unit i’s potential outcome at time t if it remained untreated in both
periods. Analogously, let Yi,t p0, 1q denote unit i’s potential outcome at time t if untreated in the
first period but exposed to treatment by the second period. In our example, Yi,t p0, 0q is county i’s
mortality rate in period t in a world in which Medicaid did not expand in its state, and Yi,t p0, 1q is
its mortality rate in a world in which Medicaid did expand in 2014.3 To simplify notation, we will
write Yi,t p0q “ Yi,t p0, 0q and Yi,t p1q “ Yi,t p0, 1q, as the potential outcomes are defined by treatment
exposure in period two (Medicaid expansion status by 2014). Nonetheless, it will be useful for
later discussions that these potential outcomes correspond to treatment sequences.
In practice, we never observe Yi,t p1q and Yi,t p0q for the the same unit. Instead, the data we
observe, Yi,t , are treated outcomes Yi,t p1q for treated units, and untreated outcomes Yi,t p0q for
untreated units, as in the following equation:

Yi,t “ p1 ´ Di qYi,t p0q ` Di Yi,t p1q. (3.1)

We additionally assume that county mortality rates were not affected by the Medicaid expansion
before Medicaid expanded, which is crucial to the validity of the DiD estimator (see, e.g., Ab-
bring and van den Berg, 2003, Malani and Reif, 2015, and Roth et al., 2023). This standard “no
anticipation” assumption ensures that we observe untreated potential outcomes before Medicaid
expansion takes effect: Yi,2013 “ Yi,2013 p0q. It also helps us define effective treatment dates. For in-
stance, if the announcement of Medicaid expansion affected mortality before its actual expansion,
“treatment” would begin when the policy was announced rather than implemented. We formally
state this assumption for completeness and maintain it throughout the paper.

3
We have implicitly introduced the stable unit treatment value assumption, which holds if the only treatment
determining county i’s mortality rate is its own. In other words, Medicaid expansion in neighboring counties
must not affect deaths in county i. If this fails, then there are effectively many different treatment variables and
counterfactuals. The potential outcomes notation and ensuing analysis would then need to account for this.

6
Assumption NA (No-Anticipation). For all treated units i and all pre-treatment periods t,
Yi,t p1q “ Yi,t p0q.
The potential outcomes define a causal effect for every unit in every time period, Yi,t p1q´Yi,t p0q.
These describe what Medicaid expansion did to mortality rates in a specific treated county or what
it would have done in a specific untreated county. This framework allows for arbitrary heterogeneity
in the effects across units and time; that is, it allows the effect of Medicaid expansion to be different
in every county and year. But it is hard to learn about this degree of rich heterogeneity without
additional strong conditions holding. Instead, DiD analyses typically seek to estimate (weighted)
averages of heterogeneous treatment effects. In particular, most DiD designs target the average
treatment effect on the treated at time t, or AT T ptq:

AT T ptq “ Eω rYi,t p1q ´ Yi,t p0q|Di “ 1s

“ Eω rYi,t |Di “ 1s ´ Eω rYi,t p0q|Di “ 1s. (3.2)

Equation (3.2) shows that AT T ptq compares (weighted) average observed post-expansion mortality
rates among treated counties (Eω rYi,t |Di “ 1s) with the (weighted) average untreated mortality
rates for the same treated counties (Eω rYi,t p0q|Di “ 1s). The second quantity is counterfactual be-
cause untreated outcomes are never observed for treated counties. Note that by the no-anticipation
assumption, AT T ptq “ 0 for all pre-treatment periods; that is, AT T p2013q “ 0 in our two-period
Medicaid example. This ensures that cross-group outcome comparisons before treatment begins
reflect untreated potential outcome gaps, which is central to the logic of DiD. Note that we abuse
notation and omit the weight index when defining AT T ’s; we do that to unclutter notation through-
out the paper.
Equation (3.2) shows that weighting enters the analysis early on, as part of the definition of
the causal parameter. In the Medicaid context, the unweighted AT T p2014q answers the question,
“What was the average causal effect of Medicaid expansion on 2014 mortality rates among the 2014
expansion state counties?” The weighted AT T p2014q answers the question, “What was the average
causal effect of Medicaid expansion on 2014 mortality rates among adults in counties in states that
expanded Medicaid in 2014?” This point interacts with other justifications for weighting, such as
improving precision. With heterogeneous treatment effects, adopting a weighting scheme designed
to improve precision in the presence of heteroskedasticity in a constant-coefficient regression model
will also change the target parameter, potentially by a lot when treatment effects are correlated
with the weights (Solon, Haider and Wooldridge, 2015). Comparing weighted and unweighted
estimates, therefore, does not show whether weighting matters for estimation or inference; these
reflect different target parameters. In our example, the population-weighted parameter is probably
more policy-relevant, but we conduct some of our empirical exercises both ways to show how
weighting can affect a given DiD result. Which parameter is “of interest” is an argument about
theoretical importance, policy relevance, and the use to which it will be put.
Other target parameters are also possible. Designs other than DiD identify different kinds

7
of average treatment effects, and some DiD methods use quantile regression (Athey and Imbens,
2006; Callaway and Li, 2019) or distribution regression (Fernández-Val, Meier, van Vuuren and
Vella, 2024a) approaches to target features of the marginal distributions of Yi,t p1q and Yi,t p0q among
treated units. We focus on identification and estimation strategies that target ATT parameters
but emphasize that the 2 ˆ 2 building block framework applies to DiD methods more broadly; see
our appendix for more discussions about distributional parameters.

3.2 Identifying assumptions: Parallel trends

A research design is a strategy—a set of assumptions—to identify and estimate specific target
parameters. Many different assumptions can identify the missing counterfactual for AT T p2014q in
the Medicaid example. For example, mean independence between Yi,2014 p0q and Di implies that the
counterfactual equals average 2014 mortality rates in non-expansion counties (Eω rYi,2014 p0q|Di “
0s). Under this assumption, which essentially entails assuming that Medicaid expansion is as
good as random, the cross-sectional mortality gap in 2014 between expansion and non-expansion
counties is the AT T p2014q. Similarly, time invariance of Yi,t p0q among expansion counties (plus the
fact that we ruled out anticipatory behavior) implies that the counterfactual equals 2013 mortality
rates in expansion counties (Eω rYi,2013 p0q|Di “ 1s). Under this assumption, which essentially rules
out non-treatment-related changes in the outcome variable, the “time trend” in average mortality
in expansion counties is the AT T p2014q.
DiD comes from an alternative assumption that identifies the relevant counterfactual even
when the average untreated potential outcome differs across treatment groups (which violates
mean independence) and changes over time (which violates time invariance). The parallel trends
assumption states that in the absence of treatment, the average outcome evolution is the same
among treated and comparison groups. For general assumptions and results, we denote time
periods by t “ 1, 2, but continue to be explicit about which years are being used when we reference
the Medicaid example.

Assumption PT (2ˆ2 Parallel Trends). The (weighted) average change of Yi,t“2 p0q from Yi,t“1 p0q
is the same between treated and comparison groups; that is,

Eω rYi,t“2 p0q|Di “ 1s ´ Eω rYi,t“1 p0q|Di “ 1s “ Eω rYi,t“2 p0q|Di “ 0s ´ Eω rYi,t“1 p0q|Di “ 0s. (3.3)

If the parallel trends assumption holds, then it is easy to construct Eω rYi,t“2 p0q|Di “ 1s from
observable quantities—that is, to identify it:
` ˘
Eω rYi,t“2 p0q|Di “ 1s “ Eω rYi,t“1 |Di “ 1s ` Eω rYi,t“2 |Di “ 0s ´ Eω rYi,t“1 |Di “ 0s . (3.4)
In the Medicaid example, assumption PT says that to calculate expansion counties’ average 2014
mortality rate in a counterfactual world without Medicaid expansion, start with their average 2013
mortality rate and add the observed change in average mortality rates in non-expansion counties.
Substituting (3.4) into the definition of AT T p2014q and replacing potential outcomes with observed

8
outcomes using (3.1) gives the 2 ˆ 2 DiD estimand, an expression for the target parameter in terms
of four estimable population averages:
“Eω rYi,2014 p1q|Di “1s “E rY
ω i,2014 p0q|D “1s
i
hkkkkkkkkkkikkkkkkkkkkj hkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkikkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkj
` ` ˘˘
AT T p2014q “ Eω rYi,2014 |Di “ 1s ´ Eω rYi,2013 |Di “ 1s ` Eω rYi,2014 |Di “ 0s ´ Eω rYi,2013 |Di “ 0s
“ pE ω rYi,2014 |Di “ 1s ´ Eω rYi,2013 |Di “ 1sq ´ loooooooooooooooooooooooooomoooooooooooooooooooooooooon
loooooooooooooooooooooooooomoooooooooooooooooooooooooon pEω rYi,2014 |Di “ 0s ´ Eω rYi,2013 |Di “ 0sq .
(weighted) average change for Di “1 (weighted) average change for Di “0
(3.5)
Equation (3.5) highlights what makes DiD so attractive. It is intuitive, it has very mild data
requirements (just four means), it answers ex post questions like “what did the treatment do?”,
and its identifying assumption can be stated precisely.
Parallel trends makes DiD distinct from causal designs that are based on statistical indepen-
dence between treatment and potential outcomes. In designs like randomized trials or instrumental
variables, the conditions—mean equalities across groups, for instance—that identify counterfac-
tuals are often a statistical consequence of the randomness induced externally (Heckman, 2000).
In contrast, parallel trends is just a restriction on untreated potential outcome trends. It does
not necessarily come from exogenous variation “outside the model.” In fact, because treatment
adoption is often chosen by economic actors or policymakers “inside the model,” parallel trends
need not hold. For this reason, DiD analyses (correctly) devote significant attention to evaluating
parallel trends.
We discuss how to generate empirical evidence about the plausibility of parallel trends below,
but new theoretical findings about treatment choice behaviors or selection mechanisms also inform
the plausibility of parallel trends.4 These results explicitly connect DiD to behaviors, they provide
grounding for stories about why a given DiD analysis is internally valid (or not), and they discipline
empirical tests of parallel trends. A full review of this literature is outside the scope of this paper,
but some helpful broad themes emerge.5 For instance, a trade-off exists between the information
the agents who choose treatment know and how they act on it, and the time-series properties of
untreated potential outcomes. At one extreme, consider someone who knows Yi,t p0q before and
after treatment and can opt into or out of treatment based on it. Parallel trends can only hold in
this case if, other than common shifts for every unit, Yi,t p0q is constant (Ghanem et al., 2022). In
our Medicaid example, neither of these conditions—that state legislatures in 2013 knew their 2014
untreated mortality rates or that untreated mortality rates would all have shifted in parallel—is
plausible. The opposite extreme is when treatment timing is random, in which case parallel trends
holds without any time-series restrictions. Yet, in this case, DiD is not necessary, and more efficient
estimators exist (Roth and Sant’Anna, 2023a). Against the political and economic backdrop of
4
In fact, some of the earliest economic research on DiD methods examined exactly these questions (Ashenfelter
and Card, 1985; Heckman and Robb, 1985).
5
For details on selection models and outcome models that are consistent with parallel trends, see Ghanem,
Sant’Anna and Wüthrich (2022) and Marx, Tamer and Tang (2024); Chabé-Ferret (2015) applies these arguments
to earnings dynamics.

9
the early 2010s, it is clear that Medicaid expansion decisions were not random.
Therefore, in realistic scenarios, parallel trends only holds if some restrictions on the way
untreated outcomes enter the treatment selection mechanism hold as well. As one example, imagine
that treatment selection depends on the permanent component of Yi,t p0q (fixed effects) but not on
shorter-term fluctuations (“shocks”). For instance, if state legislatures knew and considered their
long-run mortality levels only when making their expansion decision, they would be following this
kind of selection mechanism. Expansion and non-expansion states might then have large differences
in the permanent part of untreated outcomes, which cancel in equation (3.3), and so parallel trends
would hold if shocks to Yi,t p0q had a time invariant mean conditional on the fixed effects.6 State
legislatures, however, may have also known whether their 2013 mortality rates were especially high
or low when considering expanding Medicaid. If the expansion choice is related to these 2013
mortality shocks as well, parallel trends would hold only if stronger time-series restrictions on
Yi,t p0q hold. Ghanem et al. (2022) provide a fuller discussion of the selection/time-series trade-off
and theory-driven templates to assess parallel trends, while Marx et al. (2024) discuss economic
models that are and are not compatible with parallel trends.
Another implication of the fact that DiD does not rely on statistical independence between
Yi,t p0q and treatment status is that there is no guarantee that parallel trends holds across different
transformations of Yi,t p0q. As stated, it is simply an assumption about averages for a particular
Yi,t p0q. Roth and Sant’Anna (2023b) show that parallel trends is insensitive to functional form if
and only if it holds between groups and across the distribution of Yi,t p0q. This can only be true
if Medicaid adoption is random, the mortality distribution is constant between 2013 and 2014, or
a combination of the two cases. As these conditions are arguably ex-ante implausible, our DiD
analysis may depend on our choices to measure Yi,t in rates (deaths per 100,000) as opposed to
logs, for example. One approach to this measurement choice is to propose and evaluate a theory
that delivers it, though we recognize that this is not always possible. To assess whether cases where
parallel trends holding for one functional form come at the cost of ruling other transformations,
we recommend that researchers use the Roth and Sant’Anna’s (2023b) falsification tests for the
null that parallel trends are insensitive to functional form. In our application, we do not reject the
null that parallel trends is insensitive to functional form, with p-values above 0.80.
The interplay between treatment selection and the properties of the outcome variable char-
acterize the structural basis for a DiD analysis (see DiNardo and Lee, 2011) and engaging with
them is essential to any DiD application. While every study will have its own institutions, choices,
and outcomes to consider, a rigorous DiD analysis must provide a transparent discussion about
the reliability of the underlying identification assumptions. If parallel trends requires implausible
behavioral restrictions, one may be better off using an alternative research design.
6
Some researchers may find easier to understand these as “parallel changes” rather than “parallel trends.”
However, the use of “parallel trends” is now firmly established in the literature, and other influential work has used
“changes-in-changes” to refer to an alternative estimator to classical DiD estimation (Athey and Imbens, 2006). To
avoid confusion, we use “parallel trends” throughout this paper.

10
3.3 Estimation and inference: Four means or one regression?
Mapping the DiD estimand in equation (3.5) to the canonical 2 ˆ 2 DiD estimator follows imme-
diately from replacing population expectations with their sample analogs:

AT
z T p2014q “ pY ω,D“1,t“2014 ´ Y ω,D“1,t“2013 q ´ pY ω,D“0,t“2014 ´ Y ω,D“0,t“2013 q, (3.6)
řn 1
i“1 1tDi “ g, t “ t uωi Yi,t1
where Y ω,D“g,t“t1 “ ř n is the ω-weighted sample mean of Y for treatment
i“1 ωi 1tDi “ g, t “ t u
1

group g in period t1 . Equation (3.6) is the classic difference of two differences written in terms of
sample means. It is a direct recipe for actually estimating AT T ptq and can be read directly from
the following table of average mortality rates in 2013 and 2014 by expansion group.

Table 2: Simple 2 ˆ 2 DiD

Unweighted Averages Weighted Averages

Expansion No Expansion Gap/DiD Expansion No Expansion Gap/DiD
2013 419.2 474.0 -54.8 322.7 376.4 -53.7
2014 428.5 483.1 -54.7 326.5 382.7 -56.2
Trend/DiD 9.3 9.1 0.1 3.7 6.3 -2.6

This table reports average county-level mortality rates (deaths among adults aged 20-64 per 100,000 adults)
in 2013 (top row) and 2014 (middle row) in states that expanded adult Medicaid eligibility in 2014 (columns
1 and 4, 978 counties) and states that have not expanded by 2019 (columns 2 and 5, 1,222 counties). The
first three columns present unweighted averages and the second three columns present population-weighted
averages. Columns 1, 2, 4, and 5 in the third row show time trends in mortality between 2013 and 2014 for
each group of states. The first two rows of columns 3 and 6 show the cross-sectional gap in mortality between
expansion and non-expansion states in 2013 and 2014. The entries in red text in the bottom row show the
simple 2 ˆ 2 difference-in-differences estimates without weights (column 3) and with them (column 6)

The two across-time changes in equation (3.6) are in the third row of the table. Without
weighting, average county-level mortality rates in expansion states rose by 9.3 deaths per 100,000
and 9.1 deaths in non-expansion states, so after rounding, the DiD estimate of AT T p2014q is 0.1
deaths per 100,000. This result implies that the average treatment effect of Medicaid expansion
on mortality in 2014 among counties that are part of an expansion state was an increase of 0.1
deaths per 100,000. In contrast, the DiD result using population weights suggests that Medicaid
expansion caused a reduction of 2.6 deaths per 100,000 for the average adult in expansion states.7
The same result can be obtained as the (weighted) least squares estimate of β 2ˆ2 in the following
linear regression specification (which only uses data from t “ 2013 and t “ 2014):

Yi,t “ β0 ` β1 1tDi “ 1u ` β2 1tt “ 2014u ` β 2ˆ2 p1tDi “ 1u ˆ 1tt “ 2014uq ` εi,t , (3.7)

where β’s are unknown coefficients and εi,t is an idiosyncratic term uncorrelated with Di . To see
why, let’s focus on the unweighted case (ωi “ 1) and write each of the four means in ATz T p2q in
7
Columns 3 and 6 show cross-group gaps in average mortality in each year. These can also be used to construct
the DiD estimate by rearranging equation (3.6): pY D“1,t“2014 ´ Y D“0,t“2014 q ´ pY D“1,t“2013 ´ Y D“0,t“2013 q.

11
terms of the estimated coefficients from (3.7):

– Sample average of Yi,t in post-period for treatment group is Y D“1,t“2014 “ βp0 ` βp1 ` βp2 ` βp2ˆ2 .

– Sample average of Yi,t in pre-period for treatment group is Y D“1,t“2013 “ βp0 ` βp1 .

– Sample average of Yi,t in post-period for comparison group is Y D“0,t“2014 “ βp0 ` βp2 .

– Sample average of Yi,t in pre-period for comparison group is Y D“0,t“2013 “ βp0 .

Substituting these expressions into the definition of AT

z T p2q yields:
„ ȷ „ ȷ
AT T p2014q “ pβ0 ` β1 ` β2 ` β q ´ pβ0 ` β1 q ´ pβ0 ` β2 q ´ β0 “ βp2ˆ2 .
z p p p p2ˆ2 p p p p p

Table 3 demonstrates this equivalence for both unweighted (column 1) and weighted (column 4)
regressions. In fact, with balanced panel data, the estimate of β 2ˆ2 is numerically the same if the
regression instead contains fixed effects for each unit (columns 2 and 5) or if one regresses outcome
changes on a constant and the treatment group dummy Di (columns 3 and 6).8 Provided that one
adopts the same notion of uncertainty, standard errors also coincide; such as when one clusters the
standard errors at the county level.

Table 3: Regression 2 ˆ 2 DiD

Unweighted Weighted
Crude Mortality Rate ∆ Crude Mortality Rate ∆
(1) (2) (3) (4) (5) (6)
Constant 474.0˚˚˚ 9.1˚˚˚ 376.4˚˚˚ 6.3˚˚˚
(4.3) (2.6) (7.6) (1.1)
Medicaid Expansion -54.8˚˚˚ -53.7˚˚˚
(6.3) (11.5)
Post 9.1˚˚˚ 6.3˚˚˚
(2.6) (1.1)
Medicaid Expansion ˆ Post 0.1 0.1 0.1 -2.6˚ -2.6˚ -2.6˚
(3.7) (3.7) (3.7) (1.5) (1.5) (1.5)
County fixed effects No Yes No No Yes No
Year fixed effects No Yes No No Yes No
This table reports the regression 2 ˆ 2 DiD estimate comparing 978 counties in states that
expanded Medicaid in 2014 with 1,222 counties in states that did not expand Medicaid by
2019, using only data for the years 2013 and 2014. Columns 1-3 report unweighted regression
results, while columns 4-6 weight by county population aged 20-64 in 2013. Columns 1 and 4
report results from regressing the crude mortality rate for adults ages 20-64 on indicators for
expansion states (Treat) and post-expansion year (Post); the DiD estimate is the coefficient on
the interaction term. Columns 2 and 5 report the corresponding results for the interaction term
using county and year fixed effects. Finally, Columns 3 and 6 report the results of the long
difference in county mortality rates on a treatment indicator. Standard errors (in parentheses)
are clustered at the county level.

The equivalence between calculating a 2 ˆ 2 DiD by hand or with a regression has appealing
features. Regressions are simple to run, and they do the averaging and differencing behind the
8
When population weights vary over time, the equivalence between a by-hand DiD estimate and one that comes
from a regression with unit fixed effects no longer holds.

12
scenes. They also allow the use of statistical inference tools from ordinary least squares (OLS),
which are themselves the subject of a large econometrics literature that is particularly important
when it comes to standard error estimation (Wooldridge, 2003; Bertrand, Duflo and Mullainathan,
2004; Donald and Lang, 2007; Cameron, Gelbach and Miller, 2008; Conley and Taber, 2011; Abadie,
Athey, Imbens and Wooldridge, 2020, 2023).
Many inference procedures exist for DiD-type analyses, arising from a combination of choices
about the target parameter, details of the data structure and sampling process, and maintained
assumptions about the structure of outcomes. In practice, one needs to determine and discuss the
forms of uncertainty the standard errors are designed to capture—that is, what is (conceptually)
being resampled and what may or may not vary across those resamples. As discussed in Abadie
et al. (2020), these details come from the nature of the parameter of interest—whether the focus
is on sample-specific average treatment effects or population-level average treatment effects—and
the stochastic elements of the model that make the estimator random. Heuristically, this involves
a thought experiment (or stochastic process) hypothesized to generate the random components
of the model (or the data-generating process). Different inferential frameworks highlight different
sources of uncertainty by resampling distinct model components and treating other components
as fixed (non-random).
Inferential frameworks on two extremes help cement these concepts. Design-based frameworks
treat potential outcomes and covariates as non-random, focus on finite-population parameters (e.g.,
sample average treatment effects), and consider the allocation of treatment as the only source of the
randomness in the model (Imbens and Rubin, 2015).9 The only thing that is random and thus varies
across the hypothetical resamples from this point of view is the treatment allocation. On the other
hand, a traditional sampling-based approach to inference presumes that we independently sample
units from a superpopulation. In this case, it is customary to focus on population parameters (like
AT T p2q), treat all variables in the model as random variables, and compute standard errors by
clustering at the level in which the (hypothesized) sampling was conducted. In this framework,
every variable in the analysis—outcomes, covariates, and treatment—is randomly redrawn across
the hypothetical resamples. A drawback of the sampling approach is that sometimes, it is unnatural
to think of the data as a random sample from a well-defined population.
A third popular approach to inference—the model-based approach—is more structural and
involves taking a stand on the structure of the error component of the model (e.g., imposing a
putative model for how shocks affect outcomes and their relationship with treatment and other
variables in the model). The uncertainty reflected in this model-based setting entails a thought
experiment in which different values of these shocks and the other random variables in the model are
drawn from their joint distribution (Abadie et al., 2023). This model-based approach is common in

9
Traditionally, design-based inference procedures are justified when treatment assignment is fully random, which
is a much stronger requirement than parallel trends. See Rambachan and Roth (2024) for a discussion on design-
based inference for quasi-experimental designs, including a discussion of the Medicaid expansion.

13
econometrics, and it almost always takes the linear regression specification (or model, in this case)
as the starting point of the analysis. Although this is often convenient, it is important to note that
imposing model restrictions on the error component of the model necessarily imposes restrictions
on treatment effect heterogeneity and on the relationship between potential outcomes; see Section
5 and Appendix A of Roth et al. (2023) for a discussion. Another challenge with the model-based
approach is that it is hard to use this framework when adopting estimation strategies other than
linear regressions—for example, inverse probability weighting or doubly robust procedures—which
we will discuss in Session 4.4.
Ultimately, as the discussion above highlights, each approach has pros and cons, and discussions
about the best way to compute standard errors are complex and intrinsically context-specific. A
detailed treatment of the topic is outside the scope of this paper. It requires information about
the sampling process, the research design and target parameters, what is treated as fixed and
random, and the structure of the error components of their models (e.g., presence of spatial or
serial correlation), among other factors. We refer interested readers to Abadie et al. (2020, 2023)
and Section 5 of Roth et al. (2023) for discussions on these topics, though we also emphasize that
further methodological research in this area is warranted. For the remainder of this article, we
adopt a sampling perspective for uncertainty and cluster our standard errors at the county level.
In our context, this is compatible with treating all variables as random, including treatment groups
and potential outcomes. It also allows us to avoid (a) making time-series dependence restrictions
on potential (and realized) outcomes—as we are in a short-panel framework with a large number
of units and a fixed number of time periods—and (b) taking an explicit stand on the structure
of error components of the model, which is particularly appealing as the starting point of our
analysis is potential outcomes rather than regression models. It is also worth mentioning that
as our treatment in the empirical example is assigned at the state level, clustering at the county
level would also be compatible with treating state-specific shocks as fixed (or conditioned on) and
assessing if they lead to violations of parallel trends (Roth et al., 2023, Section 5.1). Clustering
at the state level would be justified if we were using a design-based perspective (Rambachan and
Roth, 2024), though that would require us to treat potential outcomes as fixed (which we do not do
in this paper). Our choice of inference procedure is not without controversies, and other inferential
approaches may also be rationalized.
We conclude this section by stressing that the appeal of using regressions like (3.7) to estimate
ATT in DiD designs comes from the fact that their numerical equivalence to the “by-hand” DiD
estimator (3.6), which was explicitly derived from the AT T ptq and the parallel trends assumption.
This ensures that the regression specification respects the underlying identifying assumptions and
estimates the desired target parameter. Unfortunately, the tight connection between (TWFE)
regressions and DiD designs breaks under more complex setups that are ubiquitous in practice.
We now turn to some of these issues and how approaching them from the point of view of 2 ˆ 2
building blocks can guide good econometric practices.

