In addition to offering an improved fit to external validation data, the UKPDS 82 equations are also capable of predicting an expanded set of diabetes related endpoints, in particular, secondary myocardial infarction, stroke and amputation events and ulcers.
The results from this study would suggest that the UKPDS 82 risk equations are associated with an improved fit to external validation studies compared to UKPDS 68 (MAPE of 40.0 and 49.0% for UKPDS 82 and 68, respectively) and, as expected, a modest improvement to the UKPDS study endpoints.
Over-fitted models will show a bad fit to cross-validation data since they fitted the noise, rather than the signal, and therefore are less generalizable.
When fitted to the validation dataset, the most precise model included only the variables of age and Site Index.
Calibration in the large and calibration slope are calculated by fitting to the validation data two separate regression models.
The model fitted to the validation data results in the same specification and similar variable coefficient magnitudes to those in the original model.
The calibration slope will be estimated by including the linear predictor as a single term in the model fitted to the validation cohort.
In spite of the generally good agreement of model fit statistics between analysis and validation data sets, the linear models fit to the analysis and validation data sets are not identical, i.e. some variables that are highly significant in the analysis data set do not remain so in the validation data set.
Fitting to an external validation test set revealed superior prediction efficacy for statistical moments compared to conventional model fitting, while the combination of statistical moments with ANNs presented the most efficient approach (R2 and RMSEp values of 0.922, 0.833, 0.987 and 0.443, 0.691, 0.173 for MST, MET, and MDT, respectively).
We found GAM's to perform as well as nonlinear models with respect to fit and validation statistics; however, when applied to the population of predictor data the bootstrap confidence interval was less efficient than the null model.
