Reproducing Stata defaults¶

Most "StatsPAI gives a different number" reports on replication packages come from a Stata default rather than an estimator. This page lists the defaults that decide N, the sample, or a threshold, and the StatsPAI argument that corresponds to each. Every row is checked against Stata 18 by a test in tests/reference_parity/.

Commands and options¶

Stata	StatsPAI	Notes
`ppmlhdfe y x, absorb(id ind#year) vce(cluster city#year)`	`sp.ppmlhdfe("y ~ x", data=df, absorb="id + ind#year", cluster="city#year")`	`^` (fixest) and `#` (Stata) both build interacted groups.
`ppmlhdfe` singleton and separation drops	default (`drop_singletons=True`, `separation=True`)	`model_info['n_singletons']` + `['n_separated']` equal Stata's `e(num_singletons)`. `drop_singletons=False` matches R `fixest::fepois`.
`ppmlhdfe` "(omitted)" regressors	automatic	Regressors absorbed by the fixed effects are listed in `model_info['omitted']`.
Hand-built stack + `ppmlhdfe y tp x, absorb(id#cohort year#cohort ...)`	`sp.stacked_did(df, y=..., group=..., time=..., first_treat=..., family="poisson", spec="pooled", absorb="...", control_group="notyettreated_rows")`	Later-treated units serve as controls only before their own treatment. `spec="event_study"` gives the event-time coefficients.
`reghdfe ..., vce(cluster c)`: `e(r2)`, `e(r2_a)`, `e(r2_within)`, `e(df_a)`, `e(df_a_nested)`	`sp.hdfe_ols(...).r2`, `.r2_a`, `.r2_within`, `.df_a`, `.df_a_nested`	`r2_a` charges effects nested in the cluster, as `reghdfe` does.
`winsor2 v, cuts(1 99)`	`sp.winsor(df, ["v"], cuts=(1, 99))`	Stata `_pctile` percentiles by default; `method="linear"` for numpy's.
`winsor2 v if year>=2007, by(g)`	`sp.winsor(df, ["v"], subset="year >= 2007", by="g")`	Outside `subset`, new columns are missing and `replace=True` leaves values unchanged.
`logit y x i.g` ("predicts failure perfectly")	`sp.logit("y ~ x + C(g)", data=df)`	Default `perfect_prediction="drop"`; `"keep"` keeps the rows. Same for `probit` / `cloglog`.
`psmatch2 d x, outcome(y) neighbor(1) ties ate common`	`sp.psmatch2(df, treat="d", covariates=["x"], outcome="y", common_support="minmax", ties=True, ate=True)`	PSM-DID on `_weight != .` uses `m.matched_data["_weight"].notna()`. ATU / ATE are in `m.result.model_info`.
`xtreg y d x i.ind#i.year, fe vce(cluster c)` then `psacalc` with `rmax(1.3*e(r2_a))`	`sp.oster_bounds(df, y="y", treat="d", controls=["x"], absorb="id", absorb_controls="ind#year", cluster="c", r_max="1.3*r2_a")`	Exact Oster solution. `moments=` gives the exact solution from summary statistics.
`csdid2 y, ivar(id) time(t) gvar(g)` on an unbalanced panel	`sp.callaway_santanna(..., allow_unbalanced_panel=True)`	The default (`False`) keeps only units observed in both periods of a comparison and gives a different point estimate.
`csdid` 2.0.0 or R `did::att_gt` on an unbalanced panel, at their defaults (`bal(full)`, `allow_unbalanced_panel = FALSE`)	`sp.callaway_santanna(..., balance="full")`	Both drop every unit not observed in all periods before estimating. StatsPAI's default is `balance="pair"`, the rule of `csdid` 1.8x (`bal(pair)` in 2.0.0).
`csdid y, time(t) gvar(g) cluster(c) agg(group)` without `ivar()` (rows as repeated cross-sections)	`sp.callaway_santanna(..., panel=False, clustervars="c")` then `sp.aggte(r, type="group", agg_weights="csdid", share_variance=False)`	Any row-level cluster, e.g. `city#year`. Equal to `csdid` v1, including its estimated cell-count weights.
`csdid2` with no never-treated units	`control_group="notyettreated"`	`csdid2` switches to not-yet-treated controls without saying so; StatsPAI asks for it explicitly.

Known remaining differences¶

Propensity scores from logit. StatsPAI and Stata agree to about 1e-7, which is Stata's convergence tolerance. Nearest-neighbour matching on 10^5 scores can switch a few hundred matches at that scale. Given the same score, sp.psmatch2 reproduces psmatch2 exactly.
xtreg, fe's e(r2_a) with thousands of dummies. StatsPAI counts the control dummies at their exact rank given the panel effect. On one replication, Stata's figure implied three more regressors than that rank. Pass Stata's number as r_max= when a table must be reproduced digit for digit.
csdid2's standard errors on repeated cross-sections. csdid2 (without ivar()) gives the same point estimates as csdid and StatsPAI, but smaller SEs. Its influence function for control rows is scaled by the share of the whole 2×2 subsample instead of the control cell's share. StatsPAI follows csdid v1 and R DRDID, which agree with each other. On one replication this moved a z statistic from 3.171 (csdid2) to 3.113 (csdid, StatsPAI).