Variance in VC matrix?

[KGurney]

Good morning;

A question from a total newbie regarding known-fate models in MARK. I've searched the posts on negative variance values and can't seem to find the answer I'm looking for.

I've been running known-fate models, using 3 individual covariates and 6 time intervals. The 'best' model suggests that only one of the covariates (hatch date) is informative. When I output the beta estimates for this model, the standard errors are positive for all 7 parameters.

However, when I output the variance-covariance matrix for the beta values, only one of the variance values shows up as positive; the rest are negative.

Question is: why would the variance values in the VC matrix be negative? Does this mean there is a problem with the model? The survival estimates with negative variances in the VC matrix do not all have values approaching the boundary of 1... (as per the suggestion in the manual).

Any insight is appreciated.

Thanks,
kbg

[cooch]

Good morning;

A question from a total newbie regarding known-fate models in MARK. I've searched the posts on negative variance values and can't seem to find the answer I'm looking for.

I've been running known-fate models, using 3 individual covariates and 6 time intervals. The 'best' model suggests that only one of the covariates (hatch date) is informative. When I output the beta estimates for this model, the standard errors are positive for all 7 parameters.

However, when I output the variance-covariance matrix for the beta values, only one of the variance values shows up as positive; the rest are negative.

Question is: why would the variance values in the VC matrix be negative? Does this mean there is a problem with the model? The survival estimates with negative variances in the VC matrix do not all have values approaching the boundary of 1... (as per the suggestion in the manual).

Any insight is appreciated.

Thanks,
kbg

Quick thing to try (late for a meeting) - In the output from MARK, the variance-covariance values are below the diagonal, whereas
the standardized correlation values are above the diagonal. Variances along the diagonal. Make sure you're looking at the right part of the output (you probably are, but wanted to make sure).

[sbonner]

kbg, can you post the VC matrix?

[KGurney]

Thanks for the thought - I did a double check - the negative values are indeed on the diagonal.

The VC matrix is copied below...

PAR1 PAR2 PAR3 PAR4 PAR5 PAR6 PAR7
1.6491436864 0.9323630412 -0.3511916777 0.0000000000 0.0000000000 0.3062098877 0.0097859128
-0.0105293263 -0.0514635932 -0.0528123416 0.0000000000 0.0000000000 0.0097859128 0.1593767319
-0.0027047790 -0.0132200030 -0.0135664704 0.0000000000 0.0000000000 0.0119596381 0.0030721911
-0.0033055950 -0.0161565199 -0.0165799470 0.0000000000 0.0000000000 0.0111336025 0.0028600085
-0.0030772753 -0.0150406280 -0.0154348099 0.0000000000 0.0000000000 -0.0105293263 -0.0027047790
0.1167099077 0.0142242895 0.0145970771 0.0000000000 0.0000000000 -0.0514635932 -0.0132200030
0.0142242895 0.3888053457 0.0713453248 0.0000000000 0.0000000000 -0.0528123416 -0.0135664704

Here's a copy of the Beta estimates:

Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
1: 3.7220453 0.5533623 2.6374551 4.8066355
2: 2.6223734 0.3992202 1.8399019 3.4048449
3: 3.7853034 0.7209200 2.3723002 5.1983067
4: 2.9428745 0.5211717 1.9213780 3.9643710
5: 0.9795522 0.3416283 0.3099607 1.6491437
6: -0.2897804 0.6235426 -1.5119239 0.9323630
7: -0.8815343 0.2705829 -1.4118768 -0.3511917

Not sure if that helps...

kbg

[gwhite]

What you show is NOT the VC matrix that generated the beta estimates and SE. Not sure what you are doing, but it isn't the right thing to be doing!

Gary

[KGurney]

Thanks for confirming that something is wonky - it certainly has me scratching my head!!

To get these outputs, I worked through the goose mass example in Appendix B of the online MARK book.

To output the VC matrix, I highlighted the model, went to the Output --> Specific model output --> Variance-covariances matrices --> Beta Estimates --> to a .DBF file. 'When I opened the DBF file in Excel, that is what it looked like. Honest.

Please advise me if I have done this incorrectly.

The beta estimates and SE were obtained by selecting "View estimates of beta in Notepad window".

I'm totally new to this, so like I said, I'm just working through the examples. My goal is to produce some plots of Survival as a function of Hatch date (with SE included)...

[gwhite]

I can't duplicate anything like you show, so you better send me your DBF and FPT files and tell me which model is causing this problem. But before you do that, try re-running the model to make sure that these files are not somehow contaminated.

Gary

[KGurney]

Wow.

I think I just learned a *valuable* lesson. Files can get contaminated.

I re-ran the 'problem' model as you suggested, and voila, the variances in the VC matrix and the SE estimates of beta are now jiving. Awesome.

Next time, I'll be trying a re-run FIRST!!!

Thanks so much for your help.

kbg

[cooch]

To get these outputs, I worked through the goose mass example in Appendix B of the online MARK book.

Actually, there is no 'goose' mass example in the book. I suspect you mean the simulated individual covariate data (example 4) in the book.

Suggestion for the future - if you post a question related to something in 'the book', please be as precise as possible in terms of page, (sub)section number, even paragraph if it will help. While some of us have a passing familiarity with the contents of 'the book', it has gotten sufficiently lengthy that perfect recall of the example that you might be referring to (and where it is presented in the book) is getting more difficult.

[KGurney]

Corred. I was referring to Example 4, Appendix B: pp B-19 to B-20. I modified the first formula on B-19 to include only one covariate. Then I used SAS's PROC IML to do the matrix calculations, and it all works out - now that I have the right VC matrix!

Thanks for the tip, and many thanks to Dr. White for the helping me get this problem solved...

A great day to all,
kbg