I want to do survival analysis in a situation where I expect the survival time depends on two factors:

  • Environment. Each person is in one of three environments, $E_1,E_2,E_3$. I expect that the survival time for people in $E_3$ will generally be much longer than those in $E_2$, whose survival time will generally be much longer than those in $E_1$.

  • Treatment. Each person is in either the control group (no treatment) or experimental group (they are given the treatment/intervention that we're trying to evaluate).

Both environment and treatment can be fully controlled by the experimenter: they're independent variables. In my experimental design, I have $3 \times 2 = 6$ conditions, one for each combination; each person was randomly assigned to one of the 6 conditions, and then I measured their survival time. My data is right-censored (at the same threshold for all people and all conditions).

I want to test whether the intervention increased survival time. My null hypothesis is that the survival time depends only on environment. The alternative hypothesis is that the survival time depends on both factors and specifically, the intervention increases the survival time in all 3 environments.

How can I test this? Is there a statistical hypothesis test I could use for this purpose?

Some approaches I've considered that don't seem well-suited:

  • If I only had one environment (2 conditions) and didn't have right-censoring, I could use the Mann-Whitney U (Wilcoxon rank sum). However, my data has right-censoring and more than 2 conditions.

  • If my data was normal and not right-censored, maybe ANOVA would be useful. However, my data has right-censoring and doesn't look normal.

  • Because there are 6 groups, I considered using the Kruskal-Wallis test. However, Kruskal-Wallis doesn't handle right-censoring. Also, I have the impression Kruskal-Wallis doesn't really do what I want: it sounds like it compares the null to an alternate hypothesis that at least one of the 6 groups has a higher survival time than the others, rather than that treatment systematically increases survival times in all 3 environments.

  • If I had only 2 conditions (i.e., only one environment), I could use the log-rank test. It sounds like it would be perfect, because it handles right-censored data and is non-parametric. However, I have more than 2 groups, so the log-rank test doesn't seem applicable here.

  • I could ignore the environment, and aggregate the data from the 3 experimental groups into a single group and aggregate the data from the 3 treatment groups into a single group, and then apply the log-rank test. However, I expect this might lose a lot of power, because the environment has a big effect on survival time.

Is there a better approach to this?

Perhaps one could apply log-rank to compare $E_1$-Control to $E_1$-Experimental and get a p-value, say $p_1$; compare $E_2$-Control to $E_2$-Experimental to get a p-value, $p_2$; and similarly a p-value $p_3$ for comparing $E_3$-Control vs $E_3$-Experimental; then multiply these three p-values to get a single number $p_1 p_2 p_3$, which is used as a p-value for the null hypothesis overall. Does this make sense? Is this a sound method?

  • $\begingroup$ Is there some reason why you can't use a Cox proportional hazards survival model? $\endgroup$
    – EdM
    Commented May 6, 2016 at 13:56

1 Answer 1


Your understanding of the Cox proportional hazards model is not quite correct. Yes, the standard Cox model assumes that the hazard ratios are constant over time, but it does not require any particular parametric form of the baseline hazard function over time (e.g., exponential distributions of event times). It uses the data to estimate the baseline hazard, so it's a semi-parametric rather than a fully parametric model. The tools for performing Cox regressions contain tests for the validity of the proportional-hazards assumption, and modifications are available to incorporate time dependencies.

The Cox model thus allows you to perform tests like those you would do with ANOVA or regression in other contexts: determine whether there is a significant effects of the treatment, or whether there is an interaction between environment and treatment. From the hypothesis-testing perspective, it's just another type of regression.

For a start, you might read the series of 4 articles from the British Journal of Cancer in 2003, which are freely available from PubMed Central (Part I, Part II, Part III, and Part IV). The base distribution of R contains the survival package, with both the fundamental tools and some useful vignettes that demonstrate how to use them. I expect that all statistical software packages support Cox regressions. There also are many additional web resources for further study.

  • $\begingroup$ I see what you mean. Thank you! Yup, the proportional hazards assumption will be something I definitely need to validate -- glad to hear that there are ways to test the validity of that. Thanks for the detailed explanation! $\endgroup$
    – D.W.
    Commented May 6, 2016 at 18:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.