Bayesian clinical reasoning: does intuitive estimation of likelihood ratios on an ordinal scale outperform estimation of sensitivities and specificities?
Version of Record online: 31 OCT 2008
© 2008 The Authors. Journal compilation © 2008 Blackwell Publishing Ltd
Journal of Evaluation in Clinical Practice
Special Issue: Evidence Based Medicine
Volume 14, Issue 5, pages 934–940, October 2008
How to Cite
Moreira, J., Bisoffi, Z., Narváez, A. and Van den Ende, J. (2008), Bayesian clinical reasoning: does intuitive estimation of likelihood ratios on an ordinal scale outperform estimation of sensitivities and specificities?. Journal of Evaluation in Clinical Practice, 14: 934–940. doi: 10.1111/j.1365-2753.2008.01003.x
- Issue online: 31 OCT 2008
- Version of Record online: 31 OCT 2008
- Accepted for publication: 13 December 2007
- Bayes' theorem;
- clinical reasoning;
- educational tools
Rationale: Bedside use of Bayes' theorem for estimating probabilities of diseases is cumbersome. An alternative approach based on five categories of powers of tests from ‘useless’ to ‘very strong’ has been proposed. The performance of clinicians using it was assessed.
Methods Fifty clinicians attending a course of tropical medicine estimated powers of tests and post-test probabilities using the classical vs. the categorical Bayesian approach. The estimation of post-test probability was assessed for real and dummy diseases in order to avoid the bias of previous knowledge. Accuracy of answers was measured by the difference with reference values obtained from an expert system (Kabisa).
Results Clinicians estimated positive likelihood ratios (LRs) a median of −1.07 log10 lower than Kabisa [interquartile range (IQR): −1.47; −0.80] when derived classically and −0.17 (IQR: −0.42; +0.04) when estimated categorically (P < 0.001). For negative LRs the median was +0.39 log10 higher (IQR: +0.71; +0.08) when derived classically and −0.18 log10 lower (IQR: +0.03; −0.36) when estimated categorically (P < 0.001).
Twenty (40%) disclosed not being able to calculate post-test probabilities using sensitivities and specificities. Regardless the approach post-test probabilities were overestimated both for real and dummy diseases [respectively +1.23 log10 (IQR: +0,67; +2.08) and +2.03 log10 (IQR: +0.49; +2.42)] (P = 0277), but the range was wider for the latter (P = 0.001).
Conclusions Participants were more accurate in estimating powers with a categorical approach than with sensitivities and specificities. Post-test probabilities were overestimated with both approaches. Knowledge of the disease did not influence the estimation of post-test probabilities. A categorical approach might be an interesting instructional tool, but the effect of training with this approach needs assessment.