70
rate was 0.2. This implies that the expected False Negative Rate (FNR, proportion
of false negatives, relative to the total number of unselected pairs) was 0.117. We
declared that 715 pairs had increasing means across the three stages. Of them 578
really did. The observed FDR and FNR were 0.192 and 0.113 respectively.
Selecting the pairs according to the utility function (3.34), with c/(fc +1) = 8, we
declared that 722 pairs had increasing means. Of them, 575 actually exhibited this
pattern. The expected values of the FDR and FNR were 0.206 and 0.117 respectively.
The observed values of these quantities were 0.204 and 0.116 respectively. The number
of pairs chosen by both methods was 688.
Our model is a particular case of a Dirichlet Process Mixture model. MacEach-
ern and Müller (1998) mention that if the data follows a distribution according to a
Dirichlet Process Mixture model, the predictive final distribution of a future obser-
vation matches with the posterior expected value of the density that generated the
data. Thus, the true distribution of the data can be estimated from a (simulated)
sample of future observations. This is, averages over the values of the expression
(3.17) evaluated at the values of the posterior simulated configurations ((/3*, <5*, φ, k).
We employed this method to simulate a sample of a future observations of β and
δ. In Figure 3.4, we compare the histograms of this simulated sample with the true
distribution.