---
title: "A comparison of `geex` and sandwich for robust covariance estimation"
author: "Bradley Saul"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{A comparison of geex and sandwich for robust covariance estimation}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

This examples uses the `vaccinesim` dataset from the `inferference` package to compare the estimated covariance matrix obtained from `geex` and `sandwich`.
An example $\psi$ function written in `R`. 

This function computes the score functions for a GLM.
```{r eefun, echo=TRUE}
eefun <- function(data, model){
  X <- model.matrix(model, data = data)
  Y <- model.response(model.frame(model, data = data))
  function(theta){
    lp  <- X %*% theta
    rho <- plogis(lp)

    score_eqns <- apply(X, 2, function(x) sum((Y - rho) * x))
    score_eqns
  }
}
```

Compare sandwich variance estimators to `sandwich` treating individuals as units:
```{r example1}
library(geex)
library(inferference)
mglm    <- glm(A ~ X1, data = vaccinesim, family = binomial)
estimates <- m_estimate(
  estFUN = eefun,
  data = vaccinesim,
  root_control = setup_root_control(start = c(-.35, 0)),
  outer_args = list(model = mglm))

# Compare point estimates
coef(estimates) # from GEEX
coef(mglm) # from the GLM function

# Compare variance estimates
vcov(estimates)
sandwich::sandwich(mglm)
```

Pretty darn good! Note that the `geex` method is much slower than `sandwich` (especially using `method = 'Richardson'` for `numDeriv`), but this is because `sandwich` uses the closed form of the score equations, while `geex` compute them numerically. However, `geex`'s real utility comes when you have more complicated estimating equations. Also, the analyst has the ability to code faster $\psi$ functions by optimizing their code or using `Rccp`, for example.