%\VignetteEngine{knitr::knitr}
%\VignetteDepends{ggplot2}
%\VignetteDepends{gamm4}
%\VignetteIndexEntry{Fitting Linear Mixed-Effects Models using lme4}
\documentclass[nojss]{jss}
\usepackage[T1]{fontenc}% for correct hyphenation and T1 encoding
\usepackage[utf8]{inputenc}%
\usepackage{lmodern}% latin modern font
\usepackage[american]{babel} %% for texi2dvi ~ bug
\usepackage{bm,amsmath,thumbpdf,amsfonts}%,minted}
\usepackage{blkarray}
\usepackage{array}
%% huxtable-ish stuff
%% \usepackage{adjustbox}
%% \usepackage{threeparttable}
%%
\newcolumntype{P}[1]{>{\raggedright\arraybackslash}p{#1}}
\newcommand{\matindex}[1]{\mbox{\scriptsize#1}}% Matrix index
\newcommand{\github}{Github}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\VEC}{vec}
\newcommand{\bmb}[1]{{\color{red} \emph{#1}}}
\newcommand{\scw}[1]{{\color{blue} \emph{#1}}}
\newcommand{\dmb}[1]{{\color{magenta} \emph{#1}}}
\shortcites{bolker_strategies_2013,sleepstudy,gelman2013bayesian}
\author{Douglas Bates\\University of Wisconsin-Madison\And
Martin M\"achler\\ETH Zurich\And
Benjamin M. Bolker\\McMaster University\And
Steven C. Walker\\McMaster University
}
\Plainauthor{Douglas Bates, Martin M\"achler, Ben Bolker, Steve Walker}
\title{Fitting Linear Mixed-Effects Models Using \pkg{lme4}}
\Plaintitle{Fitting Linear Mixed-Effects Models using lme4}
\Shorttitle{Linear Mixed Models with lme4}
\Abstract{%
Maximum likelihood or restricted maximum likelihood (REML) estimates
of the parameters in linear mixed-effects models can be determined
using the \code{lmer} function in the \pkg{lme4} package for
\proglang{R}. As for most model-fitting functions in \proglang{R},
the model is described in an \code{lmer} call by a formula, in this
case including both fixed- and random-effects
terms. The formula and data together determine a numerical
representation of the model from which the profiled deviance or the
profiled REML criterion can be evaluated as a function of some of
the model parameters. The appropriate criterion is optimized, using
one of the constrained optimization functions in \proglang{R}, to
provide the parameter estimates. We describe the structure of the
model, the steps in evaluating the profiled deviance or REML
criterion, and the structure of classes or types that represents such
a model. Sufficient detail is included to allow specialization of
these structures by users who wish to write functions to fit
specialized linear mixed models, such as models incorporating
pedigrees or smoothing splines, that are not easily expressible in
the formula language used by \code{lmer}.}
\Keywords{%
sparse matrix methods, linear mixed models, penalized least squares,
Cholesky decomposition} \Address{
Douglas Bates\\
Department of Statistics, University of Wisconsin\\
1300 University Ave.\\
Madison, WI 53706, U.S.A.\\
E-mail: \email{bates@stat.wisc.edu}\\
\par\bigskip
Martin M\"achler\\
Seminar f\"ur Statistik, HG G~16\\
ETH Zurich\\
8092 Zurich, Switzerland\\
E-mail: \email{maechler@stat.math.ethz.ch}\\
% URL: \url{http://stat.ethz.ch/people/maechler}\\
\par\bigskip
Benjamin M. Bolker\\
Departments of Mathematics \& Statistics and Biology \\
McMaster University \\
1280 Main Street W \\
Hamilton, ON L8S 4K1, Canada \\
E-mail: \email{bolker@mcmaster.ca}\\
\par\bigskip
Steven C. Walker\\
Department of Mathematics \& Statistics \\
McMaster University \\
1280 Main Street W \\
Hamilton, ON L8S 4K1, Canada \\
E-mail: \email{scwalker@math.mcmaster.ca }
}
\newcommand{\thetavec}{{\bm\theta}}
\newcommand{\betavec}{{\bm\beta}}
\newcommand{\Var}{\operatorname{Var}}
\newcommand{\abs}{\operatorname{abs}}
\newcommand{\bLt}{\ensuremath{\bm\Lambda_{\bm\theta}}}
\newcommand{\mc}[1]{\ensuremath{\mathcal{#1}}}
\newcommand{\trans}{\ensuremath{^\top}} % JSS wants \top
\newcommand{\yobs}{\ensuremath{\bm y_{\mathrm{obs}}}}
\newcommand*{\eq}[1]{eqn.~\ref{#1}}% or just {(\ref{#1})}
<<preliminaries,include=FALSE,cache=FALSE,message=FALSE>>=
options(width=70, show.signif.stars=FALSE,
str=strOptions(strict.width="cut"),
## prefer empty continuation for reader's cut'n'paste:
continue = " ", #JSS: prompt = "R> ", continue = "+ ",
useFancyQuotes = FALSE)
library("knitr")
library("lme4")
library("ggplot2")# Keeping default theme, nicer "on line":
#JSS theme_set(theme_bw())
library("grid")
zmargin <- theme(panel.spacing=unit(0,"lines"))
library("lattice")
library("minqa")
opts_chunk$set(engine='R',dev='pdf', fig.width=9, fig.height=5.5,
prompt=TRUE, cache=TRUE, tidy=FALSE, comment=NA, error = FALSE)
knitr::render_sweave()
@
\setkeys{Gin}{width=\textwidth}
\setkeys{Gin}{height=3.5in}
\begin{document}
A version of this manuscript has been published online in the
\emph{Journal of Statistical Software}, on Oct.\ 2015, with DOI \linebreak[3]
\texttt{10.18637/jss.v067.i01},
see \url{https://www.jstatsoft.org/article/view/v067i01/}.
\section{Introduction}
\label{sec:intro}
The \pkg{lme4} package \citep{lme4} for \proglang{R} \citep{R} provides functions to fit and
analyze linear mixed models, generalized linear mixed models and
nonlinear mixed models. In each of these names, the term ``mixed''
or, more fully, ``mixed effects'', denotes a model that incorporates
both fixed- and random-effects terms in a linear predictor expression
from which the conditional mean of the response can be evaluated. In
this paper we describe the formulation and representation of linear
mixed models. The techniques used for generalized linear and
nonlinear mixed models will be described separately, in a future
paper.
At present, the main alternative to \pkg{lme4} for mixed modeling in
\proglang{R} is the \pkg{nlme} package \citep{nlme_pkg}. The main features
distinguishing \pkg{lme4} from \pkg{nlme} are (1) more efficient
linear algebra tools, giving improved performance on large problems;
(2) simpler syntax and more efficient implementation for fitting
models with crossed random effects; (3) the implementation of profile
likelihood confidence intervals on random-effects parameters; and (4)
the ability to fit generalized linear mixed models (although in this
paper we restrict ourselves to linear mixed models). The main
advantage of \pkg{nlme} relative to \pkg{lme4} is a user interface for
fitting models with structure in the residuals (various forms of
heteroscedasticity and autocorrelation) and in the random-effects
covariance matrices (e.g., compound symmetric models). With some extra
effort, the computational machinery of \pkg{lme4} can be used to fit
structured models that the basic \code{lmer} function cannot handle
(see Appendix~\ref{sec:modularExamples}).
The development of general software for fitting mixed models remains
an active area of research with many open problems. Consequently, the
\pkg{lme4} package has evolved since it was first released, and
continues to improve as we learn more about mixed models. However, we
recognize the need to maintain stability and backward compatibility of
\pkg{lme4} so that it continues to be broadly useful. In order to
maintain stability while continuing to advance mixed-model
computation, we have developed several additional frameworks that draw
on the basic ideas of \pkg{lme4} but modify its structure or
implementation in various ways. These descendants include the
\mbox{\pkg{MixedModels}} package \citep{MixedModels} in \proglang{Julia}
\citep{Julia}, the
\pkg{lme4pureR} package \citep{lme4pureR} in \proglang{R}, and the \pkg{flexLambda}
development branch of \pkg{lme4}. The current article is largely
restricted to describing the current stable version of the \pkg{lme4}
package (1.1-7), with Appendix~\ref{sec:modularExamples} describing
hooks into the computational machinery that are designed for extension
development. The \pkg{gamm4} \citep{gamm4} and \pkg{blme} \citep{blme, blme2}
packages currently make use of these hooks.
Another goal of this article is to contrast the approach used by
\pkg{lme4} with previous formulations of mixed models. The
expressions for the profiled log-likelihood and profiled REML
(restricted maximum likelihood) criteria derived in
Section~\ref{sec:profdev} are similar to those presented in
\citet{bates04:_linear} and, indeed, are closely related to
``Henderson's mixed-model equations''~\citep{henderson_1982}.
Nonetheless there are subtle but important changes in the formulation
of the model and in the structure of the resulting penalized least
squares (PLS) problem to be solved (Section~\ref{sec:PLSpureR}). We
derive the current version of the PLS problem
(Section~\ref{sec:plsMath}) and contrast this result with earlier
formulations (Section~\ref{sec:previous_lmm_form}).
This article is organized into four main sections (Sections~\ref{sec:lFormula},
\ref{sec:mkLmerDevfun}, \ref{sec:optimizeLmer}, and \ref{sec:mkMerMod}),
each of which corresponds to one of the four largely separate modules
that comprise \pkg{lme4}. Before describing the details of each
module, we describe the general form of the linear mixed model
underlying \pkg{lme4} (Section~\ref{sec:LMMs}); introduce the
\code{sleepstudy} data that will be used as an example throughout
(Section~\ref{sec:sleepstudy}); and broadly outline \pkg{lme4}'s
modular structure (Section~\ref{sec:modular}).
\subsection{Linear mixed models}
\label{sec:LMMs}
Just as a linear model is described by the distribution of a
vector-valued random response variable, $\mc{Y}$, whose observed value
is $\yobs$, a linear mixed model is described by the distribution of
two vector-valued random variables: $\mc{Y}$, the response, and
$\mc{B}$, the vector of random effects. In a linear model the
distribution of $\mc Y$ is multivariate normal,%\begin{linenomath}
\begin{equation}
\label{eq:linearmodel}
\mc Y\sim\mc{N}(\bm X\bm\beta+\bm o,\sigma^2\bm W^{-1}),
\end{equation}
where $n$ is the dimension of the response vector, $\bm W$ is a
diagonal matrix of known prior weights, $\bm\beta$ is a
$p$-dimensional coefficient vector, $\bm X$ is an $n\times p$ model
matrix, and $\bm o$ is a vector of known prior offset terms. The
parameters of the model are the coefficients $\bm\beta$ and the scale
parameter $\sigma$.
In a linear mixed model it is the \emph{conditional} distribution of
$\mc Y$ given $\mc B=\bm b$ that has such a form,
\begin{equation}
\label{eq:LMMcondY}
( \mc Y|\mc B=\bm b)\sim\mc{N}(\bm X\bm\beta+\bm Z\bm b+\bm o,\sigma^2\bm W^{-1}), % | <- for ESS
\end{equation}
where $\bm Z$ is the $n\times q$ model matrix for the $q$-dimensional
vector-valued random-effects variable, $\mc B$, whose value we are
fixing at $\bm b$. The unconditional distribution of $\mc B$ is also
multivariate normal with mean zero and a parameterized $q\times q$
variance-covariance matrix, $\bm\Sigma$,
\begin{equation}
\label{eq:LMMuncondB}
\mc B\sim\mc N(\bm0,\bm\Sigma) .
\end{equation}
As a variance-covariance matrix, $\bm\Sigma$ must be positive
semidefinite. It is convenient to express the model in terms of a
\emph{relative covariance factor}, $\bLt$, which is a
$q\times q$ matrix, depending on the \emph{variance-component
parameter}, $\bm\theta$, and generating the symmetric $q\times q$
variance-covariance matrix, $\bm\Sigma$, according to%\begin{linenomath}
\begin{equation}
\label{eq:relcovfac}
\bm\Sigma_{\bm\theta}=\sigma^2\bLt\bLt\trans ,
\end{equation}%\end{linenomath}
where $\sigma$ is the same scale factor as in the conditional
distribution (\ref{eq:LMMcondY}).
Although Equations~\ref{eq:LMMcondY}, \ref{eq:LMMuncondB}, and
\ref{eq:relcovfac} fully describe the class of linear mixed models
that \pkg{lme4} can fit, this terse description
hides many important details. Before moving on to these details, we
make a few observations:
\begin{itemize}
\item This formulation of linear mixed models allows for a relatively
compact expression for the profiled log-likelihood of $\bm\theta$
(Section~\ref{sec:profdev}, Equation~\ref{eq:profiledDeviance}).
\item The matrices associated with random effects, $\bm Z$ and $\bLt$,
typically have a sparse structure with a sparsity pattern that
encodes various model assumptions. Sections~\ref{sec:LMMmatrix} and
\ref{sec:CSCmats} provide details on these structures, and how to
represent them efficiently.
\item The interface provided by \pkg{lme4}'s \code{lmer} function
is slightly less general than the model described by Equations~\ref{eq:LMMcondY}, \ref{eq:LMMuncondB}, and \ref{eq:relcovfac}. To
take advantage of the entire range of possibilities, one may use the
modular functions (Sections~\ref{sec:modular} and
Appendix~\ref{sec:modularExamples}) or explore the
experimental \pkg{flexLambda} branch of \pkg{lme4} on \github.
