TUTORIAL IN BIOSTATISTICS MULTIVARIABLE PROGNOSTIC MODELS: ISSUES IN DEVELOPING MODELS, EVALUATING ASSUMPTIONS AND ADEQUACY, AND MEASURING AND REDUCING ERRORS

摘要: SUMMARY Multivariable regression models are powerful tools that used frequently in studies of clinical outcomes. These can use a mixture categorical and continuous variables handle partially observed (censored) responses. However, uncritical application modelling techniques result poorly fit the dataset at hand, or, even more likely, inaccurately predict outcomes on new subjects. One must know how to measure qualities model's order avoid fitted or overfitted models. Measurement predictive accuracy be difficult for survival time data presence censoring. We discuss an easily interpretable index discrimination as well methods assessing calibration predicted probabilities. Both types should unbiasedly validated using bootstrapping cross-validation, before predictions series. some hazards present one strategy avoids many problems discussed. The described applicable all models, but particularly needed binary, ordinal, time-to-event Methods illustrated with analysis prostate cancer Cox regression. Accurate estimation patient prognosis is important reasons. First, prognostic estimates inform about likely her disease. Second, physician guide ordering additional tests selecting appropriate therapies. Third, assessments useful evaluation technologies; derived both without results given test compared incremental information provided by over what prior information.' Fourth, researcher may want estimate effect single factor (for example, treatment given) observational study which uncontrolled confounding factors also measured. Here simultaneous effects controlled (held constant mathematically if model) so interest purely estimated. An (especially ones) affect necessary