Message-ID: <78EDD2EB-9F18-4C8C-9516-A45FB0A3A325@berkeley.edu>
Date: 2010-02-10T22:16:54Z
From: Dan Rabosky
Subject: mantel & missing data
Howdy-
Any thoughts on testing associations between distance matrices with >50% missing data? A standard mantel test is inadequate, as it leads to many pairings between missing & non-missing values. For the particular problem of interest, I can't randomize the raw data used to compute the distance matrices, because it is the distances themselves that are measured (e.g., frequency of hybridization between two species).
What about randomizing the data (not row-column permutations), and using simulations with an identical missing-data structure to assess the Type I error rates? Seems like a crude method to check whether failing to account for within-row or within-column covariances is likely to be problematic, given the level of missing data, but I don't have any other ideas.
Cheers,
~Dan Rabosky