Table 2

Summary of model selection

ModelBinomial GLM (complementary log–log)AFT model (Inverse-Weibull)Cox PH model (stratified by time-since-injury)
Accuracy (%)Sensitivity (%)Specificity (%)AIC (# predictors)Accuracy (%)Sensitivity (%)Specificity (%)AIC (# Predictors)Accuracy (%)Sensitivity (%)Specificity (%)AIC (# predictors)
Models with main effects only
Null model64010035664010022436401001927
Physical examination only67 (63)35 (31)86 (82)338 (10)63 (62)3.1 (0)99 (99)2262 (10)75 (74)48 (46)92 (91)1928 (10)
All predictors76 (73)54 (49)89 (87)278 (15)74 (73)45 (43)91 (90)2164 (15)75 (73)51 (47)90 (88)1915 (14)
Models including interaction terms
All interactions92 (66)87 (48)96 (77)338 (120)84 (70)66 (46)95 (84)2167 (120)80 (62)71 (65)85 (67)1965 (105)
Stepwise selected models
All predictors stepwise77 (75)57 (53)89 (89)266 (7)73 (72)43 (43)92 (90)2150 (6)76 (73)54 (47)89 (88)1901 (5)
All interactions stepwise 78 (75) 56 (51) 91 (90) 263 (8)73 (72)43 (43)92 (90)2150 (6)75 (70)45 (46)92 (85)1900 (7)
  • Pairs of numbers in a cell refer to performance on training and test data, with the upper number showing performance on the training data, and the lower number the cross-validation performance. For AFT and bGLM models, results are only given for the model with the lowest AIC. AIC are presented with number of predictors in their model.

  • Bolded values indicate the final model used for the RDR Score.

  • AFT, accelerated failure time; AIC, Akaike information criterion; GLM, generalised linear modelling; PH, proportional hazard.