[Rd] Addendum to wishlist bug report #10931 (factanal) (PR#12754)
ulrich.keller at uni.lu
ulrich.keller at uni.lu
Tue Sep 9 13:55:11 CEST 2008
--=-hiYzUeWcRJ/+kx41aPIZ
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8bit
Hi,
on March 10 I filed a wishlist bug report asking for the inclusion of
some changes to factanal() and the associated print method. The changes
were originally proposed by John Fox in 2005; they make print.factanal()
display factor correlations if factanal() is called with rotation =
"promax". Since I got no replies, and I am really tired of my R-curious
social science colleagues asking "What, it can't even display factor
correlations?!", I made the changes myself. I would be very grateful if
they'd find their way into a release.
I corrected a small error in John Fox's code and made another change
that enables factor correlations not only for promax, but also for the
rotation methods in package GPArotation.
The changes are against R-devel, downloaded on September 9th 2008.
Changes are indicated by comments from John Fox and me. I also changed
factanal.Rd accordingly, this is commented too.
My bug report is at
http://bugs.r-project.org/cgi-bin/R/wishlist?id=10931;user=guest
John Fox's original post is at
http://tolstoy.newcastle.edu.au/R/devel/05/06/1414.html
The changed files factanal.R and factanal.Rd are attached. If there is
anything else I can do to help these changes make it into R, please let
me know.
Thanks and best regards,
Uli
--
Ulrich Keller
Université du Luxembourg
EMACS research unit
B.P. 2
L-7201 Walferdange
Luxembourg
Mail ulrich.keller at uni.lu
Phone +352 46 66 44 9 278
--=-hiYzUeWcRJ/+kx41aPIZ
Content-Disposition: attachment; filename="factanal.R"
Content-Type: text/plain; name="factanal.R"; charset="utf-8"
Content-Transfer-Encoding: 7bit
# File src/library/stats/R/factanal.R
# Part of the R package, http://www.R-project.org
#
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 2 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# A copy of the GNU General Public License is available at
# http://www.r-project.org/Licenses/
## Hmm, MM thinks diag(.) needs checking { diag(vec) when length(vec)==1 !}
## However, MM does not understand that factor analysis
## is a *multi*variate technique!
factanal <-
function (x, factors, data = NULL, covmat = NULL, n.obs = NA,
subset, na.action, start = NULL,
scores = c("none", "regression", "Bartlett"),
rotation = "varimax",
control = NULL, ...)
{
sortLoadings <- function(Lambda)
{
cn <- colnames(Lambda)
Phi <- attr(Lambda, "covariance")
ssq <- apply(Lambda, 2, function(x) -sum(x^2))
Lambda <- Lambda[, order(ssq), drop = FALSE]
colnames(Lambda) <- cn
neg <- colSums(Lambda) < 0
Lambda[, neg] <- -Lambda[, neg]
if(!is.null(Phi)) {
unit <- ifelse(neg, -1, 1)
attr(Lambda, "covariance") <-
unit %*% Phi[order(ssq), order(ssq)] %*% unit
}
Lambda
}
cl <- match.call()
na.act <- NULL
if (is.list(covmat)) {
if (any(is.na(match(c("cov", "n.obs"), names(covmat)))))
stop("'covmat' is not a valid covariance list")
cv <- covmat$cov
n.obs <- covmat$n.obs
have.x <- FALSE
}
else if (is.matrix(covmat)) {
cv <- covmat
have.x <- FALSE
}
else if (is.null(covmat)) {
if(missing(x)) stop("neither 'x' nor 'covmat' supplied")
have.x <- TRUE
if(inherits(x, "formula")) {
## this is not a `standard' model-fitting function,
## so no need to consider contrasts or levels
mt <- terms(x, data = data)
