# Corrections for logistic regression (multiple comparisons?)

I'll cut to the chase. Here's what I'm doing:

• looking at the influence of water temperature and salinity on the presence of water-borne parasites for a certain estuarine bird population
• birds are checked for parasites 3 times a year for 5 years (Spring, Summer, Fall)
• this is mark-recapture, and some birds are sampled multiple times across years and seasons
• logistic regression is for water temp and salinity predicting parasites

Since some birds are included multiple times in my dataset (some twice a year, and many in multiple years), would it be wise to correct for multiple comparisons (Boneffori?) or am I thinking about this the wrong way?

Thanks!

• You probably need to investigate some form of multi-level model (also known as mixed effects model). The Bonferroni correction is not going to help you here as far as I can see. – mdewey Jun 27 '20 at 16:29
• As mdeway says, a mixed effects model should work. You can add "bird" as a random effect. – rw2 Jun 29 '20 at 11:53
• This looks like a repeated measures design. That means incorporating the repeated measures into the statistics. – Michelle Jul 1 '20 at 10:57
• Thank you all! Mixed effects sounds like the right way to go. – Poquito Jul 10 '20 at 12:14

As other comments have pointed out Bonferroni correction is not the right way to control for dependent grouping factors. You would use such a correction if you ran multiple models and want to account for multiple testing.

Try using a mixed effects logistic regression (= generalized linear mixed model) from the lme4 package. It will allow you to account for the correlations between individual birds and their baseline parasite risk:

lme4::glmer(parasites ~ water_temp + salinity + (1|bird)


where bird gives a unique identifier for each bird and his observations throughout the seasons. A tutorial for such analysis can be found here.

An alternative would be to do time-to-event analysis, which wouldn't only evaluate whether birds have parasites but also the time they remain free from parasites. For this you should look into Cox proportional hazards models with time-varying coefficients. You would also have to format your data differently.

• Thanks so much for this succinct reply! I'll start with mixed effects logistic. – Poquito Jul 10 '20 at 12:15