Figure 9

Frequency-of-frequency plots. Monte Carlo populations of 10⁸ transcripts each were sampled from the empirical Bayes estimate obtained with the fetal or adult likelihood and either the tissue-specific independent marginals (dots), averaged-marginals (squares), or uniform (circles) prior. The uniform prior gives an idiosyncratic L-shaped profile with an abrupt lower copy-number limit (open circles). This reflects the implicit exchangeability of the unobserved classes: all have the same low probability, but because the majority of forms are not observed, their cumulative probability in the estimator is large. This is one example of how the uniform prior (or any prior obtained by a small constant correction to the observed frequencies), although 'uninformative', is overly simplistic, and leads to artifacts.