Modelling random effects for only part of the observations (in lme4)

Thierry ONKELINX wrote:

Dear Karl Ove,

(X|G) is equivalent to (1 + X|G). Or mathematically: b_0i +
b_1iX. But you need b_1iX.

The solution is to remove the random intercept (0 + X|G)

I would go for lmer(y ~ arm + (0 + b|gr2))

Great! Thanks. This works perfectly.

And I now understand what's going on in the formula too. The lme4 
syntax is rather elegant. :)

The fixed effect estimates are near identical to estimates based
on lm(y ~ arm), while the standard errors are different, which is 
exactly what would be expected. The reason they're not exactly 
identical is that one treatment group had only 2 subjects, so the 
design was not perfectly balanced. If I change it to a balanced 
design (equal number of subjects in each treatment group), the 
estimates are identical between lmer and lm (while of course the 
standard errors still differ).