Improving the Robustness of Phylogenetic Independent Contrasts: Addressing Abrupt Evolutionary Shifts with Outlier- and Distribution-Guided Correlation

Abstract

Traditional phylogenetically aware correlation methods perform well under gradual evolutionary processes. However, abrupt evolutionary shifts--or macroevolutionary jumps, characteristic of punctuated evolution--can produce extreme phylogenetically independent contrasts (PIC), leading to inflated false positives or increased false negatives in trait correlation analyses. We introduce O(D)GC (Outlier-and Distribution-Guided Correlation), a flexible workflow that identifies outliers in PICs using a distribution-free boxplot criterion and applies Spearman correlation whenever influential outliers are detected. If no outliers are detected, Pearson correlation is used--automatically for large datasets (n [≥] 30), or guided by normality testing in smaller samples. We systematically compared PIC-O(D)GC with five widely applied phylogenetic correlation methods--PIC-Pearson, PIC-MM, PGLS (phylogenetic generalized least squares), MR-PMM (multi-response phylogenetic mixed model), and Corphylo--on 322,000 simulated datasets spanning five evolutionary scenarios (two shift settings: single-trait shifts and dual-trait co-directional jumps; and three no-shift gradual evolution settings), including both fixed-depth and randomly located shifts, tested across 11 shift or noise gradients, three tree sizes (16, 128, 256 tips), and both balanced and random topologies. Overall, PIC-O(D)GC achieved error rates comparable to--or noticeably higher than--those of PIC-MM, while yielding substantially lower error rates than most alternative methods. Under no-shift conditions, it retained power similar to other methods. Analyses of three empirical datasets likewise showed that PIC-O(D)GC and PIC-MM corrected shift-induced distortions that misled conventional methods. Moreover, PIC-O(D)GC offers a conceptually simple framework and incurs markedly lower computational cost. By design, its correlation-only output provides less mechanistic detail than regression-based approaches like PGLS. However, when paired with PIC diagnostics, this outlier-guided strategy highlights evolutionary jumps, distinguishes coupled from decoupled shifts, and--via clade partitioning or tip pruning--recovers background correlations, offering biologically informative insights into how punctuated events interact with gradual trends in trait evolution.