Figure 3

Plot of the frequency of each discrete pretest probability as a function of the actual pretest probability, estimated on the validation population using two methods. A Gaussian curve was fitted to the attribute matching points (solid line) and the logistic regression points (dashed line). The median pretest probability estimate from attribute matching was 5.5% (1^st–3^rd quartiles: 1.0–10.5%), versus 24.0% (1^st–3^rd quartile 10.0–30.0%) for the logistic regression method (P < 0.001, Mann Whitney U). Because of rounding, each point could represent more than one unique match profile.