Detecting patterns of correlational selection with sampling error: A simulation study

The adoption of a multivariate perspective of selection implies the existence of multivariate adaptive peaks and pervasive correlational selection that promotes co-adaptation between traits. However, to test for the ubiquity of correlational selection in nature, we must first have a sense of how well can we estimate multivariate nonlinear selection (i.e., the γ-matrix) in the face of sampling error. To explore the sampling properties of estimated γ-matrices, we simulated inidividual traits and fitness under a wide range of sample sizes, using different strengths of correlational selection and of stabilizing selection, combined with different number of traits under selection, different amounts of residual variance in fitness, and distinct patterns of selection. We then ran nonlinear regressions with these simulated datasets to simulate γ-matrices after adding random error to individual fitness. To test how well could we detect the imposed pattern of correlational selection at different sample sizes, we measured the similarity between simulated and imposed γ-matrices. We show that detection of the pattern of correlational selection is highly dependent on the total strength of selection on traits and on the amount of residual variance in fitness. Minimum sample size needs to be at least 500 to precisely estimate the pattern of correlational selection. Furthermore, a pattern of selection in which different sets of traits contribute to different functions is the easiest to diagnose, even when using a large number of traits (10 traits), but with sample sizes in the order of 1000 individuals. Consequently, we recommend working with sets of traits from distinct functional complexes and fitness proxies less prone to effects of environmental and demographic stochasticity to test for correlational selection with lower sample sizes.