\end{itemize}
\subsection{Example}
\label{sec:sleepstudy}
Throughout our discussion of \pkg{lme4}, we will work with a data set
on the average reaction time per day for subjects in a sleep
deprivation study \citep{sleepstudy}. On day 0 the subjects had their
normal amount of sleep. Starting that night they were restricted to 3
hours of sleep per night. The response variable, \code{Reaction},
represents average reaction times in milliseconds (ms) on a series of tests given each
\code{Day} to each \code{Subject} (Figure~\ref{fig:sleepPlot}),
%
<<sleep>>=
str(sleepstudy)
@
<<sleepPlot, echo=FALSE, fig.scap="Reaction time versus days by subject", fig.cap="Average reaction time versus days of sleep deprivation by subject. Subjects ordered (from left to right starting on the top row) by increasing slope of subject-specific linear regressions.", fig.align='center', fig.pos='tb'>>=
## BMB: seemed more pleasing to arrange by increasing slope rather than
## intercept ...
xyplot(Reaction ~ Days | Subject, sleepstudy, aspect = "xy",
layout = c(9, 2), type = c("g", "p", "r"),
index.cond = function(x, y) coef(lm(y ~ x))[2],
xlab = "Days of sleep deprivation",
ylab = "Average reaction time (ms)",
as.table = TRUE)
@
% |
Each subject's reaction time increases approximately linearly
with the number of sleep-deprived days.
However, subjects also appear to vary in the slopes and intercepts
of these relationships, which suggests a model with random slopes and intercepts.
As we shall see, such a model may be fitted by minimizing the
REML criterion (Equation~\ref{eq:REMLdeviance}) using
<<outputExample,results="hide">>=
fm1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
@
% |
The estimates of the standard deviations of the random effects for the
intercept and the slope are
\Sexpr{round(sqrt(VarCorr(fm1)$Subject[1,1]), 2)} ms % $
and
\Sexpr{round(sqrt(VarCorr(fm1)$Subject[2,2]), 2)} ms/day. % $
The fixed-effects coefficients, $\betavec$, are
\Sexpr{round(fixef(fm1)[1], 1)} ms and
\Sexpr{round(fixef(fm1)[2], 2)} ms/day
for the intercept and slope. In this model, one interpretation of
these fixed effects is that they are the estimated population mean values of the
random intercept and slope (Section~\ref{sec:intuitiveFormulas}).
We have chosen the \code{sleepstudy} example because it is a
relatively small and simple example to illustrate the theory and
practice underlying \code{lmer}. However, \code{lmer} is capable of
fitting more complex mixed models to larger data sets. For example, we
direct the interested reader to \code{RShowDoc("lmerperf", package =
"lme4")} for examples that more thoroughly exercise the
performance capabilities of \code{lmer}.
\subsection{High-level modular structure}
\label{sec:modular}
The \code{lmer} function is composed of four largely independent
modules. In the first module, a mixed-model formula is parsed and
converted into the inputs required to specify a linear mixed model
(Section~\ref{sec:lFormula}). The second module uses these inputs to
construct an \proglang{R} function which takes the covariance
parameters, $\bm\theta$, as arguments and returns negative twice the log
profiled likelihood or the REML criterion
(Section~\ref{sec:mkLmerDevfun}). The third module optimizes this
objective function to produce maximum likelihood (ML) or REML
estimates of $\bm\theta$ (Section~\ref{sec:optimizeLmer}). Finally,
the fourth module provides utilities for interpreting the optimized
model (Section~\ref{sec:mkMerMod}).
\begin{table}[tb]
\centering
\begin{tabular}{lllp{2.1in}}
\hline
Module & & \proglang{R} function & Description \\
\hline
Formula module & (Section~\ref{sec:lFormula}) & \code{lFormula} &
Accepts a mixed-model formula, data, and other user inputs, and returns
a list of objects required to fit a linear mixed model. \\
Objective function module & (Section~\ref{sec:mkLmerDevfun}) &
\code{mkLmerDevfun} &
Accepts the results of \code{lFormula} and returns a function to
calculate the deviance (or restricted deviance) as a function of the
covariance parameters, $\bm\theta$.\\
Optimization module & (Section~\ref{sec:optimizeLmer}) &
\code{optimizeLmer} &
Accepts a deviance function returned by \code{mkLmerDevfun} and
returns the results of the optimization of that deviance function. \\
Output module & (Section~\ref{sec:mkMerMod}) & \code{mkMerMod} &
Accepts an optimized deviance function and packages the results
into a useful object. \\
\hline
\end{tabular}
\caption{The high-level modular structure of \code{lmer}.}
\label{tab:modular}
\end{table}
To illustrate this modularity, we recreate the \code{fm1} object by a
series of four modular steps; the formula module,
<<modularExampleFormula, tidy=FALSE, eval=FALSE>>=
parsedFormula <- lFormula(formula = Reaction ~ Days + (Days | Subject),
data = sleepstudy)
@
the objective function module,
<<modularExampleObjective, tidy=FALSE, eval=FALSE>>=
devianceFunction <- do.call(mkLmerDevfun, parsedFormula)
@
the optimization module,
<<modularExampleOptimization, tidy=FALSE, eval=FALSE>>=
optimizerOutput <- optimizeLmer(devianceFunction)
@
and the output module,
<<modularExampleOutput, tidy=FALSE, eval=FALSE>>=
mkMerMod( rho = environment(devianceFunction),
opt = optimizerOutput,
reTrms = parsedFormula$reTrms,
fr = parsedFormula$fr)
@
% |
\section{Formula module}
\label{sec:lFormula}
\subsection{Mixed-model formulas}
\label{sec:formulas}
Like most model-fitting functions in \proglang{R},
\code{lmer} takes as its first two arguments a \emph{formula}
specifying the model and the \emph{data} with which to evaluate the
formula. This second argument, \code{data}, is optional but
recommended and is usually the name of an \proglang{R} data frame.
In the \proglang{R} \code{lm} function for fitting linear models,
formulas take the form \verb+resp ~ expr+,
where \code{resp} determines the response variable and \code{expr}
is an expression that specifies the columns of the model matrix.
Formulas for the \code{lmer} function contain special random-effects
terms,
<<MMformula, eval=FALSE>>=
resp ~ FEexpr + (REexpr1 | factor1) + (REexpr2 | factor2) + ...
@
where \code{FEexpr} is an expression determining the columns of the
fixed-effects model matrix, $\bm X$, and the random-effects terms,
\code{(REexpr1 | factor1)} and \code{(REexpr2 | factor2)}, determine
both the random-effects model matrix, $\bm Z$ (Section~\ref{sec:mkZ}),
and the structure of the relative covariance factor,
$\bLt$ (Section~\ref{sec:mkLambdat}). In principle, a
mixed-model formula may contain arbitrarily many random-effects terms,
but in practice the number of such terms is typically low.
\subsection{Understanding mixed-model formulas}
\label{sec:intuitiveFormulas}
Before describing the details of how \pkg{lme4} parses mixed-model
formulas (Section~\ref{sec:LMMmatrix}), we provide an informal
explanation and then some examples. Our discussion assumes familiarity
with the standard \proglang{R} modeling paradigm
\citep{Chambers:1993}.
Each random-effects term is of the form \code{(expr | factor)}.
The expression \code{expr} is evaluated as a linear model formula,
producing a model matrix following the same rules used in standard
\proglang{R} modeling functions (e.g., \code{lm} or \code{glm}). The
expression \code{factor} is evaluated as an \proglang{R} factor. One
way to think about the vertical bar operator is as a special kind of
interaction between the model matrix and the grouping factor. This
interaction ensures that the columns of the model matrix have
different effects for each level of the grouping factor. What makes
this a special kind of interaction is that these effects are modeled
as unobserved random variables, rather than unknown fixed parameters.
Much has been written about important practical and philosophical
differences between these two types of interactions
\citep[e.g., ][]{henderson_1982,gelman2005analysis}. For
example, the random-effects implementation of such interactions can be
used to obtain shrinkage estimates of regression coefficients
\citep[e.g., ][]{1977EfronAndMorris}, or account
for lack of independence in the residuals due to block structure or
repeated measurements \citep[e.g., ][]{laird_ware_1982}.
Table~\ref{tab:formulas} provides several examples of the
right-hand-sides of mixed-model formulas. The first example,
\code{(1 | g)}, % |
is the simplest possible mixed-model formula, where each level of the
grouping factor, \code{g}, has its own random intercept. The mean and
standard deviation of these intercepts are parameters to be estimated.
Our description of this model incorporates any nonzero mean of the
random effects as fixed-effects parameters. If one wishes to specify
that a random intercept has \emph{a priori} known means, one may use
the \code{offset} function as in the second model in
Table~\ref{tab:formulas}. This model contains no fixed effects, or
more accurately the fixed-effects model matrix, $\bm X$, has zero
columns and $\bm\beta$ has length zero.
\begin{table}[tb]
\centering
\begin{tabular}{llP{1.5in}} %% see new column type for ragged right
\hline
Formula & Alternative & Meaning \\
\hline%------------------------------------------------
\code{(1 | g)} & \code{1 + (1 | g)}
& Random intercept with fixed mean. \\
\code{0 + offset(o) + (1 | g)} & \code{-1 + offset(o) + (1 | g)}
& Random intercept with \emph{a priori} means. \\
\code{(1 | g1/g2)} & \code{(1 | g1)+(1 | g1:g2)} % |
& Intercept varying among \code{g1} and \code{g2} within \code{g1}. \\
\code{(1 | g1) + (1 | g2)} & \code{1 + (1 | g1) + (1 | g2)}. &
Intercept varying among \code{g1} and \code{g2}. \\
\code{x + (x | g)} & \code{1 + x + (1 + x | g)} &
Correlated random intercept and slope. \\
\code{x + (x || g)} & \code{1 + x + (1 | g) + (0 +
x | g)}
& Uncorrelated random intercept and slope. \\
\hline
\end{tabular}
\caption{Examples of the right-hand-sides of mixed-effects model
formulas. The names of grouping factors are denoted \code{g},
\code{g1}, and \code{g2}, and covariates and \emph{a priori} known
offsets as \code{x} and \code{o}.}
\label{tab:formulas}
\end{table}
We may also construct models with multiple grouping factors. For
example, if the observations are grouped by \code{g2}, which is nested
within \code{g1}, then the third formula in Table \ref{tab:formulas}
can be used to model variation in the intercept. A common objective
in mixed modeling is to account for such nested (or hierarchical)
structure. However, one of the most useful aspects of \pkg{lme4} is
that it can be used to fit random effects associated with non-nested
grouping factors. For example, suppose the data are grouped by fully
crossing two factors, \code{g1} and \code{g2}, then the fourth formula
in Table \ref{tab:formulas} may be used. Such models are common in
item response theory, where \code{subject} and \code{item} factors are
fully crossed \citep{doran2007estimating}. In addition to varying
intercepts, we may also have varying slopes (e.g., the
\code{sleepstudy} data, Section~\ref{sec:sleepstudy}). The fifth
example in Table~\ref{tab:formulas} gives a model where both the
intercept and slope vary among the levels of the grouping factor.
\subsubsection{Specifying uncorrelated random effects}
\label{sec:uncor}
By default, \pkg{lme4} assumes that all coefficients associated with
the same random-effects term are correlated. To specify an
uncorrelated slope and intercept (for example), one may either use
double-bar notation, \code{(x || g)}, or equivalently use multiple
random-effects terms, \code{x + (1 | g) + (0 + x | g)}, as in the
final example of Table~\ref{tab:formulas}. For example, if one
examined the results of model \code{fm1} of the \code{sleepstudy} data
(Section~\ref{sec:sleepstudy}) using \code{summary(fm1)}, one would
see that the estimated correlation between the slope for \code{Days}
and the intercept is fairly low
(\Sexpr{round(attr(VarCorr(fm1)$Subject, "correlation")[2],3)}) % $
(See Section~\ref{sec:summary} below for more on how to extract the
random-effects covariance matrix.) We may use double-bar notation to
fit a model that excludes a correlation parameter:
<<uncorrelatedModel,results="hide">>=
fm2 <- lmer(Reaction ~ Days + (Days || Subject), sleepstudy)
@
Although mixed models where the random slopes and intercepts are
assumed independent are commonly used to reduce the complexity of
random-slopes models, they do have one subtle drawback. Models in
which the slopes and intercepts are allowed to have a nonzero
correlation (e.g., \code{fm1}) are invariant to additive shifts of the
continuous predictor (\code{Days} in this case). This invariance
breaks down when the correlation is constrained to zero; any shift in
the predictor will necessarily lead to a change in the estimated
correlation, and in the likelihood and predictions of the model. For
example, we can eliminate the correlation in \code{fm1} simply by
adding an amount equal to the ratio of the estimated
among-subject standard deviations multiplied by the estimated
correlation (i.e., $\sigma_{\text{\small slope}}/\sigma_{\text{\small
intercept}} \cdot \rho_{\text{\small slope:intercept}}$) to the \code{Days} variable. The use
of models such as \code{fm2} should ideally be restricted to cases where the
predictor is measured on a ratio scale (i.e., the zero point on the
scale is meaningful, not just a location defined by convenience or
convention), as is the case here.
