Handling missing probe-sets in a linear gene classifier

In multiple myeloma (a type of cancer) linear models (e.g. EMC92) can be used to estimate a patients’ prognosis. In case the model outcome exceeds a specific value t (i.e. a dichotomizing threshold) the subject is classified as highrisk. These linear models consist of a number of covariates, each contributing to the model outcome. In the case of EMC92 or UAMS70, these covariates are probe sets. This vignette deals with the situation that not all covariates are available for generating the model outcome in independent data. The method is based on redistributing the weights of the discarded covariates over the remaining covariates based on the covariance structure in the training data of that model, i.e. a reweighted model.