An Introductory Guide to Event Study Models

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Schmidheiny and Siegloch (2023) recommend end caps (binning) and note it can provide a natural identifying restriction for timing-based data structures, create a connection to distributed lag models, and model multiple events per unit.

“Schmidheiny and Siegloch (2023) recommend this approach (which they call”

• cites Schmidheiny and Siegloch (2023) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Finkelstein, Gentzkow, and Williams (2016) track Medicare patients who move across regions with different intensities of medical usage.

“For example, Finkelstein, Gentzkow, and Williams (2016) track Medicare patients who move across regions with different intensities of medical usage.”

• cites Finkelstein, Gentzkow, and Williams (2016) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Molitor (2018) examines cardiologists' patterns of practice as they move across regions.

“Molitor (2018) examines cardiologists' patterns of practice as they move across regions.”

• cites Molitor (2018) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Schmidheiny and Siegloch (2023) argue that using end caps can provide identification of event-time coefficients in a timing-based data structure, restoring ability to examine pre-event trends.

“Schmidheiny and Siegloch (2023) argue that using end caps can provide identification of the eventtime coefficients in a timing-based data structure.This would restore the ability to examine pre-event trends.”

• cites Schmidheiny and Siegloch (2023) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

They can also be used to estimate statistical moments, which in turn can be used to estimate a structural model, as in Finkelstein et al. (2022).

“They can also be used to estimate statistical moments, which in turn can be used to estimate a structural model, as inFinkelstein et al. (2022).”

• cites Finkelstein et al. (2022) sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Figure 1 shows a modest pre-event trend, a sharp drop in earnings at the time of displacement, followed by a bounce back over the next two years that levels off at an earnings decline of about $500 to $1,000 per quarter compared to the pre-event level.

“The figure shows a modest pre-event trend, but also shows a sharp drop in earnings at the time of displacement, followed by a bounce back over the next two years that levels off at an earnings decline of about $500 to $1,000 per quarter compared to the pre-event level.”

• cites Jacobson, LaLonde, and Sullivan (1993)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Using a two-way fixed effects approach helps isolate the effect of the event.

“Using this two-way fixed events approach helps to isolate the effect of the event.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

We might be hesitant about including post end caps if there is a chance that treatment effects are changing over event time.

“That section argues that we might be hesitant about including”

Classified as probably not a substantial claim.

Unit and time fixed effects control for confounding omitted variables that vary at the unit or time level.

“These serve the role of controlling for confounding omitted variables that vary at the unit or time level.”

May combine two claims.

Including as many event dummies as possible directly estimates a γ j for each event time and removes the problem of data outside the event window.

“One natural option is to create and include as many event dummies as possible.By directly estimating a γ j for each event time, this removes the problem.”

Probably not a substantial claim: may not pass S3 (load-bearing).

When an event study allows comparison of treated and untreated units, presenting actual and counterfactual average outcomes allows readers to assess whether unit types experience parallel trends during periods when treatment status is unchanging.

“For example, when the event study allows comparison of treated and untreated units, this presentation allows readers to assess whether the unit types experience parallel trends during periods when treatment status is unchanging.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

One can look at the degree of similarity between treated and control units along a number of dimensions using covariates.

“Third, one can look at the degree of similarity between treated and control units along a number of dimensions, using covariates.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Event study models are estimated on data that have a panel structure.

“Event study models are estimated on data that have a panel structure.”

Classified as probably not a substantial claim.

In the Jacobson, LaLonde, and Sullivan (1993) setting, the control group may be positively selected with regard to unobservable skill, labor force attachment, and/or job match quality.

“In addition, the control group may be positively selected with regard to unobservable skill, labor force attachment, and/or job match quality.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A trade-off of dropping observations outside the window is that loss of data can weaken statistical power.

“One possible trade-off is that the loss of data can weaken statistical power.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The trade-off between using never-treated units and discarding them for a timing-based strategy varies on a case-by-case basis.

“The trade-off between these two considerations will vary on a case-by-case basis.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Letting software automatically choose collinear variables to drop has unknown and possibly problematic implications.

“In current practice, a common approach is to let the software (like Stata) automatically choose some collinear variables to drop, with unknown and possibly problematic implications.”

Classified as probably not a substantial claim. May combine two claims.

Making the event window as wide as possible allows seeing a long path of dynamic treatment effects and gives a long window to detect troublesome pre-event patterns.

