1
$\begingroup$

I'm using a tobit model for a left censored dataset in R, and including both continuous and categorical predictor variables. I've converted the factors to indicator coding for each level. I was initially getting errors when trying to run a summary of the model output, which I discovered was due to the fact that some levels of the categorical variables have a coefficient of 0, even though there are observations for those levels. I'm setting a reference level for the category using the level with the most observations.

I can't seem to find anything about why this would be happening. This isn't survival regression where 0 coefficient is a soft model selection, is it?

If it is a form of model selection, how do I handle that given the factor itself? I know in ridge regression I could use a form of grouping to test the variable overall. Could I use a similar approach here?

$\endgroup$
1
$\begingroup$

The betas at zero were appearing only when using the {AER} Tobit function. Using the {VGAM} package with the vglm function returned values for the previously missing coefficients. I'm assuming this is a problem with some levels of the categorical variables only having a few observations. Combining those levels with others when creating the indicator variables fixed all problems with the models.

$\endgroup$

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.