14
4 Incorporating covariates into 2ˆ2 DiD
So far, we have focused on 2ˆ2 DiD designs that do not leverage any information about covariates,
but researchers frequently incorporate them into DiD analyses in one of three ways: checking
for balance in variables thought to influence Yi,t p0q, controlling for those variables in the main
estimates, and estimating treatment effect heterogeneity. For example, in the absence of Medicaid
expansion, mortality rates likely would have evolved differently in poorer and richer counties;
they certainly did before 2014 (Currie and Schwandt, 2016). Therefore, parallel trends may fail
if poverty rates differ between expansion and non-expansion counties. If they do, then one may
want to “control for” poverty rates when estimating AT T parameters. Finally, because Medicaid
expansion reached more people in higher-poverty counties, its average effects on overall mortality
may be larger there. This heterogeneity may be of interest in its own right or may be used to
assess the plausibility of the overall DiD design.
This section discusses how to use auxiliary information on covariates to evaluate parallel trends,
identify AT T parameters under potentially weaker assumptions, and study heterogeneity. These
approaches stem from viewing a DiD with covariates in terms of 2 ˆ 2 building blocks that them-
selves condition on those variables, which creates a clear link to the assumptions, estimators, and
interpretation of unconditional 2 ˆ 2 designs. We also discuss how regression estimators that
control for covariates impose extra assumptions and can fail to identify AT T parameters.

4.1 Covariate balance: Is unconditional parallel trends plausible?

Assumption PT is fundamentally untestable, as it contains an unobserved counterfactual compo-
nent. Therefore, “tests” of these assumptions are necessarily indirect and rely on other observed
variables thought to be related to untreated potential outcome trends. For example, during the
2010s, demographics and economic conditions were strongly correlated with mortality levels and
trends. If these relationships would have held in the absence of Medicaid expansion, and if expan-
sion and non-expansion counties differ in those demographic and economic characteristics, then
the parallel trends assumption (3.3) may fail to hold. Checking balance in observable determinants
of changes in Yi,t p0q is thus a common and sensible way to evaluate parallel trends.
Most DiD analyses check for balance across groups in baseline covariate levels (Eω rXi,t“1 |Di “
1s ´ Eω rXi,t“1 |Di “ 0s) or covariate trends before and after treatment (Eω r∆Xi,t“2 |Di “ 1s ´
Eω r∆Xi,t“2 |Di “ 0s, where ∆Xi,t“2 “ Xi,t“2 ´ Xi,t“1 ). We consider the following variables, Xi,t :
the percentages of a county’s population that are female, white, or Hispanic; the unemployment
rate; the poverty rate; and county-level median income (in thousands of dollars).10 The top
panel of Table 4 reports averages of these variables by group in 2013 with and without population

10
We focus on these variables for convenience. Borgschulte and Vogler (2020) use a LASSO procedure that
selects more and different covariates to include in their analysis. We replicate their findings when we follow their
methodology but diverge here for the sake of brevity.

15
weights. We also report a measure of imbalance that is comparable across variables: the normalized
difference in means between the treatment and comparison groups (Imbens and Rubin, 2015,
Chapter 14),
X ω,T ´ X ω,C
Norm. Diffω “ b ,
2 2
pSω,T ` Sω,C q{2

where X ω,T and X ω,C are the sample weighted or unweighted averages for the treatment and com-
2 2
parison groups, respectively, and Sω,T and Sω,C are the sample weighted or unweighted variances
of the covariates for the treatment and comparison group. As a general rule of thumb, values
of the normalized difference in excess of 0.25 in absolute value indicate a potentially problematic
imbalance between the two groups (Imbens and Rubin, 2015, page 277).11

Table 4: Covariate Balance Statistics

Unweighted Weighted
Variable Non-Adopt Adopt Norm. Diff. Non-Adopt Adopt Norm. Diff.
2013 Covariate Levels
% Female 49.43 49.33 -0.03 50.48 50.07 -0.24
% White 81.64 90.48 0.59 77.91 79.54 0.11
% Hispanic 9.64 8.23 -0.10 17.01 18.86 0.11
Unemployment Rate 7.61 8.01 0.16 7.00 8.01 0.50
Poverty Rate 19.28 16.53 -0.42 17.24 15.29 -0.37
Median Income 43.04 47.97 0.43 49.31 57.86 0.68
2014 - 2013 Covariate Differences
% Female -0.02 -0.02 0.00 0.02 0.01 -0.09
% White -0.21 -0.21 0.01 -0.32 -0.33 -0.04
% Hispanic 0.20 0.21 0.04 0.25 0.33 0.29
Unemployment Rate -1.16 -1.30 -0.21 -1.08 -1.36 -0.55
Poverty Rate -0.55 -0.28 0.14 -0.41 -0.35 0.05
Median Income 0.98 1.11 0.06 1.10 1.74 0.32

This table reports the covariate balance between 978 counties in states that expanded Medicaid in 2014 and
1,222 counties in states that did not expand by 2019. In the top panel, we report the averages and standardized
differences of each variable, measured in 2013, by adoption status. All variables are measured in percentage val-
ues, except for median household income, which is measured in thousands of U.S. dollars. In the bottom panel
we report the average and standardized differences of the county-level long differences between 2014 and 2013 of
each variable. We report both weighted and unweighted measures of the averages to correspond to the different
estimation methods of including covariates in a 2 ˆ 2 setting.

We find meaningful imbalance in several baseline measures. Expansion counties in 2013 were
11
As discussed in Austin (2009, Section 3.2), the normalized difference was initially proposed in the psychological
literature and is sometimes referred to as Cohen’s effect size index. In this context, normalized differences of 0.2,
0.5, and 0.8 are sometimes used to represent small, medium, and large imbalances; however, normalized differences
as small as 0.1 are sometimes considered worrisome, depending on how important the covariate in question is.
Ultimately, there is no universally accepted threshold for what value indicates important imbalances. See Ho, Imai,
King and Stuart (2007) and Austin (2009) for additional discussions.

16
whiter and had higher unemployment rates despite lower poverty and higher median income than
non-expansion counties. Because DiD uses changes in outcomes, researchers sometimes argue that
the effect of pre-treatment variables is differenced out. This logic does not hold, though, if baseline
covariates are related to untreated potential outcome trends themselves. The imbalance in the top
panel of Table 4 will lead to violations of parallel trends to the extent that counties with different
racial composition or income distributions would have had different mortality changes even without
Medicaid expansion.
Nevertheless, checks of balance in covariate changes can be informative about parallel trends
as well. The bottom panel of Table 4 reports average changes by group between 2013 and 2014 as
well as normalized differences. Many of the imbalances evident in baseline levels change, or even
flip signs, when measured in changes. Unemployment, for example, was higher in expansion states
in 2013 but fell faster. To the extent that these changes are important determinants of ∆Yi,t p0q,
then these results could suggest that Assumption PT is violated.
Why do we say “could”? A major challenge in interpreting cross-group gaps in ∆Xi,t involves
deciding which variables are truly covariates and which are mechanisms/outcomes. If an element
of Xi,t cannot be affected by the treatment, it is a (strictly exogenous) covariate, and differential
changes in exogenous covariates may indicate a PT violation. Since the treatment cannot have
caused Xi,t to change (by assumption), something else that differs across groups and over time must
have. Since little research suggests an effect of Medicaid expansion on unemployment, this may be
a good assumption. On the other hand, if Medicaid expansion can change the demographic and
economic composition of its counties, then differential changes in these variables may actually be a
consequence of the expansion itself.12 If so, then differential post-treatment changes in them would
not necessarily indicate a parallel trends violation; they could partially reflect a causal effect. As
with the plausibility of Assumption PT itself, whether something is a covariate or a mechanism is
not a data question per se. It requires context-specific knowledge about how the treatment works.

4.2 DiD with covariates: Identification under conditional parallel trends

Having detected covariate imbalance that casts doubt on Assumption PT, how should we proceed
to estimate AT T p2q? Because the imbalance documented in Table 4 suggested that unconditional
parallel trends may not hold, our goal is to develop a DiD identification strategy based on an
assumption that accounts for this imbalance. Working from a conditional parallel trends assump-
tion shows how to construct AT T p2q from 2 ˆ 2 comparisons that are each conditioned on specific
covariate values, thus addressing the imbalance problem.
Let Xi be a vector of observed determinants of changes in Yi,t p0q. Here, we purposefully
omit the time subscript on Xi because the covariates in this section can be time-invariant, such as

12
In fact, comparing mean covariate changes in expansion and non-expansion is the same as using Xi,t as the
outcome in a 2 ˆ 2 DiD estimator.

17
fixed variations or baseline values (Xi,t“1 ), or time-varying in the sense of including values from the
second period, Xi,t“2 . The empirical content of a “new” identification assumption that incorporates
Xi , henceforth conditional parallel trends (CPT) assumption, is formalized as follows.

Assumption CPT (2ˆ2 Conditional Parallel Trends). The (weighted) average change of Yi,t“2 p0q
from Yi,t“1 p0q is the same between treated and comparison units that share the same covariate
values,

Eω rYi,t“2 p0q ´ Yi,t“1 p0q|Xi , Di “ 1s “ Eω rYi,t“2 p0q ´ Yi,t“1 p0q|Xi , Di “ 0s. (4.1)

Assumption CPT has the same structure as Assumption PT but states that PT holds within
each covariate-specific stratum rather than across the whole population. With respect to the
baseline covariates in Table 4, this amounts to assuming parallel trends in Yi,t p0q between expansion
and non-expansion counties that in 2013 had the same female, white, and Hispanic shares, as well
as the same unemployment and poverty rates and median income. This does not restrict how Yi,t p0q
changes in different covariate strata, nor is not necessary to estimate these trends. Assumption
CPT only requires that covariate-specific trends are common between expansion and non-expansion
counties.
For both expectations in Assumption CPT to be well-defined for all values of Xi , there must be
both untreated and treated units in the population at each covariate value. If, for some covariate
values, there are only treated units, for example, then the right-hand side of (4.1) is undefined. A
formal statement of this assumption, called “common support” or “strong overlap,” is as follows.13

Assumption SO (Strong overlap). The conditional (weighted) probability of belonging to the

treatment group, given observed covariates Xi , which are determinants of untreated potential
outcome growth, is uniformly bounded away from zero and one. That is, for some ϵ ą 0, ϵ ă
Pω rDi “ 1|Xi s ă 1 ´ ϵ.

Assumptions CPT and SO allow us to identify the AT T p2q:

AT T p2q “ Eω rYi,t“2 p1q|Di “ 1s ´ Eω rYi,t“2 p0q|Di “ 1s

” ˇ ı
“ Eω rYi,t“2 |Di “ 1s ´ Eω Eω rYi,t“2 p0q|Xi , Di “ 1sˇDi “ 1
ˇ
ı ” ˇ ı
“ Eω rYi,t“2 |Di “ 1 ´ Eω Eω rYi,t“1 p0q|Xi , Di “ 1s ` Eω r∆Yi,t“2 p0q|Xi , Di “ 0sˇDi “ 1
ˇ
ı ” ˇ ı
“ Eω r∆Yi,t“2 |Di “ 1 ´ Eω Eω r∆Yi,t“2 |Xi , Di “ 0sˇDi “ 1 . (4.2)
ˇ

The first line restates the definition of AT T p2q, and the second line uses the law of iterated
expectations to write the counterfactual mean for the whole treatment group as an average of
counterfactual means conditional on Xi . These quantities are exactly the ones that appear in the
13
AT T p2q is still identified under conditional parallel trends if some values of Xi have only untreated observations.
We require Pω rDi “ 1|Xi s to be bounded away from zero and one to avoid irregular inference procedures; see Khan
and Tamer (2010) for additional details.

18
conditional parallel trends assumption (and the overlap condition). Therefore, as in section 3, the
third line uses Assumption CPT to rewrite the conditional counterfactuals in terms of observable
population quantities under no anticipation (Assumption NA). Equation (4.2) then uses the law
of iterated expectations again to group terms and express AT T p2q in terms of observed variables
pYi,t“2 , Yi,t“1 , Gi , Xi q; that is, it establishes that AT T p2q is nonparametrically identified under our
assumptions. This expression has a clear intuition: the AT T p2q is equal to the path of outcomes
experienced by the treated group (the term on the left) minus the average path of outcomes in the
comparison group for each value of the covariates, averaged over the treated group’s distribution
of covariates (the term on the right).

4.3 DiD estimation with covariates: TWFE

Moving from the population identification result in equation (4.2) to sample analogs is a challenge
unless the covariates are discrete and the conditional expectations themselves are easily calcula-
ble. This difficulty does not arise in an unconditonal DiD. With continuous covariates, or many
discrete ones, it may not be feasible to construct Eω r∆Yi,t“2 |Xi , Di “ 0s. Conditional DiD estima-
tion, therefore, uses additional econometric techniques to bridge this gap. We begin, however, by
discussing how regression DiD estimators that include covariates relate to the assumptions used
for identification in (4.2).
Because the TWFE specification in (3.7) recovers the AT T p2q in 2 ˆ 2 DiD setups without
covariates, it is natural to extend this logic to regressions with covariates. Indeed, this is by far the
most popular approach adopted by practitioners, arguably because it is both easy and familiar. A
typical regression specification is
1
Yi,t “ θt ` ηi ` βtreat Di,t ` Xi,t βcovs ` ei,t . (4.3)

where the unit and time fixed effects, treatment status, and covariates have already been defined,
ei,t is an error term, and βtreat is interpreted as the parameter of interest. A related specification
explicitly controls for baseline covariates by replacing Xi,t with interactions of the pre-treatment
covariates Xi,t“1 and a post-treatment dummy,

Yi,t “ θt ` ηi ` βtreat,2 Di,t ` p1tt “ 2uXi,t“1 qβcovs,2 ` ei,t , (4.4)

In Table 5, we report the OLS and weighted least squares estimates of the unconditional 2 ˆ 2
DiD estimate, βcovs from (4.4), βcovs,2 from (4.3), and their cluster-robust standard errors, using
the covariates from Table 4.

19
 Table 5: Regression 2 ˆ 2 DiD with Covariates

Unweighted Weighted
No Covs Xi,t“2013 Xi,t No Covs Xi,t“2013 Xi,t
(1) (2) (3) (4) (5) (6)
˚
Medicaid Expansion 0.12 -2.35 -0.49 -2.56 -2.56 -1.37
(3.75) (4.29) (3.83) (1.49) (1.78) (1.62)

This table reports the regression 2 ˆ 2 DiD estimate comparing 978 counties that expanded
Medicaid in 2014 with 1,222 counties that did not expand Medicaid by 2019, adjusting for
covariates (percent female, percent white, percent hispanic, the unemployment rate, the poverty
rate, and median household income). Columns 1-3 report unweighted regression results, while
columns 4-6 weight by county population aged 20-64 in 2013. Columns 1 and 4 report results
for expansion states without covariates, columns 2 and 5 adjust for the baseline levels of the
covariates in 2013, and columns 3 and 6 control for the time-varying covariate values in 2014
and 2013. Standard errors (in parentheses) are clustered at the county level.

Although only one covariate-adjusted estimate in Table 5 is (marginally) statistically significant,

the point estimates differ noticeably. In the unweighted case, adjusting for the 2013 levels of the
covariates decreases the estimated effect of Medicaid expansion on short-run mortality rates from
a point estimate of roughly 0.12 to -2.35. However, if we include their time-varying values instead,
we estimate an effect of -0.49, a large difference. We find a similar result when using weighted
regressions; while the coefficient remains constant (-2.56) when using 2013 values of the covariates,
it attenuates to -1.37 if we use (4.3).
The jump from the conditional DiD identification result in (4.2) to the TWFE estimators in
(4.3) and (4.4) skips a crucial question about βtreat or βtreat,2 : do they equal the target parameter
AT T p2q under the conditional parallel trends assumption? It turns out that the close relationship
between regression DiD, AT T p2q, and parallel trends in a design without covariates does not
hold with covariates. The issues come from exactly what kinds of covariates are effectively being
“controlled for” in these specifications and how the regression estimator combines outcome trends
for covariate sub-groups.
Note that in our two-period setup, (4.3) and (4.4) are respectively equivalent to (with some
abuse of notation),
1
∆Yi,t“2 “ α ` βtreat Di ` ∆Xi,t“2 βcovs ` ∆ei,t“2 ,
1
∆Yi,t“2 “ α ` βtreat,2 Di ` Xi,t“1 βcovs,2 ` ∆ei,t“2 .

The first thing that is clear from these representations is that because time-invariant variables drop
out of equation (4.3), a TWFE specification can account for differential trends related to baseline
covariate levels only if they enter as interactions with the post-treatment dummy as in equation
(4.4). The exact regression specification, therefore, determines the implied conditional paral-
lel trends assumption. Controlling for annual poverty rates really means controlling for poverty

20
changes, and areas that are poor are not the same as areas that are becoming poor.
Another limitation evident in (4.3) relates to “bad controls.” Whenever Xi,t“2 is affected by
the treatment, then conditioning on it (in any way) can bias estimates of the AT T p2q. If Medicaid
expansion lowered poverty rates, for example, then including 2014 poverty rates or the 2013-2014
change in poverty rates as a covariate is problematic. This echoes our discussion about testing
balance in ∆Xi,t in the sense that time-varying covariates must be unaffected by treatment in
order to interpret imbalance in their trends as a source of bias, and to be able to control for them
to address that bias. See Caetano, Callaway, Payne and Rodrigues (2022) for a discussion.
Suppose we have decided on which variables to include in a conditional parallel trends assump-
tion and whether to measure them in levels or changes. If Assumptions CPT and SO hold with
respect to this set of covariates, does βtreat recover the AT T p2q? In the DiD context, Caetano and
Callaway (2024) tackle exactly this question. They show that βtreat equals a weighted average of
“
conditional average treatment effects, defined as AT Txk p2q ” Eω Yi,t“2 p1q ´ Yi,t“2 p0q|Di “ 1, Xi “
‰
xk , with weights that may not be convex, plus three bias terms reflecting misspecification either
in the set of control variables or the fact that they are included linearly. These conclusions relate
to recent findings about the properties of regression estimators in the presence of heterogeneity in
other contexts, including instrumental variables (Mogstad and Torgovitsky, 2024), cross-sectional
designs (Angrist, 1998; Aronow and Samii, 2015; Sloczynski, 2022; Goldsmith-Pinkham, Hull and
Kolesár, 2024), and panel data (Goodman-Bacon, 2021; de Chaisemartin and D’Haultfoeuille, 2020;
Sun and Abraham, 2021; Poirier and Sloczynski, 2024).
The weighting results suggest that even when the covariates are correctly selected, measured,
and added with the correct functional form, βtreat could be negative even when AT TXi p2q is positive
for all values of covariates. Short of this extreme sign-reversal case, βtreat ’s weighting scheme gen-
erally does not yield the AT T p2q target parameter and instead puts too much weight on AT TXi p2q
for Xi ’s that are relatively uncommon among the treated group relative to the untreated group
and puts too little weight on AT TXi p2q for Xi ’s that are relatively common among the treated
group relative to the untreated group (Sloczynski, 2022; Caetano and Callaway, 2024).
Taken together, these results imply that βtreat identifies AT T p2q under the additional assump-
tion that treatment effects across covariate strata are constant. To see why, write the conditional
ATT in period two given ∆Xi,t“2 as

AT T∆Xi,t“2 p2q “ Eω rYi,t“2 p1q ´ Yi,t“2 p0q|Di “ 1, ∆Xi,t“2 s,

and note that, under Assumptions CPT and SO, it is identified by

AT T∆Xi,t“2 p2q “ Eω r∆Yi,t“2 |Di “ 1, ∆Xi,t“2 s ´ Eω r∆Yi,t“2 |Di “ 0, ∆Xi,t“2 s.

If we take (4.3) to be a correctly specified regression, then

` 1
˘ ` 1
˘
AT T∆Xi,t“2 p2q “ βtreat ` ∆Xi,t“2 βcovs ´ ∆Xi,t“2 βcovs “ βtreat .

In other words, (4.3) implicitly rules out that treatment effects can vary across covariate-strata,

21
which makes the weighting issues identified by Caetano and Callaway (2024) irrelevant to the
interpretation of βtreat . Research on the Medicaid expansion using data on mortality rates by
income, however, shows clear evidence of heterogeneous effects (Miller et al., 2021; Wyse and
Meyer, 2024).
One way to avoid these limitations would be to make (4.4) (or (4.3)) more flexible by includ-
ing interactions of the covariates with treatment group, time, and treatment-group-by-time. An
alternative possibility is to adopt a “forward-engineering” perspective (Mogstad and Torgovitsky,
2024) and derive an estimator for AT T p2q that directly leverages Assumptions NA, CPT and SO.
In some situations, the forward-engineering approach will also use regressions. Still, the target
parameter and the identifying assumptions will guide the specification we should use (and not the
other way around).

4.4 DiD estimators with covariates that target the ATT(2)

Fortunately, TWFE is not the only way to bring covariates into DiD estimation. We now discuss
alternative ways to use covariates in a DiD analysis that start from the identification result in
(4.2):
” ˇ ı
AT T p2q “ Eω r∆Yi,t“2 |Di “ 1s ´ Eω Eω r∆Yi,t“2 |Xi , Di “ 0sˇDi “ 1 .
ˇ

This equation provides an intuitive recipe for estimating AT T p2q. The first term in AT T p2q is the
same as in an unconditional 2 ˆ 2 design and can be replaced by its sample analog as in equation
(3.6): pY ω,D“1,t“2 ´ Y ω,D“1,t“1 q.
One way to obtain the second term is to first estimate the inner conditional expectation,
Eω r∆Yi,t“2 |Xi , Di “ 0s. This object is just an (unknown) function that relates average outcome
trends for untreated units to their covariates. The most common way to proceed, especially
in cases where Xi contains many variables or continuous ones, is to specify a working model,
µω,∆,D“0 pXi q, with parameters that are simple to estimate. A natural and empirically friendly
choice is a linear model, µω,∆,D“0 pXi q “ Xi1 βD“0 , whose parameters come from a (weighted)
regression of ∆Yi,t“2 on Xi in the sample of untreated units. The results of this regression describe
untreated outcome trends as a function of Xi . The fitted model then generates predicted values,
pω,∆,D“0 pXi q “ Xi1 βpD“0 , for
µ ” all units in the sample,
ˇ including
ı treated units. With these fitted
values we can estimate Eω Eω r∆Yi,t“2 |Xi , Di “ 0sˇDi “ 1 using the plug-in principle; that is,
ˇ

by replacing Eω r∆Yi,t“2 |Xi , Di “ 0s with its fitted value

řn
µ
pω,∆,D“0 pXi q, and replacing population
i“1 Diřωi µ
pω,∆,D“0 pXi q
(weighted) expectations with their sample analogs: n .
i“1 Di ωi
Putting the pieces together gives the following estimator for AT T p2q:
řn ` ˘
i“1 Di ωi ∆Yi,t“2 ´ µ pω,∆,D“0 pXi q
AT T ra p2q “
z řn . (4.5)
i“1 Di ωi

This strategy is often referred to as the regression-adjustment (RA) or outcome regression approach

22
to DiD; see, for example, Heckman, Ichimura and Todd (1997a).
We apply this strategy in our application using only the baseline covariates from Table 4, and
report the parameters βpD“0 of our working model in columns (1) and (3) of Table 6. In prac-
tice, mortality changes in non-expansion states are only weakly related to our baseline covariates.
Multiplying the weighted coefficients in Table 6 times the weighted treatment group 2013 means
in Table 4 gives a predicted change in untreated mortality rates for the average treated county of
7.2 deaths per 100,000, the second term in equation (4.5), compared to the observed (weighted)
change among untreated counties of 6.3 deaths. The observed weighted trend in mortality for
expansion counties from Table 2 is 3.7 deaths. Together, these results imply that this approach,
based only on assumptions CPT, SO, NA, the choice of baseline covariates, and a linear model for
Eω r∆Yi,t“2 |Xi , Di “ 0s, yields an estimated AT T p2014q of -3.5. This matches the formal RA DiD
estimate we report in column 4 of Table 7 (labeled as “Regression”). Column 1, however, gives an
unweighted estimate from the same procedure of -1.62.