%% <<uncorrelatedModelByShifting>>=
%% sleepstudyShift <- within(sleepstudy, {
%% Days <- Days + (24.74*0.07)/5.92 })
%% lmer(Reaction ~ Days + (Days | Subject), sleepstudyShift)
%% @
\subsection{Algebraic and computational account of mixed-model formulas}
\label{sec:LMMmatrix}
The fixed-effects terms of a mixed-model formula are parsed to produce
the fixed-effects model matrix, $\bm X$, in the same way that the
\proglang{R} \code{lm} function generates model matrices. However, a
mixed-model formula incorporates $k\ge1$ random-effects terms of the
form \code{(r | f)} as well. % |
These $k$ terms are used to produce the random-effects model matrix,
$\bm Z$ (Equation~\ref{eq:LMMcondY}; Section~\ref{sec:mkZ}), and the
structure of the relative covariance factor, $\bLt$
(Equation~\ref{eq:relcovfac}; Section~\ref{sec:mkLambdat}), which are
matrices that typically have a sparse structure. We now describe how
one might construct these matrices from the random-effects terms,
considering first a single term, \code{(r | f)}, % |
and then generalizing to multiple terms. Tables~\ref{tab:dim} and
\ref{tab:algebraic} summarize the matrices and vectors that determine
the structure of $\bm Z$ and
$\bLt$.
\begin{table}[tb]
\centering
\begin{tabular}{lll}
\hline
Symbol & Size \\
\hline
$n$ & Length of the response vector, $\mc{Y}$ \\
$p$ & Number of columns of fixed-effects model matrix, $\bm X$ \\
$q = \sum_i^k q_i$ & Number of columns of random-effects model matrix, $\bm Z$ \\
$p_i$ & Number of columns of the raw model matrix, $\bm X_i$ \\
$\ell_i$ & Number of levels of the grouping factor indices, $\bm i_i$ \\
$q_i = p_i\ell_i$ & Number of columns of the term-wise model matrix, $\bm Z_i$ \\
$k$ & Number of random-effects terms \\
$m_i = \binom{p_i+1}{2}$ & Number of covariance parameters for term
$i$ \\
$m = \sum_i^k m_i$ & Total number of covariance parameters \\
\hline
\end{tabular}
\caption{Dimensions of linear mixed models. The
subscript $i = 1, \dots, k$ denotes a specific random-effects term.}
\label{tab:dim}
\end{table}
\begin{table}[tb]
\centering
\begin{tabular}{lll}
\hline
Symbol & Size & Description \\
\hline
$\bm X_i$ & $n\times p_i$ & Raw random-effects model matrix \\
$\bm J_i$ & $n\times \ell_i$ & Indicator matrix of grouping factor indices\\
$\bm X_{ij}$ & $p_i\times 1$ & Column vector containing $j$th row of $\bm X_i$ \\
$\bm J_{ij}$ & $\ell_i\times 1$ & Column vector containing $j$th row of $\bm J_i$ \\
$\bm i_i$ & $n$ & Vector of grouping factor indices \\
$\bm Z_i$ & $n\times q_i$ & Term-wise random-effects model matrix \\
$\bm\theta$ & $m$ & Covariance parameters \\
$\bm T_i$ & $p_i\times p_i$ & Lower triangular template matrix \\
$\bm\Lambda_i$ & $q_i\times q_i$ & Term-wise relative covariance factor \\
\hline
\end{tabular}
\caption{Symbols used to describe the structure of the random-effects model matrix and the relative covariance factor. The
subscript $i = 1, \dots, k$ denotes a specific random-effects term.}
\label{tab:algebraic}
\end{table}
The expression, \code{r}, is a linear model formula that evaluates to
an \proglang{R} model matrix, $\bm X_i$, of size $n\times p_i$, called
the \emph{raw random-effects model matrix} for term $i$. A term is
said to be a \emph{scalar} random-effects term when $p_i=1$, otherwise
it is \emph{vector-valued}. For a \emph{simple, scalar}
random-effects term of the form \code{(1 | f)}, $\bm X_i$ is the % |
$n\times 1$ matrix of ones, which implies a random intercept model.
The expression \code{f} evaluates to an \proglang{R} factor, called
the \emph{grouping factor}, for the term. For the $i$th term, we
represent this factor mathematically with a vector $\bm i_i$ of
\emph{factor indices}, which is an $n$-vector of values from
$1,\dots,\ell_i$.\footnote{In practice,
fixed-effects model matrices and random-effects terms are evaluated
with respect to a \emph{model frame}, ensuring that any expressions
for grouping factors have been coerced to factors and any unused
levels of these factors have been dropped. That is, $\ell_i$, the
number of levels in the grouping factor for the $i$th random-effects
term, is well-defined.}
Let $\bm J_i$ be the $n\times \ell_i$ matrix of
indicator columns for $\bm i_i$. Using the \pkg{Matrix} package
\citep{Matrix_pkg} in
\proglang{R}, we may construct the transpose of $\bm J_i$ from a
factor vector, \code{f}, by coercing \code{f} to a `\code{sparseMatrix}'
object. For example,
<<setSeed, echo=FALSE>>=
set.seed(2)
@
<<factorToSparseMatrix>>=
(f <- gl(3, 2))
(Ji <- t(as(f, Class = "sparseMatrix")))
@
When $k>1$ we order the random-effects terms so that
$\ell_1\ge\ell_2\ge\dots\ge\ell_k$; in general, this ordering reduces
``fill-in'' (i.e., the proportion of elements that are zero in the
lower triangle of $\bLt\trans\bm Z\trans\bm W\bm Z\bLt+\bm I$ but not in
the lower triangle of its left Cholesky factor, $\bm L_{\bm\theta}$,
described below in Equation~\ref{eq:blockCholeskyDecomp}). This reduction
in fill-in provides more efficient matrix operations within the
penalized least squares algorithm (Section~\ref{sec:plsMath}).
\subsubsection{Constructing the random-effects model matrix}
\label{sec:mkZ}
The $i$th random-effects term contributes $q_i=\ell_ip_i$ columns to
the model matrix $\bm Z$. We group these columns into a matrix, $\bm
Z_i$, which we refer to as the \emph{term-wise model matrix} for the
$i$th term. Thus $q$, the number of columns in $\bm Z$ and the
dimension of the random variable, $\mc{B}$, is
\begin{equation}
\label{eq:qcalc}
q=\sum_{i=1}^k q_i = \sum_{i=1}^k \ell_i\,p_i .
\end{equation}
Creating the matrix $\bm Z_i$ from $\bm X_i$ and $\bm J_i$ is a
straightforward concept that is, nonetheless, somewhat awkward to
describe. Consider $\bm Z_i$ as being further decomposed into
$\ell_i$ blocks of $p_i$ columns. The rows in the first block are the
rows of $\bm X_i$ multiplied by the 0/1 values in the first column of
$\bm J_i$ and similarly for the subsequent blocks. With these
definitions we may define the term-wise random-effects model matrix,
$\bm Z_i$, for the $i$th term as a transposed Khatri-Rao product,
\begin{equation}
\label{eq:Zi}
\bm Z_i = (\bm J_i\trans * \bm X_i\trans)\trans =
\begin{bmatrix}
\bm J_{i1}\trans \otimes \bm X_{i1}\trans \\
\bm J_{i2}\trans \otimes \bm X_{i2}\trans \\
\vdots \\
\bm J_{in}\trans \otimes \bm X_{in}\trans \\
\end{bmatrix},
\end{equation}
where $*$ and $\otimes$ are the Khatri-Rao\footnote{Note that the
original definition of the Khatri-Rao product is more general than
the definition used in the \pkg{Matrix} package, which is the
definition we use here.} \citep{khatri1968solutions} and Kronecker
products, and $\bm J_{ij}\trans$ and $\bm X_{ij}\trans$ are row
vectors of the $j$th rows of $\bm J_i$ and $\bm X_i$. These rows
correspond to the $j$th sample in the response vector, $\mc Y$, and
thus $j$ runs from $1, \dots, n$. The \pkg{Matrix} package for
\proglang{R} contains a \code{KhatriRao} function, which can be used
to form $\bm Z_i$. For example, if we begin with a raw model matrix,
<<rawModelMatrix>>=
(Xi <- cbind(1, rep.int(c(-1, 1), 3L)))
@
then the term-wise random-effects
model matrix is,
<<KhatriRao>>=
(Zi <- t(KhatriRao(t(Ji), t(Xi))))
@
<<sanityCheck, include=FALSE>>=
## alternative formulation of Zi (eq:Zi)
rbind(
Ji[1,] %x% Xi[1,],
Ji[2,] %x% Xi[2,],
Ji[3,] %x% Xi[3,],
Ji[4,] %x% Xi[4,],
Ji[5,] %x% Xi[5,],
Ji[6,] %x% Xi[6,])
@
In particular, for a simple, scalar term, $\bm Z_i$ is exactly $\bm
J_i$, the matrix of indicator columns. For other scalar terms, $\bm
Z_i$ is formed by element-wise multiplication of the single column of
$\bm X_i$ by each of the columns of $\bm J_i$.
Because each $\bm Z_i$ is generated from indicator columns, its
cross-product, $\bm Z_i\trans\bm Z_i$ is block-diagonal consisting of
$\ell_i$ diagonal blocks each of size $p_i$.\footnote{To see this,
note that by the properties of Kronecker products we may write the
cross-product matrix $Z_i\trans Z_i$ as $\sum_{j=1}^n \bm J_{ij} \bm
J_{ij}\trans \otimes \bm X_{ij} \bm X_{ij}\trans$. Because $\bm
J_{ij}$ is a unit vector along a coordinate axis, the cross-product
$\bm J_{ij} \bm J_{ij}\trans$ is a $p_i\times p_i$ matrix of all
zeros except for a single $1$ along the diagonal. Therefore, the
cross-products, $\bm X_{ij} \bm X_{ij}\trans$, will be added to one
of the $\ell_i$ blocks of size $p_i\times p_i$ along the diagonal of
$Z_i\trans Z_i$.} Note that this means that when $k=1$ (i.e., there
is only one random-effects term, and $\bm Z_i = \bm Z$), $\bm
Z\trans\bm Z$ will be block diagonal. These block-diagonal properties
allow for more efficient sparse matrix computations
(Section~\ref{sec:CSCmats}).
The full random-effects model matrix, $\bm Z$, is constructed from
$k\ge 1$ blocks,
\begin{equation}
\label{eq:Z}
\bm Z =
\begin{bmatrix}
\bm Z_1 &
\bm Z_2 &
\hdots &
\bm Z_k \\
\end{bmatrix}.
\end{equation}
By transposing Equation~\ref{eq:Z} and substituting in Equation~\ref{eq:Zi}, we
may represent the structure of the transposed random-effects model
matrix as follows,
\begin{equation}
\label{eq:Zt}
\bm Z\trans =
\begin{blockarray}{ccccc}
\text{sample 1} & \text{sample 2} & \hdots & \text{sample } n & \\
\begin{block}{[cccc]c}
\bm J_{11} \otimes \bm X_{11} &
\bm J_{12} \otimes \bm X_{12} &
\hdots &
\bm J_{1n} \otimes \bm X_{1n} &
\text{term 1} \\
\bm J_{21} \otimes \bm X_{21} &
\bm J_{22} \otimes \bm X_{22} &
\hdots &
\bm J_{2n} \otimes \bm X_{2n} &
\text{term 2} \\
\vdots & \vdots & \ddots & \vdots & \vdots \\
\end{block}
\end{blockarray}.
\end{equation}
Note that the proportion of elements of $Z\trans$ that are structural
zeros is
\begin{equation}
\label{eq:ZtSparsity}
\frac{\sum_{i=1}^k p_i(\ell_i - 1)}{\sum_{i=1}^k p_i} \qquad .
\end{equation}
Therefore, the sparsity of $\bm Z\trans$ increases with the number of
grouping factor levels. As the number of levels is often large in
practice, it is essential for speed and efficiency to take account of
this sparsity, for example by using sparse matrix methods, when
fitting mixed models (Section~\ref{sec:CSCmats}).
\subsubsection{Constructing the relative covariance factor}
\label{sec:mkLambdat}
The $q\times q$ covariance factor, $\bLt$, is a block diagonal matrix
whose $i$th diagonal block, $\bm\Lambda_i$, is of size
$q_i,i=1,\dots,k$. We refer to $\bm\Lambda_i$ as the \emph{term-wise
relative covariance factor}. Furthermore, $\bm\Lambda_i$ is a
homogeneous block diagonal matrix with each of the $\ell_i$
lower-triangular blocks on the diagonal being a copy of a $p_i\times
p_i$ lower-triangular \emph{template matrix}, $\bm T_i$. The
covariance parameter vector, $\bm\theta$, of length $m_i
=\binom{p_i+1}{2}$, consists of the elements in the lower triangle of
$\bm T_i,i=1,\dots,k$. To provide a unique representation we require
that the diagonal elements of the $\bm T_i,i=1,\dots,k$ be
non-negative.
The template, $\bm T_i$, can be constructed from the number $p_i$
alone. In \proglang{R} code we denote $p_i$ as \code{nc}. For
example, if we set \code{nc <- 3}\Sexpr{nc <- 3}, we could create the
template for term $i$ as,
<<nc, echo=FALSE>>=
nc <- 3
@
%% sequence() is equivalent to unlist(lapply(nvec, seq_len))
%% and (?sequence) ``mainly exists in reverence to the very early history of R''
%% scw: i like sequence, and in fact i never understood why that
%% statement is there in the help file.
<<template>>=
(rowIndices <- rep(1:nc, 1:nc))
(colIndices <- sequence(1:nc))
(template <- sparseMatrix(rowIndices, colIndices,
x = 1 * (rowIndices == colIndices)))
@
Note that the \code{rowIndices} and \code{colIndices} fill the entire
lower triangle, which contains the initial values of the covariance
parameter vector, $\bm\theta$,
<<thetaFromTemplate>>=
(theta <- template@x)
@
<<thetaSanityCheck, include=FALSE>>=
lFormula(Reaction ~ (Days + I(Days^2) | Subject), sleepstudy)$reTrms$theta
@
(because the \code{@x} slot of the sparse matrix \code{template} is a
numeric vector containing the nonzero elements). This template
contains three types of elements: structural zeros (denoted
by~\code{.}), off-diagonal covariance parameters (initialized at
\code{0}), and diagonal variance parameters (initialized at \code{1}).
