Methodological Considerations When Using RAD-Seq to Discriminate Among Weakly Differentiated Fish Populations: Why Cross-Validation Procedures and Sample Sizes Matter

Molecular assignment methods, which use genetic information to ascertain population membership of individuals, hold great potential to benefit fisheries management. Such methods often employ a self-assignment component (i.e., leave-one-out or jackknifing-based cross-validation approach) in which samples are divided into training data and independent test data which are used to test the statistical power of assignment tests. However, this approach has been rarely applied to studies where large RADseq datasets (1000s of loci) are used to assign individuals to populations with weak genetic differentiation, despite being necessary to determine the number of individuals and loci required to generate appropriate genetic baselines for assignment testing. Here, we use spawners from three weakly differentiated western Lake Erie walleye subpopulations to assess how sample size in training data can affect the results of assignment tests. Our analyses showed that when the number of training individuals increased from 10 to 40 per stock, assignment performance increased to levels better than random. This finding is encouraging, as it supports the notion that, with sufficient training-sample sizes, RADseq could be used as a tool to discriminate among weakly differentiated populations and determine relative stock contributions to the fishery.