“On one hand, making the event window as wide as possible allows us to see a long path of dynamic treatment effects, and for the pre-event coefficients it gives us a long window to detect troublesome patterns.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In event studies, it is common to have data for some units before or after the event window.

“In event studies, it will be common to have data for some units that occur before or after the event window.”

May combine two claims.

In the Jacobson, LaLonde, and Sullivan (1993) setting, never-laid-off individuals may not provide a good counterfactual for treated individuals, and using this control group could make the impacts of the layoff look worse than they actually are.

“In this setting, those who are never laid off may not provide a good counterfactual for outcome for treated individuals; indeed, the use of this control group could make the impacts of the layoff look worse than they actually are.”

May combine two claims.

In Jacobson, LaLonde, and Sullivan (1993), events occur between 1980 and 1986 and outcome data are observed for 1974-1986.

“For example, in Jacobson, LaLonde, and Sullivan (1993) Jacobson, LaLonde, and Sullivan (1993) the events (job displacements) occur between 1980 and 1986, and the outcome (earnings) data are observed for the period 1974-1986.”

• cites Jacobson, LaLonde, and Sullivan (1993)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Dropping observations with event dates outside the main window keeps the specification simple and creates balance in event time in the analysis sample.

“Another possible approach is just to drop observations that have event dates outside of our main window of interest.This option keeps the specification simple and creates a balance in event time in our analysis sample (for example, all US states are observed from three periods before their event to five periods after).”

Classified as probably not a substantial claim.

To proceed with an event study model, restrictions on the parameters are needed.

“To proceed, we need some restrictions on these parameters.”

Classified as probably not a substantial claim.

Excluding unrepresentative control units can make the assumption that control units tell the counterfactual over-time changes more plausible.

“There are a range of options to have the control units (for example, individuals who did not experience job displacement) offer better counterfactuals, with the overall goal of making the assumption that”

• AssumptionThe authors state this as an assumption: In event studies with control units, the control units tell us the counterfactual over-time changes.

May combine two claims.

A pre-event trend in an event study graph suggests that the treatment may have been expected or that other factors are in play.

“a trend suggests that the treatment may have been expected or that other factors are in play.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The problems of multicollinearity compound when unit-specific time trend controls are added directly.

“The problems of multicollinearity also compound when we directly add in unit-specific time trend controls.”

Classified as probably not a substantial claim.

For pre-event periods, it is possible to check for parallel trends between control and treated units.

“Second, for the time periods before the event, it is possible to check for parallel trends between the control and treated units.”

May combine two claims.

In Jacobson et al. (1993), event-time coefficients for zero years since displacement are based on all displacements, but coefficients for five years since displacement can only be estimated for displacements in 1980 or 1981, so post-displacement coefficients are estimated off different sets of individuals.

“The event-time coefficients for”

• cites Jacobson, LaLonde, and Sullivan (1993)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

When a dip in outcomes shortly before the event is suspected or seen, the period of the dip probably should not be used as the counterfactual baseline because it is actually part of the treated period.

“In this case, we probably do not want to use the period of the dip as our counterfactual baseline because it is actually part of the treated period, even though nominally it's before treatment.”

Probably not a substantial claim: may not pass S4 (based).

In the absence of anticipation effects, model misspecification, or omitted confounding variables, the pre-event terms should not have a trend in j.

“In the absence of anticipation effects, model misspecification, or omitted confounding variables, these pre-event terms should not have a trend in j.”

Classified as probably not a substantial claim.

The restrictions imposed to identify the model are untestable, at least in part.

“these restrictions are untestable, at least in part”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In the canonical event study graph, the treatment effects line represents allowing for over-time changes in the treated group and over-time changes in the untreated group and then looking at the differences between these changes.

“In the canonical event studies graph, the treatment effects line represents allowing for over-time changes in the treated group and over-time changes in the untreated group and then looking at the differences between these changes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Including event dummies for each event time is natural and appropriate when the data structure has both treated and untreated units balanced in calendar time.

“This approach is natural and appropriate when the data structure has both treated and untreated units that are balanced in calendar time (for example, all US states are observed over the period 1980-2020).”

Classified as probably not a substantial claim.

The coefficients after the event has occurred (γj for j ≥ 0) capture the dynamic effects of the treatment as these effects manifest over time since the event.

