Using clustered covariance matrix in predict.lm()

Question

I am analyzing a dataset in which data is clustered in several groups (towns in regions). The dataset looks like:

R> df <- data.frame(x = rnorm(10), 
             


        

Answer 1

I modified the above code slightly to be more consistent with the predict function--this way you are not expected to enter values for the outcome in the newdata dataframe

predict.rob <- function(x,clcov,newdata){
if(missing(newdata)){ newdata <- x$model }
tt <- terms(x)
Terms <- delete.response(tt)
m.mat <- model.matrix(Terms,data=newdata)
m.coef <- x$coef
fit <- as.vector(m.mat %*% x$coef)
se.fit <- sqrt(diag(m.mat%*%clcov%*%t(m.mat)))
return(list(fit=fit,se.fit=se.fit))}

Answer 2

The se.fit in predict is not calculated using the vcov matrix, but using the qr decomposition and the residual variance. This goes for the vcov() function as well: it takes the unscaled cov matrix from the summary.lm() together with the residual variance, and uses those ones. And the unscaled cov matrix is - again- calculated from the QR decomposition.

So I'm afraid the answer is "no, there is no other option than to write your own function". You can't really set the vcov matrix, as it is recalculated when needed. Yet, writing your own function is rather trivial.

predict.rob <- function(x,clcov,newdata){
    if(missing(newdata)){ newdata <- x$model }
    m.mat <- model.matrix(x$terms,data=newdata)
    m.coef <- x$coef
    fit <- as.vector(m.mat %*% x$coef)
    se.fit <- sqrt(diag(m.mat%*%clcov%*%t(m.mat)))
    return(list(fit=fit,se.fit=se.fit))
}

I didn't use the predict() function to avoid unnecessary calculations. It wouldn't shorten the code too much anyway.

---

On a side note, questions like this are better asked on stats.stackexchange.com