I have a dataset that is survival data for 3 different treatments, single time point data (ie: % survived at end of treatment), with an n=10*3 for each treatment. This data is non-parametric, which I verified. So I have analysed difference with a Kruskal Wallis, which has in this instance given me a statistically significant p value. I though job done, not a complicated experiment, got a p value. Lovely....

However, a collaborator from a different discipline is pushing to reanalyse the data with a robust sandwich variance estimator, which I feel is innapropriate -not least because I have never seen such analysis used in similar experiments, but also because this is not-continuous data. Im not very familiar with this test at all, and am also a PhD student going up against a seasoned professor so feel out of my depth!

Would this be an appropriate use of the test, or did I get it right the first time? Any advice would be appreciated.

Many thanks

  • $\begingroup$ A robust sandwich variance estimator would allow you to calculate standard errors in from your ANOVA model if your data is heteroskedastic - Kruskal Wallis doesn't help with this. You need a large sample size for this. The Kruskal Wallis test is appropriate when your outcome variable is non-normal. However, if your sample size is large enough for robust standard errors to be valid, it should be large enough for the central limit theorem to take effect. So non-normality wouldn't be an issue anyway. $\endgroup$ – Great38 Mar 20 '18 at 20:48
  • $\begingroup$ What do you mean by "the data are non-parametric"? $\endgroup$ – StasK Oct 29 '18 at 16:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.