“The coefficients after the event has occurred (γ j for j ≥ 0) capture the dynamic effects of the treatment as these effects manifest over time since the event.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Using never-treated units brings in more data, usually improving statistical power and requiring fewer parameter restrictions to identify the model.

“On the other side, using never-treated units will bring in more data, usually improving statistical power and requiring fewer parameter restrictions in order to identify the model.”

• AssumptionThe authors state this as an assumption: In event studies with control units, the control units tell us the counterfactual over-time changes.

Probably not a substantial claim: may not pass S2 (checkable).

Ideally, event-time coefficients γ j should all be estimated off the same set of units.

“The main competing consideration is that we would ideally like for the eventtime coefficients γ j to all be estimated off of the same set of units.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The main risks to creating end caps arise with trending counterfactuals or trending treatment effects.

“The main risks to creating”

May combine two claims.

Pooling treatment coefficients in the tails to be equal can be problematic if there are uncontrolled-for underlying trends or if posttreatment treatment effects are trending.

“This approach can sometimes be okay, but it can be problematic if there are uncontrolled-for underlying trends or (for a posttreatment end cap) if the treatment effects themselves are trending.”

May combine two claims.

A post-event end cap forces all estimated event-time effects within the end cap to be the same; if they are truly trending, this can cause problems.

“The post-event end cap (for example,”

Probably not a substantial claim: may not pass S3 (load-bearing).

Another approach to multiple events is to adjust the definition of an event so as to have only one per unit.

“Another approach is to adjust the definition of an event so as to have only one per unit.”

Probably not a substantial claim: may not pass S2 (checkable).

If the underlying event is shared across units, this argues for clustering at a higher level.

“If the underlying event is shared across units, then this argues for clustering at a higher level.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Conley (1999) provides spatial robust standard errors as a potential way to address spatial correlation.

“for example, see Conley (1999) on spatial robust standard errors.”

• cites Conley (1999) sourceNot checked — no full text of the source

May combine two claims.

A unit-specific time trend term will be multicollinear with unit dummies, time dummies, and event-time variables, necessitating at least one more additional parameter restriction.

“A unit-specific time trend term will be multicollinear with the unit dummies, time dummies, and event-time variables.This means that it will be necessary to add (at least) one more additional parameter restriction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For the basic case, multiple events per unit can be straightforward to implement.

“For the basic case, this can be straightforward to implement.”

May combine two claims.

Driscoll and Kraay (1998) provide a method that allows for arbitrary correlations within a cluster and correlation that decays in calendar time across nearby time periods regardless of unit.

“as in Driscoll and Kraay 1998”

• cites Driscoll and Kraay 1998 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Roth (2022) notes that rounding insignificant pre-trends to zero can lead to pre-test bias.

“Roth (2022) notes that the widespread, informal practice of”

• cites Roth (2022) sourceNot checked — no full text of the source

May combine two claims.

Standard cluster-robust methods provide accurate standard errors only if the number of clusters is large enough, with no hard and fast rule for what that means.

“However, one potential concern is that standard cluster-robust methods provide accurate standard errors only if the number of clusters is”

May combine two claims.

Limiting the sample to be balanced in event time creates an imbalance in calendar time, so time dummies at the extremes are estimated off strangely selected units.

“If you limit your sample to be balanced in event time, then this creates an imbalance in calendar time.This in turn means that the time dummies at the extremes will be estimated off of strangely selected units.”

May combine two claims.

Event study specifications can ask a lot of the data, and many event study models have wide confidence intervals around each main γ j coefficient.

“With so many”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

For event study models, current practice appears to be to calculate cluster-robust standard errors with clusters defined as i-level units.

“For event study models, the current practice appears to be to calculate clusterrobust standard errors, with clusters defined as the i-level units.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Including endpoint coefficients in the graph gives a fuller picture of the model and can help diagnose problems with the specification or data.

“On the other hand, including such endpoints in the graph gives a fuller picture of the model; indeed, including them can sometimes help to diagnose problems with the specification or the data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

When an endpoint has its own dummy variable, it captures different averaging than interior terms and may appear offset from the rest of the graph.

“When the endpoint has its own dummy variable, it will capture different averaging than the”

Probably not a substantial claim: may not pass S3 (load-bearing).

The pre-event coefficient diagnostic is most appropriate when working with difference-in-difference or hybrid data structures that include never-treated units.

