A Novel Weighting-Based Approach to Cohort Replenishment in Dynamic Microsimulations

We propose a new method for generating replenishment cohorts in dynamic microsimulation models. Standard dynamic microsimulations project the future states of an initial population through the recursive application of one-step-ahead predictions. Over time, sample size declines due to attrition (e.g., mortality), and without the integration of new individuals, the projected population progressively departs from the target population structure. To preserve representativeness, replenishment cohorts must therefore be introduced at each simulation step.

Cohort replenishment is challenging because it must simultaneously (i) reflect secular trends in individual characteristics (e.g., declining smoking prevalence) and (ii) preserve the underlying correlation structure among these characteristics (e.g., the relationship between smoking and lung cancer). Existing approaches, most notably the method used in the Future Elderly Model, address these challenges but are computationally intensive and algorithmically complex.

We introduce an alternative algorithm that draws eligible donors from historical data while preserving their observed characteristics. Donor sampling weights are adjusted to match period-specific target prevalences using a procedure akin to entropy balancing (Hainmueller, 2012). The proposed method offers three key advantages: (i) efficiency, as it is substantially simpler to implement and less computationally demanding than existing approaches; (ii) accuracy, as it closely tracks specified feature trends given a sufficiently rich donor pool; and (iii) parsimony, as it avoids the need to explicitly specify trends for all variables. Instead, trends in non-targeted characteristics emerge endogenously from the imposed constraints, an especially valuable property in settings with limited longitudinal data.

We evaluate the proposed method by comparing its performance with the benchmark approach in reproducing target prevalences, preserving the joint distribution of individual characteristics, and generating plausible trends for features not explicitly constrained features.