Has there been or is there a consensus about how permutation testing should be done in multiply adjusted regression analyses? I understand the notion of "iteratively permuting the outcome variable" so as to simulate a distribution of the data according to the null hypothesis. In doing this, test statistic's distribution under the null can be obtained and used to obtain a p-value.

Suppose we are fitting the multivariate regression model:

$$E[Y|X, W] = \beta_0 + \beta_1 X + \beta_2 W$$

and we are interested in testing the hypothesis:

$$\mathcal{H}_0: \beta_1 = 0 \quad \text{vs} \quad \mathcal{H}_1: \beta_1 \ne 0$$

If one were to fit the two regression models:

$$E[Y|W] = \beta^*_0 + \beta^*_2 W$$

the $\beta_2$ may be entirely different from $\beta_2^*$ because of the causal relationship between $X$, $W$, and $Y$. So it prompts the question: what is $\beta_2$, conditional mean difference between $W$ and $Y$, supposed to be according to the null hypothesis? I am worried that permuting $Y$ throws the baby out with the bathwater so to speak.

If $W$ confounds the relationship between $X$ and $Y$, in an analysis not adjusting for $X$, $W$ should still have a causal relationship with the outcome of interest (on top of being correlated/causally related to $X$). So this suggests that under the null, $\beta_2$ should at least not be zero even if it is not $\beta_2^*$ per se though that is the most logical assumption.

However, if we randomly permute the labels of $Y$ according to the assumptions of the permutation test, the $\beta_2$ obtained in the permuted distribution is not $\beta_2$ but it's 0.

If $W$ were categorical say binary 0/1, the most logical way to obtain a consistent permutation test would be doing something like permuting $Y$, only within clusters of $W$. (call $Y^{(*)}$ the permuted $Y$) so that $E[Y|W=0] = E[Y^{(*)}|W=0]$ and $E[Y|W=1] = E[Y^{(*)}|W=1]$ and since $\text{cor}(Y^{(*)}, X) \approx 0$, and therefore $\beta_2|\mathcal{H}_0\text{ is true} = \beta_2^*$.

But there is no analogue for continuous $W$.


I think this paper might answer your question: Permutation tests for univariate or multivariate by Marti J. Anderson analysis of variance and regression http://citeseerx.ist.psu.edu/viewdoc/download?doi=

Think about two-way ANOVA and permute the rows. Then think about how to extend to the occasion of continuous variables. This might be an approach.

| cite | improve this answer | |

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.