Why sampling scheme matters: the effect of sampling scheme on landscape genetic results

Abstract

There has been a recent trend in genetic studies of wild populations where researchers have changed their sampling schemes from sampling pre-defined populations to sampling individuals uniformly across landscapes. This reflects the fact that many species under study are continuously distributed rather than clumped into obvious ''populations''. Once individual samples are collected, many landscape genetic studies use clustering algorithms and multilocus genetic data to group samples into subpopulations. After clusters are derived, landscape features that may be acting as barriers are examined and described. In theory, if populations were evenly sampled, this course of action should reliably identify population structure. However, genetic gradients and irregularly collected samples may impact the composition and location of clusters. We built genetic models where individual genotypes were either randomly distributed across a landscape or contained gradients created by neighbor mating for multiple generations. We investigated the influence of six different sampling protocols on population clustering using program STRUCTURE, the most commonly used model-based clustering method for multilocus genotype data. For models where individuals (and their alleles) were randomly distributed across a landscape, STRUCTURE correctly predicted that only one population was being sampled. However, when gradients created by neighbor mating existed, STRUCTURE detected multiple, but different numbers of clusters, depending on sampling protocols. We recommend testing for fine scale autocorrelation patterns prior to sample clustering, as the scale of the autocorrelation appears to influence the results. Further, we recommend that researchers pay attention to the impacts that sampling may have on subsequent population and landscape genetic results.

Genetic differentiation of eastern wolves in Algonquin Park despite bridging gene flow between coyotes and grey wolves

“…As noted by the authors of the Bayesian algorithms used here (Pritchard et al, 2000;Guillot et al, 2005) and recently shown by simulation (Frantz et al, 2009;Schwartz and McKelvey, 2009), deviations from random mating not caused by barriers to gene flow (that is, spatial autocorrelation and isolation by distance) and the sampling scheme can have impacts on the detection and interpretation of genetic structure. These include potential overestimation of genetic structure for data sets characterized by continuously distributed individuals and spatially autocorrelated allele frequencies (Frantz et al, 2009;Schwartz and McKelvey, 2009).…”

“…Recently, Jombart et al (2008) suggested that Bayesian clustering may be inappropriate when populations are structured across a cline, and they developed a 'spatially explicit multivariate method' or sPCA that accounted for spatial structure and genetic variability and could identify different genetic structures, including clines, without having to meet the assumptions of Bayesian approaches. The sPCA complements the Bayesian approach implemented in Geneland by identifying more cryptic spatial patterns of genetic structuring across the landscape, and accounts for spatial autocorrelation issues associated with neighbour-mating and sample distribution (Schwartz and McKelvey, 2009). …”

“…These include potential overestimation of genetic structure for data sets characterized by continuously distributed individuals and spatially autocorrelated allele frequencies (Frantz et al, 2009;Schwartz and McKelvey, 2009). sPCA explicitly incorporates spatial autocorrelation as well as variance into the clustering procedure.…”

Genetic structure in Red Junglefowl (Gallus gallus) populations: Strong spatial patterns in the wild ancestors of domestic chickens in a core distribution range

“…When a regular increase or decrease in correlation between genetic variability and geographic distance occurs (i.e., IBD), it may disrupt the resampling process and generate false‐positives in Bayesian randomization (Meirmans, 2012). With IBD, MCMC models likely fail to explain spatially explicit genetic variation (Frantz, Cellina, Krier, Schley, & Burke, 2009; Schwartz & McKelvey, 2009). There are a few biological and technical procedures to accommodate the presence of IBD, including stratified sampling (as employed in this study) (Storfer et al., 2007) and correlogram analyses to determine whether or not IBD patterns exist.…”

Clines, Clusters, and Clades in the Race Debate

“…Recently, biologists have been testing whether the statistical clustering programs used in population genetics do in fact pick out the biologically important groups in the population. For example, Schwartz and McKelvey (2009) used computer simulations to assess the clusters that STRUCTURE picks out. They found that "for models where individuals (and their alleles) were randomly distributed across a landscape, STRUCTURE correctly predicted that only one population was being sampled.…”