Leveraging population change to improve estimation of mortality in data-scarce context

In countries with incomplete vital registration, contemporary mortality estimation models such as the log-quadratic model the and Singular Value Decomposition (SVD)-Comp model often produce sizable errors, particularly at prime adult ages (20-45) and older ages (70+), where mortality varies substantially due to country-specific epidemiological profiles. This paper introduces a novel model-agnostic sorting approach that incorporates population change information, derived from decennial censuses and typically available even in data-scarce settings, to improve mortality estimation accuracy. Our method uses cosine similarity of z-score standardized population change patterns to identify the most demographically similar country-years (top 10%) in the training dataset. Using data from the Human Mortality Database and from HIV-affected countries from UN World Population Prospects, we test this approach with both log-quadratic and SVD-Comp models. Results demonstrate consistent improvements in prediction accuracy, with median RMSE for probability of dying (𝑞𝑞𝑥𝑥) decreasing by 17% (log-quadratic) and 13% (SVD-Comp), respectively. Gains are most apparent at young adult and older ages, particularly in countries with distinct mortality patterns, such as those affected by HIV. This approach offers a practical, easily implementable solution that can be integrated into existing mortality estimation workflows without additional data requirements or model restructuring.