Table 6: Outcome Regression and Propensity Score Models

Unweighted Weighted
Regression Propensity Score Regression Propensity Score
(1) (2) (3) (4)
OLS Logit OLS Logit
Constant -20.91 -10.00˚˚˚ -4.62 -8.17˚˚˚
(69.85) (1.35) (44.98) (0.01)
% Female 0.04 -0.04˚˚ -0.09 -0.19˚˚˚
(0.82) (0.02) (0.68) (0.00)
% White 0.15 0.06˚˚˚ 0.20˚ 0.04˚˚˚
(0.22) (0.00) (0.11) (0.00)
% Hispanic -0.08 -0.02˚˚˚ -0.08 -0.02˚˚˚
(0.20) (0.00) (0.08) (0.00)
Unemployment Rate 1.14 0.32˚˚˚ 0.88 0.68˚˚˚
(1.56) (0.03) (0.99) (0.00)
Poverty Rate 0.21 0.03˚ -0.13 0.11˚˚˚
(0.98) (0.02) (0.53) (0.00)
Median Income 0.09 0.08˚˚˚ -0.05 0.15˚˚˚
(0.52) (0.01) (0.24) (0.00)

This table reports the outcome regression propensity score models that enter into the estimator
from Sant’Anna and Zhao (2020) and Callaway and Sant’Anna (2021). The first two columns
report the results for unweighted regressions and the second two report results from weighted
regression models. The regression model predicts changes in the outcome variable (mortality
rates) as a function of 2013 covariate values for the 1,222 counties that do not expand Medicaid
in 2014. The propensity score model uses data on all 2,200 counties for 2013 and estimates a
logit model of an expansion indicator variable on the 2013 covariate levels. Standard errors (in
parentheses) are clustered at the county level.

To compute standard errors, we need to account for the fact that (4.5) is a two-step estimation
procedure and take into account the uncertainty associated with estimating the working model

23
µω,∆,D“0 pXi q. This is standard, though, and most statistical software automates this process; see,
for example, Sant’Anna and Zhao (2020) and Callaway and Sant’Anna (2021).

Table 7: DiD Estimates with Covariates

Unweighted Weighted
Regression IPW Doubly Robust Regression IPW Doubly Robust
Medicaid Expansion -1.62 -0.86 -1.23 -3.46 -3.84 -3.76
(4.73) (4.76) (4.61) (2.29) (3.22) (3.59)

This table reports the 2 ˆ 2 DiD estimate comparing 978 counties in states that expand Medicaid in 2014 to 1,222
counties in states that did not expand Medicaid by 2019, adjusting for 2013 covariate values using the methodologies
discussed in Sant’Anna and Zhao (2020) and Callaway and Sant’Anna (2021). The first column reports results us-
ing regression adjustment, the second column uses inverse probability weighting based on a propensity score model
using the included covariates, and the third column uses the doubly robust combination of the two approaches.
Standard errors (in parentheses) are clustered at the county level.

A linear working model for Eω r∆Yi,t“2 |Xi , Di “ 0s is a familiar choice, but not the only one.
More flexible, or even fully nonparametric, working models are possible, and the procedure is
the same: estimate the model on untreated units, get its fitted values for the covariate values
of the treated units, and then estimate the AT T p2q using (4.5). An important consideration
when choosing the working models is sample size. Large samples permit more flexible estimators
that do not sacrifice too much precision. In smaller samples, a parametric linear model may be
more appealing. Ultimately, the reliability of DiD RA estimators for AT T p2q depends on how
well µ
pω,∆,D“0 pXi q approximates Eω r∆Yi,t“2 |Xi , Di “ 0s. If the working model is misspecified—for
example by omitting relevant nonlinear terms—the resulting DiD RA estimator will be biased.
Table 6 showed that our covariates did not strongly predict untreated mortality trends and thus
do little to change our potential biased unconditional DiD estimates. However, AT T p2q can be
estimated conditional on covariates in a different way without needing to specify which variables
determine outcome trends. Instead, one can improve the comparability of the comparison group
directly by selecting a model for the conditional probability of being treated and applying an
inverse probability weighted (IPW) DiD procedure (Abadie, 2005). The logic of IPW builds on
the balance checks we conducted in Table 4: if imbalance in covariates is the source of parallel
trends violations, then adjusting the comparison group to be balanced on covariates can address
that bias. The adjustment takes the form of re-weighting the observed changes in adult mortality
rates for non-expansion counties to ensure that the expansion and non-expansion counties are
similar on covariates, thus addressing the compositional source of bias.
To implement the IPW DiD procedure and construct these “balancing weights,” we need to
model pω pXi q “ Pω pDi “ 1|Xi q, the (weighted) conditional probability of belonging to the treat-
ment group, known as the “propensity score.” IPW weights are a function of pω pXi q, and for
estimating AT T p2q, they take a form that forces the underlying weighted distribution of covariates

24
for comparison units to match the distribution for treatment units (Rosenbaum and Rubin, 1983).
Intuitively, these weights are formed so that if we find some units that were likely to be observed
in the treatment groups (based on their covariate values) but ended up in the comparison group,
we give these untreated observations “extra” weight.
Formally, we can show that, under Assumptions CPT and SO, when panel data are available,
”´ ¯ ı
AT T p2q “ E wω,D“1 pDi q ´ wω,D“0 pDi , Xi q ∆Yi,t“2 , (4.6)

where
O „ ȷ
M ωp1 ´ Dqpω pXq ωp1 ´ Dqpω pXq
wω,D“1 pDq “ Dω ErωDs, and wω,D“0 pD, Xq “ E ; (4.7)
1 ´ pω pXq 1 ´ pω pXq
see, for example, Abadie (2005) and Sant’Anna and Zhao (2020). The structure of the weights in
equation (4.7) and the way they enter equation (4.6) highlight several intuitive features of how the
IPW estimator works. First, wω,D“1 pDq is only non-zero for treated units, and wω,D“0 pD, Xq is only
non-zero for untreated units, which means the estimator subtracts a particular mean of outcome
trends for untreated units from a particular mean of outcome trends for treated units. Second,
the wω,D“1 pDq weights do not involve Xi and simply lead simply to a (ω-weighted) mean for the
treatment group. Third, the wω,D“0 pD, Xq weights are functions of Xi that, as (4.7) shows, give
increasingly more weight to untreated units with high propensity scores. This specific weighting
function is what allows IPW weights to build a comparison group whose covariate distribution
matches the treatment group. Fourth, equation (4.7) clarifies that two types of weights both enter
an IPW analysis. We already discussed how the ω weights shape the parameter of interest, parallel
trends assumptions, and estimation. These simply multiply the IPW weights which act to address
imbalance (and their product is rescaled to integrate to one within group). Finally, the appeal of
(4.6) is that if we have a better sense of how units sort into treatment than of the factors that
shape outcome trends, we may be more comfortable modeling pω pXi q than Eω r∆Yi,t“2 |Xi , Di “ 0s.
To leverage the characterization of the AT T p2q in (4.6) for estimation and inference purposes,
we need a working model for the true propensity score pω pXi q. When X’s are all discrete and
low dimensional, this is simple and does not involve functional form restrictions: create covariate-
specific strata and then, within each strata, compute the proportion of treated units, and call these
estimates π ppxk q, xk being a strata-indicator. When X’s have continuous components, or when there
are too many strata relative to the available sample size, one can adopt a flexible working model,
πω pXi q, for the propensity score. A common choice for πω pXi q is a (weighted) logistic model whose
parameters can be estimated using maximum likelihood—or an alternative estimation procedure
such as inverse probability tilting (Graham, Pinto and Egel, 2012). We follow this strategy in our
Medicaid application and report in columns (2) and (4) of Table 6 the unweighted and weighted
maximum likelihood logit coefficients from our propensity score model. Our covariates appear to
explain expansion decisions better than untreated outcome trends, suggesting that an estimation
approach based on propensity scores may change our AT T p2014q estimate more than the RA

25
approach did. We then use these logit coefficients to get the fitted values for all observations,
π
pω pXi q, that will serve as our estimates of pω pXi q.
From these fitted values, we can estimate AT T p2q by
n
1 ÿ´ ¯
AT T ipw p2q “
z w
pω,D“1 pDi q ´ w
pω,D“0 pDi , Xi q ∆Yi,t“2 , (4.8)
n i“1
where
O
n
M1 ÿ n
ωp1 ´ Dqpπω pXq 1 ÿ ωi p1 ´ Di qp
πω pXi q
w
pω,D“1 pDq “ Dω ωi D i , w
pω,D“0 pD, Xq “ . (4.9)
n i“1 1´π pω pXq n i“1 1´π pω pXi q

We report the AT T p2014q estimates and their standard errors using this IPW DiD procedure in
Table 7 (labeled as “IPW”) using both unweighted and population-weighted procedures, where we
use a logistic regression that is linear in covariates as our propensity score working model. We use
the delta method to account for the estimation uncertainty inherited in this two-step estimation
procedure when computing standard errors. The IPW DiD estimates are generally similar to the
RA DiD estimates in that we obtain negative estimates that are larger in magnitude when using
population weights. However, despite neither being statistically significant, the unweighted IPW
estimate is less than half the size of the RA estimate. This is consistent with the broad conclusion
from Table 6 that our covariates explain Medicaid expansion better than mortality trends.
IPW estimators for AT T p2q tend to be noisy when fitted propensity score estimates are too
close to 1 among untreated units, a condition related to the Assumption SO. The reason for this is
simple: when π pω pXi q is close to one among units with Di “ 0, the (estimated) inverse probability
weights wpω,D“0 pD, Xq become more volatile, as one is essentially dividing by zero. A good practice
when using IPW estimators is to check the plausibility of strong overlap using estimated propensity
scores. Figure 1 plots the propensity scores that come from the logit models in Table 6 for the two
groups of counties. Few untreated units have very high estimated propensity scores, so extreme
weighting is not a significant concern.14 In addition, propensity scores of non-expansion counties
seem to lie within the support of the expansion counties’ propensity scores, supporting strong
overlap. Trimming high- or low-propensity score observations from the sample may be warranted
when overlap is weak; for a discussion, see, for example, Crump, Hotz, Imbens and Mitnik (2009),
Sasaki and Ura (2022), and Ma, Sant’Anna, Sasaki and Ura (2023).
The reliability of these two approaches to estimating conditional DiD designs centers on se-
lecting good models for two different functions: the conditional expectation of untreated outcome
changes and the true, unknown propensity score pω pXi q. So which approach should one pick? In
practice, there are major advantages to combining both in a way that leads to estimators for the
AT T p2q that are more robust against model misspecification (Sant’Anna and Zhao, 2020). This
is the so-called doubly robust (DR) approach, sometimes referred to as the augmented inverse

14
By default, the did and DRDID R packages from Callaway and Sant’Anna (2021) and Sant’Anna and Zhao
(2018) trim untreated units with propensity scores in excess of 0.995.

26
 Figure 1: Distribution of Propensity Scores

Unweighted Weighted

Density 2

0
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Propensity Score

Expansion Counties Non−Expansion Counties

Notes: This figure shows the distribution of propensity scores using both the weighted and unweighted logit propensity score estimates
in our 2 ˆ 2 Medicaid example.

probability weighting approach; see, for instance, Sant’Anna and Zhao (2020) and Chang (2020).
The key idea of the DR DiD approach is to express the AT T p2q in terms of both ppXi q and
Eω r∆Yi,t“2 |Xi , Di “ 0s in a way that it gives provides some “protection” in case the working models
for these functions, πppXi q and µ
p∆,G“0 pXi q, are wrong. The resulting DR estimator for the AT T p2q
is consistent when either of these nuisance working models is correctly specified. If one is exactly
right, it does not matter if the other is wrong. Furthermore, if both working models are only
slightly wrong, their errors will multiply, and DR will perform (asymptotically) better than either
one alone.15 Following the steps in Sant’Anna and Zhao (2020), we can express the AT T p2q as
”´ ¯´ ¯ı
AT T p2q “ E wω,D“1 pDi q ´ wω,D“0 pDi , Xi q ∆Yi,t“2 ´ Eω r∆Yi,t“2 |Xi , Di “ 0s , (4.10)

with the weights as defined in (4.7). Notice that when we omit the Eω r∆Yi,t“2 |Xi , Di “ 0s term
from (4.10), we are back to the IPW estimand (4.6). When we omit the wω,D“0 pDi , Xi q term from
(4.10), we are back to the regression adjustment estimand (4.2).
Constructing a DR DiD estimator for the AT T p2q based on (4.6) is straightforward. Choose
a flexible working model for the propensity score and compute its fitted values, πpω pXi q. Then, a
flexible working model for the outcome evolution of untreated units and compute its fitted values
for all treated and untreated units, µ
pω,∆,D“0 pXi q, and then use the plug-in principle to estimate

15
We usually arrive at such estimands by deriving the efficient influence function; see, for example, Sant’Anna
and Zhao (2020) for a discussion in DiD setups. See also Seaman and Vansteelandt (2018) for an overview. When
we adopt nonparametric or machine-learning-based estimators for the nuisance functions, we get a different type of
double robustness, “rate double robustness,” where we can trade-off precision between the different working models.
For more details, see, Kennedy, Ma, McHugh and Small (2017) and Smucler, Rotnitzky and Robins (2019).

27
the AT T p2q by
n
1 ÿ´ ¯´ ¯
AT T dr p2q “
z w
pω,D“1 pDi q ´ w
pω,D“0 pDi , Xi q ∆Yi,t“2 ´ µ
pω,∆,D“0 pXi q , (4.11)
n i“1
with the estimated weights as defined in (4.9).
We report the AT T p2014q estimates and standard errors using the DR DiD procedure with
a linear in covariates working model for µω,∆,D“0 pXi q and (weighted) logistic regression work-
ing model that is also linear in covariates for pω pXi q in Table 7 (labeled as “Doubly Robust”).
Population-weighted results are fairly similar across our methods, but the unweighted DR DiD
estimates are about halfway between the RA and IPW results. However, an advantage of the DR
DiD procedures is that model misspecifications are arguably less of a concern.
Overall, in this section, we discussed different “forward-engineered” estimators for the AT T p2q
that respect conditional parallel trends, correctly incorporate covariates, and do not restrict treat-
ment effect heterogeneity. Between RA, IPW, and DR DiD estimators, we recommend practitioners
favor the doubly robust one compared with the RA and IPW appraoches on their own, as this pro-
cedure adds additional “protection.” When strong overlap fails, though, RA DiD estimators can
still work because they extrapolate the outcome model to obtain predicted counterfactual outcome
changes even for treated units with covariate values not observed in the comparison group. The
credibility of this extrapolation, however, rests on the accuracy of the working model outside the
support of Xi in the untreated group. If this extrapolation is not reliable, one can make the DR
approach more robust against weak overlap by trimming “extreme” propensity scores and perform-
ing a bias correction to ensure that the target parameter (the AT T p2) in our case) remains the
same. See Ma et al. (2023) for details.

4.5 Heterogeneity analysis

A different motivation for using covariates is to estimate heterogeneous treatment effects by the
values of Xit . The identification result in equation (4.2) can be altered slightly to show that
AT T p2q is simply an aggregation of covariate-specific 2 ˆ 2 DiD estimands:
” ˇ ı
AT T p2q “ Eω Eω r∆Yi,t“2 |Xi , Di “ 1s ´ Eω r∆Yi,t“2 |Xi , Di “ 0sˇDi “ 1 .
ˇ

Thus, Assumptions CPT and SO also imply that we can identify conditional ATT parameters:

AT TXi p2q ” Eω rYi,t“2 p1q ´ Yi,t“2 p0q|Di “ 1, Xi s

“ Eω r∆Yi,t“2 |Xi , Di “ 1s ´ Eω r∆Yi,t“2 |Xi , Di “ 0s.

This not only demonstrates the building block structure of conditional designs, but also connects
the strategies we discussed for estimating AT T p2q itself to the underlying treatment effect hetero-
geneity by covariate values. When this heterogeneity is of interest in its own right it can also be
targeted, identified, and estimated (at least under some additional conditions).

28
 When all covariates are discrete, estimating AT TXi p2q parameters is fairly straightforward.
One can form saturated partitions xk , k “ 1, . . . , K, and then subset the data to contain only
information from the units from the specified partition k. At this stage, the analysis is analogous
to the unconditional DiD setup (except that you need to repeat this step from all partitions of
interest), as each AT Txk p2q is identified by a (conditional on a discrete-variable) comparison of
means’ that is,
“ ‰ “ ‰
AT Txk p2q “ Eω ∆Yi,t“2 |Di “ 1, Xi “ xk ´ Eω ∆Yi,t“2 |Di “ 0, Xi “ xk .

One can estimate these AT Txk p2q’s using their sample analogs or using two-way fixed effects
regression specifications analogous to (3.7). Inference is also standard, provided that each partition
is sufficiently large. This type of exercise is commonly used to conduct heterogeneity analysis. For
example, suppose one wants to see if the effect of Medicaid expansion on adult mortality rate
varies across US census regions. Then, one would partition the data into US regions and run one
(unconditional) DiD analysis for each region, provided we have treated and untreated units in each
region.
When some covariates are continuous or there are so many partitions that each one contains
few observations, one can still identify the AT TXi p2q’s, though they are hard to estimate unless
auxiliary assumptions hold. In such cases, it is customary to identify and estimate a more ag-
gregated conditional ATT parameter than AT TXi p2q. For instance, one may want to assess if the
effect of Medicaid expansion on adult mortality rates is higher (or lower) in counties with an un-
employment rate above the median than it is in those below. Similar partitions could be made for
any other covariates. To formalize this notion of “partition specific” ATT, let PARTpXi q be some
user-specified partition of the covariate space such that PARTpXi q P t1, 2, . . . , Ku, and define
“
AT Tk p2q “ Eω Yi,t“2 p1q ´ Yi,t“2 p0q|Di “ 1, P ART pXi q “ ks.

Under Assumptions CPT and SO, it follows that

“ ‰
AT Tk p2q “ Eω ∆Yi,t“2 |Di “ 1, P ART pXi q “ k
” “ ˇ ‰ˇˇ ı
´ Eω Eω ∆Yi,t“2 Di “ 0, Xi , P ART pXi q “ k ˇDi “ 1, P ART pXi q “ k ,
ˇ

which implies that

K
ÿ
AT T p2q “ Pω pP ART pXi q “ k|Di “ 1qAT Tk p2q.
k“1

Thus, for heterogeneity analysis with covariates, one can partition the data with a user-specified
partition map PARTpXi q and then, within each of these partitions, use arguments like the ones
we used to establish (4.2) to guarantee that each partition-specific ATT is identified (which is
precisely what we did to get AT Tk p2q above). Regarding estimation and inference, one can use
regression adjustment, inverse probability weighting, or doubly robust estimators as in Section 4.4;
the difference is that these are implemented “locally” on each partition.

29
 Obtaining the overall AT T p2q is just a matter of aggregating these partition-specific ATTs
using weights equal to the relative partition size among treated units. All these parameters have
clear causal interpretations, can be used to answer different policy questions, and are identified
under the same identification assumptions already discussed. Other types of heterogeneity analysis
are also possible and even attractive. For instance, Abadie (2005) discusses how one can highlight
how the ATT varies across a subset of the covariates required for conditional parallel trends to
hold. For example, this would entail checking if the average effect of Medicaid expansion among
expansion counties varies with the 2013 poverty rate and/or median income. One can also get
the best linear approximation of these conditional ATT curves, which involves estimating fewer
parameters. These heterogeneity analyses are generally more granular than the partition-based
ones discussed above and do not require discretizing the data. They complement each other well.
We close this section with a remark on the choice of partition and the type of heterogeneity
analysis to conduct. Heterogeneity analysis has the potential to offer policymakers and researchers
novel insights about the treatment of interest and its mechanisms, opening the door for more
informed policy recommendations and targeted expansions. But how should one define the sub-
groups in which to estimate heterogeneous effects? If researchers knew the relevant partition that
policymakers care about, they could aim to estimate these particular partition-specific ATTs. But
this is rarely the case. Taking no stand about heterogeneity implies reporting unit-level effects,
which requires incredibly strong assumptions and could also lead to noisy estimates. On the other
hand, taking a too-coarse partition may mask important types of treatment effect heterogeneity.
So, it seems that a balance between these extremes is important for a policy-relevant heterogeneity
analysis. Estimating conditional ATTs across one or two covariate dimensions is useful, but it may
mix some other interactive effects. Another potential avenue is to go beyond averages and adapt
the sorted-effects procedure proposed by Chernozhukov, Fernández-Val and Luo (2018) to DiD
designs. It would also be interesting and practically relevant to extend the heterogeneity tools for
experimental data discussed in Chernozhukov, Demirer, Duflo and Fernández-Val (2023) to DiD
setups. In our view, this is an area in which applied econometrics practice would benefit from
more thorough methodological guidance.

5 DiD designs with multiple time periods

The previous sections focused on fundamental identification issues and estimation approaches for
2 ˆ 2 building blocks, but generally did not build them into anything because they targeted the
only feasible AT T ptq, AT T p2q. DiD designs with multiple periods are notably different. They can
target AT T s in each period after treatment to trace out dynamics, and they can produce cross-
group outcome trends from before treatment starts to evaluate the plausibility of parallel trends.
Multiple periods also admit the possibility of more complex treatment variation. We will focus on
staggered-timing designs in which treatment “turns on” at different times for different units and

30
stays on, as the Medicaid expansion did. It is possible, however, to use DiD methods to study
treatments that “turn off” or occur more than once, though some modifications are warranted; see,
for example, de Chaisemartin and D’Haultfoeuille (2020, 2023a).
We need to expand the notation to define the relevant concepts with multiple periods. When
treatment can happen only at one point in time (so we still have only two treatment groups with
Di equal to one or zero), the only changes we need to make are to acknowledge that time runs
from t “ 1, 2, . . . , T , and that the map between potential outcomes and observed outcomes is

Yi,t “ Di Yi,t p1q ` p1 ´ Di qYi,t p0q.

By Assumption NA, we have that Yi,t “ Yi,t p0q for all pre-treatment periods, and that, in post-
treatment periods, we observe Yi,t p0q if the group remains untreated by t “ T , and Yi,t p1q for
groups treated at the unique treatment date, t “ g.

5.1 Simple event studies p2 ˆ Tq

The term “event study” refers to estimating and reporting effects across a range of time periods
before and after treatment. A design with one treatment timing group and multiple time periods
(2 ˆ T ) is the simplest case in which to discuss event studies. We thus expand our analysis of the
2014 Medicaid expansion group to include data from 2009 to 2019. We report population-weighted
results for brevity. Figure 2 plots the time series of the weighted mortality rates for the 2014 and
post-2019 expansion counties. It is the analog of Table 2 in the sense that it presents the raw data
elements necessary to construct event study estimates. The treatment year 2014 naturally divides
the x-axis into two windows: post-treatment (2014-2019) and pre-treatment (2009-2013). An event
study constructs DiD-type estimates in both windows, but they have different interpretations.

5.1.1 Event study estimates in the post-treatment periods

The target parameters in a simple event study are the AT T ptq’s with the same definition as in
(3.5); there are just more of them to identify, estimate, and interpret than in a 2 ˆ 2 design. The
AT T ptqs in the post-treatment period, t ě g, reflect treatment effect dynamics. For example,
economic models that view health as a stock imply that AT T ptq may grow if Medicaid stimulates
health investments. Furthermore, people and institutions may take time to adjust their behavior
after Medicaid expands, suggesting a dynamic treatment effect. Event study parameters answer
these kinds of subtle questions.
Identification of each AT T ptq follows from the same arguments outlined in section 3. Note,
however, that an event study analysis requires parallel trends in every post-period, as in the
following assumption.

Assumption PT-ES (2 ˆ T Parallel Trends). The average change of Yi,t p0q from Yi,t“g´1 p0q is

31
 Figure 2: County Mortality Trends by Expansion Decision

400

380

Mortality (20−64) 360

Per 1,000

340

320

2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

Expansion Counties Non−Expansion Counties

Notes: This figure shows county population-weighted average mortality rates for adults ages 20-64 for 978 counties that expanded
Medicaid in 2014 and 1,222 counties that did not expand Medicaid by 2019 from 2009 to 2019.

the same between treated and comparison units for all post-treatment periods t ě g; that is,

Eω rYi,t p0q ´ Yi,t“g´1 p0q|Di “ 1s “ Eω rYi,t p0q ´ Yi,t“g´1 p0q|Di “ 0s @t ě g. (5.1)

Assumption PT-ES suggests that learning about long-run effects requires stronger assumptions
than learning about short-run effects. That is, to identify the average effect of Medicaid expansion
in 2019 among expansion counties, Assumption PT-ES requires parallel trends to hold in every
year from 2014 to 2019. On the other hand, if we are interested in learning only about short-run
effects—say effects up until 2015—we would require it to hold from 2014 and 2015 only.
If Assumption PT-ES holds (as well as no-anticipation), then each AT T ptq is identified by a
DiD comparison between period g ´ 1 and t, as in (3.5), and the AT T ptqs can be estimated by the
familiar comparison of four sample averages:

AT
z T ptq “ pY ω,D“1,t ´ Y ω,D“1,t“g´1 q ´ pY ω,D“0,t ´ Y ω,D“0,t“g´1 q. (5.2)

Figure 3 plots (weighted) event study estimates for the 2014 Medicaid expansion group.
AT
z T ptq’s lie to the right of the vertical dashed line. The estimate for event-time 0 (that is,
t “ g “ 2014) equals the 2 ˆ 2 results from Table 2 (´2.6). The other estimates have the same
DiD form: comparisons of cross-group changes in different post-2014 years (t) but always relative
to 2013 (g ´1). The point estimates do not suggest large mortality effects from Medicaid expansion
among expansion counties.

