Abstract

Dynamic spatial microsimulations integrate heterogeneous data sources to generate big and highly detailed synthetic microdata sets, which are then processed through complex procedures to produce forecasts for policy analysis. However, a comprehensive uncertainty analysis of such complex data and models, which involve many uncertain input factors, is computationally expensive. Design of experiments (DoE) methods, particularly orthogonal arrays (OAs), can be used to reduce this complexity, while ensuring reliable estimates of variance-based sensitivity indices. High-quality OAs may be selected prior to simulations.

On the other hand, the complex data at hand may include missing data blocks, or latent variables. In these situations, data augmentation, e.g., based on a mixed-effects model, is a suitable way to enrich the observed data. The augmented data are generated in a principled way, i.e., reflecting the existing uncertainty and being in accordance with the observations.

In this talk, both aspects, dimension reduction using OAs as well as data augmentation, will be considered jointly for the analysis of big and complex data sets of high-quality for policy analyses.