Getting started

The estimand

The building block is the group-time average treatment effect ATT(g,t): the average effect, in period t, on the cohort first treated in period g.

A staggered design produces one of these for every cohort and every period, which is more numbers than anyone reads directly. csdid estimates all of them and leaves the aggregation to you. The cells themselves stay available for inspection after estimation.

The three choices

The comparison group. Not-yet-treated units are the default, and nevertreated restricts the comparison to units that are never treated. Not-yet-treated gives you more comparisons and does not require a never-treated group to exist. The cost is an extra assumption: later-treated cohorts’ untreated paths have to be comparable over the periods they spend in the comparison group. We would check that against the pre-treatment cells before leaning on it.

The 2×2 estimator. method(dr) (default) is doubly robust: consistent if either the outcome model or the propensity score is right. method(reg) is outcome regression, and method(ipw) is inverse probability weighting. We recommend keeping the default unless you have a reason to trust one of the two models over the other.

The aggregation. estat event (by time since treatment), estat group (by cohort), estat calendar (by period), estat simple (one number).

Your first estimate

Let’s run one from the raw file to the event study, on county mortality data.

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 > 2019

bysort county_code: generate byte nyears = _N
keep if nyears == 11

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

bindquote(strict) is not optional here: the file has quoted fields containing newlines, and without it Stata splits them into extra observations. Several columns carry "NA", hence the destring ... , force. The last two lines are the whole estimation.

Two rules about the axis

gvar() is 0 for never-treated units and the first treated period otherwise; time() is 1 or more. Use the same calendar-time units for both variables, and reserve gvar() == 0 for never-treated units. If your time axis starts at or below zero, add the same constant to time() and to treated cohorts only, leaving never-treated codes at zero. This preserves calendar distances, event times, and the anticipation setting.

Next: balanced panels.