Estimate correlation in non-normally distributed (e.g. Poisson), unbalanced, repeated measures data

Question

I have a large behavioural data set, and I would like to measure pairwise correlations between several of my outcome variables, and binned categories of outcome variables. However, I have multiple 'subjects' (animal nests), with highly variable numbers of fixed observation periods between each (from 2-30+). What is the best method to test correlation, that properly models the distribution (typically Poisson or negative binomial) and the mixed design?

I have found the Rmcorr R package, but its solution to normal data is to bootstraping. Is there a rank-based method? Should I reshape my data and use a glmm with my outcome variables as predictor and predicted?

score 0 · Answer 1 · answered Mar 21 '20 at 23:46

0

I am not an expert with that, but I think the intraclass correlation coefficient is what are you looking for. Honestly, I cannot elaborate more on this, but you can start from Wikipedia to have an overview and moving towards more detailed explanations.

answered Mar 21 '20 at 23:46

s.dallapalma

131
4

Thank you, that is much appreciated, and closely related to my question. – Pierre Raynard Apr 08 '20 at 02:47

score 0 · Accepted Answer · answered Apr 08 '20 at 02:47

0

This is actually an unexpectedly challenging issue. Nakagawa and Schielzeth (2012) provides a method for some non-linear distributions with implimentation in R using the lmer4 package: http://dx.doi.org/10.1111/j.2041-210x.2012.00261.x

The method is to calculate the ratio of variance in a model with fixed effects (a fitted model), to the variance components of each random effect (additive or otherwise depending on distribution).

answered Apr 08 '20 at 02:47

Pierre Raynard

55
7

Here is some further comment: https://stats.stackexchange.com/questions/111150/calculating-r2-in-mixed-models-using-nakagawa-schielzeths-2013-r2glmm-me – Pierre Raynard May 14 '20 at 00:19

Estimate correlation in non-normally distributed (e.g. Poisson), unbalanced, repeated measures data

2 Answers2