Logistic Regression for Robust Familial DNA Analysis Amidst Population Substructure

When performing familial DNA searches, a key challenge is accounting for the unknown population background of forensic samples, which complicates the calculation of likelihood ratios that are used as a statistical measure for evaluating genealogical relationships between two individuals. This challenge is further exacerbated by population substructure, which introduces multiple potential allele frequency distributions. To address this issue, various likelihood ratio methods have been developed, including those based on averaged allele frequencies (LRLAF), as well as maximum, minimum, or average likelihood ratios (LRMAX, LRMIN, LRAVG), and Naive Bayes classification (LRNB). In this study, we introduce LRMLR (Likelihood Ratios incorporating Multinomial Logistic Regression), a new family of statistics that first classifies DNA samples into subpopulations using logistic regression and then applies the likelihood ratio framework accordingly. Our evaluation, based on three simulated populations generated from real allele frequency data, shows that LRMLR achieves comparable power in parent-child and full-sibling tests to methods such as LRNB and LRMIN. Given its robust and consistent performance across different strategies, LRMLR offers a competitive and practical alternative for conducting familial DNA searches in real-world forensic applications.