Success and failure in dynamic decision environments: understanding treatment strategies for patients with a chronic disease.

Abstract

This dissertation proposes and tests a theory explaining how people make decisions to achieve a goal in a specific task environment. The theory is represented as a computational model and implemented as a computer program. The task studied was primary care physicians treating patients with type 2 diabetes. Some physicians succeed in achieving evidence-based goals, but many fail. In a previously conducted experiment 19 physicians treated 3 simulated patients with type 2 diabetes, this was the dataset used for modeling and testing (O'Connor et al., 2009). Models were constructed to deliver care in the manner of an idealized physician. These models were tested by treating the same simulated patients that subject physicians treated. Perturbations in model processes were used to explain failure to achieve goal. These perturbations represented forms of omission bias which result in errors of under-treating patients. Each physician's dataset on each case was scored for errors and decision strategy. A mapping was developed from an error to a process perturbation. Models of each physician were constructed by: (1) selecting an idealized model that used the same decision strategy as the subject, and (2) introducing perturbations in the model based on errors that the subject committed when treating the cases. For each case each physician and corresponding model were evaluated for a goodness of fit based on similarity of error patterns committed, differences in final blood glucose values obtained, and similarity of final medications prescribed. Models made point predictions for when during the course of treatment errors would be committed: 90% of models generated the same types of errors as modeled physicians and 67% of models committed the same errors on the same visits as physicians. Based on tests of the models (theory) we found support for omission bias as a plausible explanation for agents committing errors of under-treating which prevents reaching clinical goals with type 2 diabetes patients. While the models could predict treatment actions for prescribing oral medications, it failed to accurately predict prescriptions for insulin doses.