“This tool is most appropriate when working with difference-in-difference or hybrid data structures, which include never-treated units.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Jacobson, LaLonde, and Sullivan (1993) focus on the first layoff, and subsequent layoffs are not modeled.

“For example, in the Jacobson, LaLonde, and Sullivan (1993) paper behind Figure 1, the focus is on the first layoff, and subsequent layoffs are not modeled.”

• cites Jacobson, LaLonde, and Sullivan (1993)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

Constraining the trend in the reference period event study coefficients to be zero averages across several coefficients, reduces impact of noise, respects reference period normalization, and offers a natural counterfactual interpretation.

“One restriction that accommodates the considerations above-at least for the difference-in-differences and the hybrid data structures-is to constrain the trend in the”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the Molitor (2018) design, the variable intensity of the event is given by the difference in regional patterns of practice across destination and origin locations.

“Again, the variable intensity of the event is given by the difference in regional patterns of practice across destination and origin locations.”

May combine two claims.

Key right-hand-side variables in an event study have some autocorrelation, and model error term is plausibly positively autocorrelated over time within a unit, arguing for clustering at the unit level.

“The key right-hand-side variables in an event study have some degree of autocorrelation and it is plausible to think about the model error term also being positively autocorrelated over time within a unit.Taken together, this argues for clustering at the unit level.”

Probably not a substantial claim: may not pass S4 (based).

Treatment effects themselves may be trending in time-since-treatment.

“Second, treatment effects themselves may be trending in time-since-treatment.”

Probably not a substantial claim: may not pass S3 (load-bearing).

More than one of the event-time dummies can be turned on simultaneously.

“Thus, more than one of the event-time dummies can be turned on simultaneously.”

May combine two claims.

Spatial correlation in event dates undercuts the premise of cluster-robust inference and may result in over-rejection of the null hypothesis.

“Spatial correlation in event dates undercuts this premise, and doing so may result in over-rejection of the null hypothesis.”

• AssumptionRests on an assumption: When events are the result of a political process or influenced by economic circumstances, neighboring units can be closer in event date than more distant units.
• AssumptionRests on an assumption: Economic outcomes are often spatially correlated.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Interactions can use up a lot of variation in the data, and in response, it may be useful to impose a parametric simplification on the interaction terms.

“These interactions can use up a lot of variation in the data, and in response, it may be useful to impose a parametric simplification on the interaction terms.”

Classified as probably not a substantial claim. May combine two claims.

Pooling adjacent event-time dummies strives for a balance between flexibility and statistical power, but risks obscuring features of the empirical results.

“This approach strives for a balance between flexibility and statistical power.The main risk is that the pooling might obscure features of the empirical results.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Goodman-Bacon (2018) examines the introduction of Medicaid in the 1960s across US states.

“One interesting variation is found in Goodman-Bacon (2018), who examines the introduction of Medicaid in the 1960s across US states.”

• cites Goodman-Bacon (2018) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The estimated pre-event terms can serve as a tool for diagnosing trends, often by inspecting the graph of pre-event coefficients.

“The estimated pre-event terms can serve as a tool for diagnosing trends.This is often done informally by inspecting the graph of the pre-event coefficients.”

Classified as probably not a substantial claim.

There is a recent, active, and promising literature on how event study models perform when treatment effects differ across units and when we do not know the functional form of how they differ.

“There is a recent, active, and promising literature on how event study models perform when treatment effects differ across units and when we do not know the functional form of how they differ.”

• cites Sun and Abraham (2021) sourceNot checked — no full text of the source
• cites de Chaisemartin and D'Haultfœuille (2022)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

If one of the extra parameter restrictions is equality of two event coefficients, spacing those coefficients further apart reduces the impact of statistical noise between them.

“Second, if one of your extra parameter restrictions is in the form of equality of two event coefficients, consider spacing those coefficients further apart, because the impact of any statistical”

May combine two claims.

The problem of few clusters is exacerbated when clusters are of asymmetric size or when there are very few treated units.

“The problem of few clusters is exacerbated when the clusters are of asymmetric size or when there are very few treated units.”

The quote only partly states this claim.

Issues will arise in interpreting the resulting coefficient when using a one-event-per-unit approach.

“Again, issues will arise in interpreting the resulting coefficient.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Sandler and Sandler (2014) suggest the approach of allowing multiple events per unit by turning on multiple event-time dummies.

