Data from: How should genes and taxa be sampled for phylogenomic analyses with missing data? An empirical study in iguanian lizards

Targeted sequence capture is becoming a widespread tool for generating large phylogenomic datasets to address difficult phylogenetic problems. However, this methodology often generates datasets in which increasing the number of taxa and loci increases amounts of missing data. Thus, a fundamental (but still unresolved) question is whether sampling should be designed to maximize sampling of taxa or genes, or to minimize the inclusion of missing data cells. Here, we explore this question for an ancient, rapid radiation of lizards, the pleurodont iguanians. Pleurodonts include many well-known clades (e.g. anoles, basilisks, iguanas, spiny lizards) but relationships among families have proven difficult to resolve strongly and consistently using traditional sequencing approaches. We generated up to 4,921 ultraconserved elements with sampling strategies including 16, 29, and 44 taxa, from 1,179 to ~2.4 million characters per matrix and ~30% to 60% total missing data. We then compared mean branch support for interfamilial relationships under these 15 different sampling strategies for both concatenated (maximum likelihood) and species tree (NJst) approaches (after showing that mean branch support appears to be related to accuracy). We found that both approaches had the highest support when including loci with up to 50% missing taxa (matrices with ~40–55% missing data overall). Thus, our results show that simply excluding all missing data may be highly problematic as the primary guiding principle for the inclusion or exclusion of taxa and genes. The optimal strategy was somewhat different for each approach, a pattern that has not been shown previously. For concatenated analyses, branch support was maximized when including many taxa (44) but fewer characters (1.1 million). For species-tree analyses, branch support was maximized with minimal taxon sampling (16) but many loci (4,789 of 4,921). We also show that the choice of these sampling strategies can be critically important for phylogenomic analyses, since some strategies lead to demonstrably incorrect inferences (using the same method) that have strong statistical support. Our preferred estimate provides strong support for most interfamilial relationships in this important but phylogenetically challenging group.