Estimates the effect of an intervention that different areas adopt at different dates, using an estimator that compares each adopting cohort with areas that have not yet adopted rather than with areas already treated. Plain two-way fixed effects does the latter and can return an estimate outside the range of every area's true effect; this function exists so that staggered designs are not analysed that way.
Arguments
- panel
A
lamp_panelfromlamp_panel().- outcome
The outcome column, for example
"crime_total"or a crime type key fromlamp_crime_types().- treatment
A
lamp_treatment()object of typestaggered(orevent, which is a single cohort).- estimator
"callaway_santanna","sun_abraham"or"imputation".- control_group
"never_treated"(default) or"not_yet_treated".- cluster
"area"or"force". Callaway–Sant'Anna uses the multiplier bootstrap for force clustering (the backend default of 1000 draws); set the R random seed for reproduction. Sun–Abraham uses clustered regression covariance. Force clustering is not implemented for the imputation wrapper and is rejected.- family
Scale for the outcome:
"ols_log"(default, log of one plus the outcome),"ols_ihs"or"identity".- window
Relative months to report in the dynamic aggregation.
Value
A lamp_estimate of class lamp_did_staggered. coefficients
holds the dynamic (event-study) aggregation with a rel_time column;
diagnostics$overall is the single summary effect with its standard
error; diagnostics$group_time holds the group-time effects for the
Callaway and Sant'Anna backend; and diagnostics$pretrend_p is that
backend's pre-test.
Identifying assumption
Parallel trends holds for each adopting cohort against the chosen control
group, from the period before that cohort adopts onward, and there is no
anticipation: the outcome does not move before adoption in response to it.
With control_group = "never_treated" the comparison is against areas that
never adopt, which must exist; with "not_yet_treated" it is against areas
that adopt later, which uses more data but assumes their later adoption
carries no information about their current path.
Backends
callaway_santanna(default) callsdid::att_gt()anddid::aggte(), giving one effect per cohort and period and the aggregations of them.sun_abrahamusesfixest::sunab(), an interaction-weighted estimator fitted in one regression.imputationuses thedidimputationpackage, which fits the untreated potential outcome and imputes: efficient, but it needs that package, which is in Suggests.
Counts are modelled on a transformed scale by all three backends, which is
why the function warns when a count outcome meets ols_log or ols_ihs:
the coefficients are effects on that scale.
References
Callaway, B. and Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics 225(2), 200-230.
Sun, L. and Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics 225(2), 175-199.
Borusyak, K., Jaravel, X. and Spiess, J. (2024). Revisiting event-study designs: robust and efficient estimation. Review of Economic Studies 91(6).
Examples
# small: most of this example's cost is loading the did package, which
# whichever example reaches it first has to pay
sim <- lamp_simulate(n_areas = 24, n_months = 18, design = "staggered", effect = -0.3, seed = 1)
ad <- attr(sim, "truth")$adoption
tr <- lamp_treatment(sim, "staggered", adoption = ad[!is.na(ad$adoption_month), ])
fit <- lamp_did_staggered(sim, "crime_total", tr)
#> Warning: crime_total is a count, and this backend models it on the ols_log scale.
#> ℹ Coefficients are effects on that transformed scale, not proportional effects
#> on the count; zeros are handled by the transformation, not modelled.
#> You have a balanced panel. Setting allow_unbalanced_panel = FALSE.
#> Warning: Not returning pre-test Wald statistic due to singular covariance matrix
fit
#>
#> ── streetlamp estimate: lamp_did_staggered
#> Outcome: crime_total; family: least squares on log(1 + outcome)
#> Treatment: staggered; clustered by area
#> Identifying assumption: Parallel trends by cohort against never-treated areas,
#> and no anticipation before adoption. Outcome modelled on the ols_log scale.
#> Sample: 432 area-months in 24 areas; 0 rows dropped for coverage.
#> Months since adoption Estimate Std. error 95% CI p
#> ───────────────────── ──────── ────────── ───────────────── ──────
#> -11 −0.0912 0.201 [−0.485, 0.303] 0.650
#> -10 0.165 0.218 [−0.263, 0.593] 0.451
#> -9 −0.352 0.149 [−0.644, −0.0589] 0.019
#> -8 −0.114 0.155 [−0.417, 0.189] 0.460
#> -7 −0.149 0.177 [−0.496, 0.199] 0.401
#> -6 −0.246 0.132 [−0.504, 0.0128] 0.062
#> -5 0.0282 0.141 [−0.249, 0.305] 0.842
#> -4 −0.0317 0.107 [−0.241, 0.177] 0.766
#> -3 −0.01 0.139 [−0.282, 0.262] 0.942
#> -2 −0.0386 0.181 [−0.393, 0.316] 0.831
#> -1 (reference) 0
#> 0 −0.372 0.134 [−0.634, −0.111] 0.005
#> 1 −0.412 0.136 [−0.678, −0.146] 0.002
#> 2 −0.252 0.136 [−0.518, 0.0147] 0.064
#> 3 −0.499 0.135 [−0.764, −0.233] <0.001
#> 4 −0.36 0.144 [−0.641, −0.0781] 0.012
#> 5 −0.519 0.129 [−0.771, −0.267] <0.001
#> 6 −0.242 0.107 [−0.451, −0.0324] 0.024
#> 7 −0.613 0.13 [−0.869, −0.358] <0.001
#> 8 −0.474 0.18 [−0.826, −0.123] 0.008
#> 9 −0.105 0.194 [−0.485, 0.274] 0.586
#> 10 −0.517 0.167 [−0.844, −0.19] 0.002
#> 11 −0.463 0.215 [−0.885, −0.0412] 0.031
#> Overall effect: -0.401 (standard error 0.0894)
fit$diagnostics$overall
#> $estimate
#> [1] -0.4009418
#>
#> $std_error
#> [1] 0.08936057
#>