32
 Figure 3: 2 ˆ T Event Study

Estimate (t ∈ {0, 5}) = −0.70

10 Std. Error = 1.95
Conf. Int = [−4.53, 3.12]

Treatment Effect
Mortality Per 100,000 0

−5

−10
−5 −4 −3 −2 −1 0 1 2 3 4 5
Event Time
Notes: This figure shows the population-weighted event study estimates in the 2 ˆ T case, comparing 978 counties in states that
expanded in 2014 to 1,222 counties in states that had not yet expanded by 2019. It uses the unconditional estimator from Callaway
and Sant’Anna (2021). The outcome variable is the crude mortality rate for adults ages 20-64, and the standard errors are clustered at
the county level. The point estimate is reported by the circles, and both point-wise (black) and simultaneous (red) confidence intervals
are reported with the vertical lines (see section 5.1.3).

5.1.2 Event study estimates in the pre-periods

Multiple time periods also allow for falsification/placebo tests based on DiD-type comparisons
between pre-treatment periods. The no-anticipation assumption implies that all AT T ptq’s before
time g are equal to zero, which means that a 2ˆ2 estimand between periods t “ g ´k and t “ g ´1,
k ą 1, equals a difference in weighted average untreated potential outcome trends (as they are all
from pre-treatment periods):

τ´k “ Eω rYi,t“g´k p0q ´ Yi,t“g´1 p0q|Di “ 1s ´ Eω rYi,t“g´k p0q ´ Yi,t“g´1 p0q|Di “ 0s

“ Eω rYi,t“g´k ´ Yi,t“g´1 |Di “ 1s ´ Eω rYi,t“g´k ´ Yi,t“g´1 |Di “ 0s.

The τ´k terms are usually called “differential trends” or “pre-trends,” and they appear to the left of
the vertical dashed line in Figure 3. The credibility of event study analyses rests in large part on
finding small estimates of τ´k , but Figure 3 illustrates how challenging it can be to draw conclusions
from informally looking at these pre-trends. No individual τ´k is statistically significant, so we
fail to reject the null hypothesis that the pre-trend estimates equal zero (individually or jointly).
This kind of result is often interpreted to mean that parallel trends holds. But the τ´k ’s also tend
to be positive with a mean of about 2.3, which is larger in magnitude than all post-period point
estimates except one. Sometimes, this kind of result is used to argue that parallel trends fails. Do
these results support or refute parallel trends?
Recent methodological work suggests three practical lessons about using pre-trend estimates to
assess parallel trends; examples include Bilinski and Hatfield (2018), Manski and Pepper (2018),
Kahn-Lang and Lang (2020), Roth (2022), Rambachan and Roth (2023), Dette and Schumann
(2024), and Freyaldenhoven, Hansen, Pérez-Pérez and Shapiro (2024). The most fundamental,

33
which will tend to shape language more than practice, is that Assumption PT-ES is not testable,
as it only makes restrictions on untreated potential outcomes in post-treatment periods, t ě g.
Under no-anticipation, pre-trends do measure differences in untreated outcome trends between
treated and untreated units, but they necessarily measure them in the “wrong” periods t ă g. This
does not mean parallel pre-trends are not informative; it just means they are not the same as the
parallel trends assumption in Assumption PT-ES.
When specifically are observed pre-trends uninformative about the parallel trends assumptions
required for identification? One case is when pre-trends are measured too far before treatment
starts. The conditions or shocks that jointly shape treatment decisions and untreated outcomes
may differ many periods before treatment, but this need not imply that they generate bias in
the periods after treatment. A second case, discussed in Section 3.2, is when units select into
treatment based on time-varying pre-treatment unobservables. As treatment selection depends on
pre-treatment unobservables, non-parallel pre-trends may appear as a consequence of the selection
mechanism, though parallel trends may still be plausible. Ghanem et al. (2022) show that in this
case, a necessary condition for parallel trends is that the untreated potential outcomes satisfy a
martingale property. When this martingale property does not hold, parallel trends cannot hold
either, regardless of the shape of pre-trends. In such situations, it is worthwhile to assess the
potential bias components of the DiD estimator using the “selection-aware” benchmarking tools
presented in Ghanem et al. (2022), as well as to resort to their theory-based templates for justifying
parallel trends based on selection mechanisms. Ultimately, how informative pre-trends are for the
plausibility of Assumption PT-ES is case-specific, though we would of course always prefer to have
parallel pre-trends.
When Assumption PT-ES holds in all periods (pre- and post-treatment), it is worth noting
that one can construct estimators for the AT T ptq that are more precise than those in (5.2); see,
e.g., Borusyak et al. (2024), Gardner (2021), Harmon (2024), Marcus and Sant’Anna (2021),
Wooldridge (2021), and Chen, Sant’Anna and Xie (2024). In such cases, though, one would
require parallel trends to hold pre-treatment periods, parallel pre-trends becomes directly testable
(as the model is overidentified), and τ´k could be used to assess its plausibility directly. Other
types of overidentification tests could also be used in this case to assess the validity of the DiD
model directly; see Marcus and Sant’Anna (2021) and Chen et al. (2024) for a discussion.
The second lesson is that statistical precision shapes the usefulness of pre-trend estimates. The
hypothesis tests for parallel pre-trends in Figure 3 are low-powered to detect practically important
violations (Roth, 2022, and Freyaldenhoven et al., 2024). The τ´k estimates fail to rule out flat
pre-trends, but as we shall see below, they also fail to rule out large pre-trends that would indicate
serious bias in the AT
z T ptq’s. They simply do not say very much.16 Roth (2022) discusses how

16
Another possibility is that very precisely estimated pre-trends are distinguishable from zero yet also rule out
even small violations. The magnitude of the violations also matters: a precisely estimated but small violation of
parallel trends in pre-treatment periods is “better” than an imprecisely potentially large estimated pre-trend that

34
conditioning the analysis on these kinds of low-powered tests can exacerbate biases and should be
interpreted with care.
The third lesson is that researchers can make better use of pre-trend estimates by taking a
stand on the size of plausible and/or problematic parallel trends violations. For example, Bilinski
and Hatfield (2018) propose selecting a value for differential trends that would fully explain the
estimated treatment effects and then using it instead of zero as the null hypothesis when conducting
statistical tests of the estimated pre-trends; see also Dette and Schumann (2024). Rambachan
and Roth (2023) develop inference methods for an approach that bounds the AT T ptq’s under
assumptions about the maximum size of parallel trends violations based on pre-treatment periods
(see also Manski and Pepper, 2018). One can either select a magnitude using contextual knowledge
or set it equal to a multiple of the largest period-to-period pre-trend estimate. One then constructs
an identified set that contains AT T ptq not under parallel trends but under the weaker assumption
that parallel trends is not violated by more than this pre-determined maximum in each post-period.
Importantly, Rambachan and Roth (2023) also show how to use the estimated covariance matrix of
the event study estimates to construct confidence intervals around the set.17 Given the existence
of these methods, which are theoretically grounded in how to consider violations of pre-trends,
we caution producers (and consumers) of DiD work against the common practice of using only a
simple “eye-test” for whether pre-trends differ substantially from zero.
In our application, Rambachan and Roth (2023)’s method underscores how little information
the pre-trend estimates convey. The largest one-period pre-trend in Figure 3 is between event-time
-5 and -4, when outcomes fall by roughly four deaths more in the expansion group versus the non-
expansion group. If we assume that parallel trend violations are no bigger than this, the identified
set for AT T p2014q is ´2.6 ˘ 4 “ r´6.6, 1.4s, and given the size of the pre-period standard errors,
we obtain a robust confidence interval of r´11.1, 5.1s, which spans implausibly large effects in both
directions. Assuming smaller parallel trend violations would shrink this interval, but applying the
method to subsequent event-times widens it. In general, Rambachan and Roth (2023) provide a
flexible method to use information about potential parallel trends violations drawn either from
external knowledge or the τ´k estimates, as well as the precision of the pre-trend estimates, to
gauge (statistically) the robustness of the ATz T ptq’s.
When pre-trends suggest that Assumption PT-ES fails, a way forward is to assume that it holds
only after conditioning on covariates and proceeding similarly to Section 4.4; we discuss this path
in detail in Section 5.1.4. Alternatively, one can attempt to parametrically model the violations
of parallel trends. Usually this is done by including unit-specific linear trends; see, e.g., Mora and
Reggio (2019), Wooldridge (2021, Section 7), Lee and Wooldridge (2023), and Freyaldenhoven et
al. (2024). We note, however, that the practice of using unit-specific linear trends deviates from the
does not rule out zero.
17
In cases where a measured variable accounts for the pre-trends, Freyaldenhoven, Hansen and Shapiro (2019)
develop two-stage-least-squares estimators that recover AT T ptq parameters by extrapolating the pre-period rela-
tionship between outcomes and the covariate into the post-period.

35
standard DiD procedures: it relies on alternative identification assumptions involving an explicit
parametric model for unit-specific trends. We also note that sensitivity analysis procedures that
do not rely on such models are available—Rambachan and Roth (2023) is one example—and we
encourage practitioners to consider them.

5.1.3 Estimation and aggregating across time in event-studies

The link between a 2 ˆ T event study and a series of 2 ˆ 2 DiD building blocks makes estimation
simple. The point estimates in Figure 3 are the AT T ptq estimates based on (5.2).18 An equivalent
way to obtain all the AT z T ptq’s in one step is to run a TWFE regression with time fixed effects, θt ,
unit fixed effects, ηi , and a set of interactions between the treatment group dummy and the time
dummies. Omitting the treatment interaction for t “ g ´ 1 avoids multicollinearity and fixes g ´ 1
as the baseline period for all βt estimates, which matches Assumption PT-ES. This generalizes the
TWFE regression equation for a single AT T ptq in (3.7) to
g´2
ÿ T
ÿ
Yi,t “ θt ` ηi ` βk p1tGi “ gu ¨ 1tt “ kuq ` βk p1tGi “ gu ¨ 1tt “ kuq ` εi,t . (5.3)
k“1 k“g

This regression produces identical estimates to those obtained “by hand” via (5.2): βpt “ ATz T ptq.
It also generates (point-wise) confidence intervals based on clustered standard errors, as discussed
in Section 3.3. An additional issue in event study inference, however, involves the fact that we
are now estimating many treatment effect parameters. Thus, when we compare across event study
estimates, we are conducting many hypothesis tests, and the usual normal critical values used to
construct confidence intervals do not account for these. Asymptotically correct inferences about
the entire event study curve require “inflating” critical values to perform a multiple hypothesis test
adjustment. In Figure 3, the thick black bars represent the standard pointwise confidence intervals
from clustered standard errors at the county level, while the red line shows uniform confidence
bands that cover the 95% confidence interval for the entire treatment path of the event study
coefficients after accounting for multiple testing. These are produced by default in the Callaway
and Sant’Anna (2021) statistical packages, using a multiplier bootstrap procedure to compute
critical values of the sup-t test statistic. Alternatively, one can construct these using the estimated
variance-covariate matrix of all βpt ’s paired with the Montiel Olea and Plagborg-Moller (2018)
simulation procedure. Alternative bootstrap procedures, such as the nonparametric bootstrap,
the multiplier bootstrap, and the weighted/Bayesian bootstrap, can also be used to compute the
sup-t critical values that account for multiple testing.19

18
We estimate these with the did R package from Callaway and Sant’Anna (2021).
19
The sup-t critical value governs the width of the uniform confidence band that yields simultaneous coverage
probabilities for a given confidence level (Montiel Olea and Plagborg-Moller, 2018). In our context, its main idea
is constructing asymptotically valid critical values for the entire event-study trajectory based on the maximum of
all t-statics (one for each event-time considered). This procedure avoids the conservativeness of multiple-testing
corrections such as Bonferroni’s. See Montiel Olea and Plagborg-Moller (2018) for a general discussion.

36
 A final estimation issue arises when targeting aggregations of the AT T ptq’s. For example, the
1
řT
average treatment effect in the post-period, AT Tavg “ T ´pg´1q t“g AT T ptq, is a convenient scalar
measure that improves statistical precision, especially when AT T ptq is relatively constant. The
easiest way to get an estimate of ATz T avg is just to construct it from the 2 ˆ 2 AT z T ptq building
block estimates. Standard post-estimation commands achieve this if the event study estimates
come from a regression, and newer DiD packages report this parameter automatically. A common
shortcut, however, is to run a second regression that replaces the event study dummies with the
treatment status dummy. Di,t “ 1tDi “ 1u ˆ 1tt ě gu:

Yi,t “ θt ` ηi ` β OLS Di,t ` εi,t . (5.4)

Unfortunately, the (weighted) least squares estimator βpOLS does not generally equal the AT z T avg .
OLS
The reason is that βp is equivalent to first collapsing the multiple-periods data to averages in
the post- and pre-periods and then estimating a 2 ˆ 2 DiD on the resulting means. This, in turn,
is the same as subtracting the average pre-period τ´k estimate (including the zero at time g ´ 1)
from the average post-period AT z T ptq estimate. This implies that interpreting both βt s from (5.3)
OLS
and β from (5.4) requires Assumption PT-ES to hold in every time period, not just in the
post-treatment periods.20 To the extent that the gap in mean outcomes over the whole pre-period
differs from the gap in outcomes in period t “ g ´ 1, the two summary parameters will not be
equal. In Figure 3, the two summary parameters are quite different: AT z T avg equals -0.70 while
OLS 21
βp equals -2.53.

5.1.4 Covariates in event studies

Another advantage of seeing event studies as collections of 2 ˆ 2 building blocks is that all the
tools for incorporating covariates from Section 4 immediately apply to each event study estimate.
In fact, the only difference is that instead of using “short-differences,” ∆Yi,t“2 , one would now use
“long-differences,” Yi,t ´ Yi,t“g´1 . This would imply that using the regression-adjustment procedure
would require estimating a working model for Eω rYi,t ´ Yi,t“g´1 |Di “ 0, Xi s for each time t. The
propensity score working model used to construct the IPW DiD estimate (4.6), on the other hand,
is exactly the same as in a 2 ˆ 2 analysis of the same groups. Since the DR DiD estimation
procedure builds on both RA and IPW procedures, it would involve estimating different outcome-
regression working models for each time t. We also note that the potential pitfalls of controlling
for covariates in a TWFE specification still apply with multiple periods and actually become more
complex (Caetano and Callaway, 2024).

20
The decision to estimate each AT z T ptq relative to period t “ g ´ 1 comes directly from the choice to define
PT that way. If parallel trends holds in every period, one can typically form more efficient estimators than those
discussed above. See Marcus and Sant’Anna (2021) for a discussion
21
Interestingly, this quantity is almost identical to our estimate of AT T p2014q, but rather than representing
anything reassuring, it comes from the offsetting effects of positive pre-period estimates and small post-period ones.

37
 For completeness and ease of access, we list the RA, IPW, and DR estimands for AT T ptq:
” ˇ ı
AT Tra ptq “ Eω rYi,t ´ Yi,t“g´1 |Di “ 1s ´ Eω Eω rYi,t ´ Yi,t“g´1 |Xi , Di “ 0sˇDi “ 1 ,
ˇ
”´ ¯ ı
AT Tipw ptq “ E wω,D“1 pDi q ´ wω,D“0 pDi , Xi q pYi,t ´ Yi,t“g´1 q ,
”´ ¯´ ¯ı
AT Tdr ptq “ E wω,D“1 pDi q ´ wω,D“0 pDi , Xi q Yi,t ´ Yi,t“g´1 ´ Eω rYi,t ´ Yi,t“g´1 |Xi , Di “ 0s ,

where wω,D“1 pDi q and wω,D“0 pDi , Xi q are as defined in (4.7).

Figure 4 shows weighted event study estimates using our three preferred covariate strategies:
regression adjustment, inverse propensity weighting, or doubly-robust estimation. In our case,
covariates do little to change the unadjusted estimates. Note, however, that Borgschulte and
Vogler (2020) use an IPW estimator with different covariates selected by a lasso procedure and
obtain notably stronger evidence of mortality reductions. Evidently, the exact set of covariates one
conditions on matters a great deal in this analysis, and, as we have discussed earlier, one should
attempt to include in Xi all the determinants of the change in untreated potential outcome or of
the treatment assignment.
Figure 4: 2 ˆ T Event Study with Covariates

Regression IPW Doubly Robust

20

10

Treatment Effect
Mortality Per 100,000 0

−10

−20
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5
Event Time
Notes: This figure shows event study estimates that include covariates. The outcome variable is the crude mortality rate for adults
ages 20-64, and the covariates include 2013 values of the percentage of the county population that is female, the percentage of the
county population that is white, the percentage of the county population that is Hispanic, the unemployment rate, the poverty rate,
and county-level median income. The sample includes 2,200 counties (978 in states that expanded Medicaid by 2014 and 1,222 in states
that did not expand Medicaid by 2019). The point estimate is reported by the circles, and both point-wise (black) and simultaneous
(red) confidence intervals are reported with the vertical lines. All procedures use population weights.

5.2 Staggered treatment adoption (G# ˆ T)

Viewing 2ˆT event studies as a collection of 2ˆ2 DiD building blocks makes the jump to staggered
timing designs straightforward. The key distinction is that with staggered timing, each treatment
date defines a distinct treatment group, and each of these has its own set of simple event study
parameters. New choices arise about the comparison units used to identify and estimate these
group-specific event studies, as well as about how to aggregate the estimates across timing groups.
The 2 ˆ 2 structure, as well as all the tools we have developed to evaluate parallel trends (covariate
balance and pre-trends) and to estimate (with or without weights and covariates), carry over.

38
 When treatment start dates can vary across units, we need to allow the potential outcomes, and
thus the target parameters and identifying assumptions, to reflect this richer notion of treatment.
We therefore index potential outcomes by the time treatment begins, g: Yi,t pgq; and use Yi,t p8q to
denote never-treated potential outcomes.22 We use Gi to denote each unit’s treatment date, and
with some abuse of terminology, we call units not exposed to treatment by period T the “never-
treated” group.23 Finally, we use G to represent the set of all treatment times (rows of Table 1
in our example). With these modifications, we can map potential outcomes to observed outcomes
using a generalization of (3.1):
ÿ
Yi,t “ Yi,t pgq1tGi “ gu.
gPG

With multiple treatment groups, we also need to extend our notion of no-anticipation (though
its empirical content is exactly the same).

Assumption NA-S (No-Anticipation with staggered treatment timing). For all units i that are
eventually treated and all pre-treatment periods t, Yi,t pgq “ Yi,t p8q.

Like Assumption NA, Assumption NA-S imposes that treatment effects are zero in all pre-
treatment periods as a consequence of units not acting on the potential knowledge of future
treatment dates before they are actually exposed to treatment. We maintain this assumption
throughout this section.
Finally, we assume that a “never-treated” group always exists in our staggered DiD setup. If all
units are eventually treated, we drop all the data from when the last cohort is treated, so the last-
treated cohort becomes the “never-treated” cohort, and T here denotes the number of available
periods in the subset of the data that we will use in our analysis. This is essentially without
loss of generality, because under standard DiD assumptions, we cannot identify any AT T for
periods where all units are treated. We also dropped data from units treated in the first available
period, Gi “ 1, as such treatment group does not have any pre-treatment data, preventing us from
conducting a DiD analysis.
Figure 5 plots average weighted mortality data by the year of Medicaid expansion (Gi ) and
time. As in Table 2 and Figure 2, these are all the means necessary to calculate a staggered DiD
22
Let 0s and 1s be s-dimensional vectors of zeros and ones, respectively, and denote the potential outcome
for unit i at time t if first exposed to the treatment at time g by Yi,t p0g´1 , 1T ´g`1 q, and denote by Yi,t p0T q the
outcome if untreated by time t “ T . We discussed the two-period treatment this way in section 3.1 when we defined
potential outcomes as a function of the period one and period two treatment. These treatment paths define the
potential outcomes we work with: Yi,t pgq “ Yi,t p0g´1 , 1T ´g`1 q and Yi,t p8q “ Yi,t p0T q. Writing potential outcomes
as functions of treatment paths helps with transparency regarding causal parameters of interest and the DiD design
we are in. When treatment can turn on and off, writing potential outcomes in terms of the entire path becomes
crucial to avoid “hidden” assumptions that rule out treatment effect heterogeneity and dynamics. Because of space
constraints, we do not cover these cases in this article.
23
In practice, “never-treated” really means “not observed to be treated by t “ T .” Given more data, units
untreated at T could, in many cases, take up the treatment. In fact, this is the case with the Medicaid expansion.
We use data through 2019 but include states that expanded Medicaid in 2020, 2021, and 2023 as “never treated,”
alongside states that have not expanded as of 2024.

39
estimate.

Figure 5: County Mortality Trends by Expansion Decision with Staggered Timing

450

Mortality (20−64) 400

Per 1,000

350

2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

2014 2015 2016 2019 Non−Expansion Counties

Notes: This figure shows county population-weighted average mortality rates for adults ages 20-64 from 2009 to 2019. There are 978
counties in states in the 2014 expansion group, 171 counties in the 2015 expansion group, 93 counties in the 2016 expansion group, 140
counties in the 2019 expansion group, and 1,222 counties that did not expand Medicaid by 2019.

5.2.1 Building block parameters with staggered adoption

Staggered treatment timing affects the structure of a DiD analysis because it changes the definition
of treatment. Until now, we have used counties in the 2014 expansion states as the only treatment
group represented with a single treatment dummy, Di . But Table 1 shows that as of 2019, there
were four different groups of expansion states defined by whether they expanded Medicaid in
2014, 2015, 2016, or 2019. Therefore, Di is not rich enough to capture the relevant definition of
treatment groups in staggered setups, because there are many treatment groups, not just two.
Fortunately, we can use the treatment timing notation, Gi , to define AT T parameters, parallel
trends assumptions, and estimators, just as we have done so far.
A simple 2ˆ2 DiD design had one target parameter (AT T p2q), and a 2ˆT DiD design had T ´1
of them: T ´ pg ´ 1q post-treatment parameters, AT T ptq, t ě g, and g ´ 2 pre-trend parameters.
In staggered DiD designs, each treatment group (sometimes referred to as a cohort), defined by
its treatment date g, has its own set of T ´ 1 event study parameters. We call these group-time
average treatment effects:

AT T pg, tq “ Eω rYi,t pgq ´ Yi,t p8q|Gi “ gs.

40
Each AT T pg, tq is the average treatment effect of starting treatment at period g relative to never-
starting it, at time period t, among units that actually started treatment in period g. It is simply
a set of event study parameters for treatment timing group g. The unobserved counterfactuals are
now untreated potential outcome means for each treatment group in each period, Eω rYi,t p8q|Gi “
gs.

5.2.2 Identification with staggered designs

Identifying the AT T pg, tq’s works exactly as in the previous sections because they are just group-
specific event studies. Under no anticipation, a set of parallel trends assumptions for t ě g identifies
the causal post-treatment parameters. DiD comparisons for t ă g represent differential pre-trends
in untreated potential outcomes.
The most important way that staggered DiD changes this approach is that having access to
multiple treatment groups with different treatment starting dates allows one to use alternative sets
of comparison groups. For example, our Medicaid analysis so far has used counties in states that
did not expand Medicaid by 2019 as the comparison group. For estimating, say, the AT T for the
2014 expansion group in 2015, counties in states that did not expand until 2016, 2017, or 2018
could also serve as comparison units. Choosing which comparison groups to use to identify an
AT T pg, tq is directly tied to which form(s) of parallel trends hold. Relative to simple event-studies
and especially 2 ˆ 2 setups, staggered timing creates many potential parallel trends assumptions.
Here, we discuss three types of staggered parallel trends. The first two use either the never-treated
units or any not-yet-treated groups as the comparisons for all eventually-treated groups (Callaway
and Sant’Anna, 2021), and the third option assumes that parallel trends holds in all periods
and in all groups, which is something that several DiD methods require; see de Chaisemartin
and D’Haultfoeuille (2020); Sun and Abraham (2021); Wooldridge (2021); Borusyak et al. (2024);
Harmon (2024).24

Assumption PT-GT-Nev (Parallel Trends based on never-treated groups). For every eventually
treated group g and post-treatment time period t ě g,

Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ gs “ Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ 8s.