The next step in the construction of the relative covariance factor is
to repeat the template once for each level of the grouping factor to
construct a sparse block diagonal matrix. For example, if we set the
number of levels, $\ell_i$, to two, \code{nl <- 2}\Sexpr{nl <- 2}, we
could create the transposed term-wise relative covariance factor,
$\bm\Lambda_i\trans$, using the \code{.bdiag} function in the
\pkg{Matrix} package,
<<nl, echo=FALSE>>=
nl <- 2
@
<<relativeCovarianceBlock>>=
(Lambdati <- .bdiag(rep(list(t(template)), nl)))
@
For a model with a single random-effects term, $\bm\Lambda\trans$
would be the initial transposed relative covariance factor
$\bm\Lambda_{\bm\theta}\trans$ itself.
The transposed relative covariance factor, $\bLt\trans$, that arises
from parsing the formula and data is set at the initial value of the
covariance parameters, $\bm\theta$. However, during model fitting, it
needs to be updated to a new $\bm\theta$ value at each iteration (see
Section~\ref{sec:plsI}). This is achieved by constructing a vector of
indices, \code{Lind}, that identifies which elements of \code{theta}
should be placed in which elements of \code{Lambdat},
<<relativeCovarianceBlockIndices>>=
LindTemplate <- rowIndices + nc * (colIndices - 1) - choose(colIndices, 2)
(Lind <- rep(LindTemplate, nl))
@
For example, if we randomly generate a new value for \code{theta},
<<newTheta>>=
thetanew <- round(runif(length(theta)), 1)
@
we may update \code{Lambdat} as follows,
<<relativeCovarianceBlockUpdate>>=
Lambdati@x <- thetanew[Lind]
@
Section~\ref{sec:PLSpureR} describes the process of
updating the relative covariance factor in more detail.
\section{Objective function module}
\label{sec:mkLmerDevfun}
\subsection{Model reformulation for improved computational stability}
\label{sec:stability}
In our initial formulation of the linear mixed model
(Equations~\ref{eq:LMMcondY}, \ref{eq:LMMuncondB}, and \ref{eq:relcovfac}),
the covariance parameter vector, $\bm\theta$, appears only in the
marginal distribution of the random effects (Equation~\ref{eq:LMMuncondB}).
However, from the perspective of computational stability and
efficiency, it is advantageous to reformulate the model such that
$\bm\theta$ appears only in the conditional distribution for the
response vector given the random effects. Such a reformulation allows
us to work with singular covariance matrices, which regularly arise in
practice (e.g., during intermediate steps of the nonlinear optimizer,
Section~\ref{sec:optimizeLmer}).
The reformulation is made by defining a \emph{spherical}%
\footnote{$\mathcal{N}(\bm\mu,\sigma^2\bm I)$ distributions are called
``spherical'' because contours of the probability density are
spheres.} \emph{random-effects} variable, $\mc U$, with
distribution%\begin{linenomath}
\begin{equation}
\label{eq:distU}
\mc U\sim\mathcal{N}(\bm 0,\sigma^2\bm I_q).
\end{equation}%\end{linenomath}
If we set,
% \begin{linenomath}
\begin{equation}
\label{eq:BequalsLamU}
\mc B=\bLt\mc U,
\end{equation}%\end{linenomath}
then $\mc B$ will have the desired $\mathcal{N}(\bm
0,\bm\Sigma_{\bm\theta})$ distribution (Equation~\ref{eq:LMMuncondB}).
Although it may seem more natural to define $\mc U$ in terms of $\mc
B$ we must write the relationship as in Equation~\ref{eq:BequalsLamU} to allow
for singular $\bLt$. The conditional distribution
(Equation~\ref{eq:LMMcondY}) of the response vector given the random
effects may now be reformulated as,%\begin{linenomath}
\begin{equation}
\label{eq:distYgivenU}
(\mc Y|\mc U=\bm u)\sim\mc{N}(\bm\mu_{\mc Y|\mc U=\bm u},\sigma^2\bm W^{-1}),
\end{equation}%\end{linenomath}
where
\begin{equation}
\label{eq:meanYgivenU}
\bm\mu_{\mc Y|\mc U=\bm u} = \bm X \bm\beta + \bm Z\bLt\bm u + \bm o
\end{equation} % |
is a vector of linear predictors, which can be interpreted as a
conditional mean (or mode). Similarly, we also define $\bm\mu_{\mc
U|\mc Y = \yobs}$ as the conditional mean (or mode) of the spherical
random effects given the observed value of the response vector. Note
also that we use the $\bm u$ symbol throughout to represent a specific
value of the random variable, $\mc U$.
\subsection{Penalized least squares}
\label{sec:plsMath}
Our computational methods for maximum likelihood fitting of the linear
mixed model involve repeated applications of the PLS method. In
particular, the PLS problem is to minimize the penalized weighted
residual sum-of-squares,
\begin{equation}
\label{eq:pwrss}
r^2(\bm\theta, \bm\beta, \bm u)
= \rho^2(\bm\theta, \bm\beta, \bm u)
+ \left\| \bm u \right\|^2,
\end{equation}
over $\begin{bmatrix} \bm u \\
\bm\beta \\ \end{bmatrix}$, where,
\begin{equation}
\label{eq:wrss}
\rho^2(\bm\theta, \bm\beta, \bm u) = \left\|
\bm W^{1/2}\left[\yobs - \bm\mu_{\mc Y|\mc U=\bm u} \right]\right\|^2,
\end{equation} %|
is the weighted residual sum-of-squares. This notation makes explicit
the fact that $r^2$ and $\rho^2$ both depend on $\bm\theta$,
$\betavec$, and $\bm u$. The reason for the word ``penalized'' is that
the term, $\left\| \bm u \right\|^2$, in Equation~\ref{eq:pwrss} penalizes models with
larger magnitude values of $\bm u$.
In the so-called ``pseudo-data'' approach we write the penalized
weighted residual sum-of-squares as the squared length of a block
matrix equation,
\begin{equation}
\label{eq:pwrssPseudoData}
r^2(\bm\theta,\bm\beta,\bm u) =
\left\|
\begin{bmatrix}
\bm W^{1/2}(\yobs - \bm o) \\
\bm 0
\end{bmatrix} -
\begin{bmatrix}
\bm W^{1/2}\bm Z\bm\bLt & \bm W^{1/2}\bm X \\
\bm I_q & \bm 0 \\
\end{bmatrix}
\begin{bmatrix}
\bm u \\
\bm\beta \\
\end{bmatrix}
\right\|^2.
\end{equation}
This pseudo-data approach shows that the PLS problem may also be
thought of as a standard least squares problem for an extended
response vector, which implies that the minimizing value,
$\begin{bmatrix}
\bm\mu_{\mc U|\mc Y = \yobs} \\
\widehat{\bm\beta}_{\bm\theta} \\ \end{bmatrix}$, satisfies the
normal equations,
\begin{equation}
\label{eq:normalEquations}
\begin{bmatrix}
\bLt\trans \bm Z\trans \bm W (\yobs - \bm o) \\
\bm X\trans \bm W (\yobs - \bm o) \\
\end{bmatrix} =
\begin{bmatrix}
\bLt\trans \bm Z\trans \bm W \bm Z \bLt + \bm I &
\bLt\trans \bm Z\trans \bm W \bm X \\
\bm X\trans \bm W \bm Z \bLt &
\bm X\trans \bm W \bm X \\
\end{bmatrix}
\begin{bmatrix}
\mu_{\mc U | \mc Y = \yobs} \\ \widehat{\betavec}_{\bm\theta}
\end{bmatrix} ,
\end{equation} % |
where $\mu_{\mc U | \mc Y = \yobs}$ is the conditional mean of $\mc U$
given that $\mc Y = \yobs$. Note that this conditional mean depends
on $\bm\theta$, although we do not make this dependency explicit in
order to reduce notational clutter.
The cross-product matrix in Equation~\ref{eq:normalEquations} can be
Cholesky decomposed,
\begin{equation}
\label{eq:blockCholeskyDecomp}
\begin{bmatrix}
\bLt\trans \bm Z\trans \bm W \bm Z \bLt + \bm I &
\bLt\trans \bm Z\trans \bm W \bm X \\
\bm X\trans \bm W \bm Z \bLt &
\bm X\trans \bm W \bm X \\
\end{bmatrix} =
\begin{bmatrix}
\bm L_{\bm\theta} & \bm 0 \\
\bm R\trans_{ZX} & \bm R\trans_{X} \\
\end{bmatrix}
\begin{bmatrix}
\bm L\trans_{\bm\theta} & \bm R_{ZX} \\
\bm 0 & \bm R_{X}
\end{bmatrix} .
\end{equation}
We may use this decomposition to rewrite the penalized weighted
residual sum-of-squares as,
\begin{equation}
\label{eq:pwrssIdentity}
\begin{aligned}
r^2(\bm\theta,\bm\beta,\bm u) = r^2(\bm\theta) &
+ \left\|\bm L\trans_{\bm\theta} (\bm u - \bm\mu_{\mc U|\mc Y = \yobs})
+ \bm R_{ZX} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
+ \left\|\bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2,
\end{aligned}
\end{equation} % |
where we have simplified notation by writing
$r^2(\bm\theta,\widehat{\bm\beta}_{\bm\theta},\bm\mu_{\mc U|\mc Y =
\yobs})$ as $r^2(\bm\theta)$. This is an important expression in the
theory underlying \pkg{lme4}. It relates the penalized weighted
residual sum-of-squares, $r^2(\bm\theta,\bm\beta,\bm u)$, with its
minimum value, $r^2(\bm\theta)$. This relationship is useful in the
next section where we integrate over the random effects, which is
required for maximum likelihood estimation.
\subsection{Probability densities}
\label{sec:densities}
The residual sums-of-squares discussed in the previous section can be
used to express various probability densities, which are required for
maximum likelihood estimation of the linear mixed model\footnote{These
expressions only technically hold at the observed value, $\yobs$, of
the response vector, $\mc Y$.},
\begin{align}
\label{eq:densityYgivenU}
f_{\mc Y | \mc U}(\yobs|\bm u) & =
\frac{|\bm W|^{1/2}}{(2\pi\sigma^2)^{n/2}}
\exp\left[\frac{-\rho^2(\bm\theta,\bm\beta,\bm u)}{2\sigma^2}\right] ,\\
\label{eq:densityU}
f_{\mc U}(\bm u) &=
\frac{1}{(2\pi\sigma^2)^{q/2}}
\exp\left[\frac{-\left\| \bm u \right\|^2}{2\sigma^2}\right] ,\\
\label{eq:densityYU}
f_{\mc Y , \mc U}(\yobs,\bm u) & =
\frac{|\bm W|^{1/2}}{(2\pi\sigma^2)^{(n+q)/2}}
\exp\left[\frac{-r^2(\bm\theta,\bm\beta,\bm u)}{2\sigma^2}\right] , \\
\label{eq:densityUgivenY}
f_{\mc U | \mc Y}(\bm u | \yobs) & =
\frac{f_{\mc Y , \mc U}(\yobs,\bm u)}{f_{\mc Y}(\yobs)} , \\
\end{align}
where
\begin{equation}
\label{eq:densityY}
f_{\mc Y}(\yobs) =
\int f_{\mc Y , \mc U}(\yobs,\bm u)d\bm u .
\end{equation}
The log-likelihood to be maximized can therefore be expressed as,
\begin{equation}
\label{eq:likelihood}
\mc L(\bm\theta,\bm\beta,\sigma^2 | \yobs) = \log f_{\mc Y}(\yobs).
\end{equation} % |
The integral in Equation~\ref{eq:densityY} may be more explicitly written
as,
\begin{equation}
\label{eq:densityYexpand}
\begin{aligned}
f_{\mc Y}(\yobs) =
\frac{|\bm W|^{1/2}}{(2\pi\sigma^2)^{(n+q)/2}} &
\exp\left[\frac{
-r^2(\bm\theta)
- \left\|\bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
}{2\sigma^2}\right] \\ &
\int \exp\left[\frac{
- \left\|\bm L\trans_{\bm\theta} (\bm u - \bm\mu_{\mc U|\mc Y = \yobs})
+ \bm R_{ZX} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
}{2\sigma^2}\right]
d\bm u ,
\end{aligned}
\end{equation} % |
which can be evaluated with the change of variables,
\begin{equation}
\label{eq:changeOfVariables}
\bm v = \bm L\trans_{\bm\theta} (\bm u - \mu_{\mc U|\mc Y = \yobs})
+ \bm R_{ZX} (\bm\beta - \widehat{\bm\beta}_{\bm\theta}) .
\end{equation} % |
The Jacobian determinant of the transformation from $\bm u$ to $\bm
v$ is $|\bm L_{\bm\theta} |$. Therefore we are able to write the integral
as,
\begin{equation}
\label{eq:densityYintermediate}
\begin{aligned}
f_{\mc Y}(\yobs) =
\frac{|\bm W|^{1/2}}{(2\pi\sigma^2)^{(n+q)/2}} &
\exp\left[\frac{
-r^2(\bm\theta)
- \left\|\bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
}{2\sigma^2}\right] \\ &
\int \exp\left[\frac{
- \left\|\bm v\right\|^2
}{2\sigma^2}\right]
|\bm L_{\bm\theta}|^{-1} d\bm v ,
\end{aligned}
\end{equation}
which by the properties of exponential integrands becomes,
\begin{equation}
\label{eq:densityYfinal}
\exp\mc L(\bm\theta,\bm\beta,\sigma^2 | \yobs) = f_{\mc Y}(\yobs) =
\frac{|\bm W|^{1/2}|\bm L_{\bm\theta}|^{-1}}{(2\pi\sigma^2)^{n/2}}
\exp\left[\frac{
-r^2(\bm\theta)
- \left\|\bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
}{2\sigma^2}\right] .
