Weighting the evidence: allele frequency strategies for improving LD-based Ne estimation

Accurate estimation of contemporary effective population size (Ne) is a central requirement for genetic diversity monitoring in conservation genomics. Linkage disequilibrium (LD)-based methods offer a practical single-sample solution, but their performance depends critically on SNP panel composition: rare alleles downwardly bias r² and inflate Ne estimates, a problem typically addressed by minor allele frequency (MAF) filtering. Sved et al. (2013) proposed an alternative – weighting locus pairs by allele frequency P or by P(1−P) – since intermediate-frequency alleles carry more information about LD. Waples (2023) further discussed this idea, but both approaches remain unrigorously benchmarked. We evaluated five scenarios differing in how allele frequency information enters LD-based Ne estimation, spanning MAF filtering, preferential SNP selection by P or P(1−P), and pair-level δ²-based weighting, as implemented in the software currentNe (Santiago et al. 2023). All analyses use exclusively inter-chromosomal SNP pairs (c = 0.5), ensuring valid expected r² under drift. We used Wright-Fisher forward simulations across a full factorial grid of true Ne, sample size and panel size, evaluating bias, precision and RMSE under each scenario. Results reveal a context-dependent performance trade-off. SNP selection by P(1−P) and P achieve the strongest bias reduction across the parameter space, even at small sample sizes where MAF filtering still shows substantial overestimation. Pair-level weighting provides superior precision at small Ne and low sample sizes, where rare-allele noise dominates variance. RMSE favours different strategies depending on the Ne–sample size regime: pair-level weighting tends to perform best at low S and small Ne, while SNP-selection approaches become more competitive as S and Ne increase. No single strategy universally dominates, and the optimal choice depends on the relative priority of bias versus precision in a given monitoring context.