Abstract
Statisticians have proposed a number of theories for choosing between possible experiments in order to identify which experiment is expected to be the most useful. Behavioral and eye movement research has addressed whether one or more of these theories describe human information acquisition in eye movement and behavioral information search tasks. Based on theoretical analysis and an experiment, Nelson (2005) suggested (1) that two theories of the value of information, Bayesian diagnosticity and log diagnosticity, are not credible descriptive psychological theories of the value of information, but (2) that data to date did not differentiate which of several other theories — information gain-KL distance, probability gain, and impact — best approximates human search.
We now report an experiment to test the different theories. Subjects learned the statistics of an environment involving simulated plankton. The plankton could be one of two species, which depended probabilistically on two binary features. Subjects first went through a learning phase, in which they classified randomly drawn plankton specimens as species A or B, with feedback. Once a subject consistently performed optimally, the experiment moved into the information-acquisition phase, in which the features of interest were obscured on each trial, and the subject selected a single feature to view.
Several between-subjects conditions contrasted the various theories of information search and, in every condition, the feature with higher probability gain was preferred by a majority of subjects. It appears that probability gain explains human information search much better than information gain-KL distance, impact, Bayesian diagnosticity or log diagnosticity.
funding provided by NIH 5T32MH020002-04