Top AIC model with no significant parameter estimates


I am working on a project in which my top model has normal looking standard errors, but all of the parameter estimates overlap zero, suggesting that none of the parameters are important. Why would this be? Is there a way to troubleshoot this?

If my model averaged parameter estimates all overlap zero, can I really say any of the parameters are statistically important?

I do not think I have multicollinearity, so I don't think that should be the issue. All of my other, slightly less supported models suggest that 2 parameters are important (their CIs of these 2 parameters do not overlap zero), except in my top model.

Thank you for any help you may be able to provide.


Fortran must die
You need to say what method you are using and for what type of data. It is hard to answer this question with the information provided. When you say you parameters overlap zero I assume you mean they have a confidence interval that includes zero, they are not signficant.

You say you don't think MC is an issue. Have you actually run a VIF or Tolerance test? I would think if you have a signficant model with no significant parameters this would be the issue (power might also be one).
Hello. You are correct in that I did not provide enough information. I apologize. I am running nest survival models in R using the RMark interface. The confidence intervals of my parameter estimates overlap zero, this is correct.

I do not think that multicollinearity is an issue because my variables are categorical, and I don't think they should be related at all. I do know that "perturb" in R can check for MC with categorical variables, but I haven't been able to make that run with my data.

Also, I do not know how to check for power with the models. I can look into that.


Fortran must die
I am not sure what nest survival models are. In regression (or I believe any general linear model including ANOVA) you can request VIF or tolerance and I assume the method you use would be similar. I would not assume because variables logically are unrelated they do not exibit multicolinearity - I have not seen it argued this is the case. Certainly categorical variables can exhibit multicolinearity.

There are tools such as G power that will test for power. They tend to be associated with common methods such as regression or ANOVA, but I assume you can find an equivalent tool for other methodologies. It is a good idea to test for power regardless.
Thanks for the suggestions. I can't find a power or MC test that works with the Rmark interface, so this is where I seem to be stuck.


Fortran must die
Power calculations that I have seen work with specific methods not specific software. So for example if you are using ANOVA or regression they will solve for that regardless of what interface you use. All you need is the basic information they ask for, such as effect size. You can run your methods in one software, do power calculations in the power software. For example G power.

Having said that I am not familiar with the method (nest survivial) you mention. Is nest survival a method, or are you using another method to determine if nests survive :)