[R-sig-ME] Excluding random intercepts?

[Rolf Turner]

On 07/12/12 10:43, Kirsty E. B. Gurney wrote:
> Good afternoon;
>
> I have recently come across a *general question* about linear mixed models that I can't seem to find an answer for on my own.  I am hoping that one of the braintrust on this list might be willing / able to help.
>
> Specifically, I am analyzing a dataset that describes pond temperatures for a series (n = 25 ponds, 436 measurements).  Ponds were assigned to 1 of 4 treatments, and I want to find find out if temperature was affected by treatment.
>
> Pond-level inference, in general, is not of interest, so it seemed clear that POND should be included as a random effect.  However, I have good reason to believe that the change in temperature across the study was also variable by POND, and a model that includes a random slope term has good support, as determined by AIC.
>
> However, as it turns out, the covariance parameter estimate for the intercept term in the random slope model is pretty small (i.e., I think most ponds start out at a pretty similar temperature).
>
> I am just wondering if it makes good statistical sense to run a random slope model without the random intercept term?
>
> If so, does anyone know how one would one specify this in the model structure?
>
> Many thanks in advance.

This is definitely a case of the blind leading those who need glasses,  
but ....

Doing:

set.seed(42)
fake <- 
data.frame(y=rnorm(200),temp=runif(200,10,20),pond=factor(rep(letters[1:20],10)))
fit     <- lmer(y ~ temp + (0 + temp|pond),data=fake)

*seems* to give results that are compatible with the idea of fixed 
intercept and
random slope.

Almost surely someone younger and wiser will correct me if I am rong. ( 
:-) )

     cheers,

         Rolf Turner