{smcl} {* *! version 2.0.3 30oct2026}{...} {vieweralsosee "csdid" "help csdid"}{...} {vieweralsosee "csdid postestimation" "help csdid_postestimation"}{...} {vieweralsosee "csdid_estat" "help csdid_estat"}{...} {vieweralsosee "csdid_stats" "help csdid_stats"}{...} {vieweralsosee "csdid_plot" "help csdid_plot"}{...} {vieweralsosee "csdid legacy utilities" "help csdid_legacy"}{...} {viewerjumpto "csdid 2.0.2" "csdid_whatsnew##v202"}{...} {viewerjumpto "csdid 2.0.1" "csdid_whatsnew##v201"}{...} {viewerjumpto "csdid 2.0.0" "csdid_whatsnew##v200"}{...} {viewerjumpto "Also see" "csdid_whatsnew##alsosee"}{...} {title:Title} {p2colset 5 22 24 2}{...} {p2col:{bf:csdid whatsnew} {hline 2}}What's new in csdid{p_end} {p2colreset}{...} {pstd} Version 2.0.2 is a patch to 2.0.1. Its changes appear first below. The 2.0.0 section describes what changed for someone upgrading from csdid Version 1.82, the last revision of the 1.8x line; {cmd:ssc install csdid} installs Version 1.81, whose two further differences are listed in {helpb csdid}. The command surface is deliberately the same, so most existing do-files run unchanged. The entries below summarize where they do not, and the things that are new; {help csdid##remarks_legacy:Migrating from Stata csdid Version 1.82} in {helpb csdid} lists every change with what replaces it. {marker v202}{...} {hline 8} {hi:csdid 2.0.2} {hline} {pstd} Anticipation-window comparisons use the last unaffected base period under either base-period setting, and the parallel-trends pretest excludes anticipation periods. Finite positive standard errors are retained even when the outcome is measured in small units. Two-period designs use pointwise intervals. {pstd} Sampling-weight checks use positive support after normalization. Simple and group aggregations apply both {cmd:window()} bounds in the original time units. Balanced event studies require complete observed event-time support before applying display limits. Panel balancing announces vanished cohorts and continues with the surviving comparison groups. Under {cmd:rcs}, nonmissing {cmd:ivar()} values must identify unique observations; repeated cluster codes belong in {cmd:cluster()}. {marker v201}{...} {hline 8} {hi:csdid 2.0.1} {hline} {pstd} {bf:Corrected covariate-adjusted inference in repeated cross sections.} Under {cmd:method(dr)}, the pre-period outcome-regression contribution to the influence function is corrected. Point estimates are unchanged; standard errors, covariance, confidence bands, pre-tests and aggregation inference can differ from 2.0.0. This also applies to panels estimated through {cmd:bal(none)} or {cmd:fix_weights(varying)}. Re-estimate affected results, including saved influence-function files. {pstd} {bf:Clearer diagnostics for undersized regressions.} The diagnostic identifies a failed 2x2 regression with fewer observations than regression coefficients. A square, full-rank design remains estimable. Reduce the covariate list if this diagnostic appears. {marker v200}{...} {hline 8} {hi:csdid 2.0.0} {hline} {pstd} {bf:Five defaults moved. Read these first: they change numbers.} {phang} {bf:1. Not-yet-treated is the default comparison group.} Version 1.82 compared each treated cohort with the never-treated units. Version 2.0.0 also uses every unit not yet treated in either period of the comparison. It uses more of the data, usually gives tighter standard errors, and does not depend on a never-treated group existing or being large enough to trust. {cmd:nevertreated} asks for the older behaviour. See {help csdid##opt_control:comparison-group options}. {phang} {bf:2. The base period is universal.} Version 1.82 measured each cell against the preceding observed period for pre-treatment comparisons. Version 2.0.0 uses one reference per cohort: the last observed period before {it:g - anticipation}, or {it:g}{cmd:-1} on a consecutive calendar without anticipation. Post-treatment effects are the same under either choice; only the pre-treatment cells differ, and the universal base period additionally reports its reference-period normalisation row. Use {cmd:base_period(varying)} when pre-testing, so that a violation shows up in the period where it happens rather than being carried into every later cell. {phang} {bf:3. Standard errors are bootstrapped, with simultaneous confidence bands.} Version 1.82 reported pointwise analytical standard errors unless {cmd:wboot} was asked for. Version 2.0.0 runs the multiplier bootstrap over 1,000 iterations by default and reports bands that cover the estimated effects jointly, because a staggered design produces dozens of estimates and pointwise intervals do not account for looking at all of them at once. An aggregation's overall summary effect is a single number, so it is reported with a pointwise interval. {cmd:analytical} restores analytical standard errors (an aggregation's per-effect rows still carry a simultaneous band, its critical value bootstrapped with a note, unless {cmd:pointwise} is added), and {cmd:pointwise} gives pointwise intervals from the bootstrap. Point estimates are unaffected by either. The bootstrap draws Rademacher multipliers only; Version 1.82's {cmd:wboot} drew Mammen multipliers unless told otherwise. {phang} {bf:4. Propensity scores are trimmed at .995.