\end{equation}
\subsection{Evaluating and profiling the deviance and the REML criterion}
\label{sec:profdev}
We are now in a position to understand why the formulation in
Equations~\ref{eq:LMMcondY} and \ref{eq:LMMuncondB} is particularly
useful. We are able to explicitly profile $\betavec$ and $\sigma$ out
of the log-likelihood (Equation~\ref{eq:likelihood}), to find a compact
expression for the profiled deviance (negative twice the profiled
log-likelihood) and the profiled REML criterion as a function of the
relative covariance parameters, $\bm\theta$, only. Furthermore these
criteria can be evaluated quickly and accurately.
To estimate the parameters, $\bm\theta$, $\bm\beta$, and $\sigma^2$, we
minimize negative twice the log-likelihood, which can be written as,
\begin{equation}
\label{eq:fullDeviance}
-2\mc L(\bm\theta,\bm\beta,\sigma^2 | \yobs) = % |
\log\frac{|\bm L_{\bm\theta}|^2}{|\bm W|}
+ n\log(2\pi\sigma^2)
+ \frac{r^2(\bm\theta)}{\sigma^2}
+ \frac{ \left\|\bm R_{X} (\bm\beta -
\widehat{\bm\beta}_{\bm\theta})\right\|^2}{\sigma^2} .
\end{equation}
It is very easy to profile out $\bm\beta$, because it enters into
the ML criterion only through the final term, which is zero if
$\bm\beta = \widehat{\bm\beta}_{\bm\theta}$, where
$\widehat{\bm\beta}_{\bm\theta}$ is found by solving the penalized
least-squares problem in Equation~\ref{eq:pwrssPseudoData}. Therefore we
can write a partially profiled ML criterion as,
\begin{equation}
\label{eq:partiallyProfiledDeviance}
-2\mc L(\bm\theta,\sigma^2 | \yobs) = % |
\log\frac{|\bm L_{\bm\theta}|^2}{|\bm W|}
+ n\log(2\pi\sigma^2)
+ \frac{r^2(\bm\theta)}{\sigma^2} .
\end{equation}
This criterion is only partially profiled because it still depends on
$\sigma^2$. Differentiating this criterion with respect to $\sigma^2$
and setting the result equal to zero yields,
\begin{equation}
\label{eq:residualVarianceEstimatingEquation}
0 = \frac{n}{\widehat{\sigma}_{\bm\theta}^2} -
\frac{r^2(\bm\theta)}{\widehat{\sigma}_{\bm\theta}^4},
\end{equation}
which leads to a maximum profiled likelihood estimate,
\begin{equation}
\label{eq:residualVarianceEstimate}
\widehat{\sigma}_{\bm\theta}^2 = \frac{r^2(\bm\theta)}{n}.
\end{equation}
This estimate can be substituted into the partially profiled criterion to yield the
fully profiled ML criterion,
\begin{equation}
\label{eq:profiledDeviance}
-2\mc L(\bm\theta | \yobs) = % |
\log\frac{|\bm L_{\bm\theta}|^2}{|\bm W|}
+ n\left[1 + \log\left(
\frac{2\pi r^2(\bm\theta)}{n}
\right)\right] .
\end{equation}
This expression for the profiled deviance depends only on
$\bm\theta$. Although $q$, the number of columns in $\bm Z$ and the
size of $\bm\Sigma_{\bm\theta}$, can be very large indeed, the
dimension of $\bm\theta$ is small, frequently less than 10. The
\pkg{lme4} package uses generic nonlinear optimizers
(Section~\ref{sec:optimizeLmer}) to optimize this expression over
$\bm\theta$ to find its maximum likelihood estimate.
\subsubsection{The REML criterion}
\label{sec:reml}
The REML criterion can be obtained by integrating the marginal density
for $\mc Y$ with respect to the fixed effects \citep{laird_ware_1982},
\begin{equation}
\label{eq:REMLintegral}
\int f_{\mc Y}(\yobs)d\bm\beta =
\frac{|\bm W|^{1/2}|\bm L_{\bm\theta}|^{-1}}{(2\pi\sigma^2)^{n/2}}
\exp\left[\frac{-r^2(\bm\theta)}{2\sigma^2}\right]
\int \exp\left[\frac{
- \left\|\bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta})\right\|^2
}{2\sigma^2}\right]d\bm\beta ,
\end{equation}
which can be evaluated with the change of variables,
\begin{equation}
\label{eq:REMLchangeOfVariables}
\bm v = \bm R_{X} (\bm\beta - \widehat{\bm\beta}_{\bm\theta}) .
\end{equation}
The Jacobian determinant of the transformation from $\bm\beta$ to
$\bm v$ is $|\bm R_X |$. Therefore we are able to write the
integral as,
\begin{equation}
\label{eq:REMLintegralIntermediate}
\int f_{\mc Y}(\yobs)d\bm\beta =
\frac{|\bm W|^{1/2}|\bm L_{\bm\theta}|^{-1}}{(2\pi\sigma^2)^{n/2}}
\exp\left[\frac{-r^2(\bm\theta)}{2\sigma^2}\right]
\int \exp\left[\frac{
- \left\|\bm v\right\|^2
}{2\sigma^2}\right]|\bm R_X|^{-1} d\bm v ,
\end{equation}
which simplifies to,
\begin{equation}
\label{eq:REMLintermediate}
\int f_{\mc Y}(\yobs)d\bm\beta =
\frac{|\bm W|^{1/2}|\bm L_{\bm\theta}|^{-1}|\bm R_X|^{-1}}{(2\pi\sigma^2)^{(n-p)/2}}
\exp\left[\frac{-r^2(\bm\theta)}{2\sigma^2}\right] .
\end{equation}
Minus twice the log of this integral is the (unprofiled) REML criterion,
\begin{equation}
\label{eq:REMLdeviance}
-2\mc L_R(\bm\theta,\sigma^2 | \yobs) =
\log\frac{|\bm L_{\bm\theta}|^2|\bm R_X|^2}{|\bm W|}
+ (n-p)\log(2\pi\sigma^2)
+ \frac{r^2(\bm\theta)}{\sigma^2} .
\end{equation}
Note that because $\bm\beta$ gets integrated out, the REML criterion
cannot be used to find a point estimate of $\bm\beta$. However, we
follow others in using the maximum likelihood estimate,
$\widehat{\bm\beta}_{\widehat{\bm\theta}}$, at the optimum value of
$\bm\theta = \widehat{\bm\theta}$. The REML estimate of $\sigma^2$ is,
\begin{equation}
\label{eq:REMLresidualVarianceEstimate}
\widehat{\sigma}_{\bm\theta}^2 = \frac{r^2(\bm\theta)}{n-p} ,
\end{equation}
which leads to the profiled REML criterion,
\begin{equation}
\label{eq:REMLprofiled}
-2\mc L_R(\bm\theta | \yobs) =
\log\frac{|\bm L_{\bm\theta}|^2|\bm R_X|^2}{|\bm W|}
+ (n-p)\left[1 + \log\left(\frac{2\pi
r^2(\bm\theta)}{n-p}\right)\right] .
\end{equation}
\subsection{Changes relative to previous formulations}
\label{sec:previous_lmm_form}
We compare the PLS problem as formulated in Section~\ref{sec:plsMath}
with earlier versions and describe why we use this version. What have
become known as ``Henderson's mixed-model equations'' are given as
Equation 6 of \citet{henderson_1982} and would be expressed as,
\begin{equation}
\label{eq:henderson}
\begin{bmatrix}
\bm X\trans\bm X/\sigma^2& \bm X\trans\bm Z/\sigma^2\\
\bm Z\trans\bm X/\sigma^2& \bm Z\trans\bm Z/\sigma^2 + \bm\Sigma^{-1}
\end{bmatrix}
\begin{bmatrix}
\widehat{\bm\beta}_\theta \\ \bm\mu_{\mc B|\mc Y=\yobs}
\end{bmatrix}=
\begin{bmatrix}\bm X\trans\yobs/\sigma^2\\\bm Z\trans\yobs/\sigma^2
\end{bmatrix} ,
\end{equation}
in our notation (ignoring weights and offsets, without loss of
generality). The matrix written as $\bm R$ in \citet{henderson_1982}
is $\sigma^2\bm I_n$ in our formulation of the model.
\citet{bates04:_linear} modified the PLS equations to
\begin{equation}
\label{eq:batesDebRoy}
\begin{bmatrix}
\bm Z\trans\bm Z+\bm\Omega&\bm Z\trans\bm X\\
\bm X\trans\bm Z&\bm X\trans\bm X
\end{bmatrix}
\begin{bmatrix}
\bm\mu_{\mc B|\mc Y=\yobs}\\\widehat{\bm\beta}_\theta
\end{bmatrix}=
\begin{bmatrix}\bm Z\trans\yobs\\\bm X\trans\yobs\end{bmatrix} \quad .
\end{equation}
where
$\bm\Omega_\theta=\left(\bLt\trans\bLt\right)^{-1}=\sigma^2\bm\Sigma^{-1}$
is the \emph{relative precision matrix} for a given value of
$\bm\theta$. They also showed that the profiled log-likelihood can be
expressed (on the deviance scale) as
\begin{equation}
\label{eq:batesDebRoyprofdev}
-2\mc L(\bm\theta)=
\log\left(\frac{|\bm Z\trans\bm Z+\bm\Omega|}{|\bm\Omega|}\right) +
n\left[1+\log\left(\frac{2\pi r^2_\theta}{n}\right)\right] .
\end{equation}
The primary difference between Equation~\ref{eq:henderson} and
Equation~\ref{eq:batesDebRoy} is the order of the blocks in the system matrix.
The PLS problem can be solved using a Cholesky factor of the system
matrix with the blocks in either order. The advantage of using the
arrangement in Equation~\ref{eq:batesDebRoy} is to allow for evaluation of
the profiled log-likelihood. To evaluate $|\bm Z\trans\bm
Z+\bm\Omega|$ from the Cholesky factor that block must be in the
upper-left corner, not the lower right. Also, $\bm Z$ is sparse whereas $\bm
X$ is usually dense. It is straightforward to exploit the sparsity of
$\bm Z\trans\bm Z$ in the Cholesky factorization when the block containing
this matrix is the first block to be factored. If $\bm X\trans\bm X$ is the first block to be factored, it is much more difficult to
preserve sparsity.
The main change from the formulation in \citet{bates04:_linear} to the
current formulation is the use of a relative covariance factor,
$\bLt$, instead of a relative precision matrix, $\bm\Omega_\theta$,
and solving for the mean of $\mc U|\mc Y = \yobs$ instead of the mean
of $\mc B|\mc Y = \yobs$. This change improves stability, because the
solution to the PLS problem in Section~\ref{sec:plsMath} is
well-defined when $\bLt$ is singular whereas the formulation in
Equation~\ref{eq:batesDebRoy} cannot be used in these cases because
$\bm\Omega_\theta$ does not exist.
It is important to allow for $\bLt$ to be singular because situations
where the parameter estimates, $\widehat{\bm\theta}$, produce a
singular $\bm\Lambda_{\widehat{\theta}}$ do occur in practice. And
even if the parameter estimates do not correspond to a singular
$\bm\Lambda_\theta$, it may be desirable to evaluate the estimation
criterion at such values during the course of the numerical
optimization of the criterion.
\citet{bates04:_linear} also provided expressions for the gradient of
the profiled log-likelihood expressed as
Equation~\ref{eq:batesDebRoyprofdev}. These expressions can be translated
into the current formulation. From Equation~\ref{eq:profiledDeviance} we can see
that (again ignoring weights),
\begin{equation}
\label{eq:grad}
\begin{aligned}
\nabla\left(-2\mc L(\theta)\right)
&=\nabla\log(|\bm L_\theta|^2)+\nabla \left(n\log(r^2(\bm\theta))\right)\\
&=\nabla\log(|\bLt\trans\bm Z\trans\bm Z\bLt+\bm I|)+
n\left(\nabla r^2(\bm\theta)\right)/r^2(\bm\theta)\\
&=\nabla\log(|\bLt\trans\bm Z\trans\bm Z\bLt+\bm I|)+
\left(\nabla r^2(\bm\theta)\right)/(\widehat{\sigma^2}).
\end{aligned}
\end{equation}
The first gradient is easy to express but difficult to evaluate for
the general model. The individual elements of this gradient are
\begin{equation}
\label{eq:graddet}
\begin{aligned}
\frac{\partial\log(|\bLt\trans\bm Z\trans\bm Z\bLt+\bm
I|)}{\partial \theta_i} &=
\tr\left[\frac{\partial\left(\bLt\trans\bm Z\trans\bm
Z\bLt\right)}{\partial\theta_i}
\left(\bLt\trans\bm Z\trans\bm Z\bLt+\bm I\right)^{-1}\right]\\
&=\tr\left[
\left(\bm L_\theta\bm L_\theta\trans\right)^{-1}
\left(
\bLt\trans\bm Z\trans\bm Z\frac{\partial\bLt}{\partial\theta_i} +
\frac{\partial\bLt\trans}{\partial\theta_i}\bm Z\trans\bm Z\bLt
\right)
\right] .
\end{aligned}
\end{equation}
The second gradient term can be expressed as a linear function of the
residual, with individual elements of the form
\begin{equation}
\label{eq:gradr2}
\frac{\partial r^2(\bm\theta)}{\partial\theta_i}=
-2\bm u\trans\frac{\partial\bLt\trans}{\partial\theta_i}\bm
Z\trans\left(\bm y-\bm Z\bLt\bm u-\bm X\widehat{\bm\beta}_{\bm\theta
}\right),
\end{equation}
using the results of \citet{golub_pereyra_1973}. Although we do not
use these results in \pkg{lme4}, they are used for certain model types
in the \pkg{MixedModels} package for \proglang{Julia} and do provide
improved performance.
\subsection{Penalized least squares algorithm}
\label{sec:PLSpureR}
For efficiency, in \pkg{lme4} itself, PLS is implemented in compiled
\proglang{C++} code using the \pkg{Eigen} \citep{Eigen} templated \proglang{C++}
package for numerical linear algebra. Here however, in order to
improve readability we describe a version in pure \proglang{R}.
