beta SE's

[bakpak]

I'm using the MARK nest survival module and have a question about the SEs for some of the beta estimates. My real estimates look just fine, and some of my beta estimates for the covariates look good, but others are either giving me: a. SEs of 0.0000000 which seem highly unlikely, or b. really LARGE SEs (e.g., estimate of 0.025 and SE = 4,236). I don't know if this could possibly be a small sample size issue. My smallest sample sizes are for Year A: 37 and Year B: 20. Generally the problem occurs when the covariates are 1s and 0s (treatment vs. control, or comparing between 2 years). Sometimes even the Bo estimate has SE of 0.0000000. I believe my input file looks good and my models are correct, but perhaps I wouldn't be posting here if they were

Thanks for your help!

[bacollier]

bakpak,
Saw no one had responded, so here are first thoughts based on what you wrote. Do you have separation in your data (e.g., all your successful nests were in either treatment control in one year, or you only had failures in your treatment/control in 1 year)? How many parameters are you trying to estimate for the models that don't work, and are you using interaction terms like year*treatment? What is the structure of the working versus non-working models?

Bret

[bakpak]

Hi Bret,

Thanks for the response. I don't have separation in my data based on trmt/control or anything else. My trmt/control didn't show any evidence of support. I'm not using any interaction terms. Here's an example of my beta values for a model. Parameters 2 & 3 are years A & B respectively.


LOGIT Link Function Parameters of {Bo + year + time}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1: 2.9069376 0.0000000 2.9069376 2.9069376
2: 0.2428604 0.0000000 0.2428604 0.2428604
3: 2.5633157 0.0000000 2.5633157 2.5633157
4: -0.0221675 0.0075504 -0.0369662 -0.0073687

In my input file I have the 2 years separated out as the only two groups. Then in the design matrix I specifically add two columns and have 1s & 0s in the columns to designate which year is represented. So perhaps I'm doing this wrong and I DO have extreme separation because of how I'm handling year.

Pretending there are only 8 nests I found, the design matrix for the Bo + year + time model would look like this (minus the (year)):

1 1 0 1 (Year 1)
1 1 0 2
1 1 0 3
1 1 0 4
1 0 1 5 (Year 2)
1 0 1 6
1 0 1 7
1 0 1 8

Perhaps I'm not doing time correctly and it should restart at 1-4 for the 2nd year?

Thanks very much!
Becky

[jlaake]

Becky-

Your model is over-parameterized. You should only have an intercept and one additional column for the year factor because it has 2 levels. Delete either column 2 or 3 in your design matrix. I haven't read your post in detail but this is an odd model with both a continuous time trend and year factor. At the least wouldn't you want

1 0 1
1 0 2
1 0 3
1 0 4
1 1 1
1 1 2
1 1 3
1 1 4

such that the time within year effect is the same for each year? The above model has a constant slope within year with a different intercept for each year. Yours doubled the slope in year 2. Is that what you wanted?

-jeff

[bakpak]

Thanks Jeff. The extra year column hadn't really made sense to me, but someone more experienced than myself in Mark told me to do it that way. The time makes much more sense too.

I'll see if that fixes the problems.

Cheers!
Becky

[bakpak]

For anyone interested, the model likelihood remained the same (same AICs), but the beta values were correct this time when I made Jeff's suggested changes.

Thanks again!
Becky