Skip to content

Outputs

twowayfeweights() returns a result object. Printing it shows the Stata-style report; every number in that report is also stored in the object, so you can use it in your own code.

# pip install twowayfeweights
from twowayfeweights import twowayfeweights, load_wagepan

df = load_wagepan()
res = twowayfeweights(df, Y="lwage", G="nr", T="year", D="union",
                      test_random_weights=["educ", "exper"])

print(res)        # the full report, as in Stata
res.beta          # 0.10662746555843541
res.nr_minus      # 147 cells with a negative weight
res.sum_minus     # -0.010528987131443892
res.sensibility   # 0.09686909502992004

All the statistics

attribute what it is Stata
res.beta The TWFE (or first-difference) coefficient. e(beta)
res.nr_plus Number of treated cells with a positive weight. e(M)[1,1]
res.nr_minus Number of treated cells with a negative weight. e(M)[2,1]
res.nr_weights Number of nonzero weights (nr_plus + nr_minus). e(M)[3,1]
res.sum_plus Sum of the positive weights. e(M)[1,2]
res.sum_minus Sum of the negative weights. e(M)[2,2]
res.tot_cells Number of (group, period) cells that receive the treatment. Cells whose weight is exactly zero are counted here but not in nr_weights.
res.sensibility Summary measure 1: the smallest standard deviation of the treatment effects across cells under which beta is compatible with an average effect of zero. Small values mean beta is fragile. e(lb_se_te)
res.sensibility2 Summary measure 2: the smallest standard deviation under which the treatment effect could have the opposite sign of beta in every cell. None when no weight is negative. e(lb_se_te2)
res.mat The test_random_weights table (see below). None if the option is not used. e(randomweightstest1)
res.other_treatments One entry per variable in other_treatments (see below). e(M2), e(M3)...
res.weights The weight of every (group, period) cell (see below). file saved by path()
res.M The table of counts and sums, laid out like Stata's e(M). e(M)
res.notes The notes printed before the results (e.g. a treatment that varies within cells).
res.type, res.treatment The type used and the name of the treatment variable.

The summary measures are always computed; summary_measures=True only adds them to the printed report. With other_treatments, the summary measures are not defined and are None.

All the numbers at once

res.to_dict()

{'type': 'feTR', 'beta': 0.10662746555843541, 'nr_plus': 820, 'nr_minus': 147, 'nr_weights': 967,
 'sum_plus': 1.0105289871314438, 'sum_minus': -0.010528987131443892, 'tot_cells': 1016,
 'sensibility': 0.09686909502992004, 'sensibility2': 3.175858897581741}

To put several results in one table, collect the dictionaries:

import pandas as pd

rows = []
for t in ["feTR", "feS"]:
    r = twowayfeweights(df, Y="lwage", G="nr", T="year", D="union", type=t)
    rows.append(r.to_dict())
pd.DataFrame(rows)

Counts and sums (Stata's e(M))

res.M

             N_ATTs  Sum_weights
Pos_Weights     820     1.010529
Neg_Weights     147    -0.010529
Tot             967     1.000000

Are the weights correlated with other variables? (res.mat)

With test_random_weights, each variable is regressed on the weights. res.mat is a DataFrame with one row per variable:

column meaning
Coef coefficient of the regression of the variable on the weights
SE its standard error, clustered by group
t-stat Coef / SE
Correlation correlation between the variable and the weights 
res.mat
res.mat.loc["educ", "SE"]     # one number

           Coef        SE    t-stat  Correlation
educ  -0.134455  0.071360 -1.884177    -0.118259
exper -0.201326  0.105611 -1.906288    -0.083915

Like Stata, the command does not report p-values. A two-sided p-value from the t-stat (normal approximation), if you want one:

from scipy import stats
2 * stats.norm.sf(abs(res.mat["t-stat"]))

The weight of every cell (res.weights)

A DataFrame with one row per (group, period) cell, the same content that Stata saves with path():

column meaning
Group_TWFE group
Time_TWFE period
weight the weight of the cell (0 for untreated cells)

With other_treatments, the columns are Group, Time, weight (the main treatment) and weight_others1, weight_others2... (one per other treatment).

w = res.weights
w[w["weight"] < 0]            # the cells with a negative weight
w.groupby("Time_TWFE")["weight"].sum()   # total weight by period

To save it to a file, pass path (or use pandas directly):

res = twowayfeweights(df, Y="lwage", G="nr", T="year", D="union", path="weights.csv")  # or .dta, .parquet
res.weights.to_excel("weights.xlsx")   # needs openpyxl

Other treatments (res.other_treatments)

With other_treatments, res.other_treatments is a list with one entry per other treatment, in the order given. Each entry has the same fields as the main treatment:

res = twowayfeweights(df, Y="lwage", G="nr", T="year", D="union", other_treatments=["married"])
o = res.other_treatments[0]
o.name, o.nr_plus, o.nr_minus, o.sum_plus, o.sum_minus, o.tot_cells

('married', 917, 997, 0.4827791014372501, -0.4827791014372501, 1914)

Code written for older versions

Dictionary access with the names of the previous Python version still works, e.g. res["beta"], res["n_pos"], res["n_neg"], res["sum_neg"], res["dat_result"] (the cell weights) and res["test_random_weights"]["educ"]["coef"]. print_twowayfeweights(res) prints the report.