Testing the proportional hazards assumption in medical survival studies--application to a population-based study of acute myeloid leukaemia

J Epidemiol Biostat. 1999;4(2):83-92.

Abstract

Background: In the analysis of survival data using the Cox proportional hazard model, it is assumed that the magnitude of mortality risk for a predictor variable remains proportional over time. The time-dependent linear model and the piece-wise proportional hazard model (two or four intervals) take into account the variation of the risk over the entire follow-up period.

Method: The three existing models were applied to a series of 266 patients with acute myeloid leukaemia (AML), diagnosed between 1980 and 1992 and recorded by the Registry of Hematopoietic Neoplasms in Côte d'Or, France.

Results: A non-proportional effect of age, period of diagnosis, whether the illness was primary or secondary and French-American-British (FAB) subtype was found significant. In particular, the effect of M2 versus M4-M5 subtypes was revealed by the non-proportional analyses, although this effect was non-significant using the Cox model.

Conclusions: The clinical explanation of the variation of these effects over time is discussed, for example, relating the increase over time of the positive effect of the period of diagnosis to therapeutic improvements. Confirmation of these results on an independent data set is required.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Acute Disease
  • Age Distribution
  • Analysis of Variance
  • Bias
  • Female
  • France / epidemiology
  • Humans
  • Leukemia, Myeloid / classification
  • Leukemia, Myeloid / diagnosis
  • Leukemia, Myeloid / mortality*
  • Linear Models
  • Male
  • Middle Aged
  • Population Surveillance / methods*
  • Proportional Hazards Models*
  • Registries
  • Reproducibility of Results
  • Survival Analysis*
  • Time Factors