Base rate neglect
Base rate neglect is giving too little weight to how common an event is when interpreting evidence about it. A high probability of a signal given an event does not, by itself, establish a high probability of the event given the signal.[1]
A worked example
Suppose a hypothetical batch contains 1,000 items: 10 defective and 990 sound. An inspection flags 80% of the defective items and 10% of the sound items. The resulting counts are:
| Items | Flagged | Not flagged | Total |
|---|---|---|---|
| Defective | 8 | 2 | 10 |
| Sound | 99 | 891 | 990 |
| Total | 107 | 893 | 1,000 |
Among the 107 flagged items, only eight are defective: 8/107, or about 7.5%. The 80% detection rate answers a different question: how often a defective item is flagged. The low base rate means that even the smaller false-alarm rate produces many more flagged sound items.
Information formats
Gigerenzer and Hoffrage found that presenting inference problems as frequencies drawn from a common population increased Bayesian responses in their experiments. In their first study, the proportions were 46–50% with frequency formats and 16–28% with probability formats, depending on which information was supplied. They classified responses using both answers and reasoning protocols.[1]
The arrangement of the information also matters. Barbara Mellers and Peter McGraw found benefits from frequencies for rare-event problems, but showed that probability information could outperform frequency information when the former made the relevant subsets easier to identify. Neither format consistently produced correct answers.[2]
In the worked example, all four cells belong to the same batch, so the base rate is already embodied in the counts. Sampling equal numbers of defective and sound items would produce a different sample composition; its raw counts could not establish the proportion defective in the original batch. A helpful representation therefore preserves the relevant population and subsets, as well as making the arithmetic easier.
See also
References
- ↑ 1.0 1.1 Gerd Gigerenzer and Ulrich Hoffrage, “How to Improve Bayesian Reasoning Without Instruction: Frequency Formats”, Psychological Review 102(4), 1995, pp. 684–704.
- ↑ Barbara A. Mellers and A. Peter McGraw, “How to Improve Bayesian Reasoning: Comment on Gigerenzer and Hoffrage (1995)”, Psychological Review 106(2), 1999, pp. 417–424.