if(attr(mt, "response") > 0)
stop("response not allowed in formula")
attr(mt, "intercept") <- 0
mf <- match.call(expand.dots = FALSE)
names(mf)[names(mf) == "x"] <- "formula"
mf$factors <- mf$covmat <- mf$scores <- mf$start <-
mf$rotation <- mf$control <- mf$... <- NULL
mf[[1]] <- as.name("model.frame")
mf <- eval.parent(mf)
na.act <- attr(mf, "na.action")
if (.check_vars_numeric(mf))
stop("factor analysis applies only to numerical variables")
z <- model.matrix(mt, mf)
} else {
z <- as.matrix(x)
if(!is.numeric(z))
stop("factor analysis applies only to numerical variables")
if(!missing(subset)) z <- z[subset, , drop = FALSE]
}
covmat <- cov.wt(z)
cv <- covmat$cov
n.obs <- covmat$n.obs
}
else stop("'covmat' is of unknown type")
scores <- match.arg(scores)
if(scores != "none" && !have.x)
stop("requested scores without an 'x' matrix")
p <- ncol(cv)
if(p < 3) stop("factor analysis requires at least three variables")
dof <- 0.5 * ((p - factors)^2 - p - factors)
if(dof < 0)
stop(gettextf("%d factors is too many for %d variables", factors, p),
domain = NA)
sds <- sqrt(diag(cv))
cv <- cv/(sds %o% sds)
cn <- list(nstart = 1, trace = FALSE, lower = 0.005)
cn[names(control)] <- control
more <- list(...)[c("nstart", "trace", "lower", "opt", "rotate")]
if(length(more)) cn[names(more)] <- more
if(is.null(start)) {
start <- (1 - 0.5*factors/p)/diag(solve(cv))
if((ns <- cn$nstart) > 1)
start <- cbind(start, matrix(runif(ns-1), p, ns-1, byrow=TRUE))
}
start <- as.matrix(start)
if(nrow(start) != p)
stop(gettextf("'start' must have %d rows", p), domain = NA)
nc <- ncol(start)
if(nc < 1) stop("no starting values supplied")
best <- Inf
for (i in 1:nc) {
nfit <- factanal.fit.mle(cv, factors, start[, i],
max(cn$lower, 0), cn$opt)
if(cn$trace)
cat("start", i, "value:", format(nfit$criteria[1]),
"uniqs:", format(as.vector(round(nfit$uniquenesses, 4))), "\n")
if(nfit$converged && nfit$criteria[1] < best) {
fit <- nfit
best <- fit$criteria[1]
}
}
if(best == Inf) stop("unable to optimize from these starting value(s)")
load <- fit$loadings
if(rotation != "none") {
rot <- do.call(rotation, c(list(load), cn$rotate))
# the following lines modified by J. Fox, 26 June 2005
if (is.list(rot)){
load <- rot$loadings
#the following lines changed by Ulrich Keller, 9 Sept 2008
fit$rotmat <-
if(inherits(rot, "GPArotation")) {
t(solve(rot$Th))
} else {
rot$rotmat
}
#end changes Ulrich Keller, 9 Sept 2008
}
else load <- rot
# end modifications J. Fox, 26 June 2005
}
fit$loadings <- sortLoadings(load)
class(fit$loadings) <- "loadings"
fit$na.action <- na.act # not used currently
if(have.x && scores != "none") {
Lambda <- fit$loadings
zz <- scale(z, TRUE, TRUE)
switch(scores,
regression = {
sc <- zz %*% solve(cv, Lambda)
if(!is.null(Phi <- attr(Lambda, "covariance")))
sc <- sc %*% Phi
},
Bartlett = {
d <- 1/fit$uniquenesses
tmp <- t(Lambda * d)
sc <- t(solve(tmp %*% Lambda, tmp %*% t(zz)))
})
rownames(sc) <- rownames(z)
colnames(sc) <- colnames(Lambda)
if(!is.null(na.act)) sc <- napredict(na.act, sc)
fit$scores <- sc
}
if(!is.na(n.obs) && dof > 0) {
fit$STATISTIC <- (n.obs - 1 - (2 * p + 5)/6 -
(2 * factors)/3) * fit$criteria["objective"]
fit$PVAL <- pchisq(fit$STATISTIC, dof, lower.tail = FALSE)
}
fit$n.obs <- n.obs
fit$call <- cl
fit
}
factanal.fit.mle <-
function(cmat, factors, start=NULL, lower = 0.005, control = NULL, ...)
