compar.ou

0th

Percentile

Ornstein--Uhlenbeck Model for Continuous Characters

This function fits an Ornstein--Uhlenbeck model giving a phylogenetic tree, and a continuous character. The user specifies the node(s) where the optimum changes. The parameters are estimated by maximum likelihood; their standard-errors are computed assuming normality of these estimates.

Keywords
models
Usage
compar.ou(x, phy, node = NULL, alpha = NULL)
Arguments
x
a numeric vector giving the values of a continuous character.
phy
an object of class "phylo".
node
a vector giving the number(s) of the node(s) where the parameter `theta' (the trait optimum) is assumed to change. The node(s) can be specified with their labels if phy has node labels. By default there is no change (same optimum
alpha
the value of $\alpha$ to be used when fitting the model. By default, this parameter is estimated (see details).
Details

The Ornstein--Uhlenbeck (OU) process can be seen as a generalization of the Brownian motion process. In the latter, characters are assumed to evolve randomly under a random walk, that is change is equally likely in any direction. In the OU model, change is more likely towards the direction of an optimum (denoted $\theta$) with a strength controlled by a parameter denoted $\alpha$.

The present function fits a model where the optimum parameter $\theta$, is allowed to vary throughout the tree. This is specified with the argument node: $\theta$ changes after each node whose number is given there. Note that the optimum changes only for the lineages which are descendants of this node.

Hansen (1997) recommends to not estimate $\alpha$ together with the other parameters. The present function allows this by giving a numeric value to the argument alpha. By default, this parameter is estimated, but this seems to yield very large standard-errors, thus validating Hansen's recommendation. In practice, a ``poor man estimation'' of $\alpha$ can be done by repeating the function call with different values of alpha, and selecting the one that minimizes the deviance (see Hansen 1997 for an example).

If x has names, its values are matched to the tip labels of phy, otherwise its values are taken to be in the same order than the tip labels of phy.

The user must be careful here since the function requires that both series of names perfectly match, so this operation may fail if there is a typing or syntax error. If both series of names do not match, the values in the x are taken to be in the same order than the tip labels of phy, and a warning message is issued.

Value

  • an object of class "compar.ou" which is list with the following components:
  • deviancethe deviance (= -2 * loglik).
  • paraa data frame with the maximum likelihood estimates and their standard-errors.
  • callthe function call.

Note

The inversion of the variance-covariance matrix in the likelihood function appeared as somehow problematic. The present implementation uses a Cholevski decomposition with the function chol2inv instead of the usual function solve.

References

Hansen, T. F. (1997) Stabilizing selection and the comparative analysis of adaptation. Evolution, 51, 1341--1351.

See Also

ace, compar.lynch, corBrownian, corMartins, pic

Aliases
  • compar.ou
Examples
data(bird.orders)
### This is likely to give you estimates close to 0, 1, and 0
### for alpha, sigma^2, and theta, respectively:
compar.ou(x <- rnorm(23), bird.orders)
### Much better with a fixed alpha:
compar.ou(x, bird.orders, alpha = 0.1)
### Let us 'mimick' the effect of different optima
### for the two clades of birds...
x <- c(rnorm(5, 0), rnorm(18, 5))
### ... the model with two optima:
compar.ou(x, bird.orders, node = 25, alpha = .1)
### ... and the model with a single optimum:
compar.ou(x, bird.orders, node = NULL, alpha = .1)
### => Compare both models with the difference in deviances
##     which follows a chi^2 with df = 1.
Documentation reproduced from package ape, version 2.7-3, License: GPL (>= 2)

Community examples

Looks like there are no examples yet.