[ape.git] / inst / doc / MoranI.Rnw

\documentclass[a4paper]{article}
%\VignetteIndexEntry{Moran's I}
%\VignettePackage{ape}
\usepackage{fancyvrb}
\usepackage{color}

\newcommand{\code}{\texttt}
\newcommand{\pkg}{\textsf}
\newcommand{\ape}{\pkg{ape}}
\newcommand{\ade}{\pkg{ade4}}
\newcommand{\spatial}{\pkg{spatial}}
\renewcommand{\sp}{\pkg{sp}}

\author{Emmanuel Paradis}
\title{Moran's Autocorrelation Coefficient in Comparative Methods}

\begin{document}

\maketitle

<<echo=false,quiet=true>>=
options(width=60)
@ 

This document clarifies the use of Moran's autocorrelation coefficient
to quantify whether the distribution of a trait among a set of species
is affected or not by their phylogenetic relationships.

\section{Theoretical Background}

Moran's autocorrelation coefficient (often denoted as $I$) is an
extension of Pearson product-moment correlation coefficient to a univariate
series \cite{Cliff1973, Moran1950}. Recall that Pearson's correlation
(denoted as $\rho$) between two variables $x$ and $y$ both of length $n$ is:

\begin{displaymath}
\rho = \frac{\displaystyle\sum_{i=1}^n(x_i - \bar{x})(y_i -
  \bar{y})}{\displaystyle\left[{\sum_{i=1}^n(x_i - \bar{x})^2\sum_{i=1}^n(y_i - \bar{y})^2}\right]^{1/2}},
\end{displaymath}
where $\bar{x}$ and $\bar{y}$ are the sample means of both
variables. $\rho$ measures whether, on average, $x_i$ and $y_i$ are
associated. For a single variable, say $x$, $I$ will
measure whether $x_i$ and $x_j$, with $i\ne j$, are associated. Note
that with $\rho$, $x_i$ and $x_j$ are {\em not} associated since the
pairs $(x_i,y_i)$ are assumed to be independent of each other.

In the study of spatial patterns and processes, we may logically
expect that close observations are more likely to be similar than
those far apart. It is usual to associate a {\em weight} to each
pair $(x_i,x_j)$ which quantifies this \cite{Cliff1981}. In its simplest
form, these weights will take values 1 for close neighbours, and 0
otherwise. We also set $w_{ii}=0$. These weights are sometimes
referred to as a {\em neighbouring function}.

$I$'s formula is:

\begin{equation}
I = \frac{n}{S_0}
\frac{\displaystyle\sum_{i=1}^n \sum_{j=1}^n w_{ij}(x_i - \bar{x})(x_j -
  \bar{x})}{\displaystyle\sum_{i=1}^n (x_i - \bar{x})^2},\label{eq:morani}
\end{equation}
where $w_{ij}$ is the weight between observation $i$ and $j$, and
$S_0$ is the sum of all $w_{ij}$'s:

\begin{displaymath}
S_0 = \sum_{i=1}^n \sum_{j=1}^n w_{ij}.
\end{displaymath}

Quite not so intuitively, the expected value of $I$ under the null
hypothesis of no autocorrelation is not equal to zero but given by
$I_0 = -1/(n-1)$. The expected variance of  $I_0$ is also known, and
so we can make a test of the null hypothesis. If the observed value
of $I$ (denoted $\hat{I}$) is significantly greater than $I_0$, then
values of $x$ are positively autocorrelated, whereas if $\hat{I}<I_0$,
this will indicate negative autocorrelation. This allows us to design
one- or two-tailed tests in the standard way.

Gittleman \& Kot \cite{Gittleman1990} proposed to use Moran's $I$ to
test for ``phylogenetic effects''. They considered two ways to
calculate the weights $w$:

\begin{itemize}
\item With phylogenetic distances among species, e.g., $w_{ij} =
  1/d_{ij}$, where $1/d_{ij}$ are distances measured on a tree.
\item With taxonomic levels where $w_{ij} = 1$ if species $i$ and $j$
  belong to the same group, 0 otherwise.
\end{itemize}

Note that in the first situation, there are quite a lot of
possibilities to set the weights. For instance, Gittleman \& Kot also proposed:

\[\begin{array}{ll}
w_{ij} = 1/d_{ij}^\alpha & \mathrm{if}\ d_{ij} \le c\\
w_{ij} = 0 & \mathrm{if}\ d_{ij} > c,\\
\end{array}\]
where $c$ is a cut-off phylogenetic distance above which the species
are considered to have evolved completely independently, and $\alpha$
is a coefficient (see \cite{Gittleman1990} for details).
By analogy to the use of a spatial correlogram where coefficients are
calculated assuming different sizes of the ``neighbourhood'' and then
plotted to visualize the spatial extent of autocorrelation, they
proposed to calculate $I$ at different taxonomic levels.