{
FAout <- function(Psi, S, q)
{
sc <- diag(1/sqrt(Psi))
Sstar <- sc %*% S %*% sc
E <- eigen(Sstar, symmetric = TRUE)
L <- E$vectors[, 1:q, drop = FALSE]
load <- L %*% diag(sqrt(pmax(E$values[1:q] - 1, 0)), q)
diag(sqrt(Psi)) %*% load
}
FAfn <- function(Psi, S, q)
{
sc <- diag(1/sqrt(Psi))
Sstar <- sc %*% S %*% sc
E <- eigen(Sstar, symmetric = TRUE, only.values = TRUE)
e <- E$values[-(1:q)]
e <- sum(log(e) - e) - q + nrow(S)
## print(round(c(Psi, -e), 5)) # for tracing
-e
}
FAgr <- function(Psi, S, q)
{
sc <- diag(1/sqrt(Psi))
Sstar <- sc %*% S %*% sc
E <- eigen(Sstar, symmetric = TRUE)
L <- E$vectors[, 1:q, drop = FALSE]
load <- L %*% diag(sqrt(pmax(E$values[1:q] - 1, 0)), q)
load <- diag(sqrt(Psi)) %*% load
g <- load %*% t(load) + diag(Psi) - S
diag(g)/Psi^2
}
p <- ncol(cmat)
if(is.null(start))
start <- (1 - 0.5*factors/p)/diag(solve(cmat))
res <- optim(start, FAfn, FAgr, method = "L-BFGS-B",
lower = lower, upper = 1,
control = c(list(fnscale=1,
parscale = rep(0.01, length(start))), control),
q = factors, S = cmat)
Lambda <- FAout(res$par, cmat, factors)
dimnames(Lambda) <- list(dimnames(cmat)[[1]],
paste("Factor", 1:factors, sep = ""))
p <- ncol(cmat)
dof <- 0.5 * ((p - factors)^2 - p - factors)
un <- res$par
names(un) <- colnames(cmat)
class(Lambda) <- "loadings"
ans <- list(converged = res$convergence == 0,
loadings = Lambda, uniquenesses = un,
correlation = cmat,
criteria = c(objective = res$value, counts = res$counts),
factors = factors, dof = dof, method = "mle")
class(ans) <- "factanal"
ans
}
print.loadings <- function(x, digits = 3, cutoff = 0.1, sort = FALSE, ...)
{
Lambda <- unclass(x)
p <- nrow(Lambda)
factors <- ncol(Lambda)
if (sort) {
mx <- max.col(abs(Lambda))
ind <- cbind(1:p, mx)
mx[abs(Lambda[ind]) < 0.5] <- factors + 1
Lambda <- Lambda[order(mx, 1:p),]
}
cat("\nLoadings:\n")
fx <- format(round(Lambda, digits))
names(fx) <- NULL
nc <- nchar(fx[1], type="c")
fx[abs(Lambda) < cutoff] <- paste(rep(" ", nc), collapse = "")
print(fx, quote = FALSE, ...)
vx <- colSums(x^2)
varex <- rbind("SS loadings" = vx)
if(is.null(attr(x, "covariance"))) {
varex <- rbind(varex, "Proportion Var" = vx/p)
if(factors > 1)
varex <- rbind(varex, "Cumulative Var" = cumsum(vx/p))
}
cat("\n")
print(round(varex, digits))
invisible(x)
}
print.factanal <- function(x, digits = 3, ...)
{
cat("\nCall:\n", deparse(x$call), "\n\n", sep = "")
cat("Uniquenesses:\n")
print(round(x$uniquenesses, digits), ...)
print(x$loadings, digits = digits, ...)
# the following lines added by J. Fox, 26 June 2005
if (!is.null(x$rotmat)){
tmat <- solve(x$rotmat)
R <- tmat %*% t(tmat)
factors <- x$factors
rownames(R) <- colnames(R) <- paste("Factor", 1:factors, sep="")
# the following line changed by Ulrich Keller, 9 Sept 2008
if (TRUE != all.equal(c(R), c(diag(factors)))){
cat("\nFactor Correlations:\n")
print(R, digits=digits, ...)