“Sandler and Sandler (2014) suggest this approach.”

• cites Sandler and Sandler (2014)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The impact of Medicaid introduction varied state-by-state as a function of the fraction of population receiving AFDC assistance at the time of introduction.

“The impact of introduction varied state-by-state as a function of the fraction of population that was receiving assistance from the Aid to Families with Dependent Children welfare program at the time when Medicaid was introduced.”

• cites Goodman-Bacon (2018) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable).

Controlling for unit-specific trends is suitable if pre-trends reflect trending omitted variables that could bias main treatment estimates.

“This approach is suitable if we believe that pre-trends reflect trending omitted variables that could bias the main estimates of the treatment.”

May combine two claims.

If there are too few pretreatment periods, it can be hard to distinguish between actual pre-trends and statistical noise, limiting comfort from passing a test of no visible pretrend.

“A separate difficulty is that if you have too few pretreatment periods, it can be hard to distinguish between actual pre-trends and statistical noise.This limits the comfort a researcher can take from”

May combine two claims.

Because the counterfactual controls for trends, pre-trends should look flat by construction, so we lose the basic falsification test that pre-trends provide in the basic model.

“Second, because the counterfactual now controls for these trends, our pre-trends should look flat by construction.Thus, we lose the basic falsification test that the pre-trends provide in the basic model.”

May combine two claims.

The additional restrictions are untestable, but estimated coefficients on trend terms will be affected by the restrictions placed.

“The choice of which restriction to apply is guided by the same principle as before: the additional restrictions are untestable, but estimated coefficients on the trend terms will be affected by the restrictions we place, so we want it to be as”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

In principle, the basic model could be modified to estimate sensitivity in treatment effects directly, based on the history of prior events, but this has not yet been implemented.

“In principle, the basic model could be modified to estimate sensitivity in treatment effects directly, based on the history of prior events, but I have not yet seen this implemented.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the Finkelstein, Gentzkow, and Williams (2016) design, the event is the move, and the variable intensity reflects the difference in medical usage between the destination and origin locations.

“In this case the event is the move, and the variable intensity reflects the difference in medical usage between the destination and origin locations.”

May combine two claims.

Controlling for unit trends changes the counterfactual thought experiment, so treatment effect estimates have an interpretation of outcome compared to reference period net of underlying linear trends in the counterfactual.

“First, controlling for unit trends changes the counterfactual thought experiment.Our treatment effect estimates now have an interpretation of”

May combine two claims. The quote only partly states this claim.

Spline restrictions for event study coefficients can improve statistical power but risk mischaracterizing the pattern of treatment effects if the imposed model is not flexible enough.

“However, this approach comes with some risk of mischaracterizing the pattern of treatment effects, in particular if the imposed model is not flexible enough to reflect reality.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Treating the actual event date as an observable variable and estimating treatment effects based on the date should work, so long as there are control units or enough variation in event dates.

“This approach should work, so long as there are control units or enough variation in event dates.”

May combine two claims.

The trend estimates will be estimated in the context of the restrictions; later treatment estimates may pivot as a result of imposing the restriction.

“The trend estimates will be estimated in the context of those restrictions; in particular, the later treatment estimates may”

Probably not a substantial claim: may not pass S3 (load-bearing).

Mover designs often pair with variable intensity of treatment.

“These mover designs often pair with variable intensity of treatment.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

In a setting with varying event dates, we might want to model the possibility that the dynamic treatment effects depend on the timing of adoption.

“In a setting with varying event dates, we might want to model the possibility that the dynamic treatment effects depend on the timing of adoption.”

Probably not a substantial claim: may not pass S2 (checkable).

Despite these risks, these decisions are unavoidable.

“Despite these risks, these decisions are unavoidable.”

Classified as probably not a substantial claim.

The specification error here is the assumption of treatment effects that are the same for both early and late treated units.

“The specification error here is the assumption of treatment effects that are the same for both early and late treated units.”

Probably not a substantial claim: may not pass S2 (checkable).

This raises risks of bias due to systematic (if perhaps unconscious) model selection processes, committed by either the researcher or the journal review process.

“This raises risks of bias due to systematic (if perhaps unconscious) model selection processes, committed by either the researcher or the journal review process.”

Classified as probably not a substantial claim.

