Good morning I am an oncology resident and I am struggling with some basic concepts of statistics.
In Oncology, when survival data from randomized trials (comparing the effect of an experimental drug with a control drug) are presented, Kaplan Meier curves are depicted (attached), which show the difference between the survival outcomes of the two groups.
Below the curves, an HR is shown, which allows to estimate the risk of an event for the experimental arm compared to the control arm.
In the section dedicated to the statistical analysis, it is explained that HR has been calculated through a Cox proportional hazard model.
I would have 2 clarifications to ask:
- I thought that Cox regression was not used to compare the effect of two drugs in RCTs but was rather essentially used in observational studies when we want to explore the correlation between a given independent variable and a clinical outcome (making either a univariate or a multivariate analysis in case a correction for the confounding effect of other variables is necessary).
- If cox regression is used in RCTs to compare the efficacy of two treatments (i.e. to calculate the HR), is it a univariate (with “treatment” as the only covariate) or a multivariate cox regression analysis? This is usually not specified in papers.
Best regards
3 Likes
Welcome Federico! I highly recommend @Stephen’s textbook “Statistical Issues in Drug Development”. Here is also a nice explainer on Cox vs log-rank by @f2harrell. From a terminology perspective (further reading here) what you are referring to is “multivariable” (not “multivariate”) and “univariable” (not “univariate”).
You definitely should prioritize use of multivariable Cox regression in RCTs over unadjusted models although there is certainly a common misconception in oncology that the opposite is true. In video format here is an RCT lecture tailored to oncologists. I am sure others here will provide additional educational links I am forgetting.
3 Likes
Welcome Federico! To add to Pavlos’ excellent response, Kaplan-Meier estimates are extremely overused in research. K-M, like simple proportions, was designed for the situation where every patient within a treatment group has the same outcome tendency, i.e., where risk factors do not exist so you have a homogeneous sample. This is never what actually happens. instead of using KM we should be doing two things: reporting estimates from multivariable models that are likely to fit the data generating process and include one or more parameters for treatment, and (2) report adjusted KM curves where no model is assumed for treatment effects but a model is assumed for covariate effects. See this for adjusted KM curve how-to.
2 Likes