}
}
# end additions J. Fox, 23 June 2005
if(!is.null(x$STATISTIC)) {
factors <- x$factors
cat("\nTest of the hypothesis that", factors, if(factors == 1)
"factor is" else "factors are", "sufficient.\n")
cat("The chi square statistic is", round(x$STATISTIC, 2), "on", x$dof,
if(x$dof == 1) "degree" else "degrees",
"of freedom.\nThe p-value is", signif(x$PVAL, 3), "\n")
} else {
cat(paste("\nThe degrees of freedom for the model is",
x$dof, "and the fit was", round(x$criteria["objective"], 4),
"\n"))
}
invisible(x)
}
varimax <- function(x, normalize = TRUE, eps = 1e-5)
{
nc <- ncol(x)
if(nc < 2) return(x)
if(normalize) {
sc <- sqrt(drop(apply(x, 1, function(x) sum(x^2))))
x <- x/sc
}
p <- nrow(x)
TT <- diag(nc)
d <- 0
for(i in 1:1000) {
z <- x %*% TT
B <- t(x) %*% (z^3 - z %*% diag(drop(rep(1, p) %*% z^2))/p)
sB <- La.svd(B)
TT <- sB$u %*% sB$vt
dpast <- d
d <- sum(sB$d)
if(d < dpast * (1 + eps)) break
}
z <- x %*% TT
if(normalize) z <- z * sc
dimnames(z) <- dimnames(x)
class(z) <- "loadings"
list(loadings = z, rotmat = TT)
}
promax <- function(x, m = 4)
{
if(ncol(x) < 2) return(x)
dn <- dimnames(x)
xx <- varimax(x)
x <- xx$loadings
Q <- x * abs(x)^(m-1)
U <- lm.fit(x, Q)$coefficients
d <- diag(solve(t(U) %*% U))
U <- U %*% diag(sqrt(d))
dimnames(U) <- NULL
z <- x %*% U
U <- xx$rotmat %*% U
dimnames(z) <- dn
class(z) <- "loadings"
list(loadings = z, rotmat = U)
}
--=-hiYzUeWcRJ/+kx41aPIZ
Content-Disposition: attachment; filename="factanal.Rd"
Content-Type: text/x-matlab; name="factanal.Rd"; charset="utf-8"
Content-Transfer-Encoding: 8bit
% File src/library/stats/man/factanal.Rd
% Part of the R package, http://www.R-project.org
% Copyright 1995-2007 R Core Development Team
% Distributed under GPL 2 or later
\name{factanal}
\alias{factanal}
%\alias{factanal.fit.mle}
\encoding{latin1}
\title{Factor Analysis}
\description{
Perform maximum-likelihood factor analysis on a covariance matrix or
data matrix.
