Anticipation

Callaway–Sant’Anna assumes that units do not respond before treatment starts. That assumption fails when treatment is announced in advance. A state’s Medicaid expansion is legislated well before it takes effect, and the behavior of hospitals and households alike can move in between.

If units respond k periods early, then those k periods are not clean pre-treatment periods, and using one of them as the base period contaminates every comparison drawn from it. anticipation(#) tells csdid to treat the last # pre-treatment periods as already affected and to measure from before them.

The data

Let’s build the balanced sample once. It is the same block used elsewhere on this site, and we run it twice below (once with an anticipation window and once without).

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
save "jel_balanced.dta", replace

No anticipation: the default

use "jel_balanced.dta", clear
csdid mrate, ivar(county_code) time(year) gvar(gvar) analytical
display "anticipation: " e(anticipation)
display "cells: " e(N_attgt)
estat event

The stored anticipation value is zero here (that is the default), and the run returns the full set of pre-treatment cells, so this is the specification most readers will assume you used.

Allowing one period of anticipation

use "jel_balanced.dta", clear
csdid mrate, ivar(county_code) time(year) gvar(gvar) anticipation(1) analytical
display "anticipation: " e(anticipation)
display "cells: " e(N_attgt)
estat event

Two things happen here. The base period moves one period earlier, so that every comparison is drawn from a period the cohort could not yet have responded to. The period immediately before treatment is no longer reported as a pre-treatment placebo. It is now part of the response.

Compare the event-study output from the two runs. On this consecutive calendar, the universal normalization moves from event time −1 to −2. The effect at −1 remains in the table and may reflect anticipation; it is no longer an unaffected reference. These cohorts all retain a clean base period, so the number of ATT(g,t) cells does not change. Cohorts too near the start of a shorter sample can instead be dropped.

Choosing the value

anticipation() encodes an assumption about behavior; it does not test one. We recommend setting it from what you know about the policy (when it was announced, when it was signed, when it became widely expected) instead of by trying values until the pre-trends look flat, which amounts to searching over specifications. Searching over it and keeping the value with the best-looking pre-treatment plot invalidates the pre-test you are then using to justify the design.

If you genuinely do not know, the two options we would consider are reporting the default alongside a sensitivity check at one period, or moving to a design that does not lean on the immediate pre-treatment period at all.

The cost is real, because each anticipated period removes one usable pre-treatment period. Effects are then measured from further back, and cohorts treated early enough in the sample may drop out entirely for want of a clean base period (the earliest cohorts have the fewest pre-periods to spare). We would still take the value from the institutional record, and we would not set the window to zero simply to keep those cohorts in the sample.

Interaction with the comparison group

With notyet (the default), later-treated cohorts serve as comparison units, and if those cohorts are themselves anticipating, then they are not clean comparisons either. anticipation() applies to them as well. A unit leaves the comparison group once it is within the anticipation window of its own treatment date, which can be a period earlier than the date itself. A unit can sit in the comparison group one period and be out of it the next. That bookkeeping is done for you.

use "jel_balanced.dta", clear
csdid mrate, ivar(county_code) time(year) gvar(gvar) notyet anticipation(1) analytical
estat event

The reported joint Wald pre-test still tests estimable cells before the nominal treatment date, including those inside the anticipation window. It is not a test restricted to the periods before anticipation begins. Inspect those earlier cells when assessing the no-anticipation assumption.

See Comparison groups for the choice itself, and Pre-testing for reading the pre-treatment cells that remain.