Repeated cross sections

Omitting ivar() does the same thing, and for data with no identifier that is the natural way to write it (there is nothing to declare). We provide rcs for the common case where the data do carry an identifier of some kind (a survey respondent number, a county code). You should not have to withhold a real variable in order to describe your data correctly.

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

A worked example

The JEL data are a panel. The block below constructs a repeated cross section from it by keeping one randomly chosen year per county, after which every row belongs to a different unit. Nobody would collect data this way; it is simply the shape we need for the illustration:

set seed 20240617
generate double pick = runiform()
bysort county_code (pick): keep if _n == 1

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

We built this for illustration and would not recommend it as a design, since throwing away 10 of every 11 observations (most of the sample) costs a great deal of precision. We include it because the shape of the data is what matters here, and because a genuine repeated-cross-section dataset with staggered treatment is not part of the JEL package.

When the data carry an identifier anyway

county_code still exists in the extract above, and it is a perfectly real variable (it is still a county code) that no longer identifies a unit followed over time. Declaring the structure explicitly, and keeping the identifier, gives exactly the same estimates.

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

csdid says in the header that it is not using ivar() as a panel identifier (it does not fail silently), and e(idvar) comes back empty while e(panel_mode) reads repeated-cross-section. This mirrors the R did implementation. That implementation validates the identifier and then replaces it with a row number (the same convention, arrived at for the same reason).

The identifier is not wasted here. It is exactly what you pass to cluster() if observations sharing it are correlated:

csdid mrate, ivar(county_code) time(year) gvar(gvar) rcs ///
    cluster(stfips) rseed(20250101)
display "clusters: " e(N_clusters)

Because there is nothing to balance in a cross section (each row appears once), rcs implies bal(none). Asking for bal(full) alongside it is an error, not a silent override, as is fix_weights(base), which needs the same unit to appear in more than one period. See unbalanced panels.

What to watch

Unit counts. Every row is its own cross-sectional unit. e(N_units) equals the number of observations (not the number of counties), and the standard errors are scaled accordingly. There is no within-unit differencing to remove fixed unobserved heterogeneity, so the identifying assumption is doing more work here than it does in a panel. Be careful about the covariates you condition on and about the comparison group you choose.

Covariates. They matter more in this setting, not less. Without differencing, the composition changes between waves are absorbed only by the covariate model, and we would think hard about which covariates go into it.

Next: inference.