Options¶

The Python options follow the terminology used in the paper and Stata package. Below, each supported Stata option is followed by its Python equivalent and explanation.

Main options¶

`estimator(string)` — `estimator=...`¶

Names of the estimator(s) to compute. Allowed values:

"as" — AS (Average Slope).
"was" — WAS (Weighted Average Slope).
"iv-was" — IV-WAS.

You can pass a string or a list:

estimator="was"
estimator=["as", "was"]

If multiple estimators are requested in Stata, they must be separated by a single space.

`exact_match` — `exact_match=True`¶

The DID estimators compare the outcome evolution of switchers and stayers with the same period-\((t-1)\) treatment (or instrument) value. Only valid for binary or discrete treatments. With a discrete treatment taking a large number of values, specifying this option may be undesirable: there may only be few switchers that can be matched to a stayer with the exact same period-\((t-1)\) treatment, restricting the estimation sample.

The package uses doubly robust estimation by default. When exact_match=True, regression adjustment is selected internally. RA and PS are not user-selectable estimation methods.

`order(#/####/########)` — `order=int|list`¶

When exact_match is not specified, this option specifies the polynomial orders used in the OLS regressions of \(Y_t - Y_{t-1}\) on a polynomial in \(D_{t-1}\) and/or in the logistic regressions of an indicator for \((t-1)\)-to-\(t\) switchers on a polynomial in \(D_{t-1}\).

The option allows for 1, 4, or 8 arguments (8 only allowed for IV-WAS):

Arguments	Meaning
`order=1` (default)	Same order for everything
`order=[1, 4, 3, 2]`	`[E(ΔY\|D_{t-1}), P(S_t=0\|D_{t-1}), P(S_{+t}=1\|D_{t-1}), P(S_{-t}=1\|D_{t-1})]`
`order=[1, 4, 3, 2, 1, 2, 3, 4]`	First 4: first-stage; last 4: reduced-form (IV only)

`placebo(#)` — `placebo=N`¶

The command computes placebo versions of each estimator requested. Actual estimators compare the \((t-1)\)-to-\(t\) outcome evolution of \((t-1)\)-to-\(t\) switchers and stayers with the same baseline treatment.

When N=1, placebos compare the \((t-2)\)-to-\((t-1)\) outcome evolution of \((t-1)\)-to-\(t\) switchers and stayers with the same baseline treatment, restricting attention to \((t-2)\)-to-\((t-1)\) stayers.
When N>1, placebos comparing the outcome evolutions of \((t-1)\)-to-\(t\) switchers and stayers from \(t-3\) to \(t-2\), from \(t-4\) to \(t-3\), ..., and from \(t-N-1\) to \(t-N\) are also reported, always restricting attention to stayers between those pairs of periods.

`as_vs_was` — `as_vs_was=True`¶

Shows a test that the AS and WAS are equal. Only valid when both AS and WAS estimation is requested, i.e. estimator=["as", "was"].

`controls(varlist)` — `controls=[...]`¶

Estimators with control variables rely on a conditional parallel trends assumption: the counterfactual outcome evolution of switchers had their treatment not changed is equal to the outcome evolution of stayers with the same baseline treatment and the same value of the controls.

When time-varying controls are inputted, the command compares the \((t-1)\)-to-\(t\) outcome evolution of switchers and stayers with the same baseline treatment and with the same controls at period \(t-1\).

Warning

Specifying too many control variables may lead to noisy estimators. If placebo estimators are small, insignificant, and precisely estimated without controls, including controls may not be necessary.

`weights(varname)` — `weight="..."`¶

Compute estimators weighted by varname.

`trimming(#)` — `trimming=N`¶

Integer between 0 and 100.

If AS or WAS is requested: trim observations for which the estimated \(P(S_t=0 \mid D_{t-1})\) is lower than N%.
If IV-WAS: trim observations such that \(P(SI_t=1 \mid Z_{t-1}, D_{t-1})\) is lower than N%.

Only relevant when probabilities are estimated with logit models — not compatible with exact_match.

Heterogeneous treatment effects¶

`switchers(string)` — `switchers="up"|"down"`¶

"up": estimate the AS / WAS / IV-WAS for switchers-up only (units whose treatment or instrument increases from \(t-1\) to \(t\)).
"down": estimate for switchers-down only.
Default: all switchers.

`by_fd(#)` — `by_fd=N`¶

Assess heterogeneity by the absolute value of the treatment's (or instrument's) change. For example, by_fd=5 splits switchers into 5 quintiles of \(|\Delta D_t|\) (or \(|\Delta Z_t|\)), then estimates treatment effects per subsample. If \(|\Delta D_t|\) has mass points, the command splits switchers into groups with as-equal-as-possible sizes.

`by_baseline(#)` — `by_baseline=N`¶

Same as by_fd but quantiles are built on \(D_{t-1}\) (or \(Z_{t-1}\)).

`bysort varlist:` — `by=[...]`¶

In Stata, prefix the command with bysort varlist:. In Python, pass by=["v1", "v2"]. Only time-invariant variables are allowed.

Standard-error options¶

`bootstrap(#)` — `bootstrap=N`¶

Compute bootstrap standard errors with N replications.

Use case: if the number of switchers or stayers is low, you may want to check the analytic SEs against bootstrap SEs as a diagnostic. Currently only allowed when IV-WAS is requested (asymptotic approximations are more likely to fail with weak instruments).