} Under {cmd:method(dr)} and {cmd:method(ipw)}, comparison observations whose estimated propensity score is .995 or more are dropped, without a message. Version 1.82 did not trim. {cmd:pscoretrim(1)} turns trimming off. {phang} {bf:5. Unbalanced panels are balanced once, and csdid says so.} Version 1.82 balanced each comparison separately: it said the panel was unbalanced, but not which or how many units each comparison dropped. Version 2.0.0 makes the choice explicit: {cmd:bal(full)}, the default, drops units not observed in every period, once, for all comparisons; {cmd:bal(pair)} balances each 2x2 separately, which is what Version 1.82 did; {cmd:bal(none)} keeps every unit. Whenever a mode discards observations, {cmd:csdid} reports how many, and {cmd:e(panel_mode)} records the layout it resolved to. {pstd} {bf:Three more changes move numbers, and no option undoes them.} {phang} {bf:6. notyet pre-treatment cells follow Version 1.82's asinr rule.} A comparison unit must be untreated in both periods of the comparison, as under Version 1.82's {cmd:notyet asinr}; Version 1.82's plain {cmd:notyet} used only the cohorts treated after {it:g} in pre-treatment cells. Under {cmd:notyet base_period(varying)}, pre-treatment cells and the pre-test move; post-treatment cells do not. {phang} {bf:7. long reverses the sign of Version 1.82's long pre-treatment cells.} {cmd:long} and {cmd:long2} both give Version 1.82's {cmd:long2} cells, so after {cmd:long} each pre-treatment cell and each pre-period event-study coefficient changes sign, and the coefficients move one event time earlier. {phang} {bf:8. The overall standard error of estat group} differs from Version 1.82's. The cohort effects, their standard errors and the overall effect are unchanged. {pstd} {bf:New in 2.0.0} {phang} {bf:9. No external dependencies.} Version 1.82 required {cmd:drdid} from SSC. Version 2.0.0 requires nothing beyond Stata itself: the estimation engine is Mata, ships precompiled, and runs on Windows, macOS, and Linux. {phang} {bf:10. A rewritten engine.} On the same data, with 2.0.0 asked for Version 1.82's own defaults so that both compute the same numbers, 2.0.0 runs between 10 and 308 times faster in the published comparisons, with the gap widening as the number of cohorts and periods -- and so the number of ATT(g,t) cells -- grows. {phang} {bf:11. Postestimation in the conventional forms.} {cmd:estat event}, {cmd:estat group}, {cmd:estat calendar}, {cmd:estat simple} and {cmd:estat attgt} are joined by {cmd:estat dynamic}, and every one of them takes {cmd:saving()}, so any aggregation can be written to a dataset without a separate export command. {cmd:estat tidy} and {cmd:estat glance} export the table and the header. {helpb csdid_plot} draws the figure -- also reachable as {cmd:estat plot} -- and {cmd:csdid_plot, saving()} exports the numbers behind it -- estimates, band bounds, axis values -- to draw with {helpb twoway} exactly as you want it. See {helpb csdid_estat} and {helpb csdid_plot}. {phang} {bf:12. Two diagnostics.} {cmd:csdid version} reports the version, the copy of {cmd:csdid.ado} that answered, and the engine the session is using, and changes nothing. {cmd:csdid reset} clears the session's engine decision and estimation cache, so that a csdid installed or replaced mid-session is the one that runs next. See {help csdid##support:Installation, upgrading and diagnostics}. {phang} {bf:13. A bootstrap accelerator on macOS.} The package installs a small compiled accelerator, a universal binary covering Intel and Apple-silicon machines, used for explicitly seeded Rademacher draws. It draws the same multipliers as the Mata path, leaves the same random-number state, and agrees with it to floating-point rounding. Unseeded draws, every other platform, and a Mac where it cannot load use the Mata path, and {cmd:e(bootstrap_accelerator)} and {cmd:e(bootstrap_accelerator_status)} report which path ran. {phang} {bf:14. Repeated cross sections, declared.} {cmd:rcs} says that the data are repeated cross sections while keeping an identifier variable, which {cmd:cluster()} can then use. Omitting {cmd:ivar()} still works and means the same thing. {phang} {bf:15. More is refused instead of being accepted quietly.} A panel whose shape contradicts the design is refused with a message naming the variable at fault: a unit appearing twice in a period, a {cmd:gvar()} that changes within a unit, a {cmd:cluster()} that changes within a unit. Options that Version 1.82 took and then ignored are refused too, rather than leaving inference running at settings you did not ask for. See {help csdid##remarks_behavior:Notes on specific behavior} in {helpb csdid}. {pstd} {bf:Migrating} {phang} Six legacy spellings still run, each with a message saying what it resolved to. Version 1.82 options and postestimation forms that are gone, such as {cmd:method(drimp)}, {cmd:estat pretrend} and {cmd:estore()}, are refused. The option-by-option list, with what replaces each, is at {help csdid##remarks_legacy:Migrating from Stata csdid Version 1.82} in {helpb csdid}, and installation and upgrading are covered at {help csdid##support:Installation, upgrading and diagnostics}. {hline} {marker alsosee}{...} {title:Also see} {psee} Online: {helpb csdid}, {helpb csdid_postestimation:csdid postestimation}, {helpb csdid_estat}, {helpb csdid_stats}, {helpb csdid_plot}, {helpb csgvar}, {helpb csdid_legacy:csdid legacy utilities} {p_end}