Injury risk | Cost: benefi† | ‘Treat all’ | Original model (M1) | Original model with linearity assumed for continuous predictors (M2) | Original model with continuous predictors and non-linear transformations (M3) | Original model with further included predictors that are dichotomised (M4) | Predictor selection based on rationale from literature and clinical reasoning and kept continuous with non-linear transformations (M5) | Treat none |
0.25 | 1:3 | 0.35 | 0.36 | 0.79 | 0.79 | 0.38 | 0.77 | 0.00 |
0.30 | 3:7 | 0.16 | 0.25 | 0.69 | 0.70 | 0.25 | 0.70 | 0.00 |
0.35 | 7:13 | −0.05 | 0.19 | 0.61 | 0.61 | 0.14 | 0.59 | 0.00 |
0.40 | 2:3 | −0.31 | 0.08 | 0.47 | 0.45 | 0.08 | 0.47 | 0.00 |
0.45 | 9:11 | −0.60 | 0.06 | 0.39 | 0.39 | 0.06 | 0.39 | 0.00 |
0.50 | 1:1 | −0.96 | 0.03 | 0.36 | 0.36 | 0.07 | 0.36 | 0.00 |
*The threshold probability was defined as the population risk of injury within military personnel of 0.25–0.50. The original model with continuous predictors and non-linear transformations and the newly developed model demonstrated improved net benefit (ie, resource allocation) compared with ‘treat all’ and the original model at these threshold probabilities.
†Cost:benefit reports the acceptability of performing a certain number of screens to find one patient with the outcome (eg, injury). For example at an injury risk threshold of 0.20, clinicians would be willing to perform screens on 5 patients to find one military personnel truly at risk for an injury.