multe {multe}

Multiple Treatment Effects Regression

Description

Compute contamination bias diagnostics for the partially linear (PL) regression estimator with multiple treatments. Also report four alternative estimators:

OWN

The own treatment effect component of the PL estimator.

ATE

The unweighted average treatment effect, implemented using interacted regression.

EW

Weighted ATE estimator based on easiest-to-estimate weighting (EW) scheme, implemented by running one-treatment-at-a-time regressions.

CW

Weighted ATE estimator using easiest-to-estimate common weighting (CW) scheme, implemented using weighted regression.

Usage

multe(r, treatment_name, cluster = NULL, tol = 1e-07, cw_uniform = FALSE)

Arguments

r

Fitted model, output of the lm function.

treatment_name

name of treatment variable

cluster

Factor variable that defines clusters. If NULL (or not supplied), the command computes heteroscedasticity-robust standard errors, rather than cluster-robust standard errors.

tol

Numerical tolerance for computing LM test statistic for testing variability of the propensity score.

cw_uniform

For the CW estimator, should the target weighting scheme give all comparisons equal weight (if FALSE), or should it draw from the marginal empirical treatment distribution (if TRUE)?

Value

Returns a list with the following components:

est_f

Data frame with alternative estimators and standard errors for the full sample

est_o

Data frame with alternative estimators and standard errors for the overlap sample

cb_f, cb_0

Data frame with differences between PL and alternative estimators, along with standard errors for the full, and for the overlap sample.

n_f, n_o

Sample sizes for the full, and for the overlap sample.

k_f, k_o

Number of controls for the full, and for the overlap sample.

t_f, t_o

LM and Wald statistic, degrees of freedom, and p-values for the full and for the overlap sample, for testing the hypothesis of no variation in the propensity scores.

pscore_sd_f, pscore_sd_o

Standard deviation of the estimated propensity score in the full and overlap samples.

Y, X, wgt

Vector of outcomes, treatments and weights in the overlap sample

Zm

Matrix of controls in the overlap sample

References

Paul Goldsmith-Pinkham, Peter Hull, and Michal Kolesár. Contamination bias in linear regressions. ArXiv:2106.05024, February 2024.

Examples

wbh <- fl[fl$race=="White" | fl$race=="Black" | fl$race=="Hispanic", ]
wbh <- droplevels(wbh)
r1 <- stats::lm(std_iq_24~race+factor(age_24)+female, weight=W2C0, data=wbh)
m1 <- multe(r1, treatment="race")