Easier way to get the quasi-demeaned variables

[tappek]

Hi there,
I was wondering whether you are aware of the functionality in the plm package to get the quasi-demeaned variables without doing the calcuclation yourself (as in here

qdmean/R/qdmean.R

Lines 105 to 112 in 10f15b6

	data <- data.frame(model$model, index(model), summary(model)$ercomp$theta)
	names(data) <- c(names(model$model), names(index(model)), "theta")
	# average over the predictor, given the grouping variable (both user supplied)
	data$xmean <- as.numeric(ave(data[[predictor]], data[[group]],
	FUN = function(x) mean(x, na.rm = T)))
	x.full <- as.numeric(data[[predictor]])
	# generate x.qdmean to return
	x.qdmean <- x.full - data$theta*data$xmean

)?

Just use model.matrix() on a plm model and you get the model matrix (quasi-demeaned variables of the RHS) and pmodel.repsonse for the quasi-demeaned response.

Buidling upon your example in the help page:

library(plm)
model.plm <- plm(ln_wage ~ wks_work + union, index = c("idcode", "year"),
         data = nlswork,
         model = "random", random.method = "swar", effect = "individual")

wks_work.quasidemeaned <- model.matrix(model.plm)[ , "wks_work"]
resp.demeaned  <- pmodel.response(model.plm)

Works likewise for demeaned variables of FE models.

One caveat though: for unbalanced two-way RE models, model.response and pmodel.repsonse output the original variables due to the specific estimation technique used for these models.

[andyphilips]

Thanks @tappek ! I'll forward this over to my coauthor who wrote the R portion of this.