Hi,
I have heard that the coefficient obtained under a logistic-normal model
for binary data is usually larger than that obtained in a marginal model.
What I want to know is whether the coefficient is generally larger even
for between-cluster (between-subject) effects. Because it seems to me that
between-cluster effects have same interpretation whether in a
logistic-normal or in a marginal model, and if there is systematic
difference it will suggest that even these effects have different
interpretation depending on the model.
Secondly I recognize there is usually no way of testing the underlying
assumption of normality of the random coefficients in the logistic-normal
model. Nonetheless, I feel that even if we don't assume normality, there's
no reason for the point estimates to be biased. Is this a reasonable
proposition? What I worry though are the standard errors. Is bootstrapping
over individual (rather than cluster) recommended for obtaining better
estimates of standard errors in this case?
Thanks very much for any comments.
Tim
|