confint(...,method="boot") for model with weights

[bbolker]

From Rodrigo Travitzki via r-sig-mixed-models ...

## run
## strange behaviour of confint()
#
# R version: 3.3.1
# lme4 version: 1.1-12

require(lme4)

dd=read.csv2(file("http://topicostropicais.net/bau/data.csv"))

m1<-lmer(NOTA~(1|ID_ESCOLA)+(1|ID_TURMA),data=dd,weights=PESO)

## pre conditions seems to be OK
plot(m1)
qqnorm(scale(resid(m1)),ylab="Residual quantiles",col="orange")
qqline(scale(resid(m1)),col="blue")

## reproducing error
VarCorr(m1) # Residual = 46.5197
# with "boot", sigma intervals doesn't match the estimated value 46.5197
confint(m1,method="boot",boot.type="norm")
confint(m1,method="boot",boot.type="basic")
confint(m1,method="boot",boot.type="perc")

## cause seems to be 'weigths'
m2<-lmer(NOTA~(1|ID_ESCOLA)+(1|ID_TURMA),data=dd)
VarCorr(m2)
confint(m2,method="boot")
confint(m2,method="profile") 

[bbolker]

adding to this: zero-weight observations are not taken into account when computing sigma() in general. This should be fixed or at least noted in documentation.

Exploration so far posted at https://github.com/lme4/lme4/blob/master/misc/issues/GH386.R ...

[mmaechler]

We definitely should behave as lm() here.