Assumption PT-GT-NYT (Parallel Trends based on not-yet-treated groups). For every even-
tually treated group g, not-yet-treated group g 1 and time periods t such that t ě g and g 1 ą t,

Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ gs “ Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ g 1 s.

24
While these three types of parallel trends are fairly intuitive choices, others are possible. For instance, Cengiz,
Dube, Lindner and Zipperer (2019) use a comparison group of units treated at least δ ` 1 periods after time g.
Thus, their parallel trends is tailored to this particular choice. As a result, all of their AT T pg, tq estimates from
time g to time g ` δ use the same comparison group. Marcus and Sant’Anna (2021) also discuss other alternative
parallel trends assumptions.

41
Assumption PT-GT-all (Parallel Trends for every period and group). For every treatment
groups g and g 1 and time periods t,

Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ gs “ Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ g 1 s.

Assumption PT-GT-Nev is the analog of the PT assumptions we used in the 2 ˆ T design. It

uses the never-treated units as the relevant comparison group for all eventually-treated units, and
it imposes parallel trends in post-treatment periods only. In our Medicaid application, this would
entail using the non-expansion counties as the comparison group for the 2014, 2015, 2016, and 2019
expansion groups. In addition, since Assumption PT-GT-Nev imposes parallel trends only for the
future, the farthest we can go into pre-treatment periods is t “ g ´ 1, which will serve as the only
(justifiable) baseline period. More formally, under Assumption PT-GT-Nev, it is straightforward
to show that for post-treatment periods,

AT T pg, tq “ Eω rYi,t ´ Yi,t“g´1 |Gi “ gs ´ Eω rYi,t ´ Yi,t“g´1 |Gi “ 8s, (5.5)

which shows that AT T pg, tq is identified (Callaway and Sant’Anna, 2021; Sun and Abraham, 2021).
Note that (5.5) highlights that we are essentially back to a 2 ˆ 2 design when it comes to learning
about AT T pg, tq: it leverages data from only two periods, t (post) and g ´ 1 (pre), and two
treatment groups, Gi “ g (treated) and Gi “ 8 (comparison).
Assumption PT-GT-NYT allows one to use not only the never-treated units but any group
of units that are not-yet-treated by time t. In our Medicaid example, we could now use non-
expansion counties and 2016 and 2019 expansion counties as comparison groups when estimating
AT T p2014, 2015q. Using Assumption PT-GT-NYT, Callaway and Sant’Anna (2021) has shown
that we can identify the AT T pg, tq for post-treatment periods t ě g by25

AT T pg, tq “ Eω rYi,t ´ Yi,t“g´1 |Gi “ gs ´ Eω rYi,t ´ Yi,t“g´1 |Gi ą maxtg, tus. (5.6)

Like equation (5.5), this follows a 2 ˆ 2 set up: it leverages data from only two periods, t (post)
and g ´ 1 (pre), and two treatment groups, Gi “ g (treated), and Gi ą t (comparison).26
Finally, Assumption PT-GT-all allows the use of any not-yet-treated units as a comparison
group, as well as any pre-treatment period as a baseline period. For instance, to identify the
ATT for 2015 expansion counties, we can now use the never-treated group and the 2016 and
2019 expansion groups, and use any or all of the years from 2009 to 2014 as a baseline. Using
25
de Chaisemartin and D’Haultfoeuille (2020) have derived the same results but restrict attention to instanta-
neous treatment effects—that is, AT T pg, gq’s. They do allow for treatment turning on and off, but also impose
that being exposed to a treatment today does not affect outcomes tomorrow (a no-carryover assumption). See
de Chaisemartin and D’Haultfoeuille (2023a) for some extensions.
26
We note that more general results are also possible. In fact, for any not-yet-treated group g 1 ą t, it is easy
to show that under Assumption PT-GT-NYT, for t ě g, AT T pg, tq “ Eω rYi,t ´ Yi,t“g´1 |Gi “ gs ´ Eω rYi,t ´
Yi,t“g´1 |Gi “ g 1 s. One can then flexibly combine the different comparison groups by leveraging user-specified or
efficiency-oriented weights. Chen et al. (2024) discuss how to efficiently explore all the information implied by the
identification assumptions to form semiparametrically efficient DiD estimators—that is, estimators that asymptoti-
cally enjoy the shortest possible (theoretically justified) confidence intervals without making strong functional form
or model-based assumptions related to error terms such as homoskedasticity and restrictions on serial dependence.

42
Assumption PT-GT-all, it is easy to show that, for any pre-treatment period tpre ă g and any
not-yet-treated group g 1 ą t, we can identify the AT T pg, tq for group g’s post-treatment periods
t ě g by

AT T pg, tq “ Eω rYi,t ´ Yi,tpre |Gi “ gs ´ Eω rYi,t ´ Yi,tpre |Gi “ g 1 s, (5.7)

which again maps back to the 2 ˆ 2 DiD setup, as it leverages two periods, t (post) and tpre (pre),
and two groups, Gi “ g (treated) and Gi “ g 1 (comparison). This representation makes it clear
that, in practice, one can use various pre-treatment periods and not-yet-treated comparison groups
to characterize the AT T pg, tq under Assumption PT-GT-all. We can also combine several of these
to form an AT T pg, tq estimand that uses more data (Wooldridge, 2021; Gardner, 2021; Liu, Wang
and Xu, 2024; Borusyak et al., 2024; Chen et al., 2024). An intuitive estimand that naturally
extends (5.6) and allows us to use more pre-treatment data is given by

AT T pg, tq “ Eω rYi,t ´ Y i,tďg´1 |Gi “ gs ´ Eω rYi,t ´ Y i,tďg´1 |Gi ą maxtg, tus, (5.8)
řg´1 L
where Y i,tďg´1 “ s“1 Yi,s pg´1q is the time average of group g pre-treatment periods for each unit
i; see, for example, Callaway (2023, Section 3.2), Lee and Wooldridge (2023), and the discussion
in de Chaisemartin and D’Haultfoeuille (2023b, Section 3.2.4). Although the estimand does look
different from the previous one, it too resembles a 2 ˆ 2 design: it effectively leverages two periods,
t (post) and the average of all period t ă g (pre), and two groups, Gi “ g (treated) and Gi ą t
(comparison). In practice, however, there is no general econometrics guarantee that estimators
based on (5.8) will be more precise than estimators based on (5.6); see, for instance, Harmon
(2024). This arises because it is not always optimal to weigh all pre-treatment periods equally
when forming estimators for AT T pg, tq. In fact, as discussed in Chen et al. (2024), the optimal way
(or, more formally, the semiparametric efficient way) to aggregate information across pre-treatment
periods and comparison groups under Assumption PT-GT-all depends on the correlation structure
of how the outcome changes over time across different comparison groups. And an effective way
to leverage this information consists of constructing DiD estimators for AT T pg, tq’s that efficiently
weigh several 2 ˆ 2 DiD estimators for the AT T pg, tq that use different comparison groups and
different baseline periods. These efficiency weights do not depend on additional hard-to-motivate
assumptions (e.g., homoskedasticity or restriction on the serial correlation), arise as a consequence
of the information content of the identification assumptions, and can be transparently visualized
using the tools provided Chen et al. (2024). What is perhaps most important here is that even
these more complex DiD estimators closely resemble the approach we took in the 2 ˆ 2 design.
In the end, a natural question arises: Which parallel trends assumption should one use? This
context-specific question is hard to answer, as each assumption has pros and cons. For instance,
Assumption PT-GT-all leads to more precise AT T pg, tq estimators because it uses data from
multiple pre-treatment periods and multiple comparison groups. Given that power is important
when conducting causal inference, this is appealing. On the other hand, it imposes parallel pre-

43
trends, an assumption that is not required for identification of AT T pg, tq and that we have not
imposed in 2 ˆ T DiD designs (see our discussion of equation (5.3)). If pre-trends are not parallel,
then estimates of AT T pg, tq based on Assumption PT-GT-all can be biased.
The other extreme is to make Assumption PT-GT-Nev and use only comparison groups made
up of never-treated units. This avoids compositional changes in the comparison group over time,
does not restrict pre-trends, and identifies all the AT T pg, tq’s.27 It also avoids using as a comparison
group units that may have chosen to begin treatment in a given period because of pre-treatment
outcomes, which potentially violates parallel trends. For instance, states that expanded Medicaid
in 2016 may have done so on the basis of county mortality rates from previous years, likely violating
the parallel trends assumption for this group. On the other hand, never-treated units may have
remained untreated for reasons related to trends in Yi,t p0q. Non-expansion counties may be too
different from expansion counties for them to reflect the relevant counterfactual. This could be,
in part, justified after examining the differences in covariate levels and trends between treatment
and control groups, as we saw in the discussion of Table 4. Also, depending on how widespread
treatment is, there may be too few never-treated units to obtain precise estimates.
We view Assumption PT-GT-NYT as a middle step that uses all not-yet-treated units as
a comparison group without restricting all pre-treatment trends to be parallel. It uses more
information than Assumption PT-GT-Nev, which can lead to gains in precision and helps to
incorporate covariates. While it uses less information than Assumption PT-GT-all, it is also less
susceptible to bias from violations of parallel pre-trends. For our Medicaid application, we favor
Assumption PT-GT-NYT, as we prefer not to impose parallel pre-trends from 2009 to 2014 or to
rely exclusively on comparisons to the set of states that have not expanded Medicaid as of 2024.
On the other hand, if parallel trends are not plausible for a particular group of eventually-treated
units, perhaps owing to selection based on time-varying unobservables (Ghanem et al., 2022), it is
important to remove these units from the DiD analysis to retain interpretability. Ultimately, the
plausibility of each parallel trends assumption may vary across different contexts. At the very least,
we strongly recommend that researchers clearly state the specific parallel trends assumption they
are actually imposing in their analysis to allow readers to discuss its plausibility in a scientifically
grounded manner.

5.2.3 Estimators for staggered designs without covariates

The identification results for AT T pg, tq discussed in Section 5.2.2 suggest very simple and intuitive
estimators for the AT T pg, tq. Given the estimand that comes from the chosen parallel trends as-
sumption, the estimators replace the population (weighted) expectations with their sample analogs.

27
Without never-treated units, we cannot estimate AT T pg, tq for the last observed treatment date, which shapes
the feasible target parameters. For example, in our Medicaid expansion example, without the presence of never-
treated (by 2019) counties, we would not be able to estimate the treatment effects in 2019 (i.e., AT T pg “ 2015, t “
2019q.

44
The principle is the same as in the 2 ˆ 2 setup of Section 3.3.
For example, under Assumption PT-GT-NYT, we can leverage (5.6) and form plug-in estima-
tors for AT T pg, tq using
řn řn
i“1 1tGi “ guωi pYi,t ´ Y i,t“g´1 q 1tGi ą tuωi pYi,t ´ Yi,gt“´1 q
AT
z T nyt pg, tq “ řn ´ i“1 řn . (5.9)
i“1 1tGi “ guωi i“1 1tGi ą tuωi

This simple estimator is what Callaway and Sant’Anna (2021) propose when one uses the not-yet-
treated group as the comparison group.28
When Assumption PT-GT-Nev holds, it is straightforward to build on (5.5) and estimate
AT T pg, tq by
řn řn
i“1 1tGi “ guωi pYi,t ´ Yi,t“g´1 q 1tGi “ 8uωi pYi,t ´ Yi,gt“´1 q
AT T never pg, tq “
z řn ´ i“1 řn . (5.10)
i“1 1tGi “ guωi i“1 1tGi “ 8uωi

This estimator was proposed by Callaway and Sant’Anna (2021) and Sun and Abraham (2021)
when using never-treated units as a comparison group, though Sun and Abraham (2021) arrive at
SA
this using a fully saturated regression specification and estimating the regression coefficients βg,e
with (weighted) least squares,
ÿ ÿ
SA
Yi,t “ θt ` ηi ` βg,e 1tGi “ gu1tGi ` e “ tu ` ϵi,t . (5.11)
g‰8 e‰´1

SA
It is straightforward to show that βg,e “ AT
z T never pg, g ` eq, emphasizing that (5.11) is just a way
to contrast sample means across groups and periods that respect Assumption PT-GT-Nev.
When Assumption PT-GT-all holds instead, one can construct plug-in estimators for (5.8):
řn řn
i“1 1tGi “ guωi pYi,t ´ Y i,tďg´1 q 1tGi ą tuωi pYi,t ´ Y i,tďg´1 q
AT T avg pg, tq “
z řn ´ i“1 řn .
i“1 1tGi “ guωi i“1 1tGi ą tuωi

Alternatively, Wooldridge (2021) proposed constructing estimators for AT T pg, tq based on As-
sumption PT-GT-all, using following “extended” TWFE specification:
T
ÿ ÿ
W
Yi,t “ θt ` ηi ` βg,t 1tGi “ gu1ts “ tu ` ϵi,t , (5.12)
g‰8 s“g

W W
where the βg,t ’s are estimated using (weighted) least squares. Wooldridge (2021) shows that βpg,t
consistently estimate AT T pg, tq under Assumption PT-GT-all, though we do not know the exact
W
way that βpg,t combines pre-treatment periods and not-yet-treated units, which is to say that we
W W
do not know the statistical estimand associated with βg,t . Wooldridge (2021) also shows that βpg,t
is numerically the same as the “imputation” estimators proposed by Gardner (2021), Liu et al.
(2024), and Borusyak et al. (2024) with balanced panel data (and these specifications do not have
covariates).29
28
Recently, Dube, Girardi, Jordà and Taylor (2024) show that one can also get AT T pg, tq estimates that are
equivalent to AT
z T nyt pg, tq by using local projections. One can also get similar estimators using a “stacked DiD”
procedure akin to what Fadlon and Nielsen (2021), Deshpande and Li (2019), and Cengiz et al. (2019) have
implemented.
29
Imputation procedures work in two-steps. The first step uses all untreated observations to run the re-

45
 Overall, these different AT T pg, tq estimators highlight that we can leverage our DiD expertise
built in the 2 ˆ 2 setup to estimate heterogeneous AT T pg, tq parameters. The exact estimator
we can use is shaped by which parallel trends assumption holds. In our Medicaid application, we
report in Figure 6 our AT T pg, tq estimates based on Assumption PT-GT-NYT and (5.9). As we
have four sets of counties defined by their Medicaid expansion timing, we report four sets of event
studies, one for each expansion group. For the 2014, 2016, and 2019 expansion groups, we find
that Medicaid did not lead to significant changes in adult mortality rates. For the 2015 expansion
group, adult mortality rates rose after expansion. Regarding pre-trends, Figure 6 suggests that
there may be some non-negligible pre-trends for the 2016 expansion group, though these are not
statistically different from zero.
In Section 4, we highlighted that unconditional assumptiosn such as Assumption PT-GT-
NYT may fail in our Medicaid context. One reason is that important determinants of changes
in untreated adult mortality rates are imbalanced across treatment and comparison groups. In
such cases, one should interpret the results in Figure 6 with great care. In addition, it is also
important to highlight that, as indicated in Table 1, the 2015, 2016, and 2019 expansion groups
are relatively small and represent only 6%, 2%, and 3% of the US population, respectively. Thus,
analyzing these groups separately may be “too noisy” and not representative of the overall effect
for the US population. This does not mean that these AT T pg, tq’s are not useful. They are
actually an essential part of our DiD analysis, but we may want to aggregate the cohorts to get
more informative target parameters for the overall treated population. We now turn to how to
aggregate the AT T pg, tq’s and then discuss how to incorporate covariates.

5.2.4 Aggregating group-time average treatment effects

The previous section highlighted that, in many applications, estimating all AT T pg, tq’s precisely
and attaching policy-relevant interpretations to them may be challenging. Aggregating them into
a summary treatment effect measure therefore has clear benefits: it improves precision, reduces
the number of results, and yields a parameter that averages over all treated units like the AT T p2q
identified in 2 ˆ 2 designs.
Aggregation in staggered designs involves a notion of time (either calendar time t or event-time
e “ t ´ g), a length of time (how many periods to aggregate across), and group weights (so larger
treatment groups can “matter more” than smaller ones). Given some set of weights, it is simple to

gressions Yit “ θt ` ηi ` ϵit using only data from the pi, tq pairs that satisfy this criterion, and get the fit-
ted values Ypi,t p0q “ θpt ` ηpi for all eventually-treated observations. The second step estimates AT T pg, tq by
řn
1tG “gupYi,t ´Ypi,t p0qq
AT
z T imp pg, tq “ i“1 řni 1tGi “gu . See Gardner (2021), Liu et al. (2024), and Borusyak et al. (2024) for
i“1
details. Note that the specification in (5.12) is similar to the Sun and Abraham (2021)’s specification (5.11), but
it omits the pre-treatment event-time dummies. This is justified because (5.12) imposes parallel pre-trends (As-
sumption PT-GT-all) while (5.11) does not (it effectively relies on Assumption PT-GT-Nev). Thus, in general, one
should not expect βpW to be equal to βpSA .
g,t g,t´g

46
 Figure 6: ATT(g,t)s for Each Expansion Group

2014 2015

20

−20

−40

Treatment Effect
Mortality Per 100,000 2016 2019

20

−20

−40

2009 2011 2013 2015 2017 2019 2009 2011 2013 2015 2017 2019

2014 2015 2016 2019

Notes: This figure shows the group-time ATT estimates (AT T pg, tq) in calendar time for the four groups of counties that expanded
Medicaid before 2019. Each panel uses 1,222 not-yet-treated counties as the comparison group and shows their uniform confidence
intervals at the 95% significance level. There are 978 counties in states in the 2014 expansion group, 171 counties in the 2015 expansion
group, 93 counties in the 2016 expansion group, 140 counties in the 2019 expansion group. The outcome variable is the crude mortality
rate for adults ages 20-64, and standard errors are clustered at the county level. The vertical line represents the year before Medicaid
expansion (i.e., g - 1) for the timing group. All results use population weights.

average the AT T pg, tq building blocks into many kinds of summary parameters:
ÿ
AT Taggte “ wω,g,t AT T pg, tq,
g,t

where wω,g,t is a “generic” (ω-weighted) group and time-specific (non-negative) weight that sums
up to one. Specific choices for the wω,g,t weights map to different ways to aggregate and present
interpretable causal effects in this kind of complex setting.
As highlighted in Section 5.1, an appealing feature of having access to data from multiple
periods is that we can assess how average treatment effects evolve with the time since treatment,
or event-time e “ t ´ g. Figure 6 displays event studies for each expansion group, and the
aggregation question is how to combine them into a single summary event study.
A basic observation about weighting is that we can give positive weights only to the AT T pg, tq’s
that we actually identified and estimated. Earlier treated groups have AT T pg, tq estimates for
later event-times by definition (and later treated groups have estimates for earlier pre-trends), so
it will not be possible to include every group in an aggregated event-study parameter at every
event-time. To see which AT T pg, tq’s will contribute to our event study, Figure 7 recenters our
AT T pg, tq estimates in event time instead of calendar time. That is, we plot AT T pg, g ` eq against
e for each expansion group.
We will take “vertical” weighted averages of each available AT T pg, g ` eq for each event time

47
 Figure 7: ATT(g, t) in Event Time

20

10

Treatment Effect
0
Mortality Per 100,000

−10

−20
−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5

2014 2015 2016 2019

Notes: This figure shows the group-time ATT estimates (AT T pg, tq) in relative event time for the four treatment timing groups of
counties that expanded Medicaid before 2019, using not-yet-treated units as the comparison group. The outcome variable is the crude
mortality rate for adults ages 20-64. All estimates use population weights.

e based on Figure 7. For instance, to estimate an aggregate event study in event time 0 (a
measure of instantaneous treatment effects), we would average estimates of AT T p2014, 2014q,
AT T p2015, 2015q, AT T p2016, 2016q and AT T p2019, 2019q. When we are interested in event time
1, we would now average AT T p2014, 2015q, AT T p2015, 2016q, AT T p2016, 2017q. The same logic
applies to other event times.
When constructing timing-group weights at a given event-time, it is also important to account
for group sizes so that the resulting parameters equal sensible averages of treated units. Table
1 contains all the information necessary for this. The 2014 expansion group accounts for 80% of
treated adults in the groups we consider, while the 2016 expansion group accounts for 3.5%. If we
would like our aggregate event study to be a representative summary of the dynamic effects among
treated counties, we should choose weights that are proportional to the treatment group size.
Putting these pieces together, we can formally state the exact summarized causal parameter
that highlights treatment effect dynamics in terms of event time:
„ ˇ ȷ
ˇ
AT Tes peq “ Eω AT T pG, G ` eqˇG ` e P r1, T s, G ď T
ˇ
ÿ
es
“ wω,g,e AT T pg, g ` eq, (5.13)
gă8
es
where each weight wg,e gives the share of a group G “ g among treated units that have been
exposed to treatment for exactly e periods (the groups that we have data for event time e in
Figure 7), and is formally defined as
es
wg,e “ 1tg ` e ď T uPω pG “ g|G ` e ď T, G ď T q.

Note that AT Tes peq gives the average treatment effect among the units that have been exposed to
treatment for exactly e periods, conditional on being observed having participated in the treatment

48
for that number of periods (the condition that G`e P r1, T s) and ever-participating in the treatment
by period T (G ď T ). One can also take a simple average of all available post-treatment event
times, AT Tes peq, e ě 0, and report an overall ATT measure. See Callaway and Sant’Anna (2021)
for a discussion of alternative aggregations based on calendar time and groups.
Estimating AT Tes peq is straightforward and, once again, relies on the plug-in principle: we
need to replace AT T pg, g ` eq with its sample analogs (which we already computed in Figure 6),
and use the relative adult population share of expansion group g among eventually-treated units
as estimates of the event study weights. Figure 8 reports our population-weighted estimates of
the event-study aggregation from event times -5 to 5, which respectively correspond to 5 years
before the Medicaid expansion and 5 years after the Medicaid expansion. We also report pointwise
and simultaneous confidence intervals at the 95% significance level. Overall, the results suggest
that Medicaid expansion has no effect on adult mortality rates among counties that eventually
experience a Medicaid expansion. The pre-trends are also fairly close to zero, suggesting that our
parallel trends assumption may be reasonable.

Figure 8: G ˆ T Event Study without Covariates

Estimate (t ∈ {0, 5}) = 0.03

10 Std. Error = 1.91
Conf. Int = [−3.72, 3.78]

Treatment Effect
Mortality Per 100,000
0

−5

−5 −4 −3 −2 −1 0 1 2 3 4 5
Event Time
Notes: This figure shows the event study estimates with staggered treatment timing using the doubly-robust estimation method from
Callaway and Sant’Anna (2021), using 1,222 not-yet-treated units as the comparison group. The sample sizes used to estimate ATT
parameters for each timing group are 2,200 for the 2014 group, 1,393 for the 2015 group, 1,315 for the 2016 group, and 1,362 for the
2019 group. The outcome variable is the crude mortality rate for adults ages 20-64. The point estimate is reported by the circles,
and both 95% point-wise (black) and simultaneous (red) confidence intervals are reported with the vertical lines. We also report the
simple average of all non-negative event times as a summary of the overall ATT (together with their standard errors and 95% confidence
interval). All results use population weights.

We conclude this section by stressing that the way we have constructed the event study param-
eters in Figure 8 uses all available information from Figure 6. A potential drawback of this strategy
is that we do not always use the same set of groups across all event times. Practitioners usually
refer to this as “imbalance in event time.” For instance, the 2019 expansion group contributes only
to event time e “ 0, not e “ 1 or later event times. When compositional changes are a concern,
one can impose balance in event time and estimate a balanced event study aggregation:
„ ˇ ȷ
ˇ
AT Tes,bal,re,es peq “ E AT T pG, G ` eqˇG ` e P r1, T s, G ` e P r1, T s, G ď T
ˇ

49
 ÿ es,bal
“ wg,re,es AT T pg, g ` eq, (5.14)
gPGtreat
es,bal
where balanced event-time weights wg,re,es are given by
es,bal
wg,re,es “ 1tg ` e ď T u1tg ` e ě 1uPω pG “ g|G ` e P r1, T s, G ` e P r1, T s, G ď T q.
es,bal
Although intimidating, wg,re,es just measures the relative size of a particular treatment group
that was kept in the balanced data. We can interpret AT Tes,bal,re,es peq as the average group-time
average treatment effect among units whose event time is equal to e and is observed to participate
in the treatment for at least e periods, and have at least e available pre-treatment periods (if e is
negative).30

5.2.5 Estimators for staggered designs with covariates

As the discussions in Section 5.2.2 made it clear, we can view the staggered DiD setups as a
collection of simpler 2 ˆ 2 DiD building blocks. A benefit of this interpretation is that, if parallel
trends holds only after conditioning on the covariates that determine untreated potential outcome
changes, we can easily leverage all the results discussed in Section 4 to identify, estimate and
make inference about the AT T pg, tq’s, using regression adjustment, inverse probability weighting,
or doubly robust methods.
Of course, to proceed in this manner, we would need to adopt conditioned on covariates versions
of Assumption PT-GT-Nev, PT-GT-NYT or PT-GT-all, as well as impose an overlap condition.
Given that all these are fairly similar to each other, here we state only an extension of Assumption
PT-GT-NYT and a strong overlap condition that can be used for all three cases.

