GOF:is the most general model always the most parameterized?

[simone77]

I have often read that Goodness-of-fit should always be assessed on the most general model (e.g. Ch. 4 & Ch. 5 from MARK manual). If a most general model fits data adequately, then also a nested model should fit data well, so, for instance, if model {Phi(t) p(t)} fits well, also {Phi(.) p(.)} does.
On the other hand, I have read here* that even they analysed data by using a multi-event approach on three underlying states (death or transient, alive male, alive female) they tested the GOF of the simple CJS model, i.e. a model not accounting for sex-states but only for alive or death. Authors comment on this: “On the basis of GOF test results, we calculated a variance inflation factor to be used in the models. This is conservative as the goodness-of-fit test is for a more general model with stronger assumptions (e.g. no sex or age differences)”.
I must be missing the point because on one side I find this makes sense but on the other CJS model is clearly less parameterized with respect to the models they are running that do account for sex.
It seems the most general model is not necessarily the most parameterized. I would appreciate very much if someone might shed some light on this.

*Genovart, M., Pradel, R., & Oro, D. (2012). Exploiting uncertain ecological fieldwork data with multi-event capture-recapture modelling: an example with bird sex assignment. Journal of animal ecology, 81, 970–977

[simone77]

Hi all,

It seems I am not too lucky when posting questions in this section. However, I do believe this is a very important section of this forum because it allows to post topics of general interest in the capture-recapture framework.

As it is often the case, a sort of trade-off between parsimony and reality seems to work when you choose a global model to be GOF tested: on one hand a very realistic but too much parameterized general model is probable to be uninformative as most of tests do not run appropriately, on the other the goodness of fit of a very simple model might result bad because you are not including important variables (for instance time) that do have an effect. Depending on the choice you have done you can get very different results and the rest of your analysis can follow quite different paths.

For this reason, in this context, I very would like if someone might give an opinion on the question I asked in the previous post (or suggest some specific reference).