}
\usage{
factanal(x, factors, data = NULL, covmat = NULL, n.obs = NA,
subset, na.action, start = NULL,
scores = c("none", "regression", "Bartlett"),
rotation = "varimax", control = NULL, \dots)
}
\arguments{
\item{x}{A formula or a numeric matrix or an object that can be
coerced to a numeric matrix.}
\item{factors}{The number of factors to be fitted.}
\item{data}{An optional data frame (or similar: see
\code{\link{model.frame}}), used only if \code{x} is a formula. By
default the variables are taken from \code{environment(formula)}.}
\item{covmat}{A covariance matrix, or a covariance list as returned by
\code{\link{cov.wt}}. Of course, correlation matrices are covariance
matrices.}
\item{n.obs}{The number of observations, used if \code{covmat} is a
covariance matrix.}
\item{subset}{A specification of the cases to be used, if \code{x} is
used as a matrix or formula.}
\item{na.action}{The \code{na.action} to be used if \code{x} is
used as a formula.}
\item{start}{\code{NULL} or a matrix of starting values, each column
giving an initial set of uniquenesses.}
\item{scores}{Type of scores to produce, if any. The default is none,
\code{"regression"} gives Thompson's scores, \code{"Bartlett"} given
Bartlett's weighted least-squares scores. Partial matching allows
these names to be abbreviated.}
\item{rotation}{character. \code{"none"} or the name of a function
to be used to rotate the factors: it will be called with first
argument the loadings matrix, and should return a list with component
\code{loadings} giving the rotated loadings, or just the rotated loadings.}
\item{control}{A list of control values,
\describe{
\item{nstart}{The number of starting values to be tried if
\code{start = NULL}. Default 1.}
\item{trace}{logical. Output tracing information? Default \code{FALSE}.}
\item{lower}{The lower bound for uniquenesses during
optimization. Should be > 0. Default 0.005.}
\item{opt}{A list of control values to be passed to
\code{\link{optim}}'s \code{control} argument.}
\item{rotate}{a list of additional arguments for the rotation function.}
}
}
\item{\dots}{Components of \code{control} can also be supplied as
named arguments to \code{factanal}.}
}
\details{
The factor analysis model is
\deqn{x = \Lambda f + e}{ x = Lambda f + e}
for a \eqn{p}--element row-vector \eqn{x}, a \eqn{p \times k}{p x k}
matrix of \emph{loadings}, a \eqn{k}--element vector of \emph{scores}
and a \eqn{p}--element vector of errors. None of the components
other than \eqn{x} is observed, but the major restriction is that the
scores be uncorrelated and of unit variance, and that the errors be
independent with variances \eqn{\Phi}{Phi}, the
\emph{uniquenesses}. Thus factor analysis is in essence a model for
the covariance matrix of \eqn{x},
\deqn{\Sigma = \Lambda^\prime\Lambda + \Psi}{Sigma = Lambda'Lambda + Psi}
There is still some indeterminacy in the model for it is unchanged
if \eqn{\Lambda}{Lambda} is replaced by \eqn{G\Lambda}{G Lambda} for
any orthogonal matrix \eqn{G}. Such matrices \eqn{G} are known as
\emph{rotations} (although the term is applied also to non-orthogonal
invertible matrices).
If \code{covmat} is supplied it is used. Otherwise \code{x} is used if
it is a matrix, or a formula \code{x} is used with \code{data} to
construct a model matrix, and that is used to construct a covariance
matrix. (It makes no sense for the formula to have a response,
and all the variables must be numeric.) Once a covariance matrix is found or
calculated from \code{x}, it is converted to a correlation matrix for
analysis. The correlation matrix is returned as component
\code{correlation} of the result.
The fit is done by optimizing the log likelihood assuming multivariate
normality over the uniquenesses. (The maximizing loadings for given
uniquenesses can be found analytically: Lawley & Maxwell (1971,
p. 27).) All the starting values supplied in \code{start} are tried
in turn and the best fit obtained is used. If \code{start = NULL}
then the first fit is started at the value suggested by
\enc{Jöreskog}{Joreskog} (1963) and given by Lawley & Maxwell
(1971, p. 31), and then \code{control$nstart - 1} other values are
tried, randomly selected as equal values of the uniquenesses.
The uniquenesses are technically constrained to lie in \eqn{[0, 1]},
but near-zero values are problematical, and the optimization is
done with a lower bound of \code{control$lower}, default 0.005
(Lawley & Maxwell, 1971, p. 32).
Scores can only be produced if a data matrix is supplied and used.
The first method is the regression method of Thomson (1951), the
second the weighted least squares method of Bartlett (1937, 8).
Both are estimates of the unobserved scores \eqn{f}. Thomson's method
regresses (in the population) the unknown \eqn{f} on \eqn{x} to yield
\deqn{\hat f = \Lambda^\prime \Sigma^{-1} x}{hat f = Lambda' Sigma^-1 x}
and then substitutes the sample estimates of the quantities on the
right-hand side. Bartlett's method minimizes the sum of squares of
standardized errors over the choice of \eqn{f}, given (the fitted)
\eqn{\Lambda}{Lambda}.
