Subgroup identification in a cardiovascular outcome trial

Specialeforsvar ved Nadia Okholm

Titel: Subgroup identification in a cardiovascular outcome trial

Abstract: This thesis examines a subgroup identification problem based on data from a clinical outcome trial conducted by Novo Nordisk. The subgroup identification problem consists of finding potential subgroups in the overall population with an enhanced treatment effect. For this purpose, we use the Subgroup Identification based on Differential Effect Search (SIDES) procedure and conduct the analyses based on an imputed dataset. Three analyses are carried out using the SIDES procedure. The first analysis examines the various results obtained using different versions of the SIDES procedure and investigates the reproducibility of the results obtained from the training dataset on the test dataset. No evidence of an enhanced treatment effect in any of the returned subgroups is obtained on both the training dataset and the test dataset. The second and third analysis rely on integrated features of SIDES controlling the type I error rate, utilise the SIDES procedure including a data-dependent screening rule and are based on the full analysis dataset. The second analysis includes all biomarkers but none of the subgroups returned show a substantial enhanced treatment effect. The third analysis includes only the prognostic biomarkers associated with risk of cardiovascular disease obtained by applying the survival tree algorithm to the full analysis dataset. This analysis indicates that logalbcrea, the log-transformed version of the urinary albumin creatinine ratio, is both prognostic and predictive but further examination indicates that logalbcrea is only prognostic and the apparent predictive ability of logalbcrea is caused by an unfortunate split with no clinical explanation. Therefore, based on the results obtained from the three analyses, the treatment effect in the overall population seems to be rather homogeneous. A simulation study shows that all SIDES procedures deteriorate when introducing weak positive correlation between the predictive biomarkers and the noise biomarkers. Moreover, the simulation study shows that the SIDESbase procedure has difficulties finding the subgroup with the largest treatment effect when the treatment effect is less pronounced.

Vejledere: Helle Sørensen
                   Jesper Madsen, Novo Nordisk
Censor:     Bo Martin Bibby, Aarhus Universitet