Information representation, problem format, and mental algorithms in probabilistic reasoning.

摘要

The purpose of this thesis was to better understand when people use and neglect base rate information. In four experiments, university students made probabilistic judgements for inferential reasoning tasks that were modelled after Tversky and Kabneman's taxi-cab problem (Tversky & Kahneman, 1982). In all, four theoretical perspectives were used to test predicted outcomes with observed responses and normative expectations. Experiment 1 found no differences between the trial-by-trial (on-line) format (Baker, Mercier, Vallée-Toumngeau, Frank, & Pan, 1993) and the traditional word problem (off-line) format of Tversky and Kahneman. Here the pattern of judgements was nearly identical across conditions and yet different from normative expectations. Both formats showed a strong sensitivity to base rates and the individuating information. This was contrary to the predictions based on the heuristics and biases program (Kahneman, Slovic, & Tversky, 1982) and also contrary to the predictions made by the frequentist approach (Gigerenzer, 1998; Gigerenzer & Hoffrage, 1995). Experiments 2 and 3 were designed to test the maxims of conversational conventions (Grice, 1975; Hilton, 1995). Experiment 2 was run entirely on-line and found no differences in mean probability judgements between the different conditional (individuating) sources of information. Participants relied equally heavily on a human or non-human source and also were sensitive to the changes made in base rate information. Experiment 3 was run entirely off-line and specifically tested the attribute of ambiguity. When the individuating source of information was very ambiguous, then judgements were based almost solely on the base rate information. None of the theories tested to this point were able to accurately predict the patterns of judgements made in all these conditions. What the results of these three experiments did suggest was that how the information is presented (e.g., probabilities vs. frequencies) and the context in which it is presented (i.e., how clearly the information is presented) are both important. These two factors combined suggest that a more general cognitive mechanism, one that is a function of the complexity of the judgement task and the amount of cognitive work that is required to make the judgement could best account for the data. This was tested and supported in Experiment 4. Normative responses can be elicited in judgement tasks that present only the essential probabilistic information in a format that is clear, unambiguous, and that requires few mental operations.