Assumption CPT-GT-NYT (Conditional Parallel Trends based on not-yet-treated groups).

For every eventually treated group g, not-yet-treated group g 1 , time periods t such that t ě g and
g 1 ą t, and every covariate value Xi ,

Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ g, Xi s “ Eω rYi,t p8q ´ Yi,t´1 p8q|Gi “ g 1 , Xi s.

Assumption SO-GT (Strong overlap with staggered adoption). For every group g P G, the
conditional (weighted) probability of belonging to a treatment group g, given observed covariates
Xi that are determinants of untreated potential outcome growth, is uniformly bounded away from
zero and one. That is, for some ϵ ą 0 and for every group g P G, ϵ ă Pω rGi “ g|Xi s ă 1 ´ ϵ.

Using these identifying assumptions and building on the results in Sections 4 and 5.2.2, we
can follow the arguments in Callaway and Sant’Anna (2021) and establish that the post-treatment
AT T pg, tq’s are identified by the regression-adjusted, inverse-probability-weighted, and doubly-

30
This discussion assumes that there are no “holes” in event-times for each treatment group. It is straightforward
to adjust the interpretation to those more complicated cases, as the same logic can be applied.

50
robust estimands given by31
” ˇ ı
AT Tra pg, tq “ Eω rYi,t ´ Yi,t“g´1 |Gi “ gs ´ Eω Eω rYi,t ´ Yi,t“g´1 |Xi , Gi ą tsˇGi “ g ,
ˇ
”´ ¯ ı
AT Tipw pg, tq “ E wω,G“g pGi q ´ wω,g,t pGi , Xi q pYi,t ´ Yi,t“g´1 q ,
”´ ¯´ ¯ı
AT Tdr pg, tq “ E wω,G“g pGi q ´ wω,g,t pGi , Xi q Yi,t ´ Yi,t“g´1 ´ Eω rYi,t ´ Yi,t“g´1 |Xi , Gi ą ts ,

where pwω,G“g pGi q and wω,g,Gąt pGi , Xi q are the analogs of the weights in (4.7) and are defined as
M
wω,G“g pGq “ ω1tG “ gu Erω1tG “ gus,
O „ ȷ
ω1tG ą tu1tG ‰ gupω,g,t pXq ω1tG ą tu1tG ‰ gupω,g,t pXq
wω,g,t pG, Xq “ E ,
1 ´ pω,g,t pXq 1 ´ pω,g,t pXq
and pω,g,t pXq “ Eω r1tGi “ gu|X, 1tGi “ gu ` 1tGi ą tu “ 1s denote the (weighted) probability of
belonging to the group g given covariates X and that the unit belongs to either to group g—the
treated group for the AT T pg, tq of interest—or the not-yet-treated group Gi ą t— the comparison
group.
Estimating the AT T pg, tq’s follows exactly as in Section 4, and event study aggregations follow
from the arguments in Section 5.2.4. Figure 9 reports event study summary estimates incorporating
covariates into the Medicaid analysis. Like Figure 8, it suggests that Medicaid expansion had
no effect on adult mortality among counties that expanded Medicaid by 2019. Given the point
estimates and uniform confidence interval, we can be reasonably confident that the treatment
effects are not greater than 6 or less than 11 deaths per 100,000 adults for the 6 year period
following Medicaid expansion.

5.3 Limitations of TWFE regressions

Our framework emphasizes building an estimator from 2 ˆ 2 components, each of which targets
a well-defined AT T parameter under a specific parallel trends assumption. The most common
estimator for staggered designs, a TWFE regression, comes instead from extending convenient
estimation tools that work well in the 2 ˆ 2 case. A TWFE specification that estimates a summary
treatment effect parameter is:

Yi,t “ θt ` ηi ` β twf e Di,t ` ei,t . (5.15)

In this section, we abstract from weights.

A major breakthrough in recent DiD research has been to demonstrate two potentially large
problems with β twf e (de Chaisemartin and D’Haultfoeuille, 2020; Goodman-Bacon, 2021; Sun and

31
Wooldridge (2021) propose alternative estimators for the AT T pg, tq that incorporate covariates and their
interactions with group and time dummies into (5.12). Although Borusyak et al. (2024) and de Chaisemartin and
D’Haultfoeuille (2020) also allow for covariates in their estimation procedure, they affect the outcome levels and
not their changes over time. Thus, to some extent, these procedures do not build on conditional parallel trends like
Assumption CPT-GT-NYT.

51
 Figure 9: G ˆ T Event Study with Covariates
20
Estimate (t ∈ {0, 5}) = −2.35
Std. Error = 4.11
Conf. Int = [−10.40, 5.70]
10

Treatment Effect 0
Mortality Per 100,000

−10

−20

−5 −4 −3 −2 −1 0 1 2 3 4 5
Event Time
Notes: This figure shows the event-study estimates with staggered treatment timing, using the doubly-robust estimation method from
Callaway and Sant’Anna (2021). The outcome variable is the crude mortality rate for adults ages 20-64, and the covariates include the
percentage of the county population that is female, the percentage of the county population that is white, the percentage of the county
population that is Hispanic, the unemployment rate, the poverty rate, and county-level median income. The point estimate is reported
by the circles, and both 95% point-wise (black) and simultaneous (red) confidence intervals are reported with the vertical lines. We also
report the simple average of all non-negative event times as a summary of the overall ATT (together with their standard errors and
95% confidence interval). All results use population weights.

Abraham, 2021; Borusyak et al., 2024). The primary issue comes from the fact that TWFE
implicitly uses already-treated comparison groups. Even if PT holds for all groups and all periods,
the resulting estimand can actually put negative weight on certain AT T pg, tq parameters. The only
way TWFE avoids the problem is if treatment effects do not change over time, a strong additional
assumption.
To isolate this issue, consider a setting with two time periods and three groups: a group that
enters treatment in the first period (Gi “ 1), a group that becomes treated in the second time
period (Gi “ 2), and a never treated group (Gi “ 8). This is a staggered design because group
1 and group 2 are treated at different times, but because there are only two time periods, we can
re-write the TWFE specification as

∆Yi,2 “ ∆θt ` β twf e ∆Di,2 ` ∆ei,2 .

Because ∆Di,2 takes only two values—1 for units whose treatment status increases (Gi “ 2), and
0 for units whose treatment status does not change (Gi “ 1 and Gi “ 8)—the TWFE estimand
is the following simple comparison of means:
β twf e “ Er∆Yi,2 |∆Di,2 “ 1s ´ Er∆Yi,2 |∆Di,2 “ 0s
´ ¯ ´ ¯
“ Er∆Yi,2 |Gi “ 2s ´ Er∆Yi,2 |Gi “ 8s p1 ´ w1 q ` Er∆Yi,2 |Gi “ 2s ´ Er∆Yi,2 |Gi “ 1s w1 , (5.16)
p1
where w1 “ p1 `p 8
and pg “ P pG “ gq is group g’s share of units. The TWFE coefficient, in
this case, is a weighted average of two DiD terms that use the already-treated units or the never-
treated units as comparisons. We have already discussed how under PT between group Gi “ 2 and
never-treated units, the first term in the βptwf e decomposition, Er∆Yi,2 |Gi “ 2s ´ Er∆Yi,2 |Gi “ 8s,
equals AT T p2, 2q. But what about the second term with the already treated comparison group?

52
It turns out that under parallel trends and no-anticipation, if we add and subtract different terms,
this type of estimand generally equals a combination of treatment effects for both groups:
Er∆Yi,2 |Gi “ 2s ´ Er∆Yi,2 |Gi “ 1s “ ErYi,2 |Gi “ 2s ´ ErYi,1 |Gi “ 2s
´ ¯
´ ErYi,2 |Gi “ 1s ´ ErYi,1 |Gi “ 1s
“ ErYi,2 p2q ´ Yi,2 p8q|Gi “ 2s ` ErYi,2 p8q ´ Yi,1 p8q|Gi “ 2s
´ ¯
´ ErYi,2 p1q ´ Yi,2 p8q|Gi “ 1s ´ ErYi,1 p1q ´ Yi,1 p8q|Gi “ 1s
´ ErYi,2 p8q ´ Yi,1 p8q|Gi “ 1s
´ ¯
“ AT T p2, 2q ´ AT T p1, 2q ´ AT T p1, 1q
` ErYi,2 p8q ´ Yi,1 p8q|Gi “ 2s ´ ErYi,2 p8q ´ Yi,1 p8q|Gi “ 1s
´ ¯
“ AT T p2, 2q ´ AT T p1, 2q ´ AT T p1, 1q ,

where the first equality follows from the linearity of expectations; in the second, we add and
subtract averages of untreated potential outcomes and explore no-anticipation, and the observation
rules that Yi,t pgq is observed for units with Gi “ g; in the third equality we rearrange terms; and
the last equality follows from a parallel trends for units in group G “ 1 and G “ 2.
The DiD estimand with an already treated comparison group thus equals:

β twf e “ AT T p2, 2q ` AT T p1, 1qw1 ´ AT T p1, 2qw1 . (5.17)

The AT T p1, 2q building block receives negative weight in the overall TWFE estimand unless there
are no treatment effect dynamics (AT T p1, 1q “ AT T p1, 2q). In this example, the problem is easy
to fix: drop the always-treated units and target AT T p2, 2q. With multiple periods, however, the
problem is more complex.
There are two primary results regarding TWFE for general staggered timing designs. One is a
decomposition of the TWFE estimator. Goodman-Bacon (2021) expresses the TWFE estimator as
a weighted average of all possible 2 ˆ 2 DiD comparisons between pairs of groups and time periods
during which one group enters treatment and the other does not. The terms in his decomposition
are not two-period AT T pg, tq-type estimators; they aggregate over the relevant pre- and post-
periods just as β OLS from (5.4) does. They include many comparisons to already-treated units
like (5.16). This mechanical decomposition always has strictly positive weights that are larger for
larger groups and for groups treated closer to the middle of the panel, which have a larger variance
of Di,t conditional on the fixed effects. This result shows how TWFE estimators actually function
and creates a clear link to the estimation approaches we outlined above for summary parameters.
Both are averages of 2 ˆ 2 DiD terms, but they differ in which comparisons they use and how they
aggregate.
The second result is a decomposition of the TWFE estimand, which clarifies why the TWFE
estimator is not guaranteed to identify a desirable parameter in staggered designs, even if par-
allel trends holds. The reason is their use of already-treated comparisons like the ones in (5.17)
(Goodman-Bacon, 2021; de Chaisemartin and D’Haultfoeuille, 2020; Borusyak et al., 2024; Imai

53
 and Kim, 2021; Strezhnev, 2018; Sun and Abraham, 2021). This will tend to bias β twf e away from
the sign of AT Tavg , and β twf e can even have the opposite sign of AT Tavg (Baker et al., 2022). It
may appear that a more flexible regression specification could solve this problem, but Sun and
Abraham (2021) show that a TWFE event study specification suffers from a similar bias when the
dynamics of the AT T pg, tq’s differ across cohorts. Moreover, the “variance-weighting” feature of
OLS means that β twf e has non-intuitive weights even when AT T pg, tq “ AT T pgq.
While TWFE remains common, it has well-understood, potentially serious, and easily remedied
problems, and we do not recommend using it. In many cases, especially those with many untreated
units or minimal treatment effect dynamics, TWFE estimates may be similar to those derived from
the theoretically grounded estimators discussed above. The only way to be sure, however, is to
estimate both. In that case, it is unclear why a researcher would not report the estimates from a
procedure motivated by a desirable target parameter and a credible PT assumption.

6 Conclusion
The starting point of this paper was a 2 ˆ 2 DiD design that researchers have been using for almost
200 years. The end point was a design with five treatment groups, 11 years of data, six covariates,
three types of parallel trends assumptions, and four estimation techniques. Our fundamental
message is that without understanding how complex designs are built up from simpler ones, it is
exceedingly difficult to navigate all the empirical tools now available for DiD designs. This lesson
applies not only to the design details we considered here—weighting, covariates, and staggered
designs—but to any DiD design.
The forward-engineering philosophy we followed in this paper suggests a set of steps that
researchers can follow in any DiD study:

Step 1. Define target parameters. Adopt a potential outcomes notation that fits the study’s specific
setting and use it to define causal target parameters that answer the study’s motivating
question. Building block causal parameters usually aggregate across units using (conditional)
weighted averages, and summary target parameters aggregate across the building blocks.
This step fixes the study’s goals in terms of causal quantities and facilitates comparisons
with related studies.

Step 2. State (formally) the identification assumptions. DiD studies leverage parallel trends assump-
tions, but they also rely on no-anticipation and, in some cases, overlap conditions, or more.
Be explicit about which form of these assumptions is required for identification in the study.
Engage with the theoretical arguments necessary for them to hold and generate appropriate
empirical evidence, such as pre-treatment differential trends, that can falsify or (indirectly)
support their plausibility.

54
Step 3. Determine the appropriate estimation method. In some DiD designs, estimation is as simple
as replacing population expectations with sample means. In others, such as conditional DiD
designs, estimation involves choosing econometric techniques (e.g., a regression adjustment,
inverse probability weighting, or doubly robust procedure) to map theoretical quantities to
estimable sample quantities. Each of these strategies relies on additional modeling restric-
tions that should be stated clearly.

Step 4. Discuss sources of uncertainty. Statistical inference procedures for DiD designs stem from
basic assumptions about where randomness comes from in a given design. Some researchers
may adopt a sampling approach to inference, whereas others may be more comfortable with
a design-based perspective. It is important to discuss what variables of the model are be-
ing treated as fixed and what variables are considered random, as well as to use inference
techniques that are compatible with the model structure and assumptions.

Step 5. Estimate. Steps 1-4 provide a specific structure for using data to estimate the causal param-
eters of interest.

Step 6. Conduct sensitivity analysis. A clear statement of the identification and estimation assump-
tions also facilitates a clear statement of what violations of those assumptions might mean.
No study is robust to all the ways its assumptions may fail, but a good study should be
robust against likely violations of plausible magnitudes. Combine context-specific knowledge
about how the assumptions from Step 2 might be violated, and by how much, with the
structure of the estimator from Step 3 to evaluate how much the DiD estimates vary if the
key identification assumptions are not exactly true.

Step 7. Conduct heterogeneity analysis. Sometimes aggregate parameters defined in Step 1 mask
important heterogeneity, in which case the forward-engineering approach simply suggests
targeting sub-group parameters as well. This can include variation in parameters over time,
between groups of units with different characteristics, or across different sources of treatment
variation. Be clear about which types of heterogeneous effects are relevant and how they are
identified and estimated.

Step 8. Keep learning. DiD is not the only or the best research design in all settings; it is just
one of many causal inference techniques. If the assumptions required for a DiD analysis
appear implausible ex-ante or are refuted by evidence or non-robustness in practice, then
explore different designs. If existing DiD methods do not provide enough guidance, then use
a forward engineering approach to deduce what advances would help.

Some researchers may still prefer to use standard regression tools to conduct DiD studies.
The properties and pitfalls of some popular regression specifications are now well understood, and
one can easily explain how this choice fits with (and perhaps satisfies) the steps above. But using

55
simple regressions in any DiD-type setting is an implicit choice to reverse-engineer a research design
from the statistical method, rather than forward-engineer a reliable estimator from a substantive
question and transparent assumptions. Ultimately, important questions and credible identification
strategies should guide DiD analyses (regression-based or not), not the other way around.
Although this paper is by no means an exhaustive guide to DiD practice, the eight steps above
are a rigorous framework for tackling all the DiD topics that we did not cover. The Appendix briefly
discusses DiD methods for (a) treatments that turn on and off over time, (b) continuous and multi-
valued treatments, (c) triple differences, (d) distributional parameters, and (e) repeated cross-
section or unbalanced panel data. While each of these designs differs from what we covered in the
main text, a forward-engineering approach that moves from defining parameters and assumptions,
to settling on estimation and inference techniques, to probing robustness, applies equally to all
of them. While the specifics of any given DiD analysis may change across research questions,
treatment variables, econometric techniques, and data structures, the principles by which one can
conduct reliable and transparent causal inference stay the same.32

32
There are other DiD topics of interest that we do not cover, including fuzzy DiD and instrumented DiD designs
(de Chaisemartin and D’Haultfoeuille, 2018; Miyaji, 2024), nonlinear DiD models (Wooldridge, 2023; Tchetgen Tch-
etgen, Park and Richardson, 2024), issues related to few clusters (Roth et al., 2023, Section 5), and situations with
multiple treatments (de Chaisemartin and D’Haultfoeuille, 2023a; Yanagi, 2023). We also do not cover some meth-
ods that address violations of parallel trends (Freyaldenhoven et al., 2019; Arkhangelsky, Athey, Hirshberg, Imbens
and Wager, 2021; Callaway and Karami, 2023; Imbens, Kallus and Mao, 2021), nor do we examine setups that
impose as-good-as-random treatment timing (Athey and Imbens, 2022; Roth and Sant’Anna, 2023a; Arkhangelsky,
Imbens, Lei and Luo, 2024).

56
 List of Acronyms
Acronym Definition
2ˆ2 Two-Group Two-Time-Periods DiD
2ˆT Two-Group T -Time-Periods DiD
ACA Affordable Care Act
ATT Average Treatment Effect on the Treated
CPT Conditional Parallel Trends
CPT-GT-NYT Conditional Parallel Trends Based on Not-Yet-Treated Groups
DiD Difference-in-Differences
DR Doubly Robust
ETWFE Extended Two-Way Fixed Effects
IPW Inverse Probability Weighted
NA No Anticipation
NA-S No Anticipation with Staggered Treatment Timing
OLS Ordinary Least Squares
PT Parallel Trends
PT-ES Parallel Trends Event Study
PT-GT-all Parallel Trends for Every Period and Group
PT-GT-Nev Parallel Trends Based on Never-Treated Groups
PT-GT-NYT Parallel Trends Based on Not-Yet-Treated Groups
RA Regression Adjustment
SO Strong Overlap
SO-GT Strong Overlap With Staggered Adoption
TWFE Two-Way Fixed Effects

57
A Some additional DiD-related procedures
This section discusses some important DiD-related topics that we did not cover in our main text.
These discussions are short by design, and we focus on providing the main ideas related to chal-
lenges and solutions specific to the problem. We abstract from weights and use Er¨|¨s to denote
(conditional) expectations.

A.1 Setups with treatment turning on and off

Our main text focuses on setups where treatment remains in place from the period it begins until
the end of the sample period, but in practice, some treatments turn on and off over time. This is
the setting tackled by de Chaisemartin and D’Haultfoeuille (2020, 2023a), Imai, Kim and Wang
(2023), and Liu et al. (2024).
To tackle this problem from first principles, we need to augment the potential outcomes to reflect
the richer notion of treatment sequences. Following Robins (1986), let Yi,t pdq denote the potential
outcome for unit i at time t if this unit received the T -dimensional treatment sequence d P t0, 1uT .
For simplicity, let’s say that T “ 3 and that no unit is treated in the first period. In this case, we
have four treatment sequences (or histories), which define four potential outcomes for each unit:
Yi,t p0, 0, 0q, Yi,t p0, 0, 1q, Yi,t p0, 1, 0q and Yi,t p0, 1, 1q. We then define treatment groups by treatment
sequences: G “ d0 ” p0, 0, 0q (never-treated), G “ d1 ” p0, 0, 1q (treated in the third period), G “
d2 ” p0, 1, 1q (treated in the second and third period), and G “ d3 ” p0, 1, 0q (treated only in the
second period). In general, we would have as many groups as we have different (realized) treatment
sequences. Recall that in a staggered timing design with an absorbing treatment, treatment timing
fully characterizes a treatment sequence.
Once potential outcomes and groups are well-defined, one can move to parameters of interest.
We proceed similarly to the staggered treatment setup in Section 5.2 and consider group-and-time
specific ATTs as building blocks, except that groups are now based on more complex treatment
sequences. Let 0 denote a T-dimensional vector of zeros. One intuitive building block parameter
on which to base a DiD analysis is

AT T pd, tq “ ErYt pdq ´ Yt p0q|G “ ds,

the average treatment effect at time period t of being exposed to treatment sequence d instead of
never being exposed to treatment, among units that received treatment sequence d.33
Next, one needs to establish identification for the parameters, and propose appropriate estima-
tors and inference procedures. Following similar arguments to those in Section 5.2, a DiD approach
to this problem would involve imposing a parallel trends assumption (potentially conditional on
covariates) and a no-anticipation assumption to establish that the AT T pd, tq’s are identified. If

33
One could also adopt alternative building blocks not discussed here, such as the average effect of treatment
lasting one period longer or a treatment spell of a given length beginning one period later.

58
each treatment group is sufficiently large, one could proceed in a similar fashion as the staggered
setup, comparing average outcome paths for a given sequence with the average outcome path for
never-treated (or not-yet-treated) units. One can also aggregate these different AT T pd, tq to form
different summary parameters.
In practice, however, it is often the case that the number of treatment groups is large and each
group is small. This essentially creates a “curse of dimensionality” problem: there are too many
building block parameters defined for too-small groups to be estimated reliably. In such cases,
additional assumptions that limit treatment effect dynamics (or how past treatments affect future
outcomes) are often imposed, and different aggregated summary parameters are usually targeted.
We provide a brief overview of several different solutions that have been proposed to address this
issue.
de Chaisemartin and D’Haultfoeuille (2020) impose a “no-carryover” assumption that implies
that past treatments do not affect future outcomes; which is to say that treatment effects in a given
period last only during that period. With such an assumption (in addition to parallel trends and a
no-anticipation assumption), they propose DiD estimators for an instantaneous average treatment
effects parameter by comparing currently treated units with untreated units. Imai et al. (2023)
adopt a similar approach, though they impose a limited-carryover assumption where treatments
may last for ℓ periods (with ℓ specified by the researcher). They then propose estimators for an
average treatment effect of switching into treatment in period t among units that experience the
policy change in period t, and share the same treatment history over the previous k periods; see Liu
et al. (2024) for a related procedure. Finally, de Chaisemartin and D’Haultfoeuille (2023a) avoid
making assumptions related to carryover effects and extend the DiD framework in de Chaisemartin
and D’Haultfoeuille (2020) to allow for treatment effect dynamics. The way they proceed is to
first “staggerize” treatment sequences according to first-time of treatment exposure, compute a
staggered DiD procedure for this “intention-to-treat” type parameter, and normalize them by a DiD
estimate based on the number of treated periods. A potential challenge with de Chaisemartin and
D’Haultfoeuille’s (2023a) approach is the interpretability of their proposed summary parameter,
though we should acknowledge that this is a complex setup.
One important takeaway is that comparing these DiD procedures that allow for treatments to
turn on and off may be challenging, as they target different causal parameters of interest, and
practitioners should be aware of the different assumptions and limitations. We refer the reader to
de Chaisemartin and D’Haultfoeuille (2023b) and Liu et al. (2024) for additional discussions on
these types of DiD estimators.

