This package implements a correlation structure for the R package nlme for Clarke's maximum likelihood population effects model (Clarke et al. 2002). This is useful, e.g. to construct regressions on distance matrices with nonlinearity and multiple random effects. More generally, this is a model for symmetric, relational data where the row and column effects are random and are integrated from the likelihood.
Initially, this repository housed a hacky script created for a single application (which can still be found here). The old implementation is much slower, does not scale well to large datasets, and is only still extant for the sake of reproducibility. The easiest installation of the current version is via the package devtools;
devtools::install_github("nspope/corMLPE")
Missing pairwise comparisons are allowed (e.g. one does not need a full set of pairwise measurements, as was the case for prior versions of this package), as are multiple observations from the same pair. However, observations that are self comparisons are not allowed.
The corStruct object allows a single grouping factor; for example to model isolation by distance within several species (where there are no pairwise measurements between species), an appropriate model might be
lme(genetic.distance ~ geographic.distance, random = ~geographic.distance|species,
correlation=corMLPE(form=~pop1+pop2|species), data=my.data)
In this case pop1 and pop2 are numerical labels for the populations, such that each observation corresponds to a pair of populations.
With only a single species, use gls,
gls(genetic.distance ~ geographic.distance, correlation=corMLPE(form=~pop1+pop2), data=my.data)
The package currently interfaces with lme, nlme, gls, and gamm (from package mgcv). The motivation for implementing this model in nlme was to leverage pre-existing machinery for nonlinear models, heteroskedasticity, and additive models. Look at Bill Peterman's package ResistanceGA[https://github.com/wpeterman/ResistanceGA] for a nice, lme4-based implementation.
Note that unlike the results presented in Clarke's paper, nlme/lme will return GLS standard errors rather than the OLS standard errors. If OLS standard errors are desired for some reason, see function MLPE() in the package.
You can reach me at nspope@utexas.edu if you have any questions/comments/complaints. Modify as you see fit.
Clarke et al. 2002. Confidence Limits for Regression Relationships between Distance Matrices: Estimating Gene Flow with Distance. Journal of Agricultural, Biological, and Environmental Statistics 7: 361-372.