gof with covariates?

[AndyT]

I have a question about gof with a data set with individual covariates.

I have a dataset with 10 capture occasions, two groups, and one individual covariate. Here's a sample line from the .inp file:

/*1*/ 1000100100 1 0 3.425;

My most general model Phi(g*t)p(g*t) has a deviance of 2400.485 When I did the Bootsrap GOF the mean deviance was 539.030. So obviously some evidence of lack of fit. (c hat = 4.45)

I also formatted the input file with the exact same data, but without the covariate, so with 10 capture occasions and two groups:

/*1*/ 1000100100 1 0;

The most general model Phi(g*t)p(g*t) has a deviance of 619.9475. I did a bootsrap GOF on this model and the mean deviance was 539.560. So apparently this model fits the data quite well (c hat = 1.149)

But these two models should be exactly the same, can anyone tell me which c hat I should use?

I also performed a median c hat test on Phi(g*t)p(g*t) with the data set with the covariates and got c hat = 1.334.

I am thoroughly confused; if anyone can help, it would be greatly appreciated!

[gwhite]

Andy:
If you check your 2 runs, you will see that the -2log likelihood for the saturated model changed drastically. With a set of data that includes individual covariates, the saturated model is computed based on each INDIVIDUAL. In contrast, when you take out the individual covariates, the saturated model is based on the observed ENCOUNTER HISTORIES. Hence, the reason for the difference you observed. Sounds to me like you do not have a GOF issue, because I would generally want to only evaluate GOF with the g*t model without individual covariates. The fit of the individual covariate model should only be better than one without.
Gary

[AndyT]

Thank you very much Gary; I hadn't thought of that.

So would you recommend using the c hat estimate from the g*t model without individual covariates as an adjusted c hat (1.149)? Or is it not worth adjusting c hat with a value that close to 1?

Thanks again,

Andy

[gwhite]

Andy:
I would probably just use the 1.149 value that you obtained from the g*t model, recognizing that "important" individual covariates probably improve the fit a bit more. The impact of c = 1.149 will be negligable, so it doesn't make much difference whether you use it or just c = 1.

Gary

[sixtystrat]

I am trying to run a bootstrap GOF with a global model (g*t) with known fate data. I have covariates in the data set but do not have any specified in the global model. When I try to run the boostrat GOF, I get a warning that no covariates are allowed and when I hit the run button to start the simulations, nothing happens, it just sits there. Any ideas? THANKS!
Joe

[cooch]

I am trying to run a bootstrap GOF with a global model (g*t) with known fate data. I have covariates in the data set but do not have any specified in the global model. When I try to run the boostrat GOF, I get a warning that no covariates are allowed and when I hit the run button to start the simulations, nothing happens, it just sits there. Any ideas? THANKS!
Joe

Since the data type is different, this should (probably) be in its own thread, but...

Check section 16.8 of Chapter 16.

[sixtystrat]

Thanks Evan. That's what I thought but I wanted to make sure the thinking had not changed on that. I am trying to convince and AE that a GOF cannot be done with known-fate data. He gives a couple of references where it was done by other authors, but I'll have to convince him that it was inapropriate.
Joe