Mitigating the Error Rate of an IPF-Based Population Synthesis Approach by Incorporating more Heterogeneity into the Initial Seed

In this paper, we propose a bi-level procedure for population synthesis to obtain good estimates of both the marginal and joint distributions. As a fitting-based algorithm, the Iterative Proportional Fitting (IPF) approach is capable of providing accurate synthetic population estimates from the marginal distributions perspective. The algorithm iteratively recalculates the different weights of the k-way contingency table until the deviation between the simulated and observed marginal distributions is minimized. Besides, the initial boundaries, referred to as the seed, are preserved in terms of proportions with respect to the full sample. This means that although the different values are updated along with the iterations, the correlation structure of the underlying population remains unaffected. In this paper, a hybrid model is proposed. Its originality resides in the incorporation of heterogeneity into the initial seed by using a Hidden Markov Model (HMM) (1). The enriched initial seed is then fitted to a set of stable marginal distributions. This new hybrid approach is very interesting as it for instance can be calibrated by an unlimited number of PUMS, and separate information can be merged. The estimation of the correlation structure (synthetic seed) is improved compared to the standard seed (stemming from a micro-sample) thanks to the HMM. In this regard, the results show that the coupling of IPF with the HMM method provides better estimates, i.e. decreased RMSE of 79.16% for 1% sample size, from both the marginal and the joint distributions points of view. This implies that enriching the micro sample is absolutely necessary before fitting with any aggregate marginal distributions.