staggeredGMM 0.2.0

Breaking changes

• The beck_banks dataset has been removed, together with its licence note. The examples and the vignette now illustrate the always-treated error and estimation without a never-treated group on sim_panel.

Bug fixes

• When never-treated units are coded 0 and period 0 lies inside the observed window, gmm_staggered() now stops with an error. Previously 0 was read both as a treated cohort, with cohort-by-time effects estimated for it, and as the never-treated control group. Code never-treated units with any value after the last period instead.
• When a covariate has no observed value for any unit of a treated cohort, gmm_staggered() now stops with an error naming the covariate and the cohort. Previously every estimate was returned as NA, with a misleading warning that no weighted step could be completed.

Documentation and output

• gmm_j_test() is now explicitly a test of unconditional parallel trends. Its placebo moments were never covariate-adjusted; after a fit with covar, the printed result now warns that it does not test the covariate-conditional parallel trends assumption, summary() of such a fit says the same, and the returned object records scope and covar.

staggeredGMM 0.1.0

First release.

Interface

• Single exported estimator gmm_staggered(), replacing the earlier gmm_staggered_I(), gmm_staggered_II() and gmm_staggered_III(). The covariance model is chosen with weighting, one of "pooled_toeplitz", "cohort_toeplitz" or "unrestricted" (GMM-T, GMM-HT and GMM-U in the paper).
• Argument names follow the did package convention: yname, tname, idname, gname.
• never_treated = NULL detects a never-treated group automatically, replacing has_nt = FALSE.
• Results are an S3 object of class staggered_gmm with print, summary, coef, vcov and confint methods.
• New gmm_j_test(): the serial-correlation robust over-identification test of parallel trends and no anticipation, in full-set and local-pre-window forms.

Estimation

• Both aggregate targets are reported: CW (treated-observation weighting, as used throughout the paper) and EW (cohort-equal weighting).
• Cohort-by-time effects with no clean comparison are reported as NA with identified = FALSE, never as an estimated zero, and are excluded from the parameter vector so that partial identification no longer disables the efficient weighting for every other cell.
• Unbalanced panels are supported: autocovariances use pairwise-complete denominators, and the variance of a group-period mean accounts for the overlap between the unit sets contributing at each pair of periods. Both reduce exactly to the complete-data formulas when the panel is balanced.
• Convergence and solve success are reported separately, with a termination reason and the final parameter change. max_iter now defaults to 100: the iteration can converge linearly at a slow rate, and sim_panel under weighting = "cohort_toeplitz" needs 86 steps.
• Failure to complete any weighted step always warns, including under weighting = "unrestricted", because the returned estimates are then the identity-weighted seed rather than a GMM estimate under the requested weighting. The warning reports how many independent moment directions the estimated weight supported, against the number of effects to identify. Merely failing to converge after a weighted step did run continues not to warn under "unrestricted", where it is routine.
• Units whose cohort falls after the last observed period are used as clean controls rather than discarded.

Validation

• A shared validator checks every documented requirement before estimation and reports the offending rows, units, periods or cohorts. Column types are rejected rather than coerced.
• Units treated at or before the first observed period now raise an error. They have no pre-treatment period, so none of their effects is identified.

Data

• beck_banks documentation now records its CC BY 4.0 licence and the thirteen always-treated states it contains.