Section~\ref{sec:CSCmats} provides greater detail on the techniques
and concepts for computational efficiency, which is important in cases
where the nonlinear optimizer (Section~\ref{sec:optimizeLmer})
requires many iterations.
The PLS algorithm takes a vector of covariance parameters,
$\bm\theta$, as inputs and returns the profiled deviance (Equation
\ref{eq:profiledDeviance}) or the REML criterion
(Equation \ref{eq:REMLprofiled}). This PLS algorithm consists of four main
steps:
\begin{enumerate}
\item Update the relative covariance factor (Section~\ref{sec:plsI}).
\item Solve the normal equations (Section~\ref{sec:plsII}).
\item Update the linear predictor and residuals (Section~\ref{sec:plsIII}).
\item Compute and return the profiled deviance (Section~\ref{sec:plsIV}).
\end{enumerate}
PLS also requires the objects described in Table~\ref{tab:inputsPLS},
which define the structure of the model. These objects do not get
updated during the PLS iterations, and so it is useful to store
various matrix products involving them (Table~\ref{tab:constantPLS}).
Table~\ref{tab:updatePLS} lists the objects that do get updated over
the PLS iterations. The symbols in this table correspond to a version
of \pkg{lme4} that is implemented entirely in \proglang{R} (i.e., no
compiled code as in \pkg{lme4} itself). This implementation is called
\pkg{lme4pureR} and is currently available on \github\
(\url{https://github.com/lme4/lme4pureR/}).
\begin{table}[tb]
\centering
\begin{minipage}{\textwidth}
\begin{tabular}{p{1.8in}llp{1.6in}}
\hline
Name/description & Pseudocode & Math & Type \\
\hline
Mapping from covariance parameters to relative covariance factor &
\code{mapping} & & function \\
Response vector & \code{y} & $\yobs$ (Section~\ref{sec:LMMs}) & double vector \\
Fixed-effects model matrix & \code{X} & $\bm X$ (Equation~\ref{eq:LMMcondY}) & double dense\footnote{In previous versions of \pkg{lme4} a sparse $X$ matrix, useful for models with categorical fixed-effect predictors with many levels, could be specified; this feature is not currently available.}
matrix \\
Transposed random-effects model matrix & \code{Zt} & $\bm Z\trans$
(Equation~\ref{eq:LMMcondY}) &
double sparse matrix \\
Square-root weights matrix & \code{sqrtW} & $\bm W^{1/2}$
(Equation~\ref{eq:LMMcondY}) & double diagonal matrix \\
Offset & \code{offset} & $\bm o$ (Equation~\ref{eq:LMMcondY}) & double vector \\
\hline
\end{tabular}
\renewcommand{\footnoterule}{%
\kern -3pt
\hrule width \textwidth height 0pt
\kern 2pt
}
\end{minipage}
\caption{Inputs into a linear mixed model.}
\label{tab:inputsPLS}
\end{table}
\begin{table}[tb]
\centering
\begin{tabular}{ll}
\hline
Pseudocode & Math \\
\hline
\code{ZtW} & $\bm Z\trans\bm W^{1/2}$ \\
\code{ZtWy} & $\bm Z\trans\bm W \yobs$ \\
\code{ZtWX} & $\bm Z\trans\bm W \bm X$ \\
\code{XtWX} & $\bm X\trans\bm W \bm X$ \\
\code{XtWy} & $\bm X\trans\bm W \yobs$ \\
\hline
\end{tabular}
\caption{Constant symbols involved in penalized least squares.}
\label{tab:constantPLS}
\end{table}
\begin{table}[tb]
\centering
\begin{tabular}{p{1.7in}llp{1.5in}}
\hline
Name/description & Pseudocode & Math & Type \\
\hline
Relative covariance factor & \code{lambda} & $\bLt$
(Equation~\ref{eq:relcovfac})
& double sparse lower-triangular matrix \\
Random-effects Cholesky factor & \code{L} & $\bm L_{\bm\theta}$
(Equation~\ref{eq:blockCholeskyDecomp})
& double sparse triangular matrix \\
Intermediate vector in the solution of the normal equations &
\code{cu} & $\bm c_u$ (Equation~\ref{eq:cu}) & double vector \\
Block in the full Cholesky factor & \code{RZX} & $\bm R_{ZX}$
(Equation~\ref{eq:blockCholeskyDecomp}) & double dense matrix \\
Cross-product of the fixed-effects Cholesky factor & \code{RXtRX}
& $\bm R_X\trans\bm R_X$ (Equation~\ref{eq:RX}) & double dense matrix \\
Fixed-effects coefficients & \code{beta} & $\bm\beta$
(Equation~\ref{eq:LMMcondY}) & double vector \\
Spherical conditional modes & \code{u} & $\bm u$ (Section~\ref{sec:stability}) & double vector \\
Non-spherical conditional modes & \code{b}
& $\bm b$ (Equation~\ref{eq:LMMcondY}) & double vector \\
Linear predictor & \code{mu} & $\bm\mu_{\mc Y|\mc U=\bm u}$
(Equation~\ref{eq:meanYgivenU}) & double vector \\
Weighted residuals & \code{wtres} & $\bm W^{1/2} (\yobs - \mu)$ &
double vector \\
Penalized weighted residual sum-of-squares & \code{pwrss} &
$r^2(\bm\theta)$ (Equation~\ref{eq:pwrssIdentity}) &
double \\
Twice the log determinant random-effects Cholesky factor &
\code{logDet}
& $\log|\bm L_{\bm\theta}|^2$ & double \\
\hline
\end{tabular}
\caption{Quantities updated during an evaluation of the linear mixed model
objective function.}
\label{tab:updatePLS}
\end{table}
\subsubsection{PLS step I: Update relative covariance factor}
\label{sec:plsI}
The first step of PLS is to update the relative covariance factor,
$\bLt$, from the current value of the covariance parameter vector,
$\bm\theta$. The updated $\bLt$ is then used to update the random-effects
Cholesky factor, $\bm L_{\bm\theta}$
(Equation~\ref{eq:blockCholeskyDecomp}). The mapping from the covariance
parameters to the relative covariance factor can take many forms, but
in general involves a function that takes $\bm\theta$ into the values
of the nonzero elements of $\bLt$.
If $\bLt$ is stored as an object of class `\code{dgCMatrix}' from the
\pkg{Matrix} package for \proglang{R}, then we may update $\bLt$ from
$\bm\theta$ by,
<<PLSupdateTheta, eval = FALSE>>=
Lambdat@x[] <- mapping(theta)
@
where \code{mapping} is an \proglang{R} function that both accepts and
returns a numeric vector. The nonzero elements of sparse matrix
classes in \pkg{Matrix} are stored in a slot called \code{x}.
In the current version of \pkg{lme4} (v. 1.1-7) the mapping from
$\bm\theta$ to $\bLt$ is represented as an \proglang{R} integer
vector, \code{Lind}, of indices, so that
<<PLSupdateThetaWithLind>>=
mapping <- function(theta) theta[Lind]
@
The index vector \code{Lind} is computed during the model setup and
stored in the function's environment. Continuing the example from
Section~\ref{sec:mkLambdat}, consider a new value for \code{theta},
<<exampleNewTheta>>=
thetanew <- c(1, -0.1, 2, 0.1, -0.2, 3)
@
To put these values in the appropriate elements in \code{Lambdati},
we use \code{mapping},
<<exampleLindUpdate>>=
Lambdati@x[] <- mapping(thetanew)
Lambdati
@ %
This \code{Lind} approach can be useful for extending the capabilities
of \pkg{lme4} by using the modular approach to fitting mixed models.
For example, Appendix~\ref{sec:zeroSlopeSlope} shows how to use
\code{Lind} to fit a model where two random slopes are uncorrelated,
but both slopes are correlated with an intercept.
The mapping from the covariance parameters to the relative covariance
factor is treated differently in other implementations of the
\code{lme4} approach to linear mixed models. At the other extreme, the
\code{flexLambda} branch of \pkg{lme4} and the \pkg{lme4pureR} package
provide the capabilities for a completely general mapping. This added
flexibility has the advantage of allowing a much wider variety of
models (e.g., compound symmetry, auto-regression). However, the
disadvantage of this approach is that it becomes possible to fit a
much wider variety of ill-posed models. Finally, if one would like
such added flexibility with the current stable version of \pkg{lme4},
it is always possible to use the modular approach to wrap the
\code{Lind}-based deviance function in a general mapping function
taking a parameter to be optimized, say $\phi$, into $\bm\theta$.
However, this approach is likely to be inefficient in many cases.
The \code{update} method from the \pkg{Matrix} package efficiently
updates the random-effects Cholesky factor, $\bm L_{\bm\theta}$, from a
new value of $\bm\theta$ and the updated $\bLt$.
<<updateL, eval=FALSE>>=
L <- update(L, Lambdat %*% ZtW, mult = 1)
@
The \code{mult = 1} argument corresponds to the addition of the
identity matrix to the upper-left block on the left-hand-side of
Equation~\ref{eq:blockCholeskyDecomp}.
\subsubsection{PLS step II: solve normal equations}
\label{sec:plsII}
With the new covariance parameters installed in $\bLt$, the next step
is to solve the normal equations (Equation~\ref{eq:normalEquations}) for the
current estimate of the fixed-effects coefficients,
$\widehat{\betavec}_{\bm\theta}$, and the conditional mode, $\mu_{\mc
U|\mc Y = \yobs}$. We solve these equations using a sparse Cholesky
factorization (Equation~\ref{eq:blockCholeskyDecomp}). In a complex model
fit to a large data set, the dominant calculation in the evaluation of
the profiled deviance (Equation~\ref{eq:profiledDeviance}) or REML criterion
(Equation~\ref{eq:REMLprofiled}) is this sparse Cholesky factorization
(Equation~\ref{eq:blockCholeskyDecomp}). The factorization is performed in two
phases; a symbolic phase and a numeric phase. The symbolic phase, in
which the fill-reducing permutation $\bm P$ is determined along with
the positions of the nonzeros in $\bm L_{\bm\theta}$, does not depend
on the value of $\bm\theta$. It only depends on the positions of the
nonzeros in $\bm Z\bLt$. The numeric phase uses $\bm\theta$ to
determine the numeric values of the nonzeros in $\bm
L_{\bm\theta}$. Using this factorization, the solution proceeds by the
following steps,
\begin{align}
\bm L_{\bm\theta}\bm c_u&=\bm P\bLt\trans\bm Z\trans\bm W\bm y\label{eq:cu}\\
\bm L_{\bm\theta}\bm R_{ZX}&=\bm P\bLt\trans\bm Z\trans\bm W\bm X\label{eq:RZX}\\
\bm R_X\trans\bm R_X&=\bm X\trans\bm W\bm X-\bm R_{ZX}\trans\bm R_{ZX}\label{eq:RX}\\
\left(\bm R_X\trans\bm R_X\right)\widehat{\bm\beta}_{\bm\theta}&=
\bm X\trans\bm W\bm y-\bm R_{ZX}\bm c_u\label{eq:betahat}\\
\bm L\trans\bm_{\bm\theta} \bm P\bm u&=\bm c_u-\bm R_{ZX}\widehat{\bm\beta}_{\bm\theta}\label{eq:tildeu}
\end{align}
which can be solved using the \pkg{Matrix} package with,
<<PLSsolve, eval = FALSE>>=
cu[] <- as.vector(solve(L, solve(L, Lambdat %*% ZtWy,
system = "P"), system = "L"))
RZX[] <- as.vector(solve(L, solve(L, Lambdat %*% ZtWX,
system = "P"), system = "L"))
RXtRX <- as(XtWX - crossprod(RZX), "dpoMatrix")
beta[] <- as.vector(solve(RXtRX, XtWy - crossprod(RZX, cu)))
u[] <- as.vector(solve(L, solve(L, cu - RZX %*% beta,
system = "Lt"), system = "Pt"))
@
Notice the nested calls to \code{solve}. The inner calls of the first
two assignments determine and apply the permutation matrix
(\code{system = "P"}), whereas the outer calls actually solve the
equation (\code{system = "L"}). In the assignment to \code{u[]}, the
nesting is reversed in order to return to the original permutation.
\subsubsection{PLS step III: Update linear predictor and residuals}
\label{sec:plsIII}
The next step is to compute the linear predictor, $\mu_{\mc Y| \mc U}$
(Equation \ref{eq:meanYgivenU}), and the weighted residuals with new values for
$\widehat{\betavec}_\theta$ and $\mu_{\mc B | \mc Y = \yobs}$. In
\pkg{lme4pureR} these quantities can be computed as,
<<PLSupdateResp, eval = FALSE>>=
b[] <- as.vector(crossprod(Lambdat, u))
mu[] <- as.vector(crossprod(Zt, b) + X %*% beta + offset)
wtres <- sqrtW * (y - mu)
@
where \code{b} represents the current estimate of $\mu_{\mc B | \mc Y
= \yobs}$.
\subsubsection{PLS step IV: Compute profiled deviance}
\label{sec:plsIV}
Finally, the updated linear predictor and weighted residuals can be
used to compute the profiled deviance (or REML criterion), which in
\pkg{lme4pureR} proceeds as,
<<PLScalculateObjective, eval = FALSE>>=
pwrss <- sum(wtres^2) + sum(u^2)
logDet <- 2*determinant(L, logarithm = TRUE)$modulus
if (REML) logDet <- logDet + determinant(RXtRX,
logarithm = TRUE)$modulus
attributes(logDet) <- NULL
profDev <- logDet + degFree * (1 + log(2 * pi * pwrss) - log(degFree))
@
The profiled deviance consists of three components: (1)
log-determinant(s) of Cholesky factorization (\code{logDet}), (2) the
degrees of freedom (\code{degFree}), and the penalized weighted
residual sum-of-squares (\code{pwrss}).
