Hi,
I'm seeing poor convergence of a model using the binomial family. In one
case, I have a binomial response for which I use the 'weights' argument.
In a second case, I expand the binomial response to Bernoulli trials and
refit. Theoretically, they should be the same model. However, that does
not appear to be the case because the residual scale for the binomial
case is unbelievably large. Could someone (Prof. Bates) please help
diagnose this problem?
And thank you so much for such wonderful software. I truly appreciate
the effort.
Thanks,
--sundar
tmp <- read.csv(url("http://sdorairaj.googlepages.com/tmp.csv"))
tmp[2:4] <- lapply(tmp[2:4], factor)
library(lme4)
fit <- lmer(1 - y ~ z + (1 | A) + (1 | A:B), tmp,
binomial("cloglog"), weights = as.numeric(w))
## see below for expand.bin
tmp2 <- expand.bin(tmp, "x", "w")
fit2 <- lmer(x ~ z + (1 | A) + (1 | A:B), tmp2, binomial("cloglog"))
fit3 <- glm(1 - y ~ z, binomial("cloglog"), tmp, w)
fit4 <- glm(x ~ z, binomial("cloglog"), tmp2)
cbind(sapply(list(fit, fit2), fixef),
sapply(list(fit3, fit4), coef))
# [,1] [,2] [,3] [,4]
#(Intercept) -2.8227711 -3.1247457 -2.8227711 -2.8227711
#z2 -0.2680992 -0.2694113 -0.2680992 -0.2680992
#z3 0.3109447 0.3233432 0.3109447 0.3109447
expand.bin <- function (data, x, n) {
char.x <- x
char.n <- n
x <- data[[x]]
n <- data[[n]]
i <- rep(seq(nrow(data)), n)
data <- data[i, , drop = FALSE]
expand <- function(z) c(rep(0, diff(z)), rep(1, z[1]))
x <- apply(cbind(x, n), 1, expand)
data[[char.x]] <- if(is.matrix(x)) c(x) else unlist(x)
row.names(data) <- seq(nrow(data))
data
}
> sessionInfo()
R version 2.5.0 (2007-04-23)
i386-pc-mingw32
locale:
LC_COLLATE=English_United States.1252;LC_CTYPE=English_United
States.1252;LC_MONETARY=English_United
States.1252;LC_NUMERIC=C;LC_TIME=English_United States.1252
attached base packages:
[1] "stats" "graphics" "grDevices" "utils" "datasets" "methods"
[7] "base"
other attached packages:
lme4 Matrix lattice
"0.99875-0" "0.99875-1" "0.15-6"