Problems with Parameter estimates from a Huggins Model

[ssaps33]

I ran a Huggin's closed model on data that consists of 20 encounters and two groups: positive for disease and negative for disease. I have a total of 106 individuals that were marked over this period. I just wanted to ask the simple question of whether the detection probabilities differ between the two disease groups.

These are the top models I ran:

Model Model Description AICc ΔAICc Model Weight Parameters Deviance
Mb p(.), c(.) 1380.933 0 0.99995 4 1239.442
Mt p(t)=c(t) 1400.907 19.9743 0.00005 40 1185.858
M0 p(.)=c(.) 1424.331 43.3981 0 2 1286.854

I want to estimate the parameters of p and c for each disease group and came up with these:

Parameter Estimate Standard Error Lower Upper
p for Pos 0.0702935 0.0226816 0.036883 0.129889
p for Neg 0.0000002 9.38E-05 1.75E-224 1
c for Pos 0.0799395 0.0105325 0.061581 0.10317
c for Neg 0.1674681 0.0182633 0.134654 0.206372

I have rerun this model because p for the negative group doesn't seem right. The lower CI and standard error seem very low. I followed the steps in the manual and am not sure what went wrong. I've tried to look through the forum to see if anyone has had this problem. Can anyone suggest something I might have not done properly?

Thank you very much for your help.

[cooch]

I want to estimate the parameters of p and c for each disease group and came up with these:

Parameter Estimate Standard Error Lower Upper
p for Pos 0.0702935 0.0226816 0.036883 0.129889
p for Neg 0.0000002 9.38E-05 1.75E-224 1
c for Pos 0.0799395 0.0105325 0.061581 0.10317
c for Neg 0.1674681 0.0182633 0.134654 0.206372

What are 'these', and how did you 'come up' with them? Are they model-averaged interval specific values? Are they the estimated probabilities of being encountered at least once over the entire sampling period?

Further, assuming you built your models correctly, then regardless of what these estimates represent, they're pretty low -- the rule of thumb is that for encounter probabilities <0.15, there will be very strong limits on your inference.

I have rerun this model because p for the negative group doesn't seem right. The lower CI and standard error seem very low. I followed the steps in the manual and am not sure what went wrong. I've tried to look through the forum to see if anyone has had this problem. Can anyone suggest something I might have not done properly?

Your concern with the CI for p for negative individuals is somewhat misplaced, since your estimate for the parameter is effectively zero (0.0000002). Meaning (if correct), that you have almost no chance of initial encounter with a negative individual. How big was your sample (for each 'treatment' group), and what was M(t+1) for each group? If the estimate really is 0.0000002 (which is basically zero), then it wouldn't surprise me that your CI is 'weird' (short of doing the calculations by hand - which requires M(t+1), hard to say if you have the correct UCI and LCI).

[ssaps33]

Thank you for your reply. I have been in the field so was unable to reply sooner.

The estimates I got were from "viewing estimates of real parameters" of my best model (p(.),c(.)) which had a model weight of 0.99995. These were initial capture and recapture over the entire sampling period of 20 days.

The sample size for the positive group was 56 and the negative group was 50 with a total of 106 marked individuals.

I also noticed my dervived estimates seemed off: N-hat for positive was 72.989 and N-hat for negative was 10173123.

While conducting this experiment and through personal observation, recapture of marked individuals was high as well as initial capture. So it doesn't make sense why the estimate for initial capture is so low.

I am new to MARK and am still learning. If you need more information on my data please let me know. I will try to rebuild the models to make sure I have built them correctly.

Thanks again!

[cooch]

Thank you for your reply. I have been in the field so was unable to reply sooner.

The estimates I got were from "viewing estimates of real parameters" of my best model (p(.),c(.)) which had a model weight of 0.99995. These were initial capture and recapture over the entire sampling period of 20 days.

If your best model is model M(0) (i.e., p(.)c(.)), then this is often diagnostic of real problems with your data -- assuming of course that you've set everything up correctly.

The sample size for the positive group was 56 and the negative group was 50 with a total of 106 marked individuals.

I also noticed my dervived estimates seemed off: N-hat for positive was 72.989 and N-hat for negative was 10173123.

Given that p for the negative group is estimated at near 0, this isn't surprising. Remember, the canonical estimator for N is M(t+1)/p. Given the denominator in the expression is vanishingly small, this will massively bias any estimate of N.

While conducting this experiment and through personal observation, recapture of marked individuals was high as well as initial capture. So it doesn't make sense why the estimate for initial capture is so low.

Which would strongly suggest that you either didn't set up your .inp file correctly, or you had some problems building the models, or both.

Try posting the .inp file for the negative individuals.