From Nuisance to Signal: Leveraging Close Relatives in Biobank-Scale Demographic Inference

Biobank-scale datasets now routinely include hundreds of thousands to millions of individuals, and as sample sizes grow, close relatives become increasingly prevalent. The convention in population genetics has been to remove close relatives prior to inference, effectively treating them as a nuisance parameter. However, the consequences of this practice for demographic inference, and specifically for estimates of recent effective population size ( $N_e$ ), have not been rigorously evaluated. Here, we benchmark IBDNe and HapNe-IBD, two widely-used methods for inferring recent $N_e$ from identity-by-descent (IBD) segments, under a range of demographic histories and relative sampling schemes. We show that when individuals are randomly ascertained, retaining all relatives produces the least biased $N_e$ estimates; in contrast, removing even second-degree relatives inflates recent $N_e$ and induces oscillatory artifacts that "ripple", leading to biased estimates up to ten generations into the past. We demonstrate that this ripple effect arises because close relatives contribute IBD segments that are assigned by the model to a range of ancestral ages beyond their true TMRCA, meaning their removal creates signal deficits across multiple generations simultaneously. We further show that deliberately oversampling close relatives produces severe downward bias in recent $N_e$ . To support these analyses, we develop an open-source IBD simulation pipeline using msprime that generates realistic IBD segments under arbitrary demographic histories and Wright-Fisher pedigrees. We provide practical guidelines for IBD simulation schemes incorporating pedigrees and argue that, in the biobank era, retaining close relatives is generally the best practice for IBD-based $N_e$ inference.