Missing covariate data -- How does Presence deal?

[mtingley]

Another question:

Statistically, how does Presence deal with missing survey-specific covariate data for sampling occasions when you have occupancy data?

Presence gives you a warning that there are missing covariate data when you load in the data, but I'm wondering how this affects running models.

My assumption would be that if you run a model using a covariate with missing data, then those sampling occasions with no covariate would be non-informative in deriving parameter values. However, I have some old notes from Darryl that instruct that in such a situation you should just take out the occupancy data if you don't have complete covariate data for it. In my data set I have one covariate where I only have estimates of it for about 70% of sampling occasions and I really don't want to eliminate 30% of my occupancy data just so I can include it in my models.

Thanks for your thoughts.

[darryl]

I always get worried when people start quoting my own stuff back to me, but in this case that advice still holds. If you have a missing survey-specific covariate value in your data file, then when you try fitting a model with that covariate PRESENCE will ignore the corresponding detection data. In fact, it should always ignore that detection data even if that covariate is not in the model you're fitting, otherwise your data is changing between models so AICs', likelihood etc are not comparable for models with/without that covariate.

This problem of how to deal with missing covariate values occurs in any regression-type analysis, not just occupancy modelling. I'm not sure what the latest and greatest approaches are generally, but if you're willing to define a model for your covariate values as well (eg normally distributed with some mean and variance) then I think I know how you might proceed in either a maximum likelihood (not so sure about implementation though) or Bayesian (dead easy in something like WinBUGS) approaches.

Cheers
Darryl

[jhines]

For single-season data, missing survey covariates cause the occupancy data for that site/survey to be treated as missing for the estimation of p. I haven't thought about missing site covariates, but the entire history should be discarded in those cases. At the moment, PRESENCE will give nonsensical results if this happens.

For multi-season data, missing survey covariates cause the portion of the history for a particular season to be treated as missing if survey covariates are missing. This applies only to models where p is a function of the missing covariate.

Darryl's comment about data changing between models is a good one, but PRESENCE doesn't prevent you from running models with a missing covariate and without the covariate in the same results/AIC table.

Jim