\section{Implementation in \ape}

From version 1.2-6, \ape\ has functions \code{Moran.I} and
\code{correlogram.formula} implementing the approach developed by Gittleman \&
Kot. There was an error in the help pages of \code{?Moran.I}
(corrected in ver.\ 2.1) where the weights were referred to as
``distance weights''. This has been wrongly interpreted in my book
\cite[pp.~139--142]{Paradis2006}. The analyses below aim to correct
this.

\subsection{Phylogenetic Distances}

The data, taken from \cite{Cheverud1985}, are the log-transformed
body mass and longevity of five species of primates:

<<>>=
body <- c(4.09434, 3.61092, 2.37024, 2.02815, -1.46968)
longevity <- c(4.74493, 3.3322, 3.3673, 2.89037, 2.30259)
names(body) <- names(longevity) <- c("Homo", "Pongo", "Macaca", "Ateles", "Galago")
@ 

The tree has branch lengths scaled so that the root age is one. We
read the tree with \ape, and plot it:

<<fig=true>>=
library(ape)
trnwk <- "((((Homo:0.21,Pongo:0.21):0.28,Macaca:0.49):0.13,Ateles:0.62)"
trnwk[2] <- ":0.38,Galago:1.00);"
tr <- read.tree(text = trnwk)
plot(tr)
axisPhylo()
@ 

We choose the weights as $w_{ij}=1/d_{ij}$, where the $d$'s is the
distances measured on the tree:

<<>>=
w <- 1/cophenetic(tr)
w
@ 
Of course, we must set the diagonal to zero:

<<>>=
diag(w) <- 0
@ 
We can now perform the analysis with Moran's $I$:

<<>>=
Moran.I(body, w)
@ 

Not surprisingly, the results are opposite to those in
\cite{Paradis2006} since, there, the distances (given by
\code{cophenetic(tr)}) were used as weights. (Note that the argument
\code{dist} has been since renamed \code{weight}.\footnote{The older
  code was actually correct; nevertheless, it has been rewritten, and
  is now much faster. The documentation has been clarified.
  The function \code{correlogram.phylo}, which computed
  Moran's $I$ for a tree given as argument using the distances among
  taxa, has been removed.}) We can now conclude for a slighly
significant positive phylogenetic correlation among body mass values
for these five species.

The new version of \code{Moran.I} gains the option \code{alternative}
which specifies the alternative hypothesis (\code{"two-sided"} by
default, i.e., H$_1$: $I \ne I_0$). As expected from the above result, we divide the $P$-value
be two if we define H$_1$ as $I > I_0$:

<<>>=
Moran.I(body, w, alt = "greater")
@ 

The same analysis with \code{longevity} gives:

<<>>=
Moran.I(longevity, w)
@ 

As for \code{body}, the results are nearly mirrored compared to
\cite{Paradis2006} where a non-significant negative phylogenetic
correlation was found: it is now positive but still largely not
significant.

\subsection{Taxonomic Levels}

The function \code{correlogram.formula} provides an interface to
calculate Moran's $I$ for one or several variables giving a series of
taxonomic levels. An example of its use was provided in
\cite[pp.~141--142]{Paradis2006}. The code of this function has been
simplified, and the graphical presentation of the results have been improved.

\code{correlogram.formula}'s main argument is a formula which is ``sliced'',
and \code{Moran.I} is called for each of these elements. Two things
have been changed for the end-user at this level:

\begin{enumerate}
\item In the old version, the rhs of the formula was given in the
  order of the taxonomic hierarchy: e.g.,
  \code{Order/SuperFamily/Family/Genus}. Not respecting this order
  resulted in an error. In the new version, any order is accepted, but
  the order given it is then respected when plotted the correlogram.
\item Variable transformations (e.g., log) were allowed on the lhs of
  the formula. Because of the simplification of the code, this is no
  more possible. So it is the responsibility of the user to apply any
  tranformation before the analysis.
\end{enumerate}

Following Gittleman \& Kot \cite{Gittleman1990}, the autocorrelation at a higher level
(e.g., family) is calculated among species belonging to the same
category and to different categories at the level below (genus).
To formalize this, let us write the different levels as
$X^1/X^2/X^3/\dots/X^n$ with $X^n$ being the lowest one (\code{Genus} in the
above formula):

\begin{displaymath}
\begin{array}{l}
\left.\begin{array}{ll}
w_{ij}=1 & \mathrm{if}\ X_i^k = X_j^k\ \mathrm{and}\ X_i^{k+1} \ne X_j^{k+1}\\
w_{ij}=0 & \mathrm{otherwise}\\
\end{array} \right\} k < n
\\\\
\left.\begin{array}{ll}
w_{ij}=1 & \mathrm{if}\ X_i^k = X_j^k\\
w_{ij}=0 & \mathrm{otherwise}\\
\end{array} \right\} k = n
\end{array}
\end{displaymath}
This is thus different from the idea of a ``neighbourhood'' of
different sizes, but rather similar to the idea of partial correlation
where the influence of the lowest level is removed when considering
the highest ones \cite{Gittleman1990}.

