Article #3
Newsletter list 


Previous Article 

Next Article 


Information Selection Dissertation and Brunswik Chapter

Michael Doherty
Bowling Green, OH

The following is a an abstract of a dissertation currently being written up by John R. Leach, at Bowling Green State University:

"Information Selection in a Simulated Medical Diagnosis Task: The Effects of External Representations and Completely Natural Sampling"

Gigerenzer (1994) argued that cognitive biases could be made to disappear if tasks and problems were presented to subjects in formats analogous to natural sampling, rather than probability formats.

Several researchers (Gigerenzer and Hoffrage, 1995; Cosmides and Tooby, 1994; Betsch, Biel, Eddelbuttel, & Mock, 1998; Gigerenzer & Todd, 1999) have studied the accuracy of subjects on Bayesian inference problems presented in either a frequency format (i.e., analogous to natural sampling) or a probability format.

Results generally showed that frequency formats elicited higher proportions of solutions consistent with Bayesian solutions than did the probability formats. It was assumed that frequency formats correspond to natural sampling.

Gigerenzer and Hoffrage (1995) expressly stated that sequential acquisition of information by updating event frequencies without artificially fixing the marginal frequencies is what they refer to as natural sampling. However, the tasks they used did not have subjects sequentially update frequencies. Subjects were simply presented with summary information about events and outcomes.

In the present study, subjects sequentially sampled a task environment in a "completely natural" style. They were presented with a fictitious medical diagnosis scenario. This scenario offered two competing hypotheses about symptoms associated with two different fictitious diseases. The study was designed to determine if completely natural sampling is different than natural sampling as conceptualized by Gigerenzer (i.e., frequency formatted summary information), and if subsequent diagnostic and probability judgments differ. Diagnoses and judgments of probability were analyzed in relation to external representation and style of sampling (i.e., natural, completely natural, and controlled).

Subjects exposed to completely natural sampling were 100 percent accurate in diagnosing the disease. However, in a subsequent pseudodiagnosticity task, fewer than 50 percent were able to identify the information needed to calculate the Bayesian probability. Additionally, subjects in all conditions had trouble calculating probabilities associated with the diagnosis. Fewer than 30 percent reported the exact Bayesian probability. The results are consistent with the proposition that the cognitive processes involved in global judgments and diagnoses differ from those involved in analytical reasoning.

Ryan Tweney and I have just agreed to write a chapter on Brunswik for a book dealing with the history of thinking and reasoning. Given that I will be fully retired in two weeks, this is just the sort of scholarly project I should find of great interest in retirement.

Contact Michael Doherty

Previous Article                Next Article

Newsletter list