Attributive Causal Modeling: Quantifying Human Health Risks Caused by Toxoplasmosis from Open System Production of Swine

Abstract

This is the first of two chapters that apply predictive analytics to two very different risk prediction problems. As in the previous two chapters, the challenge in this one is to estimate human health risks from a pathogen in swine using a combination of plausible conservative estimates of relevant risk factors and probabilistic simulation. However, our focus now shifts to predicting how risks would change if some fraction of swine were shifted from totally confined production systems to more humane open systems. Predicting how interventions change risk requires a causal model, as discussed in Chap. 1. As in Chaps. 5 and 6, a simple product-of-factors framework is again suitable (see Eq. 7.5). Instead of the terms describing propagation of changes along successive links in a causal chain, with the change in the quantity at each step being equal to a sensitivity or slope factor times the change in its predecessor, many of the factors in this chapter are estimated attribution fractions. These describe the fraction of relevant deaths or illnesses per year in the population due to (i.e., attributed to) and caused by infection with a foodborne pathogen; the fraction of them that are attributed specifically to pork consumption, and so forth. Unlike the attributable risk estimates or attributable fractions criticized in Chap. 2, which were derived purely from statistical associations, in this application the causal agent of disease, T. Gondii, is known and can be measured. Predictions for effects of interventions are therefore grounded in causal attribution calculations that can be compared to available data on prevalence and infectivity of the relevant causal agent. Chapter 8 will then turn to a pure prediction problem: how well the entries in one column in a table (indicating in vivo carcinogenicity of chemicals, or lack of it, in rodents) can be predicted from entries in other columns, representing results of relatively inexpensive high-throughput screening (HTS) assays. No causal model is required for this task: predictive analytics algorithms alone suffice.