Figure 1.

The Latent Dirichlet Allocation (LDA) estimates well the true number of groups (boxplots) and the ${\theta }_{\mathit{lk}}$ parameter values (scatterplots). Results from all 10 datasets in each simulation setting are displayed simultaneously, based on LDA with $\gamma$ set to 0.1. Top and bottom panels display results for three and 10 groups, respectively. Boxplots in panels (a) and (f) show the estimated number of groups (i.e., the number of groups deemed not to be superfluous), revealing that LDA can estimate well the true number of groups (K*) except for datasets with few locations (L), few species (S) but many groups (i.e., 100L 50S 10K*). Scatterplots (panels b–e and g–j) reveal that the ${\theta }_{\mathit{lk}}$ parameters can also be well estimated but there is considerable noise for datasets with few species but many groups (panels g and h). A 1:1 line and a linear regression line were added for reference (blue and red lines, respectively)