Heteroskedastic consistent covariance matrix with univariate regression [R] -

how can calculate heteroskedastic consistent covariance matrix (hccm) univariate regression (i.e., ols regression 1 regressor , no intercept)?

back story: have fixed effect panel regression, 1 regressor. can either least squares dummy variable (lsdv) approach

lm.ser.yr <- lm(realwage ~ placebo.ser + factor(incyear) + 0, data = cps) 

or can demean lhs @ year level , regress

cps <- ddply(cps, .(incyear), transform, realwage.dm.yr = realwage - mean(realwage)) lm.ser.yr <- lm(realwage.dm.yr ~ placebo.ser + 0, data = cps) 

but there problems both.

• with lsdv approach , enough dummies lm object gets huge (1 gb each). because of large qr matrix have keep pass vcovhc() calculate hccm, stops on "can't allocate vector of size" errors (or lot of paging).

• with demeaned within estimator works great, can't calculate hccm because neither vcovhc() or hccm() have handle no intercept univariate lm object (i.e., spits na, best can tell because without intercept , dummy variables, residuals farther 0 mean).

is there solution short of aggressively managing memory and/or moving cloud?

vcovhc() plm package seems handle lm() models without intercepts fine:

library(plm) df <- data.frame('a'=rnorm(1000), 'b'=rnorm(1000)) mod <- lm(a ~ b -1, df) vcovhc(mod) 

edit: here's code calculate them manually.

library(matrix) e2 = mod$residuals ^ 2 x = model.matrix(mod) n = nrow(x) bread = solve(crossprod(x)) <- matrix(data=0, ncol=n, nrow=n, sparse=true) diag(i) <- e2 salami <- t(x) %*% %*% x  v = bread %*% salami %*% bread