Risk expression using likelihood ratios and natural frequencies in Bayesian inference tasks-a preregistered randomized-controlled crossover trial.

BACKGROUND

To make reasonable future medical decisions, medical students need to be sufficiently educated to interpret diagnostic tests. Natural frequencies are considered the gold standard for understanding single diagnostic test results. However, they may be less suitable in situations involving sequential diagnostic testing. We test whether odds and likelihood ratios (odds/LR) may serve as a viable alternative in these situations.

METHODS

In our preregistered randomized-controlled crossover trial, we recruited 167 medical students and 162 psychology students. The proportion of correctly calculated positive predictive values of a single (PPV) and two sequential diagnostic tests (sPPV) was the primary, the subjective comprehensibility of the information the secondary outcome.

RESULTS

The proportion of correct PPVs was significantly higher in the natural frequency (36.2%) compared to the odds/LR format (21.6%), OR 2.41. Conversely, the proportion of correct sPPVs was significantly higher in the odds/LR (10.6%) compared to the natural frequency format (4.9%), OR 2.73. Participants indicated a higher subjective comprehension of test statistics phrased as natural frequencies (Mdn = 19) than as odds/LR (Mdn = -15), r = .61.

CONCLUSION

Teaching Odds/LR next to natural frequencies potentially improves medical students' understanding of PPV and may enhance their ability to make future diagnostic decisions.

TRIAL REGISTRATION

https://doi.org/10.17605/OSF.IO/F3297 .