Design Principles for Micro Models

Abstract

Most applied microsimulation models are hard-coded for a specific national dataset and application. The dream of using a generic modeling package instead of programming each model from the ground up has not yet materialized. Until then, several experiences from such specific programs should be useful in order to shorten and improve development cycles. In this chapter, we present design principles that gradually have emerged from our efforts to create agent-based dynamic microsimulation models. We discuss advantages of distinguishing between the internal logic of the model (kernel), the tools potentially useful for several models, and the user interface. The main part of the chapter suggests design principles regarding parameter input, matrix input, equation evaluation, result aggregation, biography aggregation, memory allocation, random number generation, handling of sets, random choice between many alternatives, primary and secondary attributes, parallel execution, and twins or equations. For almost any model, we prefer storing the entire population in core memory as this results in no file access from inner loops during execution. Several of the suggestions point towards favoring basic data structures and algorithms available in any general computer language rather than extensive use of “modern” object classes and method wrappers. For example, it is far more memory efficient to store population attributes in indexed arrays as compared to separate object instances for each individual. In addition to using the discussed design principles while constructing new intrinsic national and regional simulation models, we would emphasize their use in further efforts to create generic high-level software for dynamic microsimulation.