The bibliography is missing.

This is very relevant for me now as I’m working on splice variant differential expression with biological replicates with @schen5. We’ve set up pretty traditional posterior predictive checks and have also found that the variational inference works pretty well.

We show using a publicly available data set where nearly 10% of differentially abundant transcripts had fold change inflated by the presence of outliers.

What’s are you considering an outlier in this situation? The choice of 2.5% and 97.5% interval seems arbitrary, especially as you expect non-outliers in that range with hundreds of data points.

I’ve been heavily influenced by @andrewgelman not to think in terms of outliers at all, but in terms of building some kind of measurement error model. I’ve also been influenced by Andrew not to think in terms of FDR, which has this implicit quantization threshold, but in terms of posterior coverage. That is, if the method says the differential expression is in a given interval 50% of the time, how often is it in that interval? This requires setting the intervals to test (50% is nice for power), but not setting a threshold.

I fear in practice that the biologists don’t care about any of this and are just using all this for ranking hypotheses to take into the wet lab.

one outlier for one randomly selected sample, characterised by a 10-10 quantile distance

I don’t know what that means, either. What’s a quantile distance?

As the amount of draws from the posterior probability distribution needed to define extreme quantiles of a distribution grow exponentially with the number of samples and the false pos- itive rate,

I glanced through the paper and don’t know what is supposed to be growing exponentially as a function of what. The amount of data required to estimate a quantile q in (0, 1) is roughly 1/(min(q, 1 - q)) times the amount of data you need to estimate a median.