To repeat the analyses on the \code{carnivora} data set, we first
log$_{10}$-transform the variables mean body mass (\code{SW}) and the
mean female body mass (\code{FW}):

<<>>=
data(carnivora)
carnivora$log10SW <- log10(carnivora$SW)
carnivora$log10FW <- log10(carnivora$FW)
@ 
We first consider a single variable analysis (as in \cite{Paradis2006}):

<<fig=true>>=
fm1.carn <- log10SW ~ Order/SuperFamily/Family/Genus
co1 <- correlogram.formula(fm1.carn, data = carnivora)
plot(co1)
@ 

A legend now appears by default, but can be removed with \code{legend
= FALSE}. Most of the appearance of the graph can be customized via
the option of the plot method (see \code{?plot.correlogram} for
details). This is the same analysis than the one displayed on Fig.~6.3
of \cite{Paradis2006}.

When a single variable is given in the lhs in
\code{correlogram.formula}, an object of class \code{"correlogram"} is
returned as above. If several variables are analysed simultaneously,
the object returned is of class \code{"correlogramList"}, and the
correlograms can be plotted together with the appropriate plot method:

<<fig=true>>=
fm2.carn <- log10SW + log10FW ~ Order/SuperFamily/Family/Genus
co2 <- correlogram.formula(fm2.carn, data = carnivora)
print(plot(co2))
@ 

By default, lattice is used to plot the correlograms on separate
panels; using \code{lattice = FALSE} (actually the second argument,
see \code{?plot.correlogramList}) makes a standard graph superimposing
the different correlograms:

<<fig=true>>=
plot(co2, FALSE)
@ 

The options are roughly the same than above, but do not have always
the same effect since lattice and base graphics do not have the same
graphical parameters. For instance, \code{legend = FALSE} has no
effect if \code{lattice = TRUE}.

\section{Implementation in \ade}

The analysis done with \ade\ in \cite{Paradis2006} suffers from the
same error than the one done with \code{Moran.I} since it was also
done with a distance matrix. So I correct this below:

\begin{Schunk}
\begin{Sinput}
> library(ade4)
> gearymoran(w, data.frame(body, longevity))
\end{Sinput}
\begin{Soutput}
class: krandtest 
Monte-Carlo tests
Call: as.krandtest(sim = matrix(res$result, ncol = nvar, byr = TRUE), 
    obs = res$obs, alter = alter, names = test.names)

Test number:   2 
Permutation number:   999 
Alternative hypothesis: greater 

       Test         Obs   Std.Obs Pvalue
1      body -0.06256789 2.1523342  0.001
2 longevity -0.22990437 0.3461414  0.414

other elements: NULL
\end{Soutput}
\end{Schunk}

The results are wholly consistent with those from \ape, but the
estimated coefficients are substantially different. This is because
the computational methods are not the same in both packages. In \ade,
the weight matrix is first transformed as a relative frequency matrix with
$\tilde{w}_{ij} = w_{ij}/S_0$. The weights are further transformed with:

\begin{displaymath}
p_{ij} = \tilde{w}_{ij} - \sum_{i=1}^n\tilde{w}_{ij}\sum_{j=1}^n\tilde{w}_{ij},
\end{displaymath}
with $p_{ij}$ being the elements of the matrix denoted as $P$. Moran's
$I$ is finally computed with $x^\mathrm{T}Px$. In \ape, the weights
are first row-normalized:

\begin{displaymath}
w_{ij} \Big/ \sum_{i=1}^n w_{ij},
\end{displaymath}
then eq.~\ref{eq:morani} is applied.

Another difference between both packages, though less
important, is that in \ade\ the weight matrix is forced to be
symmetric with $(W+W^\mathrm{T})/2$. In \ape, this matrix is assumed
to be symmetric, which is likely to be the case like in the examples above.

\section{Other Implementations}

Package \sp\ has several functions, including
\code{moran.test}, that are more specifically targeted to the analysis
of spatial data. Package \spatial\ has the function \code{correlogram}
that computes and plots spatial correlograms.

\section*{Acknowledgements}

I am thankful to Thibaut Jombart for clarifications on Moran's $I$.

\bibliographystyle{plain}
\bibliography{ape}

\end{document}