Construction, validation and updating of a prognostic model for kidney graft survival

The construction, validation and updating of a prognostic model for kidney graft survival is reported using data from the Eurotransplant database. First, a model is constructed for data from transplantations in the period 1984 to 1987. The model is later updated for the 1988 1990 data. The first data set was randomly split into a training set (two-thirds of the data) and a validation set (one-third). To prevent overfitting empirical Bayes estimation of the transplantation centre effect was employed. After that, the validation set was used for fine-tuning by shrinkage. For updating with the 1988 1990 data parametric models were used after suitable transformation of the time axis; it appeared that survival had slightly improved. This necessitated a correction of the parameters in the exponential model. Correctness of the model was checked by extension to a Weibull model. The lack of fit was statistically significant, but practically ignorable. Recommendations are made to place less emphasis on the selection of variables and cut-off points, and more emphasis on the fine-tuning of the prognostic model by means of low-dimensional parametric models in independent data sets.