3
$\begingroup$

As in this post, my understanding of the standard IPTW estimator is that it is $$\frac{1}{n} \sum_i \frac{Y_i T_i}{\hat{e}_i} - \frac{Y_i (1 -T_i)}{1 - \hat{e}_i}$$

This can straightforwardly be shown to unbiasedly estimate the Average Treatment Effect (ATE) when the propensity score model is correctly specified, as this paper does. However, in equation 2, section 2.3, that same paper specifies the IPTW estimator differently: $$\sum_i \frac{Y_i T_i}{\hat{e}_i} (\sum_i \frac{T_i}{\hat{e}_i})^{-1} - \sum_i \frac{Y_i (1 -T_i)}{1 -\hat{e}_i} (\sum_i \frac{1 - T_i}{1 - \hat{e}_i})^{-1}$$

I've run some simulations:

expit <- function(x) exp(x) / (1 + exp(x))

p <- 0.05
n <- 1e4
x <- cbind(rnorm(n), rexp(n), rbeta(n, 5, 5))
# t <- rbinom(n, 1, p)
t <- rbinom(n, 1, expit(x%*%c(-3,1,0)))
y <- 5*t + x%*%c(4,2,6) + rnorm(n)

prop <- glm(t ~ x[,1] + x[,2] + x[,3], family = binomial)
ps <- predict(prop, type = 'response')

iptw0 <- sum(y[t == 0] / (1 - ps[t == 0]))/n
iptw0
mean(y[t == 0])
weird0 <- sum(y[t == 0] / (1 - ps[t == 0]))/sum(1/(1 - ps[t == 0]))
weird0
iptw1 <- sum(y[t == 1] / ps[t == 1])/n
iptw1
mean(y[t == 1])
weird1 <- sum(y[t == 1] / ps[t == 1])/sum(1/ps[t == 1])
weird1

iptw1 - iptw0
weird1 - weird0

and both estimators appear to do perform well. Can anyone reconcile the two or provide references for the latter "weird" estimator?

$\endgroup$
4
$\begingroup$

The first estimator is the Horvitz-Thompson estimator. The second estimator is the Hájek Estimator --- basically you are normalizing/stabilizing the weights to gain efficiency (but it's not guaranteed you will gain efficiency). You can read more about them in survey sampling books.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.