If \code{x} is a formula then the standard NA-handling is applied to
the scores (if requested): see \code{\link{napredict}}.
}
\value{
An object of class \code{"factanal"} with components
\item{loadings}{A matrix of loadings, one column for each factor. The
factors are ordered in decreasing order of sums of squares of
loadings, and given the sign that will make the sum of the loadings
positive.}
\item{uniquenesses}{The uniquenesses computed.}
\item{correlation}{The correlation matrix used.}
\item{criteria}{The results of the optimization: the value of the
negative log-likelihood and information on the iterations used.}
\item{factors}{The argument \code{factors}.}
\item{dof}{The number of degrees of freedom of the factor analysis model.}
\item{method}{The method: always \code{"mle"}.}
%Modification by Ulrich Keller, Sept 9 2008
\item{rotmat}{The rotation matrix if relevant.}
%End modicication by Ulrich Keller, Sept 9 2008
\item{scores}{If requested, a matrix of scores. \code{napredict} is
applied to handle the treatment of values omitted by the \code{na.action}.}
\item{n.obs}{The number of observations if available, or \code{NA}.}
\item{call}{The matched call.}
\item{na.action}{If relevant.}
\item{STATISTIC, PVAL}{The significance-test statistic and P value, if
if can be computed.}
}
\note{
There are so many variations on factor analysis that it is hard to
compare output from different programs. Further, the optimization in
maximum likelihood factor analysis is hard, and many other examples we
compared had less good fits than produced by this function. In
particular, solutions which are Heywood cases (with one or more
uniquenesses essentially zero) are much often common than most texts
and some other programs would lead one to believe.
}
\references{
Bartlett, M. S. (1937) The statistical conception of mental factors.
\emph{British Journal of Psychology}, \bold{28}, 97--104.
Bartlett, M. S. (1938) Methods of estimating mental
factors. \emph{Nature}, \bold{141}, 609--610.
\enc{Jöreskog}{Joreskog}, K. G. (1963)
\emph{Statistical Estimation in Factor Analysis.} Almqvist and Wicksell.
Lawley, D. N. and Maxwell, A. E. (1971) \emph{Factor Analysis as a
Statistical Method.} Second edition. Butterworths.
Thomson, G. H. (1951) \emph{The Factorial Analysis of Human Ability.}
London University Press.
}
\seealso{
\code{\link{print.loadings}},
\code{\link{varimax}}, \code{\link{princomp}},
\code{\link[datasets]{ability.cov}}, \code{\link[datasets]{Harman23.cor}},
\code{\link[datasets]{Harman74.cor}}
}
\examples{
# A little demonstration, v2 is just v1 with noise,
# and same for v4 vs. v3 and v6 vs. v5
# Last four cases are there to add noise
# and introduce a positive manifold (g factor)
v1 <- c(1,1,1,1,1,1,1,1,1,1,3,3,3,3,3,4,5,6)
v2 <- c(1,2,1,1,1,1,2,1,2,1,3,4,3,3,3,4,6,5)
v3 <- c(3,3,3,3,3,1,1,1,1,1,1,1,1,1,1,5,4,6)
v4 <- c(3,3,4,3,3,1,1,2,1,1,1,1,2,1,1,5,6,4)
v5 <- c(1,1,1,1,1,3,3,3,3,3,1,1,1,1,1,6,4,5)
v6 <- c(1,1,1,2,1,3,3,3,4,3,1,1,1,2,1,6,5,4)
m1 <- cbind(v1,v2,v3,v4,v5,v6)
cor(m1)
factanal(m1, factors=3) # varimax is the default
factanal(m1, factors=3, rotation="promax")
# The following shows the g factor as PC1
prcomp(m1)
## formula interface
factanal(~v1+v2+v3+v4+v5+v6, factors = 3,
scores = "Bartlett")$scores
## a realistic example from Barthlomew (1987, pp. 61-65)
utils::example(ability.cov)
}
\keyword{multivariate}
--=-hiYzUeWcRJ/+kx41aPIZ--
More information about the R-devel
mailing list