beta se = 0?

[ked]

Hi all, I'm doing a pretty standard CJS analysis through RMark, looking at survival in birds through time, in different age classes, and with some weather covariates as well. In some of my model runs, most of the beta se's are 0. The corresponding real parameter estimates seem reasonable, but I'm wondering if this indicates an underlying problem? My apologies if this is covered somewhere in the manual, but, if so, I still haven't found it. Thanks in advance for any guidance!

[jlaake]

Typically that means you have created an over-specified model. In other words some columns of the resulting design matrix are linearly dependent on others. As an example, if I created a model such as ~sex+male where sex was a 2 factor variable of F/M and male was 1 if it was a male. In that case 2 columns of the DM would be identical if F was the intercept. That is a silly example you wouldn't try but if you look at the formula it will be something like that but maybe more subtle.

--jeff

[aswea]

Hi, I had this problem last year. There might be some helpful information under Carl Scharz's reply to my post where he explained a bit about how RMark creates the design matrix-- sometimes you get redundancy when you were not expecting it. The topic was: RMark different number of beta parameters for group variable and his reply was on Wed Dec 09, 2009 4:14 pm. Also, there is a section in the RMark appendix on not fully crossed designs (page C-58 of the 9th edition; heading=More complex examples) that could help if you are doing age models.

[ked]

Thanks for your help! It appears I still haven't fully figured out when it's appropriate to remove the intercept, as well as the difference between inserting "-1" in the model formula and "remove.intercept=TRUE". I need to keep reading/experimenting, but thanks again for pointing me in the right direction!

[jlaake]

remove.intercept=TRUE will always remove the intercept. ~-1+... is a short-hand but it will not work if you also have an additve factor variable in the formula. For example, ~-1+age:time will work but ~-1+sex +age:time will not because it will simply create a M and F columns. Presumably you are doing this because you have incompletely crossed factor variables and are fitting interactions as in the example.

Have you read the workshop notes posted on phidot? I believe I have a section on this issue. Maybe the above will help as well.

--jeff