\subsection{Sparse matrix methods}
\label{sec:CSCmats}
In fitting linear mixed models, an instance of the PLS problem
(\ref{eq:normalEquations}) must be solved at each evaluation of the
objective function during the optimization
(Section~\ref{sec:optimizeLmer}) with respect to $\bm\theta$. Because
this operation must be performed many times it is worthwhile
considering how to provide effective evaluation methods for objects
and calculations involving the sparse matrices associated with
random-effects terms (Sections~\ref{sec:LMMmatrix}).
The \pkg{CHOLMOD} library of \proglang{C} functions
\citep{Chen:2008:ACS:1391989.1391995}, on which the sparse matrix
capabilities of the \pkg{Matrix} package for \proglang{R} and the
sparse Cholesky factorization in \proglang{Julia} are based, allows
for separation of the symbolic and numeric phases. Thus we perform
the symbolic phase as part of establishing the structure representing
the model (Section~\ref{sec:lFormula}). Furthermore, because \pkg{CHOLMOD}
functions allow for updating $\bm L_{\bm\theta}$ directly from the
matrix $\bLt\trans\bm Z\trans$ without actually forming $\bLt\trans\bm
Z\trans\bm Z\bLt+\bm I$ we generate and store $\bm Z\trans$ instead of
$\bm Z$ (note that we have ignored the weights matrix, $\bm W$, for
simplicity). We can update $\bLt\trans\bm Z\trans$ directly from
$\bm\theta$ without forming $\bLt$ and multiplying two sparse
matrices. Although such a direct approach is used in the
\pkg{MixedModels} package for \proglang{Julia}, in \pkg{lme4} we first
update $\bLt\trans$ then form the sparse product $\bLt\trans\bm
Z\trans$. A third alternative, which we employ in \pkg{lme4pureR}, is
to compute and save the cross-products of the model matrices, $\bm X$
and $\bm Z$, and the response, $\bm y$, before starting the
iterations. To allow for case weights, we save the products $\bm
X\trans\bm W\bm X$, $\bm X\trans\bm W\bm y$, $\bm Z\trans\bm W\bm X$,
$\bm Z\trans\bm W\bm y$ and $\bm Z\trans\bm W\bm Z$ (see
Table~\ref{tab:constantPLS}).
We wish to use structures and algorithms that allow us to take a new
value of $\bm\theta$ and evaluate the $\bm L_\theta$
(Equation~\ref{eq:blockCholeskyDecomp}) efficiently. The key to doing so
is the special structure of $\bLt\trans\bm Z\trans\bm W^{1/2}$. To
understand why this matrix, and not its transpose, is of interest we
describe the sparse matrix structures used in \proglang{Julia} and in
the \pkg{Matrix} package for \proglang{R}.
Dense matrices are stored in \proglang{R} and in \proglang{Julia} as a
one-dimensional array of contiguous storage locations addressed in
\emph{column-major order}. This means that elements in the same
column are in adjacent storage locations whereas elements in the same
row can be widely separated in memory. For this reason, algorithms
that work column-wise are preferred to those that work row-wise.
Although a sparse matrix can be stored in a \emph{triplet} format,
where the row position, column position and element value of the nonzeros
are recorded as triplets, the preferred storage forms for actual
computation with sparse matrices are compressed sparse column (CSC) or
compressed sparse row \citep[CSR,][Chapter~2]{davis06:csparse_book}. The
\pkg{CHOLMOD} (and, more generally, the \pkg{SuiteSparse} package of \proglang{C} libraries; \citealt{SuiteSparse})
uses the CSC storage format. In this format the nonzero elements in
a column are in adjacent storage locations and access to all the
elements in a column is much easier than access to those in a row
(similar to dense matrices stored in column-major order).
The matrices $\bm Z$ and $\bm Z\bLt$ have the property that the number
of nonzeros in each row, $\sum_{i=1}^k p_i$, is constant. For CSC
matrices we want consistency in the columns rather than the rows,
which is why we work with $\bm Z\trans$ and $\bLt\trans\bm Z\trans$
rather than their transposes.
An arbitrary $m\times n$ sparse matrix in CSC format is expressed as
two vectors of indices, the row indices and the column pointers, and a
numeric vector of the nonzero values. The elements of the row
indices and the nonzeros are aligned and are ordered first by
increasing column number then by increasing row number within column.
The column pointers are a vector of size $n+1$ giving the location of
the start of each column in the vectors of row indices and nonzeros.
%% This para is __extra__ in comparison to JSS (2015):
Because the number of nonzeros in each column of $\bm Z\trans$, and in
each column of matrices derived from $\bm Z\trans$ (such as
$\bLt\trans\bm Z\trans\bm W^{1/2}$) is constant, the vector of
nonzeros in the CSC format can be viewed as a dense matrix, say $\bm
N$, of size $\left(\sum_{i=1}^n p_i\right)\times n$. We do not need
to store the column pointers because the columns of $\bm Z\trans$
correspond to columns of $\bm N$. All we need is $\bm N$, the dense
matrix of nonzeros, and the row indices, which are derived from the
grouping factor index vectors $\bm i_i,i=1,\dots,k$ and can also be
arranged as a dense matrix of size $\left(\sum_{i=1}^n
p_i\right)\times n$. Matrices like $\bm Z\trans$, with the property
that there are the same number of nonzeros in each column, are
sometimes called \emph{regular sparse column-oriented} (RSC) matrices.
\section{Nonlinear optimization module}
\label{sec:optimizeLmer}
The objective function module produces a function which implements the
penalized least squares algorithm for a particular mixed model. The
next step is to optimize this function over the covariance parameters,
$\bm\theta$, which is a nonlinear optimization problem. The
\pkg{lme4} package separates the efficient computational linear
algebra required to compute profiled likelihoods and deviances for a
given value of $\bm\theta$ from the nonlinear optimization algorithms,
which use general-purpose nonlinear optimizers.
One benefit of this separation is that it allows for experimentation
with different nonlinear optimizers. Throughout the development of
\pkg{lme4}, the default optimizers and control parameters have changed
in response to feedback from users about both efficiency and
convergence properties. \pkg{lme4} incorporates two built-in
optimization choices, an implementation of the Nelder-Mead simplex
algorithm adapted from Steven Johnson's \pkg{NLopt} \proglang{C} library
\citep{NLopt} and a wrapper for Powell's BOBYQA algorithm, implemented
in the \pkg{minqa} package \citep{minqa_pkg} as a wrapper around
Powell's original \proglang{Fortran} code \citep{Powell_bobyqa}. More generally,
\pkg{lme4} allows any user-specified optimizer that (1) can work with
an objective function (i.e., does not require a gradient function to
be specified), (2) allows box constraints on the parameters, and (3)
conforms to some simple rules about argument names and structure of
the output. An internal wrapper allows the use of the \pkg{optimx}
package \citep{optimx_pkg}, although the only optimizers provided via
\pkg{optimx} that satisfy the constraints above are the \code{nlminb}
and \code{L-BFGS-B} algorithms that are themselves wrappers around the
versions provided in base \proglang{R}. Several other algorithms from Steven
Johnson's \pkg{NLopt} package are also available via the \pkg{nloptr}
wrapper package (e.g., alternate implementations of Nelder-Mead and
BOBYQA, and Powell's COBYLA algorithm).
This flexibility assists with diagnosing convergence problems -- it
is easy to switch among several algorithms to determine whether the
problem lies in a failure of the nonlinear optimization stage, as
opposed to a case of model misspecification or unidentifiability or a
problem with the underlying PLS algorithm. To date we
have only observed PLS failures, which arise if
$\bm X\trans\bm W\bm X-\bm R_{ZX}\trans\bm R_{ZX}$ becomes singular
during an evaluation of the objective function, with badly scaled
problems (i.e., problems with continuous predictors that take a
very large or very small numerical range of values).
The requirement for optimizers that can handle box constraints stems
from our decision to parameterize the variance-covariance matrix in a
constrained space, in order to allow for singular fits. In contrast to
the approach taken in the \pkg{nlme} package \citep{nlme_pkg}, which goes to some
lengths to use an unconstrained variance-covariance parameterization
(the \emph{log-Cholesky} parameterization;
\citealt{pinheiro_unconstrained_1996}), we instead use the Cholesky
parameterization but require the elements of $\bm\theta$ corresponding
to the diagonal elements of the Cholesky factor to be non-negative.
With these constraints, the variance-covariance matrix is singular if
and only if any of the diagonal elements is exactly zero.
Singular fits are common in practical data-analysis situations,
especially with small- to medium-sized data sets and complex
variance-covariance models, so being able to fit a singular model is
an advantage: when the best fitting model lies on the boundary of a
constrained space, approaches that try to remove the constraints by
fitting parameters on a transformed scale will often give rise to
convergence warnings as the algorithm tries to find a maximum on an
asymptotically flat surface \citep{bolker_strategies_2013}.
In principle the likelihood surfaces we are trying to optimize over
are smooth, but in practice using gradient information in optimization
may or may not be worth the effort. In special cases, we have a
closed-form solution for the gradients
(Equations~\ref{eq:grad}--\ref{eq:gradr2}), but in general we will have to
approximate them by finite differences, which is expensive and has
limited accuracy. (We have considered using automatic
differentiation approaches to compute the gradients more efficiently,
but this strategy requires a great deal of additional machinery, and
would have drawbacks in terms of memory requirements for large
problems.) This is the primary reason for our switching to
derivative-free optimizers such as BOBYQA and Nelder-Mead in the
current version of \pkg{lme4}, although as discussed above
derivative-based optimizers based on finite differencing are available
as an alternative.
There is most likely further room for improvement in the nonlinear
optimization module; for example, some speed-up could be gained by
using parallel implementations of derivative-free optimizers that
evaluated several trial points at once \citep{klein_nelder_2013}. In
practice users most often have optimization difficulties with poorly
scaled or centered data -- we are working to implement appropriate
diagnostic tests and warnings to detect these situations.
\section{Output module}
\label{sec:mkMerMod}
Here we describe some of the methods in \pkg{lme4} for exploring
fitted linear mixed models (Table~\ref{tab:methods}), which are
represented as objects of the \proglang{S}4 class `\code{lmerMod}'. We
begin by describing the theory underlying these methods
(Section~\ref{sec:outputTheory}) and then continue the
\code{sleepstudy} example introduced in Section~\ref{sec:sleepstudy}
to illustrate these ideas in practice.
\subsection{Theory underlying the output module}
\label{sec:outputTheory}
\subsubsection{Covariance matrix of the fixed-effect coefficients}
\label{sec:vcov}
In the \code{lm} function, the variance-covariance matrix of the
coefficients is the inverse of the cross-product of the model matrix, times
the residual variance \citep{Chambers:1993}. The inverse cross-product
matrix is computed by first inverting the upper triangular matrix
resulting from the QR decomposition of the model matrix, and then
taking its cross-product,
\begin{equation}
\label{eq:varUBeta}
\Var_{\bm\theta, \sigma}\left(
\begin{bmatrix}
\bm\mu_{\mc U|\mc Y=\yobs} \\ \hat\betavec
\end{bmatrix}
\right) =
\sigma^2
\begin{bmatrix}
\bm L\trans_{\bm\theta} & \bm R_{ZX}\\
\bm 0 & \bm R_X
\end{bmatrix}^{-1}
\begin{bmatrix}
\bm L_{\bm\theta}& \bm 0\\
\bm R_{ZX}\trans & \bm R_X\trans
\end{bmatrix}^{-1}.
\end{equation} % |
Because of normality, the marginal covariance matrix of $\hat\betavec$ is
just the lower-right $p$-by-$p$ block of
$\Var_{\bm\theta, \sigma}\left(
\begin{bmatrix}
\bm\mu_{\mc U|\mc Y=\yobs} \\ \hat\betavec
\end{bmatrix}
\right)$.
This lower-right block is
\begin{equation}
\label{eq:varBeta}
\Var_{\bm\theta, \sigma}(\hat\betavec) = \sigma^2 \bm R_X^{-1} (\bm R_X\trans)^{-1},
\end{equation}
which follows from the Schur complement identity \citep[Theorem 1.2]{zhang2006schur}. This
matrix can be extracted from a fitted \code{merMod} object as,
<<vcovByHand>>=
RX <- getME(fm1, "RX")
sigma2 <- sigma(fm1)^2
sigma2 * chol2inv(RX)
@
which could be computed with \pkg{lme4} as \code{vcov(fm1)}.
The square-root diagonal of this covariance matrix contains the
estimates of the standard errors of fixed-effects coefficients. These
standard errors are used to construct Wald confidence intervals with
\code{confint(., method = "Wald")}. Such confidence intervals are
approximate, and depend on the assumption that the likelihood profile
of the fixed effects is linear on the $\zeta$ scale
(Section~\ref{sec:profile}).
\subsubsection{Profiling}
\label{sec:profile}
\newcommand{\devprof}{-2 {\mc L}_p}
The theory of likelihood profiles is straightforward: the deviance (or
likelihood) profile, $\devprof()$, for a focal model parameter $P$ is
the minimum value of the deviance conditioned on a particular value of
$P$. For each parameter of interest, our goal is to evaluate the
deviance profile for many points -- optimizing over all of the
non-focal parameters each time -- over a wide enough range and with
high enough resolution to evaluate the shape of the profile (in
particular, whether it is quadratic, which would allow use of Wald
confidence intervals and tests) and to find the values of $P$ such
that $\devprof(P)=-2 {\mc L}(\widehat{P}) + \chi^2(1-\alpha)$, which
represent the profile confidence intervals. While profile confidence
intervals rely on the asymptotic distribution of the minimum deviance,
this is a much weaker assumption than the log-quadratic likelihood
surface required by Wald tests.