A.2 DiD setups with continuous or multi-valued treatments

Our paper focuses on binary treatments, but many treatments take multiple values or are even con-
tinuous. A number of recent papers have studied this particular type of treatment design. These

59
include Callaway, Goodman-Bacon and Sant’Anna (2021, 2024); de Chaisemartin, D’Haultfoeuille,
Pasquier and Vazquez-Bare (2024a); and de Chaisemartin, D’Haultfœuille and Vazquez-Bare
(2024b). Here we focus on a two-period setting in which no unit is treated in period one and
some units receive a treatment with varying intensities (or doses) in period two. Most of the
key results that distinguish multi-valued from binary treatments are evident with two periods
(Callaway et al., 2024).
We now need to define potential outcomes that reflect varying treatment intensity. We denote
Yi,t p0, dq as potential outcomes for unit i in period t if they are untreated in period one and receive
treatment dosage d in period two. As we focus on setups where all units are untreated in period
one, we simplify notation and index potential outcomes by treatment intensity in period two; that
is, Yit pdq “ Yi,t p0, dq. An important feature is that d is not restricted to t0, 1u and can take on
richer treatment intensities instead. We denote the treatment dosage for unit i as Di in period
two and stress that in this context, our notion of the treatment group is tied to units’ treatment
dosage: groups are defined by their treatment dosage in period 2.
A multi-valued treatment defines several different types of causal parameters that may be of
interest. For instance, dose-specific average treatment effect parameters such as

AT T pd|d1 q “ ErYt“2 pdq ´ Yt“2 p0q|D “ d1 s and AT Epdq “ ErYt“2 pdq ´ Yt“2 p0qs,

reflect the average effect of dose d relative to no treatment. Here AT T pd|d1 q is the average treatment
effect for units that experienced dose d1 ; when d1 “ d, it is the AT T among units that received
dose d. On the right side, AT Epdq is defined analogously, except that it is the effect on the overall
population. Of course, one can also aggregate these dose-specific parameters to form more precisely
estimable summary quantities; see, e.g., Callaway et al. (2021).
The two treatment effect parameters above provide average treatment effects in levels, and so
one reason why they vary could be because d itself varies. To account for the differences in d, one
may be interested in “per-dosage” effects:
AT T pd|d1 q AT Epdq
AT Tpd pd|d1 q “ and AT Epd pdq “ .
d d
One can also aggregate these parameters across dosages to analyze ErAT Tpd pD|Dq|D ą 0s, an av-
erage treatment effect among treated (or, more generally, among switchers). One can also consider
weighted averages of these to learn about ErAT T pD|Dq|D ą 0s{ ErD|D ą 0s; see de Chaisemartin
et al. (2024a) for a general discussion about such target parameters.
Finally, researchers are often interested in the causal effect of a marginal increment in the dose.
This notion is the average causal response (ACR), similar to Angrist and Imbens (1995), defined
as follows (when the dose is absolutely continuous):
1 ˇ
ˇ 1 ˇ
ˇ
BAT T pl|d q B ErY t“2 plq|D “ d s BAT Epdq B ErYt“2 pdqs
ACRT pd|d1 q “ ˇ “ ˇ and ACRpdq “ “ .
Bl l“d
ˇ Bl ˇ
l“d Bd Bd
Here ACRT pd|dq equals the derivative of the average potential outcome in period two for units

60
that received dose d evaluated at d—this is equivalent to the derivative of AT T pl|dq with respect
to l, evaluated at l “ d. We can interpret ACRpdq analogously.34
The relevant questions pertain to (a) what assumptions are needed to impose to identify these
parameters, (b) how to estimate and make inferences about these parameters of interest once
identification is established, (c) how to summarize treatment effect heterogeneity across doses
to generate interpretable aggregated causal parameters, and (d) whether traditional regression
specifications based on TWFE recover a sensible and easy-to-understand causal parameter of
interest. These questions are addressed in detail by Callaway et al. (2021) and de Chaisemartin et
al. (2024a).
Callaway et al. (2021) highlight how, when no units are treated in period one, identification
and estimation of AT T pd|dq’s (or their functionals) follows the binary case. They propose flexible
nonparametric estimators for the AT T pd|dq curve—the relationship between outcome changes
(minus the average change for untreated units) and the dose d, making it possible to visualize and
make inference about treatment effect heterogeneity across dosages. They also propose estimators
that aggregate across dosage values and can be more precisely estimated. The identification of
causal response parameters or ATE-type parameters, however, requires a stronger version of parallel
trends that holds for potential outcomes at non-zero treatment doses. Under these strong parallel
trends and no anticipation assumptions, they discuss estimation and inference procedures for the
ACR curves and their summary measures.35
de Chaisemartin et al. (2024a) consider the setup where units are already exposed to different
levels of treatment in period one. They discuss how one can identify causal quantities that gener-
alize AT Tpd pd|d1 q to this more complex setup when (a) a sizable number of units do not change
treatment dosage over time (stayers), and (ii) there is no-carryover from past treatment to future
outcomes. They propose estimation and inference procedures for aggregated parameters akin to
ErAT Tpd pD|Dq|D ą 0s and ErAT T pD|Dq|D ą 0s{ ErD|D ą 0s.
Lastly, these papers target different causal parameters, put more emphasis on different DiD
designs, and, therefore, should be viewed as complements rather than substitutes. In our view, DiD
with continuous treatment is another area in which more methodological research is warranted.
See Callaway et al. (2021) and de Chaisemartin et al. (2024a) for a more thorough discussion of
many other cases.

34
For discrete treatments, ACR’s are defined in a similar way but with a slightly different notation to accom-
modate the discreteness of d: ACRT pdj |dk q “ ErYt“2 pdj q ´ Yt“2 pdj´1 q|D “ dk s, and ACRpdj q “ ErYt“2 pdj q ´
Yt“2 pdj´1 qs.
35
Interestingly, they also show that commonly used TWFE regression specifications are too rigid to lead to easy-
to-interpret causal parameters of interest. In fact, they show that one can provide several different decompositions
of the TWFE treatment coefficient depending on the specific causal parameter being used as a building block for the
analysis, though every decomposition considered by them has some issues related to negative-weighting, additional
“bias” terms, or non-interpretable weights that can distort inference. They emphasize that all this can be easily
resolved by adopting the forward-engineering approach.

61
A.3 Triple differences
The causal interpretation of DiD estimates depends on the plausibility of their identification as-
sumptions, which involve a no-anticipation and a parallel trends condition. In some applications,
however, these assumptions may not hold—for example, when the trends of average untreated out-
comes among men and women vary across treatment groups. In these cases, a common empirical
practice is to attempt to model these violations of parallel trends directly or to conduct sensitivity
analysis (Freyaldenhoven et al., 2024; Rambachan and Roth, 2023). In some specific treatment de-
signs in which treatment is rolled out to different units or groups (e.g., states), but is targeted to a
specific subset (partition) of the population (e.g., women), it is possible to relax DiD-type parallel
trends so that partition-specific and group-specific violations of parallel trends are allowed. Such
setups are often referred to as “triple differences” (DDD). Since its introduction by Gruber (1994),
DDD has become very popular among empirical researchers—see Olden and Møen (2022) for doc-
umentation. In this section, we provide a brief overview of the target parameters and identifying
assumptions in DDD. We also highlight that, contrary to conventional wisdom, DDD procedures
cannot generally be expressed as the difference between two DiD, especially when parallel trends
assumptions only hold after conditioning on covariates or when treatment adoption is staggered.
This discussion borrows heavily from Ortiz-Villavicencio and Sant’Anna (2025).
We start our analysis by discussing potential outcomes and treatment design. As we focus
on binary treatments (with potential staggered adoption), the potential outcome is the same as
discussed in the main text, with Yi,t pgq denoting the potential outcome for unit i in time t if first
exposed to treatment in period g. In DDD setups, a unit i is exposed to treatment in period t if
(i) it belongs to a group (e.g., state) that enabled treatment in period g and t is a post-treatment
period, t ě g, and (ii) it belongs to the subset of the population that qualifies (or is eligible) for
treatment (e.g., women). Let S P S Ď t2, ..., T u Y t8u denote the time each group (e.g., state)
enables the policy/treatment, with the notion that S “ 8 if the policy is not enabled in the
observed time frame. We also denote the partition of the population that (eventually) qualifies
for the treatment by Q with Qi “ 1 if unit i is (eventually) eligible for treatment and Qi “ 0
otherwise. With these notations, we can define the treatment groups Gi according to the first time
a unit i is exposed to treatment; that is, Gi “ Si if Qi “ 1 and Gi “ 8 if Qi “ 0.36
Similar to standard DiD designs, DDD is interested in the AT T pg, tq-type parameters discussed
in Section 5.2.1. Given the particular structure of the DDD problem, we can write AT T pg, tq’s as

AT T pg, tq ” ErYi,t pgq ´ Yi,t p8q|Gi “ gs “ ErYi,t pgq ´ Yi,t p8q|Si “ g, Qi “ 1s,

to stress that it measures the average treatment effect at time period t of first being exposed
to treatment in period g versus not being exposed to treatment, among units that are actually
exposed to treatment in period g, i.e., units that are in groups that the policy was first enabled
36
Note that when all units are eligible for treatment, we have Gi “ Si , which gets us back to a (staggered) DiD
setup.

62
in period g and that qualify for treatment. One can also analyze aggregations of these AT T pg, tq
parameters to form causal summary parameters that can be more precisely estimated and highlight
treatment effect heterogeneity in some specific directions. This would follow the exact same steps
as we discussed in Section 5.2.4, once again highlighting the importance of our forward-engineering
approach.
Identifying these causal parameters involves a no-anticipation assumption and a (conditional)
parallel trends assumption. Assumption NA-S can be recycled here, as DDD has the same empirical
content as DiD when it comes to no-anticipation. The parallel trends assumption, though, needs
to be adjusted as an empirical appeal of DDD is to allow for violations of Assumption PT-GT-all,
or the other PT variations discussed in Section 5.2. Here, we consider a variation of Assumption
PT-GT-all that holds only after conditioning on covariates and allows for some partition-specific
and group-specific non-parallel trends.

Assumption DDD-PT-GT-all (DDD-Parallel Trends for every period and group). For every
group s and s1 and time periods t, with probability one,

E rYt p8q ´ Yt´1 p8q|S “ s, Q “ 1, Xs ´ E rYt p8q ´ Yt´1 p8q|S “ s, Q “ 0, Xs

“
E rYt p8q ´ Yt´1 p8q|S “ s1 , Q “ 1, Xs ´ E rYt p8q ´ Yt´1 p8q|S “ s1 , Q “ 0, Xs .

When there are only two periods, t “ 1, 2, and two groups, S P t2, 8u, and covariates play
no role in terms of identification—that is, Assumption DDD-PT-GT-all holds without X (or,
equivalently, with X “ 1 for all units)—Olden and Møen (2022) show that one can identify
AT T p2, 2q as the difference of two DiD estimands:

AT T p2, 2q “ E rYt“1 ´ Yt“1 |S “ 2, Q “ 1s ´ E rrYt“1 ´ Yt“1 |S “ 2, Q “ 0s

` ˘
´ E rYt“1 ´ Yt“1 |S “ 8, Q “ 1s ´ E rrYt“1 ´ Yt“1 |S “ 8, Q “ 0s .

Estimation and inference would be straightforward, as one could use the analogy principle or a
two-way fixed effects regression with triple interactions—see Olden and Møen (2022) for details.
Ortiz-Villavicencio and Sant’Anna (2025) show that DDD estimands cannot be written as the
difference of two DiD estimands when covariates are important for identification, or when treatment
adoption is staggered over time and one wants to use not-yet-treated units as a comparison group
(as is commonly done in DiD setups). They show how ignoring these considerations and proceeding
as if DDD were indeed just a difference of two DiDs can lead to severely biased estimates for the
AT T pg, tq’s. Ortiz-Villavicencio and Sant’Anna (2025) also show how one can avoid these issues by
adopting a forward-engineering approach to the DDD problem. They propose regression-adjusted,
inverse probability weighting, and doubly robust estimators for DDD setups that can reliably
recover AT T pg, tq and their associated summary parameters under mild assumptions. The paper
discusses using multiple comparison groups to generate more precise estimates than simply using

63
a single comparison group. Relatedly, Strezhnev (2023) discusses several limitations of common
two-way fixed effects regression specifications commonly used for DDD analysis.
Sometimes researchers use the term “triple differences” to mean different things and often use
different identification assumptions to estimate these different quantities. Caron (2025) discusses
using a triple difference strategy to estimate treatment effect heterogeneity. We recommend that
practitioners be transparent about target parameters, research designs, and identification assump-
tions to allow the research community to understand the goals and the differences between DDD
procedures.

A.4 Distributional DiD procedures

Our paper focuses on learning about average treatment effects in various DiD setups. However,
approaches that embrace heterogeneity can also target quantities that describe heterogeneity other
than average treatment effect parameters. In some settings, researchers may want more information
about the distributional impacts of treatment participation. For instance, if a policymaker faces two
different labor market programs with very similar average effects on earnings, they may prefer the
one that potentially has a higher impact on the lower tail of the income distribution. Difference-
in-Differences-type strategies can also be used to identify, estimate, and make inferences about
various distributional features of the outcome of interest. This area has received a substantial
amount of methodological consideration by econometricians in recent years; see Athey and Imbens
(2006), Bonhomme and Sauder (2011), Callaway, Li and Oka (2018), Callaway and Li (2019), Roth
and Sant’Anna (2023b), Ghanem, Kédagni and Mourifié (2023), Fernández-Val, Meier, van Vuuren
and Vella (2024b), and references therein. For some empirical literature using distributional DiD
procedures, see Meyer, Viscusi and Durbin (1995), Finkelstein and McKnight (2008), and Cengiz
et al. (2019), among many others.
An analysis of distributional quantities does not require different potential outcomes notation
from Section 5.2; it just targets functionals of the potential outcome distributions other than their
means. The first thing to notice is that there are several types of distributional causal parameters
in the treated group that one may care about. The unique feature of them is that they are all
functionals of FYt pgq|G“g pyq “ PpYt pgq ď y|G “ gq and FYt p8q|G“g pyq ” PpYt p8q ď y|G “ gq.
Examples of such functionals include distributional treatment effects in time period t among units
first treated in period g (denominated in probability units),

DT T py|g, tq “ FYt pgq|G“g pyq ´ FYt p8q|G“g pyq,

quantile treatment effects in time period t among units first treated in period g (denominated in
outcome units),

QT T pτ |g, tq “ FY´1
t pgq|G“g
pτ q ´ FY´1
t p8q|G“g
pτ q,

where FY´1
t pgq|G“g
pτ q “ inf ty : FYt pgq|G“g pyq ě τ u denotes the τ -quantile of Yt pgq among units in

64
group G “ g, and FY´1 t p8q|G“g
pτ q is defined analogously. Other functionals related to inequality
measures can also be obtained; see Firpo and Pinto (2016) for a discussion on this topic.
To make inferences about these different causal parameters, one needs to identify FYt p8q|G“g pyq
and FYt pgq|G“g pyq. Identification of FYt pgq|G“g pyq is usually non-controversial, as we can use data
from units in group G “ g to learn about the distribution of Yt pgq. The main challenge is re-
lated to how to learn the counterfactual distribution FYt p8q|G“g pyq from the data. This is where
different DiD-type procedures differ, as each paper in this literature relies on different and of-
ten non-nested identification assumptions that allow them to identify FYt p8q|G“g pyq. Given the
space constraints, we do not provide explicit and detailed discussion about how these different
DiD-related distributional procedures function. However, all distributional DiD estimators share
our forward-engineering approach; they clearly state their identification assumptions and target
parameters and then provide estimators that recover well-defined causal quantities. We also note
that most distributional DiD methodological papers focus on two-period and two-group setups.
However, it is straightforward to build similar arguments to those in Section 5.2 to extend the de-
signs to more general settings, which is again another benefit of the forward-engineering approach
to causal inference.
We close this section by noting that there exist other types of distributional parameters of
interest related to the distribution of the treatment effects in period t among the units in group
g, PpYt pgq ´ Yt p8q ď y|G “ gq. In general, such causal quantities cannot be point identified, as
discussed in Heckman, Smith and Clements (1997b), Fan and Yu (2012) and Callaway (2021).
However, often one can still partially identify such policy-relevant parameters under different re-
strictions. We refer the reader to Callaway (2021) for a more detailed discussion of this topic.

A.5 Repeated cross-sections and unbalanced panel data

An appealing feature of DiD procedures is that, although helpful, a balanced panel is not a re-
quirement for DiD analyses, which can also be deployed with repeated cross-sectional data or
unbalanced panels. Indeed, as discussed in Section 3.3 and made explicit in equation 3.6, the
2 ˆ 2 building block in unconditional DiD analyses involves only averages that are group and time
specific, and does not require the same unit to be observed in all periods. As discussed in Callaway
and Sant’Anna (2021), the same applies to unconditional staggered adoption setups, and one need
not enforce a balanced panel even within each subset of the data used to estimate the AT T pg, tq
building blocks. One caveat is that the interpretation of the parameter of interest may change,
which we discuss more below.
When covariates are available and play an important role in the plausibility of the identifica-
tion assumptions, the differences between DiD with a balanced panel and repeated cross-sections
(or unbalanced panel) are subtle, can be practically important, and are often not discussed in
methodological papers. The gist of the problem relates to potential compositional changes over

65
time. Most DiD papers that rigorously discuss repeated cross-section setups, including Abadie
(2005), Sant’Anna and Zhao (2020), and Callaway and Sant’Anna (2021), rule out compositional
changes by assuming that the joint distribution of covariates and treatment groups is invariant
over time, a stationarity-type assumption. However, this may not be warranted in empirical ap-
plications, and erroneously imposing this additional assumption can lead to biases (Hong, 2013;
Sant’Anna and Xu, 2023). On the other hand, when this stationarity assumption is justified and
correctly used, the gains in power when conducting inference for DiD parameters can be noticeable
(Sant’Anna and Xu, 2023). In what follows, we use the 2 ˆ 2 setup to explain how compositional
changes can complicate the analysis and why ruling it out leads to a gain in precision.
To see how issues related to compositional changes affect the analysis, let us first assume that
there are no compositional changes and that the stationarity assumption is valid. In this case, the
average treatment effect on the treated in period two (post-treatment) can be written as

AT T p2q ” ErYi,t“2 p1q|Di “ 1s ´ ErYi,t“2 p0q|Di “ 1s

“ ErYi,t“2 p1q|Di “ 1, Ti,t“2 “ 1s ´ ErYi,t“2 p0q|Di “ 1, Ti,t“2 “ 1s
“ ErYi p1q|Di “ 1, Ti,t“2 “ 1s ´ ErYi p0q|Di “ 1, Ti,t“2 “ 1s, (A.1)

where Ti,t is an indicator if unit i is observed in period t, Yi pdq “ Ti,t“2 Yi,t“2 pdq ` Ti,t“1 Yi,t“1 pdq
is the potential outcome for unit i, Di “ 1tGi “ 2u is a treatment group dummy that equals one
if a unit is first treated in period two and zero if it is untreated in both periods. We also set Xi
to be a vector of (pre-treatment) covariates. Note that even here, we already use the stationarity
condition that the joint distribution of pDi , Xi q is invariant to Ti,t“2 to move from the first to the
second line and establish (A.1).
To identify AT T p2q it is often constructive to first establish the identification of its conditional-
on-covariates analog; that is, the conditional ATT in period two among units with covariates Xi ,
AT TXi p2q. This is exactly how we proceeded in Section 4.2. Under the stationarity condition, and
similarly to (A.1), we can express this quantity as37

AT TXi p2q ” ErYi,t“2 p1q|Di “ 1, Xi s ´ ErYi,t“2 p0q|Di “ 1, Xi s

“ ErYi p1q|Di “ 1, Xi , Ti,t“2 “ 1s ´ ErYi p0q|Di “ 1, Xi , Ti,t“2 “ 1s. (A.2)

Next, we have to establish the identification of this quantity. As expected, we will again
use conditional parallel trends, no-anticipation, and overlap assumptions. The no-anticipation
condition used here is the same as the one in the main text. The conditional parallel trends
and overlap assumptions need to be modified, as we now work with multiple partitions of the
data depending on treatment status and the period a unit is observed. In this sense, we modify
Assumptions CPT and SO to the following related, but different, assumptions. These modifications
are warranted regardless of whether compositional changes are present; this step is instead tied to

37
To guarantee that all the conditional expectations in (A.2) are well-defined, we need an overlap condition that
guarantees that P pTi,t“2 “ 1, Di “ 1|Xi q ą 0. We discuss this below.

66
data structure.38

Assumption CPT-RCS (2 ˆ 2 Conditional Parallel Trends with repeated cross-sections). We

assume that, with probability one,

ErYi,t“2 p0q|Xi , Di “ 1, Ti,t“2 “ 1s ´ Eω rYi,t“1 p0q|Xi , Di “ 1, Ti,t“1 “ 1s

“ (A.3)
ErYi,t“2 p0q|Xi , Di “ 0, Ti,t“2 “ 1s ´ ErYi,t“1 p0q|Xi , Di “ 0, Ti,t“1 “ 1s.

Assumption SO-RCS (Strong overlap with repeated cross-sections). For some ϵ ą 0 and every
pd, sq P t0, 1u ˆ t0, 1u, ϵ ă P rDi “ d, Ti,t“2 “ s|Xi s ă 1 ´ ϵ.

With this modification, we can now show that when Assumptions NA, CPT-RCS, and SO hold,
the conditional ATT parameter AT TXi p2q is identified by39

AT TXi p2q “pErYi |Di “ 1, Ti,t“2 “ 1, Xi s ´ ErYi |Di “ 1, Ti,t“1 “ 1, Xi sq

´ pErYi |Di “ 0, Ti,t“2 “ 1, Xi s ´ ErYi |Di “ 0, Ti,t“1 “ 1, Xi sq. (A.4)

This step provides a methodological justification to estimate the AT TXi p2q’s using four conditional
expectations that use only the available data. In addition, it highlights that, under the stationarity
assumption, once we learn the AT TXi p2q’s, we can aggregate them using the covariate distribution
of treated units available from both time periods to get the AT T p2q. More formally, AT T p2q is
identified by

AT T p2q “ ErAT TXi p2q|Di “ 1s

“ ErAT TXi p2q|Di “ 1, Ti,t“2 “ 1sP pTi,t“2 “ 1|Di “ 1q
` ErAT TXi p2q|Di “ 1, Ti,t“1 “ 1sP pTi,t“1 “ 1|Di “ 1q,

where AT TXi p2q is given by (A.4). This is the second place where ruling out compositional changes
and imposing the stationarity assumption helps: it allows you to use covariates from treated units
in the entire dataset to recover AT T p2q. The fact that you can pool data across all periods to
learn about AT T p2q translates to gains in power, as formally discussed by Sant’Anna and Zhao
(2020) and Sant’Anna and Xu (2023). The third place where the stationarity condition affects the
analysis is in the characterization of how “the most precise” (regular and asymptotically linear)
estimator for the AT T p2q should look. This point relates to the semi-parametric efficiency bound
and the construction of efficient (and doubly robust) estimators. As these points are slightly more
technical, we refer the readers to Sant’Anna and Zhao (2020) and Sant’Anna and Xu (2023) for
more details.
38
In setups where we rule out compositional changes and impose the stationarity condition that the joint distribu-
tion of pDi , Xi q is invariant to Ti,t“2 , we may not need to modify Assumption CPT. We do it here for transparency
purposes.
39 n
We also require the assumption that the pooled repeated cross-section data tYi , Di , Xi , Ti,t“2 , Ti,t“1 ui“1 is iid,
though this is fairly standard and uncontroversial; see, for instance, Abadie (2005) and Sant’Anna and Zhao (2020,
Assumption 1). We maintain this condition as an assumption throughout this section.

67
 Overall, ruling out compositional changes over time allows for the pooling of information across
the entire dataset, which has an impact on the definition of target parameters and leads to more
precise inference procedures. But what happens when this condition fails? How does this affect
the analysis?
First, this matters for the definition of the treatment effect of interest. In setups where the
sampling varies across periods, we do not have a single notion of AT T p2q. Instead, we need to
accommodate the fact that the AT T p2q may vary across units sampled from different periods.
Thus, when we do not rule out compositional changes, we must be explicit about the treated
subpopulation that we are interested in. It is common to focus on the average treatment effect in
period two among treated units that are also sampled in period two, that is,40

AT T p2|Tt“2 “ 1q ” ErYi,t“2 p1q|Di “ 1, Ti,t“2 “ 1s ´ ErYi,t“2 p0q|Di “ 1, Ti,t“2 “ 1s

“ ErYi p1q|Di “ 1, Ti,t“2 “ 1s ´ ErYi p0q|Di “ 1, Ti,t“2 “ 1s. (A.5)

Although (A.5) has the same statistical estimand as (A.1)—that is, the formulas on the right-
hand side of the equation coincide—it has a very different interpretation. It is the AT T p2q among
units sampled in period 2, and is not an “overall” AT T p2q. One may think this difference is
merely cosmetic, but as discussed below, this has implications for constructing estimands when
covariates are important for identification. Where covariates do not play an important role, it is
simply a matter of changing the interpretation of your reported estimates (which also applies to
unconditional staggered setups, to be clear).
When covariates do play an important role, however, and when Assumptions CPT-RCS,
NA and SO-RCS hold, the conditional ATT parameter among units sampled in period two,
AT TXi p2|Tt“2 q is identified by

AT TXi p2|Tt“2 q “ pErYi |Di “ 1, Ti,t“2 “ 1, Xi s ´ ErYi |Di “ 1, Ti,t“1 “ 1, Xi sq

´pErYi |Di “ 0, Ti,t“2 “ 1, Xi s ´ ErYi |Di “ 0, Ti,t“1 “ 1, Xi sq, (A.6)

which, in turn, implies that AT T p2|Tt“2 “ 1q is identified by

AT T p2|Tt“2 “ 1q “ ErAT TXi p2|Tt“2 q|Di “ 1, Ti,t“2 “ 1s. (A.7)

Several remarks are worth making. First, the statistical estimand in (A.6) is the same as when
one rules out compositional changes as in (A.4), suggesting that, once again, what changes in
this step is the interpretation. However, these interpretative issues have direct consequences on
the appropriate method for aggregating across covariate values. As clearly stated in (A.7), in
the presence of potential compositional changes, one is not allowed to pool information across
periods to identify (and also estimate and make inference) about AT T p2|Tt“2 “ 1q. As discussed

40
One may also be interested in the ATT in period two among treated units that are sampled in period one.
The arguments required to establish (point) identification of this parameter differ from those we use here. A main
challenge is that we do not observe ErYi,t“2 p1q|Di “ 1, Ti,t“1 “ 1s, and the parallel trends assumption we leverage
does not involve treated potential outcomes.

