A Two-stage, Fitted Values Approach to Activity Matching

Accurate and rich representations of constituent actor populations are a critical component of agent-based models. Such populations are designed so that demographic, behavioral, procedural, and geographic characteristics of the synthetic population jointly reflect the available information about the true population. Information about the attributes to be mimicked in the synthetic population is often derived from survey samples of the real actors of interest - such as firms in a market or households in a city. This additional information is then mapped to individual actors in such a way that each actor in the population represents one sample from the joint distribution of all assigned attributes. These actors then interact according to rules, which are functions of their attributes. Thus, accurate attribute matching is necessary to ensure that model outputs are meaningful. In real applications, behavioral surveys often yield complex data types, such as daily activity schedules or action sequences, for which it is difficult to conceive of adequate conditional models of behavior that could be used to generate new behavioral data as a function of covariates. Here we propose a method for assigning behavioral templates to synthetic agents from a set of survey templates. Our method first maps the complex behavioral data to a reduced-dimension Euclidean space, then estimates conditional models in this space. We then make predictions in Euclidean space for synthetic actors and assign them the template schedule that minimizes the distance to the predicted value. By employing a two-step process, we also ensure that within-household dependence structures are maintained in the synthetic population. We illustrate the method with an application to a synthetic representation of households in the state of Israel and demonstrate superior ability to generate accurate joint distributions between demographic characteristics and behavioral activity relative to the standard behavior assignment method.