Regression commonality analyses on hierarchical genetic distances

Abstract

International audienceLandscape genetics is emerging as an important way of supporting decision-making in landscape management, in response to the deterioration of matrix permeability due to habitat loss and fragmentation. In line with unremitting methodological developments in landscape genetics, a new analytical procedure was recently proposed as a way of evaluating the effects of landscape gradients on genetic structures. This procedure is based on the computation of inter-individual hierarchical genetic distances (HGD), a metric of genetic differentiation taking into account the hierarchical structure in populations as inferred from clustering algorithms. HGD can be used as dependent variables in multivariate regressions to assess the effects of various landscape predictors on spatial patterns of genetic differentiation. However, multicollinearity may obscure the interpretation of multivariate regressions. We illustrate how regression commonality analyses (CA), a detailed variance partitioning procedure that can be used to deal with multicollinearity issues, can thoroughly improve our understanding of landscape connectivity when HGD are used as a dependent variable, with the red deer Cervus elaphus as an empirical example. Using logistic regression commonality analyses on HGD, we showed that semi-natural open areas, transportation infrastructures and, to a lesser extent, urban areas and rivers, were associated with an increase in hierarchical genetic differentiation in red deer. Regressions based on HGD provided detailed results that could not have been obtained with regressions based on standard genetic distances, with notably additional insights as to the possible influence of linear features such as roads and highways on landscape connectivity. Furthermore, CA helped identify synergistic associations among variables as well as suppressors, thus resolving inconsistencies among hierarchical levels and revealing spurious correlations that may have gone unnoticed in the course of classical regression analyses. We thus recommend the use of regression commonality analysis on hierarchical genetic distances as a promising statistical tool for landscape geneticists

Contribution of spatial heterogeneity in effective population sizes to the variance in pairwise measures of genetic differentiation

“…Whatever the proxy used for K , mr showed limited unique contribution to the variance in measures of genetic differentiation, with values ranging from 3.4 to 7.8% (Table ). This variability in unique contributions of mr stemmed from collinearity with distance‐based metrics of genetic drift, as revealed by common contributions C (Prunier et al., ): indeed, the highest unique contribution of mr ( U = 7.8%) was also associated with the highest negative common contribution ( C = −4.8%), indicating statistical suppression, a situation responsible for an artificial boost in both the regression coefficient and its significance (Paulhus et al., ; Prunier et al., ). The observed variability in model fits (ranging from 4% to 44.2%) thus mostly ensued from the variability in SHNe metrics’ unique contributions to the variance in F st .…”

Contribution of spatial heterogeneity in effective population sizes to the variance in pairwise measures of genetic differentiation

“…Whatever the proxy used for K, mr showed limited unique contribution to the variance in measures of genetic differentiation, with values ranging from 3.4 to 7.8% (Table 1). This variability in unique contributions of mr stemmed from collinearity with distance-based metrics of genetic drift, as revealed by common contributions C (Prunier et al 2015): indeed, the highest unique contribution of mr (U = 7.8%) was also associated with the highest negative common contribution (C = -4.8%), indicating statistical suppression, a situation responsible for an artificial boost in both the regression coefficient and its significance (Paulhus et al 2004;Prunier et al 2017). The observed variability in model fits (ranging from 4% to 44.2%) thus mostly ensued from the variability in SHNe metrics" unique contributions to the variance in Fst.…”

“…For each situation, 10,000 genetic datasets were simulated with set to 0, m randomly picked from a uniform distribution ranging from 0.0001 to 0.3 and N max randomly picked from a uniform distribution ranging from 100 to 1000. (Paulhus et al 2004;Prunier et al 2017).…”

Patterns of gene flow across multiple anthropogenic infrastructures: Insights from a multi-species approach

“…We finally compiled the ten best runs using CLUMPP (Jakobsson and Rosenberg, 2007) to obtain individual or population Q-values. Each individual or population was assigned to the cluster for which the Q-value was higher than 0.6 (Prunier et al, 2017). We then repeated the analysis for each inferred cluster separately until no more structure was found in the data.…”