Aggregations

ATT(g,t) is a table, and an aggregation is what turns that table into an answer. Which one you want depends on the question you are asking. They answer different questions – about exposure, about cohorts, about calendar time – so a difference between two of them is information rather than a problem to be reconciled.

The data

The runs below use the county mortality panel from the JEL-DiD replication package.

import delimited using ///
    "https://raw.githubusercontent.com/pedrohcgs/JEL-DiD/50f4f18/data/county_mortality_data.csv", ///
    clear varnames(1) bindquote(strict) stringcols(_all)
destring deaths population_20_64 year yaca county_code stfips unemp_rate poverty_rate, ///
    replace force
generate double mrate = 100000 * deaths / population_20_64
drop if missing(mrate) | population_20_64 <= 0
generate int gvar = yaca
replace gvar = 0 if missing(gvar) | gvar > 2016
bysort county_code: generate byte nyears = _N
keep if nyears == 11

csdid mrate, ivar(county_code) time(year) gvar(gvar) rseed(20250101)

estat event         // by time since treatment (event study)
estat group         // by cohort
estat calendar      // by calendar period
estat simple        // one overall number

All four aggregation commands read the same table of cohort-period estimates. They differ only in how they average its rows (one row per cohort and period). Note that all four come from one estimation, so you can look at every one without paying for the bootstrap again!

	answers
`event` / `dynamic`	how the effect evolves with exposure
`group`	whether cohorts treated at different times differ
`calendar`	whether the effect differs by period
`simple`	a single summary, weighted by cohort size

We recommend reporting at least the event study and one overall number, the event study first. The two together let a reader see whether the single summary is hiding a trend across exposure. None of these aggregations changes the underlying estimates; they change only which weighted average of those estimates you are reading. Disagreement among them tells you about heterogeneity in the data. It does not mean that one of them is wrong.

Event-time windows

At long event times only the early-treated cohorts contribute (by construction), so a trend across event time can be composition, not dynamics.

We have two tools for this, and they are not interchangeable.

csdid_stats, type(dynamic) window(-3 3)     // restrict the event-time range
csdid_stats, type(dynamic) balance(1)       // only cohorts observed 1+ periods post

window() truncates the range shown; nothing is re-estimated. balance() is the stronger instrument. It restricts the aggregation to cohorts observed for a common number of post-treatment periods, so the composition of the sample is held fixed across event time. If the event-study shape changes materially under balance(), the original shape was partly composition. We would then report the balanced version alongside the unbalanced one.

min_e() and max_e() are synonyms for the two bounds of window(), and balance_e() for balance(). We list both spellings because you will meet both in other people’s code (and in ours).

Missing cells

Some cells cannot be estimated: too few units, no overlap, or a 2×2 comparison that failed outright. The header of the estimation output tells you when this has happened. By default an aggregation containing a missing cell is itself missing, which is loud on purpose. No average is taken over whatever cells happened to survive. na_rm drops the missing cells and averages the rest, which is sometimes the right thing to do:

csdid_stats, type(dynamic) na_rm

Use it deliberately. It changes the estimand from “the average over these cells” to “the average over the cells that worked”. That is a different parameter, not a cleaner estimate of the same one.

That change is not visible in the reported number. If you use it, we recommend saying so in the notes to the table, together with how many cells were dropped and why they failed. A reader who is told that three cells were dropped for lack of overlap will read the remaining average rather differently from one who is told nothing at all.

Next: unbalanced panels.