GOF and lack of data

[annaren]

Before I start I want to state that I have read thoroughly the Gentle Intro To MARK and although very very helpful after serious reading and thinking I am still stuck on the following problem.

In trying to use either UCARE or RELEASE to test the GOF of my data they reveal that for test 2 and for test 3 SM there is insufficent data and thus I obviously have insufficient power to use a full time dependent model.
So I have two thing I would like to clarrify:
1) Test 3.SR - only significant for 1 of the 4 groups so should I incorporate transience for all groups in the model, only the one or none at all?
2) Is this an acceptable was to proceed with the GOF: Make a simplier model, i.e no time effect in p (test 2 had insufficient data) and test the GOF using the median chat test. Adjust the model accordingly if c hat <3 and proceed with the analyses

Many thanks to anyone who can help my progress with my analyses confidently.
Anna

[cooch]

Before I start I want to state that I have read thoroughly the Gentle Intro To MARK and although very very helpful after serious reading and thinking I am still stuck on the following problem.

In trying to use either UCARE or RELEASE to test the GOF of my data they reveal that for test 2 and for test 3 SM there is insufficent data and thus I obviously have insufficient power to use a full time dependent model.
So I have two thing I would like to clarrify:
1) Test 3.SR - only significant for 1 of the 4 groups so should I incorporate transience for all groups in the model, only the one or none at all?
2) Is this an acceptable was to proceed with the GOF: Make a simplier model, i.e no time effect in p (test 2 had insufficient data) and test the GOF using the median chat test. Adjust the model accordingly if c hat <3 and proceed with the analyses

Many thanks to anyone who can help my progress with my analyses confidently.
Anna

As per section 5.9.1 (current edition) there is no cardinal rule saying your general model needs to be fully time-dependent. Your general model is simply the most general one you can fit to your data. If RELEASE is failing, then the chances are very high your data are too sparse for a fully time-dependent model. So, don't bother trying to fit one.

Use a different general model, and some alternative approach (like median c-hat) to estimate c-hat. Remember thought that the median c-hat approach (like others) assumes that the lack of fit is due to overdispersion - if in fact your lack of fit is structural (general model not -adequate for the data), then the c0hat estimated using the median c-hat approach is not particularly relevant. This is all discussed in the current version of chapter 5 (if you haven't downloaded/printed the chapter in <4 weeks from today, your version is out of date).

[annaren]

Thanks so much Evan for your reply adn for informing me of the new version of chapter 5. This is an exceedingly useful chapter.

Could you give me any advice on the first question in my message as I still cant find an answer to it:
1) Test 3.SR - only significant for 1 of the 4 groups so should I incorporate transience for all groups in the model, only the one or none at all?