ohines/plmed

R-package for partially linear mediation

plmed

R-package for partially linear mediation

To install:

devtools::install_github("ohines/plmed")
library(plmed)

To see documentation

?plmed
?G_estimation
?TTS

To uninstall:

remove.packages(plmed)

plmed's Issues

Does IF.glm() work for any family and link (especially non-canonical link)?

IF.glm <- function(object,x=object$x){ 
  NROW(object$y)*(x*object$prior.weights*(object$y-object$fitted.values))%*%chol2inv(object$qr$qr)
}

The chol2inv(object$qr$qr) seems magic, can you elaborate, please? Thank you!

Under the canonical link, the estimating equation is x{y-mu(x^t b)}, hence the influence function (by the argument of Z-estimation) is [E{mu'(x^t b) x x^t}]^-1 {y-mu(x^t b)}x. My naive implementation is this:

glm_IF <- function(glm_obj) {
  # assume glm_obj fitted with x = True
  (glm_obj$x * glm_obj$prior.weights * (glm_obj$y - glm_obj$fitted.values)) %*%
    solve(crossprod(glm_obj$family$mu.eta(glm_obj$linear.predictors) * glm_obj$x, glm_obj$x) / NROW(glm_obj$y))
}

By the way, I just happened to google with "glm influence function" and your repo pops up. I'm also dealing with some two-step estimation which requires the first-step influence function. I thought it would be convenient to use the glm object.