68
in Sant’Anna and Xu (2023), ignoring these issues and pooling data from all periods in the presence
of compositional changes leads to a bias that is important to be aware of. We refer the reader to
Sant’Anna and Xu (2023) for a discussion related to unbalanced panels and also on a discussion
about doubly robust and semiparametric efficient DiD estimators under compositional changes. We
are not aware of any papers that formally extend the discussion in Sant’Anna and Xu (2023) to
staggered DiD designs. Still, this extension is surely possible by following our forward-engineering
approach to DiD.
We close this section by highlighting that, in practice, it is possible to test for composi-
tional changes by comparing the estimates from estimators that impose it and those that do
not. Sant’Anna and Xu (2023) discuss Hausman-type tests in the two-period setting, though one
can extend those to more general setups. We also highlight that when it comes to DiD setups
with staggered adoption, some equivalence results discussed in Section 5.2 no longer hold with
repeated cross-sections or unbalanced panel data. For instance, the Sun and Abraham (2021)
regression-based strategy to estimate AT T pg, tq’s using (5.11) no longer coincides with Callaway
and Sant’Anna’s (2021) estimators using the analog of (5.10):

AT
z T never pg, tq “ pY G“g,t ´ Y G“g,t“g´1 q ´ pY G“8,t ´ Y G“8,t“g´1 q,

where Y G“a,t“s is the sample mean of Y among units that belong in group G “ a and are observed
in period t “ s. In fact, it is unclear exactly what estimand is being recovered when one uses
(5.11) with an unbalanced panel. If one replaces unit fixed effects with treatment group dummies
in (5.11), such equivalence is restored, though we suspect that many practitioners do not use
this alternative specification. In general, we caution against extrapolating from a well-motivated
regression specification that was studied under one specific setup to another related but inherited
different framework. This practice has led to many issues in DiD, which can be fully avoided by
adopting the forward-engineering approach discussed in this paper.

References
Abadie, Alberto, “Semiparametric Difference-in-Difference Estimators,” The Review of Economic Stud-
ies, 2005, 72, 1–19.
, Susan Athey, Guido W. Imbens, and Jeffrey M. Wooldridge, “Sampling-Based versus Design-
Based Uncertainty in Regression Analysis,” Econometrica, January 2020, 88 (0), 265–296.
, , , and Jeffrey Wooldridge, “When Should You Adjust Standard Errors for Clustering?,” The
Quarterly Journal of Economics, 2023, 138 (1), 1–35.
Abbring, Jaap H. and Gerard J. van den Berg, “The nonparametric identification of treatment
effects in duration models,” Econometrica, 2003, 71 (5), 1491–1517.
Angrist, Joshua D., “Estimating the Labor Market Impact of Voluntary Military Service Using Social
Security Data on Military Applicants,” Econometrica, 1998, 66 (2), 249–288.

69
 and Guido W. Imbens, “Two-Stage Least Squares Estimation of Average Causal Effects in Models
with Variable Treatment Intensity,” Journal of the American Statistical Association, 1995, 90 (430),
431–442.
Arkhangelsky, Dmitry, Guido W. Imbens, Lihua Lei, and Xiaoman Luo, “Design-Robust Two-
Way-Fixed-Effects Regression For Panel Data,” Quantitative Economics, 2024, 15 (4).
, Susan Athey, David A. Hirshberg, Guido W. Imbens, and Stefan Wager, “Synthetic
Difference-in-Differences,” American Economic Review, 2021, 111 (12), 4088–4118.
Aronow, P. M. and Cyrus Samii, “Does regression produce representative estimates of causal effects?,”
American Journal of Political Science, 2015, 60 (1), 250–267.
Ashenfelter, Orley C. and David Card, “Using the longitudinal structure of earnings to estimate the
effect of training programs,” The Review of Economics and Statistics, 1985, 67 (4), 648–660.
Athey, Susan and Guido Imbens, “Design-based Analysis in Difference-In-Differences Settings with
Staggered Adoption,” Journal of Econometrics, 2022, 226 (1), 62–79.
and Guido W. Imbens, “Identification and Inference in Nonlinear Difference in Differences Models,”
Econometrica, 2006, 74 (2), 431–497.
Austin, Peter C, “Balance diagnostics for comparing the distribution of baseline covariates between
treatment groups in propensity-score matched samples,” Statistics in Medicine, 2009, 28 (25), 3083–
3107.
Baker, Andrew C, David F Larcker, and Charles CY Wang, “How much should we trust staggered
difference-in-differences estimates?,” Journal of Financial Economics, 2022, 144 (2), 370–395.
Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, “How Much Should We Trust
Differences-in-Differences Estimates?,” Quarterly Journal of Economics, February 2004, 119 (1), 249–
275.
Bilinski, Alyssa and Laura A Hatfield, “Nothing to see here? Non-inferiority approaches to parallel
trends and other model assumptions,” 2018. Working Paper.
Black, Bernard, Alex Hollingsworth, Leticia Nunes, and Kosali Simon, “Simulated power anal-
yses for observational studies: An application to the Affordable Care Act Medicaid expansion,” Journal
of Public Economics, 2022, 213, 104713.
Bonhomme, Stéphane and Ulrich Sauder, “Recovering Distributions in Difference-in-Differences
Models: a Comparison of Selective and Comprehensive Schooling,” Review of Economics and Statistics,
2011, 93 (May), 479–494.
Borgschulte, Mark and Jacob Vogler, “Did the ACA Medicaid expansion save lives?,” Journal of
Health Economics, 2020, 72, 102333.
Borusyak, Kirill, Xavier Jaravel, and Jann Spiess, “Revisiting Event Study Designs: Robust and
Efficient Estimation,” Review of Economic Studies, 2024, 91 (6), 3253–3285.
Caetano, Carolina and Brantly Callaway, “Difference-in-Differences when Parallel Trends Holds
Conditional on Covariates,” arXiv:2406.15288, 2024.
, , Stroud Payne, and Hugo Sant’Anna Rodrigues, “Difference in differences with time-varying
covariates,” 2022. Working Paper.

70
Callaway, Brantly, “Bounds on distributional treatment effect parameters using panel data with an
application on job displacement,” Journal of Econometrics, 2021, 222 (2), 861–881.
, “Difference-in-Differences for Policy Evaluation,” in Klaus F. Zimmermann, ed., Handbook of Labor,
Human Resources and Population Economics, Cham: Springer International Publishing, 2023, pp. 1–61.
and Pedro HC Sant’Anna, “Difference-in-differences with multiple time periods,” Journal of Econo-
metrics, 2021, 225 (2), 200–230.
and Sonia Karami, “Treatment effects in interactive fixed effects models with a small number of time
periods,” Journal of Econometrics, 2023, 233 (1), 184–208.
and Tong Li, “Quantile Treatment Effects in Difference in Differences Models with Panel Data,”
Quantitative Economics, 2019, 10 (4), 1579–1618.
, Andrew Goodman-Bacon, and Pedro H. C. Sant’Anna, “Difference-in-Differences with a Con-
tinuous Treatment,” arXiv:2107.02637 [econ], 2021.
, , and , “Event Studies with a Continuous Treatment,” AEA Papers and Proceedings, May 2024,
114, 601–605.
, Tong Li, and Tatsushi Oka, “Quantile Treatment Effects in Difference in Differences Models under
Dependence Restrictions and with Only Two Time Periods,” Journal of Econometrics, 2018, 206 (2),
395–413.
Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller, “Bootstrap-Based Improvements
for Inference with Clustered Errors,” Review of Economics and Statistics, August 2008, 90 (3), 414–427.
Caron, Laura, “Triple Difference Designs with Heterogeneous Treatment Effects,” arXiv:2502.19620,
2025.
Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer, “The Effect of Minimum
Wages on Low-Wage Jobs,” The Quarterly Journal of Economics, August 2019, 134 (3), 1405–1454.
Centers for Disease Control and Prevention, “Vital Statistics Data,” https://www.cdc.gov/nchs/
nvss/index.htm 2024. Accessed: 2024-09-17.
Chabé-Ferret, Sylvain, “Analysis of the bias of Matching and Difference-in-Difference under alternative
earnings and selection processes,” Journal of Econometrics, 2015, 185 (1), 110–123.
Chang, Neng-Chieh, “Double/debiased machine learning for difference-in-differences,” Econometrics
Journal, 2020, 23, 177–191.
Chen, Xiaohong, Pedro H. C. Sant’Anna, and Haitian Xie, “Efficient Difference-in-Differences and
Event Study Estimators,” Working Paper, 2024.
Chernozhukov, Victor, Iván Fernández-Val, and Ye Luo, “The Sorted Effects Method: Discovering
Heterogeneous Effects Beyond Their Averages,” Econometrica, 2018, 86 (6), 1911–1938.
, Mert Demirer, Esther Duflo, and Iván Fernández-Val, “Fisher-Schultz Lecture: Generic Ma-
chine Learning Inference on Heterogenous Treatment Effects in Randomized Experiments, with an
Application to Immunization in India,” arXiv: 1712.04802, 2023, pp. 1–81.
Conley, Timothy and Christopher Taber, “Inference with “Difference in Differences” with a Small
Number of Policy Changes,” Review of Economics and Statistics, February 2011, 93 (1), 113–125.

71
Crump, Richard K, V Joseph Hotz, Guido W Imbens, and Oscar A Mitnik, “Dealing with
limited overlap in estimation of average treatment effects,” Biometrika, 2009, 96 (1), 187–199.
Currie, Janet and Hannes Schwandt, “Mortality Inequality: The Good News from a County-Level
Approach,” Journal of Economic Perspectives, May 2016, 30 (2), 29–52.
, Henrik Kleven, and Esmée Zwiers, “Technology and Big Data Are Changing Economics: Mining
Text to Track Methods,” AEA Papers and Proceedings, 2020, 110, 42–48.
de Chaisemartin, Clément and Xavier D’Haultfoeuille, “Fuzzy Differences-in-Differences,” The
Review of Economic Studies, 2018, 85 (2), 999–1028.
and , “Two-way fixed effects estimators with heterogeneous treatment effects,” American Economic
Review, 2020, 110 (9), 2964–2996.
and , “Difference-in-Differences Estimators of Intertemporal Treatment Effects,” 2023. Working
Paper.
and , “Two-way fixed effects and differences-in-differences with heterogeneous treatment effects: a
survey,” Econometrics Journal, 2023, Forthcoming.
, , Félix Pasquier, and Gonzalo Vazquez-Bare, “Difference-in-Differences for Continuous Treat-
ments and Instruments with Stayers,” arXiv:2201.06898, 2024.
, Xavier D’Haultfœuille, and Gonzalo Vazquez-Bare, “Difference-in-Difference Estimators with
Continuous Treatments and No Stayers,” AEA Papers and Proceedings, May 2024, 114, 610–613.
Deshpande, Manasi and Yue Li, “Who Is Screened Out? Application Costs and the Targeting of
Disability Programs,” American Economic Journal: Economic Policy, November 2019, 11 (4), 213–48.
Dette, Holger and Martin Schumann, “Testing for Equivalence of Pre-Trends in Difference-in-
Differences Estimation,” Journal of Business & Economic Statistics, 2024, 42 (4).
DiNardo, John and David S. Lee, “Chapter 5 - Program Evaluation and Research Designs,” in Orley
Ashenfelter and David Card, eds., Handbook of Labor Economics, Vol. 4, Elsevier, 2011, pp. 463–536.
Donald, Stephen G. and Kevin Lang, “Inference with Difference-in-Differences and other Panel Data,”
Review of Economics and Statistics, 2007, 89 (2), 221–233.
Dube, Arindrajit, Daniele Girardi, Òscar Jordà, and Alan M Taylor, “A Local Projections
Approach to Difference-in-Differences,” Working Paper 31184, National Bureau of Economic Research
2024.
Fadlon, Itzik and Torben Heien Nielsen, “Family Labor Supply Responses to Severe Health Shocks:
Evidence from Danish Administrative Records,” American Economic Journal: Applied Economics, July
2021, 13 (3), 1–30.
Fan, Yanqin and Zhentao Yu, “Partial Identification of Distributional and Quantile Treatment Effects
in Difference-in-Differences Models,” Economics Letters, 2012, 115 (3), 511–515.
Fernández-Val, Iván, Jonas Meier, Aico van Vuuren, and Francis Vella, “Distribution Regression
Difference-in-Differences,” arXiv preprint arXiv:2409.02311, 2024.
, , , and , “Distribution Regression Difference-In-Differences,” arXiv:2409.00123, 2024.
Finkelstein, Amy and Robin McKnight, “What Did Medicare Do? The Initial Impact of Medicare on
Mortality and Out-of-Pocket Medical Spending,” Journal of Public Economics, 2008, 92 (7), 1644–1668.

72
Firpo, Sergio and Cristine Pinto, “Identification and Estimation of Distributional Impacts of In-
terventions Using Changes in Inequality Measures,” Journal of Applied Econometrics, 2016, 31 (3),
457–486.
Freyaldenhoven, Simon, Christian Hansen, and Jesse M Shapiro, “Pre-Event Trends in the Panel
Event-Study Design,” American Economic Review, 2019, 109 (9), 3307–3338.
, , Jorge Pérez-Pérez, and Jesse M. Shapiro, “Visualization, identification, and estimation
in the linear panel event-study design,” in “Advances in Economics and Econometrics: Twelfth World
Congress” 2024. Forthcoming.
Gardner, John, “Two-stage differences in differences,” Working Paper, 2021.
Ghanem, Dalia, Désiré Kédagni, and Ismael Mourifié, “Evaluating the Impact of Regulatory
Policies on Social Welfare in Difference-in-Difference Settings,” arXiv:2306.04494, 2023.
, Pedro H. C. Sant’Anna, and Kaspar Wüthrich, “Selection and parallel trends,”
arXiv:2203.09001, 2022.
Goldsmith-Pinkham, Paul, Peter Hull, and Michal Kolesár, “Contamination Bias in Linear Re-
gressions,” American Economic Review, 2024, Forthcoming.
Goodman-Bacon, Andrew, “Difference-in-differences with variation in treatment timing,” Journal of
Econometrics, 2021, 225 (2), 254–277.
Graham, Bryan, Cristine Pinto, and Daniel Egel, “Inverse Probability Tilting for Moment Condition
Models with Missing Data,” The Review of Economic Studies, 2012, 79 (3), 1053–1079.
Gruber, Jonathan, “The Incidence of Mandated Maternity Benefits,” American Economic Review, June
1994, 84 (3), 622–641.
Harmon, Nikolaj A., “Difference-in-Differences and Efficient Estimation of Treatment Effects,” Working
Paper, 2024.
Heckman, James and Richard Robb, “Alternative methods for evaluating the impact of interventions,”
in James Heckman and Burton Singer, eds., Longitudinal Analysis of Labor Market Data, Cambridge:
Cambridge University Press, 1985, pp. 156–246.
Heckman, James J., “Causal Parameters and Policy Analysis in Economics: A Twentieth Century
Retrospective,” The Quarterly Journal of Economics, 2000, 115 (1), 45–97.
, Hidehiko Ichimura, and Petra Todd, “Matching as an Econometric Evaluation Estimator: Ev-
idence from Evaluating a Job Training Programme,” The Review of Economic Studies, 1997, 64 (4),
605–654.
, Jeffrey Smith, and Nancy Clements, “Making The Most Out Of Programme Evaluations and
Social Experiments: Accounting For Heterogeneity in Programme Impacts,” The Review of Economic
Studies, 1997, 64 (4), 487––535.
Ho, Daniel E, Kosuke Imai, Gary King, and Elizabeth A Stuart, “Matching as nonparametric
preprocessing for reducing model dependence in parametric causal inference,” Political Analysis, 2007,
15 (3), 199–236.
Hong, Seung-Hyun, “Measuring the effect of Napster on recorded music sales: difference-in-differences
estimates under compositional changes,” Journal of Applied Econometrics, 2013, 28 (2), 297–324.

73
Imai, Kosuke and In Song Kim, “On the use of two-way fixed effects regression models for causal
inference with panel data,” Political Analysis, 2021, 29 (3), 405–415.
, , and Erik H. Wang, “Matching Methods for Causal Inference with Time-Series Cross-Sectional
Data,” American Journal of Political Science, 2023, 67 (3), 587–605.
Imbens, Guido, Nathan Kallus, and Xiaojie Mao, “Controlling for Unmeasured Confounding
in Panel Data Using Minimal Bridge Functions: From Two-Way Fixed Effects to Factor Models,”
arXiv:2108.03849, 2021.
Imbens, Guido W and Donald B Rubin, Causal inference in statistics, social, and biomedical sciences,
Cambridge university press, 2015.
Kahn-Lang, Ariella and Kevin Lang, “The Promise and Pitfalls of Differences-in-Differences: Reflec-
tions on 16 and Pregnant and Other Applications,” Journal of Business and Economic Statistics, 2020,
38 (3), 613–620.
Kennedy, Edward H, Zongming Ma, Matthew D McHugh, and Dylan S Small, “Non-parametric
methods for doubly robust estimation of continuous treatment effects,” Journal of the Royal Statistical
Society. Series B (Statistical Methodology), 2017, 79 (4), 1229–1245.
Khan, Shakeeb and Elie Tamer, “Irregular Identification, Support Conditions, and Inverse Weight
Estimation,” Econometrica, 2010, 78 (6), 2021–2042.
Lechner, Michael, “The Estimation of Causal Effects by Difference-in-Difference Methods,” Foundations
and Trends in Econometrics, 2011, 4, 165–224.
Lee, Soo Jeong and Jeffrey M. Wooldridge, “A Simple Transformation Approach to Difference-in-
Differences Estimation for Panel Data,” Working Paper, 2023. Available at SSRN: http://dx.doi.org/
10.2139/ssrn.4516518.
Liu, Licheng, 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, 2024, 68 (1),
160–176.
Ma, Yukun, Pedro H. C. Sant’Anna, Yuya Sasaki, and Takuya Ura, “Doubly Robust Estimators
with Weak Overlap,” arXiv:2304.08974, 2023.
Malani, Anup and Julian Reif, “Interpreting pre-trends as anticipation: Impact on estimated treatment
effects from tort reform,” Journal of Public Economics, 2015, 124, 1–17.
Manski, Charles F. 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, 2018, 100
(2), 232–244.
Marcus, Michelle and Pedro H. C. Sant’Anna, “The role of parallel trends in event study settings:
An application to environmental economics,” Journal of the Association of Environmental and Resource
Economists, 2021, 8 (2), 235–275.
Marx, Philip, Elie Tamer, and Xun Tang, “Parallel Trends and Dynamic Choices,” Journal of
Political Economy Microeconomics, 2024, 2 (1), 129–171.
Meyer, Bruce, W. Kip Viscusi, and David Durbin, “Workers’ Compensation and Injury Duration:
Evidence from a Natural Experiment,” The American Economic Review, 1995, 85 (3), 322–340.

74
Miller, Sarah, Norman Johnson, and Laura R Wherry, “Medicaid and mortality: new evidence
from linked survey and administrative data,” The Quarterly Journal of Economics, 2021, 136 (3), 1783–
1829.
Miyaji, Sho, “Instrumented Difference-in-Differences with Heterogeneous Treatment Effects,”
arXiv:2405.12083, 2024.
Mogstad, Magne and Alexander Torgovitsky, “Chapter 1 - Instrumental variables with unobserved
heterogeneity in treatment effects,” in Christian Dustmann and Thomas Lemieux, eds., Handbook of
Labor Economics, Vol. 5, Elsevier, 2024, pp. 1–114.
Mora, Ricardo and Iliana Reggio, “Alternative diff-in-diffs estimators with several pretreatment pe-
riods,” Econometric Reviews, 2019, 38 (5), 465–486.
Olden, Andreas and Jarle Møen, “The triple difference estimator,” The Econometrics Journal, 2022,
25 (3), 531–553.
Olea, José Luis Montiel and Mikkel Plagborg-Moller, “Simultaneous confidence bands: Theory,
implementation, and an application to SVARs,” Journal of Applied Econometrics, 2018, 33 (7), 943–964.
Ortiz-Villavicencio, Marcelo and Pedro H. C. Sant’Anna, “Better Understanding Triple Differences
Estimators,” 2025. Working paper.
Poirier, Alexandre and Tymon Sloczynski, “Quantifying the Internal Validity of Weighted Esti-
mands,” arXiv:2404.14603, 2024.
Rambachan, Ashesh and Jonathan Roth, “A More Credible Approach to Parallel Trends,” Review
of Economic Studies, 2023, 90 (5), 2555–2591.
and , “Design-Based Uncertainty for Quasi-Experiments,” arXiv:2008.00602, 2024.
Robins, James, “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, 1986,
7, 1393–1512.
Rosenbaum, Paul R. and Donald B. Rubin, “The Central Role of the Propensity Score in Observa-
tional Studies for Causal Effects,” Biometrika, April 1983, 70 (1), 41–55.
Roth, Jonathan, “Pretest with caution: Event-study estimates after testing for parallel trends,” Ameri-
can Economic Review: Insights, 2022, 4 (3), 305–322.
and Pedro H. C. Sant’Anna, “Efficient Estimation for Staggered Rollout Designs,” Journal of
Political Economy Microeconomics, 2023, 1 (4), 669–709.
and , “When Is Parallel Trends Sensitive to Functional Form?,” Econometrica, 2023, 91 (2), 737–747.
, , Alyssa Bilinski, and John Poe, “What’s Trending in Difference-in-Differences? A Synthesis of
the Recent Econometrics Literature,” Journal of Econometrics, 2023, 235 (2), 2218–2244.
Rubin, Donald, “Estimating Causal Effects of Treatments in Randominzed and Nonrandomized Studies,”
Journal of Educational Psychology, 1974, 66 (5), 688–701.
Sant’Anna, Pedro H. C. and Jun B. Zhao, “Doubly Robust Difference-in-Differences Estimators,”
November 29 2018. Unpublished Manuscript.
and Jun Zhao, “Doubly Robust Difference-in-Differences Estimators,” Journal of Econometrics, 2020,
Forthcoming.

75
 and Qi Xu, “Difference-in-Differences with Compositional Changes,” arXiv:2304.14256, 2023.
Sasaki, Yuya and Takuya Ura, “Estimation and inference for moments of ratios with robustness against
large trimming bias,” Econometric Theory, 2022, 38 (1), 66–112.
Seaman, Shaun R and Stijn Vansteelandt, “Introduction to Double Robust Methods for Incomplete
Data,” Statistical Science, 2018, 33 (2), 184–197.
Semmelweis, Ignaz, Etiology, Concept and Prophylaxis of Childbed Fever, The University of Wisconsin
Press, 1983.
Sloczynski, Tymon, “Interpreting OLS Estimands When Treatment Effects Are Heterogeneous: Smaller
Groups Get Larger Weights,” The Review of Economics and Statistics, 2022, 104 (3), 501—-509.
Smucler, Ezequiel, Andrea Rotnitzky, and James M. Robins, “A unifying approach for doubly-
robust ℓ1 regularized estimation of causal contrasts,” arXiv:1904.03737, 2019.
Snow, John, On the Mode of Communication of Cholera, London: John Churchill, 1855.
Solon, Gary, Steven J. Haider, and Jeffrey M. Wooldridge, “What Are We Weighting For?,” The
Journal of Human Resources, 2015, 50 (2), 301–316.
Sommers, Benjamin Daniel and Arnold M Epstein, “US governors and the Medicaid expansion—no
quick resolution in sight,” New England Journal of Medicine, 2013.
Strezhnev, Anton, “Semiparametric weighting estimators for multi-period differencein-differences de-
signs,” 2018. Working Paper.
, “Decomposing Triple-Differences Regression under Staggered Adoption,” arXiv:2307.02735, 2023.
Sun, Liyang and Sarah Abraham, “Estimating dynamic treatment effects in event studies with het-
erogeneous treatment effects,” Journal of Econometrics, 2021, 225 (2), 175–199.
Tchetgen, Eric J. Tchetgen, Chana Park, and David B. Richardson, “Universal Difference-in-
Differences for Causal Inference in Epidemiology,” Epidemiology, 2024, 35 (1), 16–22.
Wooldridge, Jeffrey M., “Cluster-Sample Methods in Applied Econometrics,” American Economic
Review P&P, 2003, 93 (2), 133–138.
, “Two-way fixed effects, the two-way mundlak regression, and difference-in-differences estimators,” 2021.
Working Paper.
, “Simple Approaches to Nonlinear Difference-in-Differences with Panel Data,” Econometrics Journal,
2023, Forthcoming.
Wyse, Angela and Bruce Meyer, “Saved by Medicaid: New Evidence on Health Insurance and Mor-
tality from the Universe of Low-Income Adults,” Working Paper, 2024.
Yanagi, Takahide, “An Effective Treatment Approach to Difference-in-Differences with General Treat-
ment Patterns,” arXiv:2212.13226, 2023.

76