Error using Loo with moment matching

[avehtari]

@Rich

If the number of influential data points is only a relatively small fraction of the dataset, can you report the fraction being low and leave them as they are? Or is that considered poor practice, even if transparent? I guess I’m thinking of a situation where there are only a small number of possible problematic datapoints, but kfold cross-validation takes 30 hours per model or longer, so it could get rather time consuming (depending on the number of models). Or would you recommend just getting on with kfold anyway, as there is no sensible alternative?

It depends. If someone’s life depends on this it’s better to make everything as carefully as possible. If the difference between models is very big anyway and no-one’s life depends on this, then you may argue that there is only a small probability that the error in elpd_loo estimate would be big.