`seed(#)` — `seed=N`¶

Set the bootstrap seed for reproducibility.

`cluster(varlist)` — `cluster="..."`¶

Cluster standard errors at the level of the supplied variable.

Advanced options¶

`other_treatments(varlist)` — `other_treatments=[...]`¶

Control for other treatments that may also change over the panel. See de Chaisemartin and D'Haultfœuille (2021) for further details.

`noextrapolation` — `noextrapolation=True`¶

Keep only switchers whose period-\((t-1)\) treatment (or instrument) is between the minimum and maximum values of stayers' period-\((t-1)\) treatment, enforcing the overlap condition.

`cross_fitting(#)` — `cross_fitting=K`¶

Perform \(K\)-fold cross-fitting for the doubly-robust estimator. With K=2, the command splits the sample into two halves, fits nuisance functions on one and estimates the parameter on the other, then swaps roles. The final point estimate is a weighted average.

See Section 3.3 of de Chaisemartin et al. (2025).

`on_placebo_sample` — `on_placebo_sample=True`¶

Compute the treatment-effect estimator on the subsample where the first placebo estimator is computed. The resulting estimator remains valid if the first treatment lag affects the outcome.

TWFE comparison¶

`twfe(twfe_suboptions)` — `twfe=True | {dict}`¶

Compare to a TWFE estimator from regressing \(Y_{i,t}\) on \(D_{i,t}\) and unit + year fixed effects. With IV-WAS, a 2SLS-TWFE is used, with \(Z_{i,t}\) as the instrument.

The command displays a table with the difference between the requested estimator and the TWFE estimator, the \(p\)-value of the test of the difference, and the corresponding CI. By default, 100 bootstrap replications. Increase with bootstrap=N. For replicability use seed=N.

You must specify either full_sample or same_sample.

Suboption	Description
`same_sample`	Estimate the TWFE regression on the same sample as `did_multiplegt_stat`. Use this when the command does not use all time periods (e.g., a period \(p\) has no switchers, so the pair \((p-1, p)\) is dropped).
`full_sample`	Estimate the TWFE regression on the full sample.
`percentile`	Use the percentile bootstrap for \(p\)-values and CIs (default is normal approximation).

Python:

twfe={"same_sample": True}
twfe={"full_sample": True, "percentile": True}

Display¶

`disaggregate` — `disaggregate=True`¶

Show estimated AS / WAS / IV-WAS effects for each pair of consecutive time periods, on top of the aggregated effects. By default, only aggregated effects are shown.

`graph_off` and `bys_graph_off`¶

Stata's did_multiplegt_stat displays a graph by default. In Python, no graph is shown unless you explicitly call model.plot(), so these flags are not needed.

Backend selection (Python-only)¶

These options have no Stata equivalent — they control which Python regression engine is used for the OLS / logit nuisance estimations.

`asinstata=True | False`¶

Value	Behaviour
`False` (default)	scikit-learn `LinearRegression` / `LogisticRegression`. Modern, fast.
`True`	statsmodels OLS + from-scratch Newton-Raphson logit replicating Stata's `logit, asis` to ~1e-7 relative error.

`iv_method="manual" | "linearmodels" | "econtools"`¶

IV regression backend (used in the bootstrap-based TWFE comparison for IV-WAS):

"manual" (default) — two-step OLS (no external dependencies).
"linearmodels" — uses linearmodels.iv.IV2SLS with proper IV standard errors. Install with pip install "did-multiplegt-stat[linearmodels]".
"econtools" — uses econtools.metrics.ivreg. Install with pip install econtools.

`model_deltay`, `model_stayer`¶

Pluggable custom nuisance models. Must implement sklearn-style .fit(X, y) and .predict(X) (regressor) or .predict_proba(X) (classifier).

from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier

model = DIDMultiplegtStat(
    model_deltay=RandomForestRegressor(n_estimators=100),
    model_stayer=RandomForestClassifier(n_estimators=100),
)

Each fit() call deep-copies the template, so the original object is never mutated.

Cross-validation for polynomial order¶

`cross_validation=dict | None`¶

If specified, order is selected by \(K\)-fold cross-validation. Pass a dict with any of:

Key	Default	Meaning
`algorithm`	`"kfolds"`	Currently only kfolds is implemented.
`tolerance`	`0.01`	Stop when the relative MSE improvement falls below this.
`max_k`	`5`	Maximum polynomial order to try.
`kfolds`	`5`	Number of folds.
`seed`	`0`	RNG seed.
`same_order_all_logits`	`False`	If `True`, use the same order for all logits.

model = DIDMultiplegtStat(
    cross_validation={"max_k": 4, "kfolds": 10, "seed": 42},
)

Options¶

Main options¶

estimator(string) — estimator=...¶

exact_match — exact_match=True¶

order(#/####/########) — order=int|list¶

placebo(#) — placebo=N¶

as_vs_was — as_vs_was=True¶

controls(varlist) — controls=[...]¶

weights(varname) — weight="..."¶

trimming(#) — trimming=N¶

Heterogeneous treatment effects¶

switchers(string) — switchers="up"|"down"¶

by_fd(#) — by_fd=N¶

by_baseline(#) — by_baseline=N¶

bysort varlist: — by=[...]¶