2
$\begingroup$

I am running a intercept model with intercepts varying across people in R. My independent variables are all numeric variables.

My question is a general one: I saw already that it can happen that my random intercept turns insignificant when I add explanatory variables.

Does anybody know how to interpret this?

$\endgroup$
  • $\begingroup$ How many observations per engineer? Also, how many explanatory variables? Do they vary at the "engineer" level? $\endgroup$ – probabilityislogic Feb 25 '15 at 13:03
1
$\begingroup$

in the meantime I probably found the answer to my question: I didn't center the variables and so the intercept changed quite a bit. After centering the variables the effect is gone and the intercept doesn't change too much anymore.

$\endgroup$
1
$\begingroup$

It should indicate simply that the mean proportion of successes does not vary much across the 15 subjects in your study. The fact that the variance of the random intercept drop when you add your continuous predictors (the independent variables) might suggest for example that these are not fully balanced across subjects. If that is the case, the variability that was attributed to the random-effects (in the model with only the random intercept) is now explained by the fixed-effects (in the model with all the explanatory variables).

In principle having only 15 random-effects levels should not be a problem, even for maximum-likelihood procedures (e.g. package lme4). However, if the variance of your random-effects is exactly zero, it is possible that your model is somehow overfitted, or that in your case 15 levels are too few (these nice simulations here illustrates the problem). If that is the case, check on this page, for example they suggest to use some informative prior on the random-effects variance.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.