Arguments¶

from twowayfeweights import twowayfeweights

res = twowayfeweights(data, Y, G, T, D, type="feTR", D0=None, summary_measures=False,
                      controls=None, weights=None, other_treatments=None,
                      test_random_weights=None, path=None)

Variables are always given by their column name in data. Where an argument accepts several variables, you can pass one name ("educ") or a list (["educ", "exper"]).

Required¶

argument	Stata	what it is
`data`	the dataset in memory	A pandas DataFrame with one row per observation. Other data frames must be converted first, e.g. `df.to_pandas()` for polars.
`Y`	1st variable	The outcome. For `fdTR` and `fdS`, the first difference of the outcome.
`G`	2nd variable	The group identifier (person, county, firm...). Numbers, strings or categories.
`T`	3rd variable	The time period. Must be sortable (years, months, 1, 2, 3...).
`D`	4th variable	The treatment. For `fdTR` and `fdS`, the first difference of the treatment.

Options¶

argument	Stata	default	what it does
`type`	`type()`	`"feTR"`	Which regression and which assumptions. See below.
`D0`	5th variable	`None`	The treatment level (not differenced). Required with `type="fdTR"`, ignored otherwise.
`summary_measures`	`summary_measures`	`False`	Print the two sensitivity measures. They are always computed and stored, whatever this option.
`controls`	`controls()`	`None`	Control variables included in the regression.
`weights`	`weight()`	`None`	A variable of analytic weights. Must not be negative.
`other_treatments`	`other_treatments()`	`None`	Other treatment variables included in the regression. Only with `type="feTR"`. The output then shows the weights attached to each treatment.
`test_random_weights`	`test_random_weights()`	`None`	Variables regressed on the weights, to see whether the weights are correlated with things that may drive the treatment effect (e.g. education, age).
`path`	`path()`	`None`	Save the weight of every (group, period) cell to a file: `.csv`, `.dta` (Stata) or `.parquet` (needs the `pyarrow` package).

The four types¶

`type`	regression	assumptions
`"feTR"`	outcome on treatment, with group and period fixed effects	common trends
`"feS"`	same as `feTR`	common trends, and each group's treatment effect does not change over time
`"fdTR"`	first difference of the outcome on first difference of the treatment, with period fixed effects	common trends (needs `D0`)
`"fdS"`	same as `fdTR`	common trends, and treatment effects stable over time

Use the fe types when you estimated a fixed effects regression, and the fd types when you estimated a first-difference regression. Use an S type if you are willing to assume that each group's treatment effect is the same in every period.

What the data can look like¶

Several observations per (group, period) cell are allowed, like individual data inside counties. If the treatment or a control varies within a cell, it is replaced by its cell average, and the output starts with a note saying so, as in Stata.
Unbalanced panels and gaps are fine.
Missing values: rows with a missing outcome, group, period, treatment or control are dropped, following the same rules as Stata.

Errors¶

error	when
`TypeError`	`data` is not a pandas DataFrame
`KeyError`	a column name is not in `data`
`ValueError`	unknown `type`; `fdTR` without `D0`; `other_treatments` with a type other than `feTR`; negative weights; no observations left after dropping missing values