Missing EH occasions and GOF-testing

[Miina Kovanen]

Thank you for keeping up this forum – very useful to my learning and understanding the literature. Now I’m not sure if I’m working around the following GOF-problem or just dealing with data that is all too sparse, would you please comment on this.

I have recaptures data (CJS) with missing occasions (some sites were not sampled at each occasion). In this case, 532 individuals with all 5 occasions (years), but 107 with 3 occasions and 144 with only 2 occasions. The data and candidate models include individual covariates.

If I would use time-structured encounter histories, some individuals would have dots in their encounter histories, and the data would include both left and right censoring. However, from previous discussions I learned to remove the time structure from the EH:s and rather treat it as a covariate if needed, so that I can group the data according to the missing occasions, and set p=0 for missing occasions by using an individual constraint. I also learned that GOF-testing should be done in data without individual covariates. I hope I’m getting it right.

For examples from previous threads, see also viewtopic.php?t=23&highlight=missing+occasions
viewtopic.php?f=1&t=1541&p=4361&hilit=censor%2A+c+hat#p4361

My questions:
1) Is it reliable to use time dependent model phi(t)p(t) as a global model and run GOF testing with it in the exactly same data set but without any covariates – also without the covariate for constraining p=0 for the missing occasions? Here, missing occasions would be coded as zeros in EH. Does GOF testing then miss the possible problem of sparse data, or is it okay to think that a model constraining p=0 for some occasions and adding the covariate for the constraint should only improve the fit compared to the global model?
2) Or, if I only use the individuals with complete histories in GOF testing, how to estimate if it’s reliable also for the whole data set? In this case, quite many individuals do have missing occasions.

I tried testing GOF in RELEASE with the model phi(t)p(t). Many of the contingency tables (especially in TEST 3.sm and TEST 2) had zero cell values, and one the test (TEST.2c3) indicated violation in assumptions also by Fisher’s exact. Would for parametric bootstrapping method work better, or is the data just too sparse to analyze as survival histories?

Thank you for your time, Miina