I'd like to test the hypothesis that there is a monotonic relationship between two variables, without assuming a specific model. What is the most robust (i.e. lowest probability of type-II error) way to do this?

I can think of a few options:

  • use a linear model of untransformed data. It'll be robust enough, even if I don't think the true relationship is linear.

  • look at rank-transformed data, e.g. with Spearman's rank correlation coefficient

  • use some kind of resampling approach, in which the order of the dependent variable is randomly shuffled. I'm not sure what statistic to compare in this approach.

Is there a fairly standard approach to this problem?

  • $\begingroup$ Based on your suggestions, I'm guessing your data points are independent, and that the null hypothesis you want to test is of no association, i.e. a completely flat relationship. Linear regression with "robust" (i.e. heteroskedasticity-consistent) standard errors provides a Wald test for what you want, that is valid and efficient in large samples. It requires no assumption of true linearity. $\endgroup$ – guest Nov 28 '12 at 18:15

Spearman's or Kendall's correlations are the standard way to do this.


Could you use some sort of generalize additive model, where the dependent variables are relatedto the predictors as a smooth function, like done in the gam() function in R


Your Answer

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

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