Page 1 of 1

Altering design matrix for 'real' covariate

PostPosted: Fri Sep 24, 2021 1:04 pm
by pat_14
Hi all,

I'm attempting to add a linear constraint to my recovery probability estimates in a dead recovery (Brownie) model using MARK GUI. I have 17 sampling intervals, and want survival to be time-invariant, but recovery probabilities to be constrained by a 'real' covariate I've supplied.

I've reconstructed the design matrix, but do not think I have done it correctly. I've had trouble uploading a picture of my design matrix so I'll try and describe it:

The design matrix has three columns (S_intercept, f_intercept, f_cov) and of course the Parm. For my S_intercept column, I've placed 1's for the first 16 rows pertaining to my survival estimates, and zeros for the rest of the rows. My f_intercept column contains the opposite configuration (zeros for the first 16 rows, and 1s for the next 17 rows). The f_cov column contains zeros for the first 16 rows pertaining to the survival estimates, and the 'real' data for the next 17 rows pertaining to the recovery probabilities.

Does this sound even remotely correct? Any suggestions would be greatly appreciated!

Re: Altering design matrix for 'real' covariate

PostPosted: Fri Sep 24, 2021 1:22 pm
by cooch
Sounds correct. At some point, you're likely going to want to generate a plot of 'recovery probability' versus the environmental covariate. This can be done in a fashion within MARK -- averaging over models. In fact, you do this by 'tricking' MARK into treating the environmental covariates as individual covariates (section 11.8.2 in Chapter 11).

Re: Altering design matrix for 'real' covariate

PostPosted: Fri Sep 24, 2021 2:58 pm
by pat_14
Thanks! Surprisingly, the model with the added covariate had lower weight than the other four pre-defined models. I suppose that tells me I need to find a better covariate!

Re: Altering design matrix for 'real' covariate

PostPosted: Fri Sep 24, 2021 6:08 pm
by cooch
pat_14 wrote:Thanks! Surprisingly, the model with the added covariate had lower weight than the other four pre-defined models. I suppose that tells me I need to find a better covariate!


Uh, no. You (the biologist) should think hard about the set of covariates that you (said biologist) think are relevant, or interesting, or important. You construct a prior a candidate model set - including models with and without the covariate(s) - and use some modality to evaluate support over the models for (or not) the covariates. You then draw your conclusions. Model building should not be an exercise in hunting around for a covariate that 'does better' (say, by reducing AIC relative to other models). Queue concerns about 'data dredging'. If you look through a large enough set of covariates, you will by random chance along likely find one that seems to do better -- but that could be entirely spurious, and beyond simple interpretation.