Model Averaging with Covariates

[TGrant]

First of all, let me make sure I understand correctly. If I run covariate.predictions on a marklist with models that have covariates, it will compute the model averaged estimates at the mean of the covariate values (unless I specify covariate values), right?

This is what I did, but I am having trouble because I have an unusual dummy (design) covariate structure that produces a lot of singular betas. (It produces the same real estimates as the model in PRESENCE I am trying to duplicate however.) I tried to run model-averaging with drop=FALSE

Code: Select all

Model.ests.4.all=covariate.predictions(LIBL.models4, data=cov1.df, drop = FALSE)

and it seems to work for the estimates and SE's but the CI's have 0 width for parameter 2-5:

Parm Index estimate SE LCL UCL
1 0.730372229 0.073582331 0.565658267 0.849267234
2 1.08E-05 4.24E-05 1.08E-05 1.08E-05
3 0.003440508 0.006970964 0.003440508 0.003440508
4 0.626230224 0.339959521 0.626230215 0.626230225
5 0.358793485 0.183570311 0.358793487 0.358793485
6 0.708071 0.046150174 0.610267243 0.789785307
7 0.714957549 0.045729741 0.617694767 0.795661347

Parameters 2-3 are the at the boundary, but I've looked at the models with the most weight for parameters 4-5 and they have good CIs. So I'm not sure why the model averaging isn't working.

Is there a way to get this to work with all my singular betas?

It may be possible to construct a covariate structure with fewer singular betas, but there are always going to be some.

[jlaake]

You are correct that it is computing the predictions at the average covariate value.

Not sure why you are getting singular betas due to your design. Can you fix parameters to avoid the problem?

--jeff

[TGrant]

There are no real parameters I can fix. If I could fix betas, there are some structural zeros among the betas I could fix to zero. Or maybe I could set those betas as separate real ests and fix those. I'm not sure, I would have to go back and think about how it's set up again. But besides those there are still some that are singular. I think I could recode the dummy variables to have fewer betas - there are 4 betas per epsilon right now - and I think I would still get the same real ests.

Am I not going to be able to model average unless I get rid of singular betas?

Tyler