On The Effect of Polygenic Risk Scores and other Non-Genetic Factors on Life Insurance Underwriting Decisions

Abstract

This study investigates the effect of polygenic risk scores and other variables on underwriting decisions in life insurance using machine learning. Using simulated data of 10,000 individuals, we developed logistic regression and random forest models to analyze the impact of each variable. Traditional variables and their limitations were described in order to contextualize the original methods for life insurance underwriting. Other genetic methods and their limitations excluding polygenic risk scores were outlined with the intention of comparing an emerging concept with various traditional genetic analytical methods. Data was generated and processed using machine learning techniques to ensure reasonable results. However, our multicollinearity analysis revealed important limitations: the synthetic PRS showed multicollinearity with other variables (VIF > 10) and correlations with family history variables that were artifacts of the data generation process. These analyses demonstrated that polygenic risk scores have a significant effect on life insurance underwriting decisions by enhancing the accuracy of the risk-prediction models. Feature importance charts showed that the more accurate model (the random forest, as confirmed by comparison tables) gave greater weight to variables related to polygenic risk scores and their interactions. However, the model also considered several traditional variables with substantial weight, including premium cost, underlying condition, age, and health risk score. This research offers a valuable perspective into the relationship between polygenic risk score and individual characteristics, highlighting the value of integrating genetic variables with traditional variables, while acknowledging the challenges of implementing synthetic PRS. These findings provide evidence supporting the potential application of polygenic risk scores in risk-prediction models, though real-world implementation would require addressing the methodological limitations identified. By increasing the accuracy of risk-prediction models, current conflicts such as adverse selection and information asymmetry could potentially be resolved with the addition of certain policies and proper implementation of true genetic risk scores.