Decision making with incomplete information

Abstract

Decision makers are continuously required to make choices in environments with incomplete information. This dissertation sought to understand and, ultimately, support the wide range of decision making strategies used in environments with incomplete information. The results showed that the standard measure of incomplete information as total information, is insufficient for understanding and supporting decision makers faced with incomplete information. The distribution of information was shown to often be a more important determinant of decision making performance. Two new measures of the distribution of incomplete information were introduced (option imbalance and cue balance) and tested across three computer simulations of 18 variations of decision making strategies within hundreds of environments and millions of decision tasks with incomplete information, and one human-subjects study. The simulations were powered by a new general linear model of decision making which can efficiently and transparently model a wide range of strategies beyond the traditional set in the literature. Of the many potential mediators of the relationship between the distributions of incomplete information and performance, only the strategies' estimates of missing information were significant in the computational studies. Accurate estimates resulted in total information being the only meaningful determinant of accuracy while inaccurate estimates resulted in low option imbalance and high cue balance causing high accuracy. The simulation results were partially contradicted by a study in which human decision makers with accurate estimates were affected by option imbalance and cue balance in the same manner as inaccurate estimates – suggesting that some distributions might simply be difficult regardless of the estimates. These results argued that decision support should modify the presentation of information away from difficult distributions. These arguments were codified as heuristic information acquisition and restriction rules which, when tested, increased accuracy without probability and cue weight information.