Risk assessment, computer learning, diagnosis and Bayes: a commentary

✍ Scribed by Lenard I. Dalgleish

• Publisher: John Wiley and Sons
• Year: 1998
• Tongue: English
• Weight: 140 KB
• Volume: 7
• Category: Article
• ISSN: 0952-9136
• DOI: 10.1002/(sici)1099-0852(199805/06)7:3<189::aid-car487>3.0.co;2-0

✦ Synopsis

Computer Learning, Diagnosis and Bayes: A Commentary T he two previous papers approach the issue of how information is combined to form risk assessments. However they do it from very dierent perspectives. Little and Rixon (1998) consider how factors in child abuse cases are combined to form a risk assessment. Kemp, Kemp, Evans, Murray, Guildea, Dunstan and Sibert (1998) consider how information on the number of bruises and the referral source is combined to estimate the probability that a diagnosis of physical abuse would be correct. This commentary will outline both similarities and dierences and highlight underlying issues raised by these papers. These issues include:

. modelling and representing the relationship between information and outcome;

. prescriptive and descriptive approaches to human judgment and decision making;

. the nature of risk and uncertainty;

. thresholds for action.

Little and Rixon (1998) descriptively model the combination of information in forming an assessment of the amount of risk in a case by looking for patterns among the factors and risk assessments over a small set of 20 cases. An algorithm, ID3, ®nds the smallest decision tree that best captures the patterns. Two factors were found that classi®ed most of the cases. They state that applying this model yields consistent assessment of new cases and provides a consistent structure to compare one case to another. This implies a predictive purpose although they also explicitly say that their emphasis is on re¯ection rather than prediction. However, why model the past if not to use the model to assist in future assessments.

How well does the algorithm, ID3, work using two factors only? Applying the rules from their Figure 3 yields 6 cases where the algorithm predicts a risk value dierent from the