Individual covariate effect constrained by age

[chris_oosthuizen]

Hello,

I am interested in the long term survival consequences of variation in body mass at birth, with birth mass used as an individual covariate in the analyses. I am using E-Surge 1.8.9.
Model selection proceeded by first selecting a good structural model, i.e., a model that does not include any covariate, as a reference model, followed by fitting ultra-structural models including covariate effects.

The most parsimonious reference model for survival was (in Gemaco MDL):
[a(1,2)+g(1).a(3:12)+g(2).a(3:5,6:12)]
where group is females and males respectively.

To evaluate covariate effects, I defined the above MDL as a shortcut called ‘base’
Covariate models were subsequently fitted in the following way: For example:

1) [base+a(1).xind(1)] where only age 1 survival is influenced by birth mass

2) [base + a(1:2).xind(1)] where age 1 and 2 survival are similarly influenced by birth mass.

Using this MDL, I get 5 beta values (intercept terms) for survival: a(1), a(2), g(1).a(3:12), g(2).a(3:5) and g(2).a(6:12) and one beta value for the covariate effect.

[I assume everything is correct up to this point, please correct if this is not the case]

My question the following: Is it valid to test for longer lasting effects of birth mass, e.g. from age 1 up to age 4, using this particular structural model, i.e. [base + a(1:4).xind(1)]?

For this model, I would then use the g(1).a(3:12) beta value in reconstituting beta values to survival probabilities for females ages 3 and 4 (estimates will be the same for age 3 and 4 and scaled with birth mass). For females ages older than 4 I will reconstitute the g(1).a(3:12) beta value, but not scale this single estimate by birth mass.

The same would apply for g(2) (=males)

Would this be correct?
Thanks!

[CHOQUET]

In the last case, I would have used a different intercept for age 3 and 4 than for age 5 and more.

[chris_oosthuizen]

Thanks Rémi, I thought that may be the case.
But my question then becomes: Is it possible to do this (use a different intercept for age 3 and 4 than for age 5 and more) without changing the structural ('base') model? (i.e. by using a different MDL for the covariate part of the model)
Ideally, for comparative purposes, I would like to use the same structural (age effect) model on which I test the duration of body mass effects, but I am not sure that this will be possible.

Suggestions are welcome
Chris