Multiple regression on (dis)similarity matrices

This page is a stub
This page is under construction. If you would like to contribute to this endpoint, please let us know through our user forum!
== incomplete ==

The main idea...

A combination of Mantel correlation and multiple regression, multiple regression on distance matrices (MRM; Manly, 1986; Smouse et al., 1986; Legendre et al., 1994) allows a regression-type analysis of two or more (dis)similarity matrices, using permutations to determine the significance of the coefficients of determination.  One matrix must contain (dis)similarities calculated from response data, such as OTU abundances, and the other matrices must contain (dis)similarities calculated from explanatory data (e.g. environmental parameters or space). MRM has been used as a method to disentangle the influence of space and environmental factors in ecological data (Lichstein, 2006).

Results and interpretation

The results of MRM are largely comparable to the Mantel test and MLR. Users are directed to those endpoints for further information. 

Key assumptions
  • An appropriate (dis)similarity coefficient must be chosen for each data-type analysed.
  • The regression analysis need not be linear. Non-parametric and non-linear methods may be used, however, the assumptions of these methods must also be met.
  • All hypotheses must be in terms of the (dis)similarities between objects (see Legendre 2005, Tuomisto & Ruokolainen, 2006, Legendre et al. 2008).
  • Confidence interval estimation can be severely affected by extreme (dis)similarities (i.e. approaching 0 or 1) between non-identical objects.
  • Be sure to consider whether a raw data approach is more appropriate to your questions (see Legendre 2005, Legendre et al. 2008)
  • An appropriate permutational scheme must be chosen. The (dis)similarity values themselves are interdependent and thus not truly exchangeable. One approach is to permute the objects in the raw response data and, after each permutation, calculate a new (dis)similarity matrix for re-analysis. Permutation of residuals may also be considered.
  • As raised by Manly (1986), the impact of transformations on (dis)similarity matrices can have large effects on regression coefficients and whether the transformation has been 'successful' is difficult to determine.
  • Spatial relationships should be modelled with care: Euclidean, polynomial, or other models are required in different circumstances. For spatial analysis, consider also principal coordinates of neighbour matrices (PCNM).