Prognostic models: What the statistician wants the clinician to know

Abstract

Prognostic model building is a process that begins much earlier than data analysis and ends later than when a model is reached. It requires careful delineation of a clinical question, methodical planning of the approach and attentive exploration of the data before attempting model building. Once following these important initial steps, the researcher may postulate a model to describe the process of interest and build such model. Once built, the model will need to be checked, validated and the exercise may take the researcher back a few steps - for instance, to adapt the model to fit a variable that displays a 'curved' pattern - to then return to check and validate the model again. To interpret and report the results it is vital to relate the output to the original question, to be transparent in the methodology followed and to understand the limitations of the data and the approach.

Keywords: Cox proportional hazards; Logistic regression; Prognosis; Statistical modelling.