Stata-to-Python parity¶

This page maps the supported Stata options in did_multiplegt_stat to Python.

How to read this page¶

Stata column shows the exact option name as it appears in did_multiplegt_stat.ado.
Python kwarg is what you pass to either DIDMultiplegtStat(...) or the functional did_multiplegt_stat(...).
Notes flag any semantic differences or wrapper behaviour.

Positional arguments¶

Stata	Python	Notes
`Y` (1st varlist token)	`Y="..."`	Outcome
`G` (2nd)	`ID="..."`	Unit identifier
`T` (3rd)	`Time="..."`	Period
`D` (4th)	`D="..."`	Treatment
`Z` (5th, optional)	`Z="..."`	Instrument (only with `estimator="iv-was"`)

Estimator selection¶

Stata	Python	Notes
`estimator(as)`	`estimator="as"`
`estimator(was)`	`estimator="was"`
`estimator(iv-was)`	`estimator="iv-was"`	Requires `Z`.
`estimator(as was)`	`estimator=["as", "was"]`	Both at once.

Python does not expose RA/PS method selection. It uses doubly robust estimation by default and activates regression adjustment internally only with exact_match=True.

Polynomial order¶

Stata	Python
`or(1)`	`order=1`
`or(1 4 3 2)`	`order=[1, 4, 3, 2]`
`or(1 4 3 2 1 2 3 4)`	`order=[1, 4, 3, 2, 1, 2, 3, 4]`

Sample-restriction options¶

Stata	Python	Notes
`exact_match`	`exact_match=True`
`noextrapolation`	`noextrapolation=True`
`switchers(up)` / `switchers(down)`	`switchers="up"` / `"down"`
`on_placebo_sample`	`on_placebo_sample=True`	Cannot be combined with `placebo` or `iv-was`.

Heterogeneity options¶

Stata	Python	Notes
`bysort g:` (prefix)	`by=["g"]`	Only time-invariant variables allowed.
`by_fd(K)`	`by_fd=K`	Quantile-bin switchers by `\|ΔD\|`.
`by_baseline(K)`	`by_baseline=K`	Quantile-bin by `D_{t-1}`.
`disaggregate`	`disaggregate=True`
`as_vs_was`	`as_vs_was=True`	Requires both `as` and `was`.

Controls / weights / cluster¶

Stata	Python	Notes
`controls(varlist)`	`controls=["v1", "v2"]`
`weights(varname)`	`weight="v"`	Renamed — Stata is plural, Python is singular.
`cluster(varlist)`	`cluster="v"`	Python accepts only a single var.
`other_treatments(varlist)`	`other_treatments=["v1", "v2"]`

Inference¶

Stata	Python
`placebo(N)`	`placebo=N`
`bootstrap(N)`	`bootstrap=N`
`seed(N)`	`seed=N`
`trimming(N)`	`trimming=N` (1–100, percentage)
`cross_fitting(K)`	`cross_fitting=K`

TWFE comparison¶

Stata uses a string suboption: twfe(same_sample percentile). Python uses a dict.

Stata	Python
`twfe(same_sample)`	`twfe={"same_sample": True}`
`twfe(full_sample)`	`twfe={"full_sample": True}`
`twfe(percentile)`	`twfe={"percentile": True}`
`twfe(same_sample percentile)`	`twfe={"same_sample": True, "percentile": True}`

twfe=True alone defaults to the normal-approximation, ambiguous-sample variant — for new code we recommend always passing an explicit dict.

Cross-validation¶

Stata uses a string suboption: cross_validation(kfolds tolerance(0.01) max_k(5)). Python uses a dict:

cross_validation={
    "algorithm": "kfolds",
    "tolerance": 0.01,
    "max_k": 5,
    "kfolds": 5,
    "seed": 0,
    "same_order_all_logits": False,
}

Display options¶

Stata	Python
`graph_off`	(no plot is drawn unless you call `.plot()`)
`bys_graph_off`	(same as above)

The Stata did_multiplegt_stat always prints a graph by default; the Python version only plots on explicit .plot() because libraries should not draw GUI windows from a fit call.

Python-only options¶

Python	Default	Meaning
`asinstata`	`False`	If `True`, use Stata-faithful regressions (statsmodels OLS + custom Newton-Raphson logit).
`iv_method`	`"manual"`	IV backend in the TWFE comparison: `"manual"`, `"linearmodels"`, or `"econtools"`.
`model_deltay`	`None`	Custom sklearn-style regressor for \(E[\Delta Y \mid D_{t-1}, S=0]\).
`model_stayer`	`None`	Custom sklearn-style classifier for \(P(S=0 \mid D_{t-1})\).

Reproducing Stata numbers exactly¶

If you need byte-level parity with the Stata ado-file:

res = did_multiplegt_stat(df, Y, ID, Time, D, Z=Z,
                          asinstata=True,          # critical
                          # ... other options matching the Stata call
                          )

With asinstata=True and identical inputs, agreement is typically:

Point estimates: ≲ 1e-7 relative error.
Standard errors: ≲ 1e-6 relative error.
Bootstrap CIs: not deterministic across runtimes — the MT19937-64 RNG is bit-equivalent to Stata, but pandas operations applied to bootstrap-resampled data can reorder rows differently across machines.

For the bootstrap, you can also import Stata's fold IDs via cf_folds_file= and guarantee an exact CF replication.