Comparison of allogeneic stem cell transplantation and non-transplant approaches in elderly patients with advanced myelodysplastic syndrome: optimal statistical approaches and a critical appraisal of clinical results using non-randomized data.

Allogeneic stem cell transplantation (ASCT) from related or unrelated donors may cure patients with myelodysplastic syndromes (MDS), a heterogeneous group of clonal stem cell disorders. We analysed 384 elderly patients (55-69 years) with advanced MDS who received either ASCT (n=247) and were reported to The European Group for Blood and Marrow Transplantation (EBMT) or a non -transplant approach (n=137) reported to the Düsseldorf registry. Besides an attempt to answer the question of "comparison", the purpose of this work is to explain the difficulties in comparing a non-transplant with a transplant cohort, when death before transplant is likely but unknown and the selection of patients for transplant is based on assumptions. It shows which methods are almost always biased and that even the most sophisticated approaches crucially rely on clinical assumptions. Using the most appropriate model for our data, we derive an overall univariate non-significant survival disadvantage for the transplant cohort (HR: 1.29, p = 0.11). We show that such an "average" hazard ratio is however misleading due to non-proportionality of the hazards reflecting early treatment related mortality, the occurring of which is logically correlated with the interval between diagnosis and transplant creating a disproportional drop in the (reconstructed) survival curve of the transplanted patients. Also in multivariate analysis (correcting for age > 60 (HR: 1.4, p = 0.02) and abnormal cytogenetics (HR: 1.46, p = 0.01)), transplantation seems to be worse (HR: 1.39, p = 0.05) but only in the (incorrect but commonly applied) model without time varying covariates. The long term (time depending) hazard ratio is shown to be virtually 1 and overall survival is virtually identical in both groups. Nonetheless no conclusion can be reached from a clinical point of view without assumptions which are by their very nature untestable unless all patients would be followed from diagnosis.