The challenge in cohort studies is how best to deal with cohort, age, and time (of survey) effects.

“The challenge here is how best to deal with cohort, age, and time (of survey) effects.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

This contrasts with our usual interpretation of pre-trends as indicating anticipation effects or different underlying trends in untreated potential outcomes.

“This contrasts with our usual interpretation of pre-trends as indicating anticipation effects or different underlying trends in untreated potential outcomes.”

May combine two claims.

The multicollinearity is due to the fact that event time, calendar time, and event date are connected by j = t - E_i, and that event dates can be defined by unit dummies.

“This multicollinearity is due to the fact that event time, calendar time, and event date are connected by j = t -E i , and that event dates can be defined by unit dummies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Our data might not naturally reflect those weights, perhaps especially when we are combining data from different-sized datasets.

“However, our data might not naturally reflect those weights, perhaps especially when we are combining data from different-sized datasets.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This theme is addressed for ordinary least squares in Angrist (1998) and Sloczynski (2022); for one way fixed effects models in Gibbons, Suárez Serrato, and Urbancic (2019) and Miller, Shenhav, and Grosz (2021); and for difference-in-difference in Goodman-Bacon (2021a), Callaway and Sant'Anna (2021), de Chaisemartin and D'Haultfœuille (2020), and Borusyak, Jaravel, and Spiess (2022).

“This theme is addressed for ordinary least squares inAngrist (1998) andSloczynski (2022); for one way fixed effects models in Gibbons, Suárez Serrato, and Urbancic (2019) andMiller, Shenhav, and Grosz (2021); and for difference-in-difference in Goodman-Bacon (2021a), Callaway and Sant'Anna (2021), de Chaisemartin and D'Haultfœuille (2020), andBorusyak, Jaravel, and Spiess (2022).”

• cites Angrist (1998) sourceNot checked — no full text of the source
• cites Sloczynski (2022)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Gibbons, Suárez Serrato, and Urbancic (2019)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

This approach could in principle also be used in situations that use only ever-treated units.

“This approach could in principle also be used in situations that use only ever-treated units.”

May combine two claims.

Sun and Abraham (2021) show that an auxiliary regression can calculate the implied weights and propose an alternative estimation method that works to recover a target average treatment effect.

“They show that an auxiliary regression can calculate the implied weights and also propose an alternative estimation method that works to recover a target average treatment effect.”

• cites Sun and Abraham (2021) sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

In producing an overall estimate, we might want to make sure that each cohort is weighted proportional to its population, because the thought experiment of the model centers on the cohorts.

“A final issue is that in producing an overall estimate, we might want to make sure that each cohort is weighted proportional to its population, because the thought experiment of the model centers on the cohorts.”

Probably not a substantial claim: may not pass S2 (checkable).

Behind the attractive interpretive graph, researchers are necessarily making decisions.

“But behind that attractive interpretive graph, researchers are necessarily making decisions.”

Probably not a substantial claim: may not pass S2 (checkable).

Typically these weights might not correspond to our common-sense intuitions or to our desired weighting.

“Typically these weights might not correspond to our common-sense intuitions or to our desired weighting.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The basic event study specification requires fixed effects for cohort.

“The basic event study specification requires fixed effects for cohort.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Duflo (2001) studies the life-course impact of childhood exposure to school availability in Indonesia based on district and year of birth, and Bailey, Sun, and Timpe (2021) examine the long run impacts of childhood exposure to Head Start in the United States, with treatment based on county and year of birth.

“This situation can arise in cohort studies: for example Duflo (2001) studies the life-course impact of childhood exposure to school availability in Indonesia based on district and year of birth, and Bailey, Sun, and Timpe (2021) examine the long run impacts of childhood exposure to Head Start in the United States, with treatment based on county and year of birth.”

• cites Duflo (2001) sourceNot checked — no full text of the source
• cites Bailey, Sun, and Timpe (2021)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

There is no avoiding the fact that the analysis becomes complicated here.

“There is no avoiding the fact that the analysis becomes complicated here.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Ashenfelter's models were not presented in the now-traditional event study graphical format, but his table's results have an event study framing, including showing a pretraining drop in earnings.

“Ashenfelter's models were not presented in the nowtraditional event study graphical format, but his table's results have an event study framing, including showing a pretraining drop in earnings.”

• cites Ashenfelter (1978) sourceNot checked — no full text of the source