Thanks so much for your help. I’ve rescaled the response variable to be bounded by 0 and 1 as you suggested. This worked well for a smaller dataset, but I am now struggling to get this to converge with a much larger dataset.

The similarity scores were originally pairwise Euclidean distances (which I’ve now converted to similarity scores bounded by 0 and 1). Because I’m only analysing a subset of the distance matrix, the lowest score in the dataset is 0.008 and the highest is 0.79 (but values of 0 and 1 are now theoretically possible). I have tried modelling these similarity scores using a beta distribution, but the posterior predictive checks don’t look particularly good. The distribution for the observed values is narrower/possibly more skewed:

Is this likely to because the beta distribution is not appropriate, or something to do with the way my model is parameterised?

Further details about the dataset/model can be found here https://discourse.mc-stan.org/t/multi-membership-random-effect-has-low-ess/32401

huge apologies for cross-posting…I realised afterwards I maybe should have replied here rather than making a new topic but couldn’t find a way to remove my other post.