using GOF diagnosic to improve model fit or data dredging?

[Eric Janney]

If Goodness of Fit testing indicates significant failure of the equal catchability assumption (i.e., test 2)in your data, some people (e.g. UCARE user's manual) advocate incorporating a "time since marking" effect into the global model. So, you build such a model (e.g., {Phi(sex*t) p(sex*t*TSM)} and it has much lower deviance (indicating better fit) and substantially lower AIC value than original global model: Phi(sex*t) p(sex*t). Then, upon further inspection of the individual contingency tables for each group and each occasion, you discover that the results of GOF test 2 are only significant in 2-3 occasions out of a twelve year study. There appears to be no easy biological explanation for why some years have really large Chi-Square values and others don't. For example, males only have large chi-square values in occasions 5,8, while females have large values in 5 and 10. Also, the directionality of the chi square statistic for these years is not consistent. Sometimes observed value is greater than the expected and other times it is less. So, is it mining/snooping to use the diagnostic information obtained from the individual chi-square tests to build a model that only incorporates the "time since marking" effect on the occasions that have significant chi-square values? This seems like a good way to reduce the # of parameters from the full TSM model but also seems pretty ad hoc. Any thoughts?