Properties of alignment methods in discrete time dynamic microsimulation models

Alignment is a critical calibration technique in microsimulation, ensuring individual-level transitions aggregate to known macro-targets. While indispensable for updating populations to match demographic projections or macroeconomic forecasts, the statistical properties of various alignment algorithms remain under-researched. This paper provides a systematic evaluation of alignment methods for discrete-time models to guide researchers in method selection.

We categorize existing methods based on a robust taxonomy: variable type (continuous vs. discrete), the nature of the outcome (exact match vs. stochastic), and time-step logic (continuous vs discrete). Building on the work of Stephenson (2018), we demonstrate that most contemporary alignment methods can be unified under a single constrained optimization framework, differing only in their choice of objective functions. This unification allows us to establish previously unrecognized relationships between seemingly disparate techniques.

Furthermore, we derive closed-form expressions and Taylor series approximations for computationally intensive iterative methods. These mathematical shortcuts allow modellers to reduce simulation run-times significantly while converting exact alignment processes into stochastic ones without losing desirable statistical properties. We then study the impact of these methods on the predicted distributions and covariance of outcomes between simulation runs.

Our analysis reveals that common techniques, such as Sorting by Predicted Probability (SBP) and Sorted By Difference between predicted Probability and a random number (SBD), are undesirable for microsimulation modelling as they can introduce biases and path dependence into the simulation outcomes. In contrast, parameter variation methods (i.e. recalibration of model coefficients such as intercept shifting) offer modellers control over the impact of alignment and better interpretability. Furthermore, methods like logit scaling, variance-weighted additive scaling for probabilities, and additive scaling for continuous variables offer greater robustness due to their foundations in probability theory. We conclude by providing a decision matrix for modellers, weighing the trade-offs between computational efficiency, distribution preservation, and variance reduction.