An additional challenge in profiling arises when we want to compute
profiles for quantities of interest that are not parameters of our PLS
function. We have two problems in using the deviance function defined
above to profile linear mixed models. First, a parameterization of the
random-effects variance-covariance matrix in terms of standard
deviations and correlations, or variances and covariances, is much
more familiar to users, and much more relevant as output of a
statistical model, than the parameters, $\bm\theta$, of the relative
covariance factor -- users are interested in inferences on variances
or standard deviations, not on $\bm\theta$. Second, the fixed-effect
coefficients and the residual standard deviation, both of which are
also of interest to users, are profiled out (in the sense used above)
of the deviance function (Section~\ref{sec:profdev}), so we have to
find a strategy for estimating the deviance for values of $\bm\beta$
and $\sigma^2$ other than the profiled values.
To handle the first problem we create a new version of the deviance
function that first takes a vector of standard deviations (and
correlations), and a value of the residual standard deviation,
maps them to a $\bm\theta$ vector, and then computes the PLS as
before; it uses the specified residual standard deviation rather than
the PLS estimate of the standard deviation
(Equation~\ref{eq:residualVarianceEstimate}) in the deviance calculation.
We compute a profile likelihood for the fixed-effect parameters, which
are profiled out of the deviance calculation, by adding an offset to
the linear predictor for the focal element of $\bm\beta$. The
resulting function is not useful for general nonlinear optimization
-- one can easily wander into parameter regimes corresponding to
infeasible (non-positive semidefinite) variance-covariance matrices --
but it serves for likelihood profiling, where one focal parameter is
varied at a time and the optimization over the other parameters is
likely to start close to an optimum.
In practice, the \code{profile} method systematically varies the
parameters in a model, assessing the best possible fit that can be
obtained with one parameter fixed at a specific value and comparing
this fit to the globally optimal fit, which is the original model fit
that allowed all the parameters to vary. The models are compared
according to the change in the deviance, which is the likelihood ratio
test statistic. We apply a signed square root transformation to this
statistic and plot the resulting function, which we call the
\emph{profile zeta function} or $\zeta$, versus the parameter
value. The signed aspect of this transformation means that $\zeta$ is
positive where the deviation from the parameter estimate is positive
and negative otherwise, leading to a monotonically increasing function
which is zero at the global optimum. A $\zeta$ value can be compared
to the quantiles of the standard normal distribution. For example, a
95\% profile deviance confidence interval on the parameter consists of
the values for which $-1.96 < \zeta < 1.96$. Because the process of
profiling a fitted model can be computationally intensive -- it
involves refitting the model many times -- one should exercise
caution with complex models fit to large data sets.
The standard approach to this computational challenge is to compute
$\zeta$ at a limited number of parameter values, and to fill in the
gaps by fitting an interpolation spline. Often one is able to invert
the spline to obtain a function from $\zeta$ to the focal parameter,
which is necessary in order to construct profile confidence intervals.
However, even if the points themselves are monotonic, it is possible
to obtain non-monotonic interpolation curves. In such a case,
\pkg{lme4} falls back to linear interpolation, with a warning.
\newcommand{\alphamax}{\alpha_{\textrm{\small max}}}
\newcommand{\pstep}{\epsilon}
The last part of the technical specification for computing profiles is
deciding on a strategy for choosing points to sample. In one way or
another one wants to start at the estimated value for each parameter
and work outward either until a constraint is reached (i.e., a value of
0 for a standard deviation parameter or a value of $\pm 1$ for a
correlation parameter), or until a sufficiently large deviation from
the minimum deviance is attained. \pkg{lme4}'s profiler chooses a
cutoff $\phi$ based on the $1-\alphamax$ critical value of the
$\chi^2$ distribution with a number of degrees of freedom equal to the
\emph{total} number of parameters in the model, where $\alphamax$ is
set to 0.05 by default. The profile computation initially adjusts the
focal parameter $p_i$ by an amount $\pstep=1.01 \hat p_i$ from its ML
or REML estimate $\hat p_i$ (or by $\pstep=0.001$ if $\hat p_i$ is
zero, as in the case of a singular variance-covariance model). The
local slope of the likelihood profile $(\zeta(\hat
p_i+\pstep)-\zeta(\hat p_i))/\pstep$ is used to pick the next point to
evaluate, extrapolating the local slope to find a new $\pstep$ that
would be expected to correspond to the desired step size $\Delta
\zeta$ (equal to $\phi/8$ by default, so that 16 points would be used
to cover the profile in both directions if the log-likelihood surface
were exactly quadratic). Some fail-safe testing is done to ensure
that the step chosen is always positive, and less than a maximum; if a
deviance is ever detected that is lower than that of the ML deviance,
which can occur if the initial fit was wrong due to numerical
problems, the profiler issues an error and stops.
%% while scaling based on the approximate variance-covariance matrix
%% might be more accurate in some cases, it may also lead to problems
%% when the log-likelihood surface is far from quadratic
\subsubsection{Parametric bootstrapping}
\label{sec:pb}
To avoid the asymptotic assumptions of the likelihood ratio test, at
the cost of greatly increased computation time, one can estimate
confidence intervals by parametric bootstrapping -- that is, by
simulating data from the fitted model, refitting the model, and
extracting the new estimated parameters (or any other quantities of
interest). This task is quite straightforward, since there is already
a \code{simulate} method, and a \code{refit} function which
re-estimates the (RE)ML parameters for new data, starting from the
previous (RE)ML estimates and re-using the previously computed model
structures (including the fill-reducing permutation) for efficiency.
The \code{bootMer} function is thus a fairly thin wrapper around a
\code{simulate}/\code{refit} loop, with a small amount of additional
logic for parallel computation and error-catching. (Some of the ideas
of \code{bootMer} are adapted from \code{merBoot} \citep{merBoot}, a
more ambitious framework for bootstrapping \pkg{lme4} model fits which
unfortunately seems to be unavailable at present.)
\subsubsection{Conditional variances of random effects}
\label{sec:condVar}
It is useful to clarify that the conditional covariance concept in
\pkg{lme4} is based on a simplification of the linear mixed model. In
particular, we simplify the model by assuming that the quantities,
$\betavec$, $\bLt$, and $\sigma$, are known (i.e., set at their
estimated values). In fact, the only way to define the conditional
covariance is at fixed parameter values. Our approximation here is
using the estimated parameter values for the unknown ``true''
parameter values. In this simplified case, $\mc U$ is the only
quantity in the statistical model that is both random and unknown.
Given this simplified linear mixed model, a standard result in
Bayesian linear regression modeling \citep{gelman2013bayesian}
implies that the conditional distribution of the spherical random
effects given the observed response vector is Gaussian,
\begin{equation}
\label{eq:condGaussian}
(\mc U | \mc Y = \yobs) \sim \mathcal{N}(\mu_{\mc U | \mc Y = \yobs},
\widehat{\sigma}^2 \bm V),
\end{equation}
where
\begin{equation}
\label{eq:condVar}
\bm V = \left(\bm\Lambda_{\widehat{\bm\theta}}\trans\bm Z\trans \bm W
\bm Z\bm\Lambda_{\widehat{\bm\theta}} + \bm I_q \right)^{-1} =
\left(\bm L_{\widehat{\bm\theta}}^{-1}\right)\trans
\left(\bm L_{\widehat{\bm\theta}}^{-1}\right)
\end{equation}
is the unscaled conditional variance and
\begin{equation}
\label{eq:condMean}
\mu_{\mc U | \mc Y = \yobs} = \bm V \bm\Lambda_{\widehat{\bm\theta}}\trans\bm Z\trans\bm W
\left(\yobs-\bm o - X\widehat{\betavec} \right)
\end{equation} % |
is the conditional mean/mode. Note that in practice the inverse in
Equation~\ref{eq:condVar} does not get computed directly, but rather an
efficient method is used that exploits the sparse structures.
The random-effects coefficient vector, $\mathbf{b}$, is often of
greater interest. The conditional covariance matrix of $\mc B$
may be expressed as
\begin{equation}
\label{eq:condVarB}
\widehat{\sigma}^2 \bm\Lambda_{\widehat{\bm\theta}} \bm V \bm\Lambda_{\widehat{\bm\theta}}\trans.
\end{equation}
<<condVarExample, echo=FALSE, eval=FALSE>>=
s2 <- sigma(fm1)^2
Lambdat <- getME(fm1, "Lambdat")
Lambda <- getME(fm1, "Lambda")
Zt <- getME(fm1, "Zt")
Z <- getME(fm1, "Z")
y <- getME(fm1, "y")
X <- getME(fm1, "X")
beta <- getME(fm1, "beta")
L <- getME(fm1, "L")
(V <- crossprod(solve(L, system = "L")))[1:2, 1:2]
(s2 * Lambda %*% V %*% Lambdat)[1:2, 1:2]
attr(ranef(fm1, condVar = TRUE)$Subject, "postVar")[ , , 1] # should be same as Lam %*% V %*% Lamt
(U <- as.vector((V %*% Lambdat %*% Zt %*% (y - X %*% beta))))
getME(fm1, "u")
@
\subsubsection{The hat matrix}
\label{sec:hat}
The hat matrix, $\bm H$, is a useful object in linear model
diagnostics \citep{cook1982residuals}. In general, $\bm H$ relates
the observed and fitted response vectors, but we specifically define
it as,
\begin{equation}
\left(\bm\mu_{\mc Y|\mc U=\bm u} - \bm o\right) =
\bm H \left(\yobs - \bm o\right).
\end{equation}
To find $\bm H$ we note that
\begin{equation}
\left(\bm\mu_{\mc Y|\mc U=\bm u} - \bm o\right) =
\begin{bmatrix}
\bm Z\bm\Lambda & \bm X \\
\end{bmatrix}
\begin{bmatrix}
\mu_{\mc U | \mc Y = \yobs} \\ \widehat{\betavec}_{\bm\theta}
\end{bmatrix}.
\end{equation}
Next we get an expression for $\begin{bmatrix}
\mu_{\mc U | \mc Y = \yobs} \\ \widehat{\betavec}_{\bm\theta}
\end{bmatrix}$ by solving the normal equations
(Equation~\ref{eq:normalEquations}),
\begin{equation}
\begin{bmatrix}
\mu_{\mc U | \mc Y = \yobs} \\ \widehat{\betavec}_{\bm\theta}
\end{bmatrix} =
\begin{bmatrix}
\bm L\trans_{\bm\theta} & \bm R_{ZX} \\
\bm 0 & \bm R_{X}
\end{bmatrix}^{-1}
\begin{bmatrix}
\bm L_{\bm\theta} & \bm 0 \\
\bm R\trans_{ZX} & \bm R\trans_{X} \\
\end{bmatrix}^{-1}
\begin{bmatrix}
\bLt\trans \bm Z\trans \\
\bm X\trans \\
\end{bmatrix} \bm W (\yobs - \bm o).
\end{equation} % |
By the Schur complement identity \citep{zhang2006schur},
\begin{equation}
\begin{bmatrix}
\bm L\trans_{\bm\theta} & \bm R_{ZX} \\
\bm 0 & \bm R_{X}
\end{bmatrix}^{-1} =
\begin{bmatrix}
\left(\bm L\trans_{\bm\theta}\right)^{-1} &
- \left(\bm L\trans_{\bm\theta}\right)^{-1} \bm R_{ZX} \bm R_{X}^{-1}\\
\bm 0 & \bm R_{X}^{-1}
\end{bmatrix}.
\end{equation}
Therefore, we may write the hat matrix as
\begin{equation}
\label{eq:hatComputational}
\bm H = (\bm C_L\trans \bm C_L + \bm C_R\trans \bm C_R),
\end{equation}
where,
\begin{equation}
\label{eq:CL}
\bm L_{\bm\theta} \bm C_L = \bLt\trans \bm Z\trans\bm W^{1/2}
\end{equation}
and,
\begin{equation}
\label{eq:CR}
\bm R\trans_{X} \bm C_R = \bm X\trans\bm W^{1/2} - \bm R\trans_{ZX}\bm C_L.
\end{equation}
The trace of the hat matrix is often used as a measure of the
effective degrees of freedom \citep[e.g., ][]{vaida2005conditional}.
Using a relationship between the trace and vec operators, the trace of
$\bm H$ may be expressed as,
\begin{equation}
\label{eq:traceHat}
\tr(\bm H) = \left\| \VEC(\bm C_L) \right\|^2
+ \left\| \VEC(\bm C_R) \right\|^2.
\end{equation}
\subsection{Using the output module}
\label{sec:outputPractice}
\begin{table}
\begin{tabular}{ll}
\hline
Generic (Section) & Brief description of return value \\
\hline
\code{anova} (\ref{sec:modComp}) &
Decomposition of fixed-effects contributions \\
& or model comparison. \\
\code{as.function} &
Function returning profiled deviance or REML criterion. \\
\code{coef} &
Sum of the random and fixed effects for each level. \\
\code{confint} (\ref{sec:CI}) &
Confidence intervals on linear mixed-model parameters. \\
\code{deviance} (\ref{sec:summary}) &
Minus twice maximum log-likelihood. \\
& (Use \code{REMLcrit} for the REML criterion.) \\
\code{df.residual} &
Residual degrees of freedom. \\
\code{drop1} &
Drop allowable single terms from the model. \\
\code{extractAIC} &
Generalized Akaike information criterion \\
\code{fitted} &
Fitted values given conditional modes (Equation~\ref{eq:meanYgivenU}). \\
\code{fixef} (\ref{sec:summary}) &
Estimates of the fixed-effects coefficients, $\widehat{\bm\beta}$ \\
\code{formula} (\ref{sec:uncor}) &
Mixed-model formula of fitted model. \\
\code{logLik} & % logLik
Maximum log-likelihood. \\
\code{model.frame} & % model.frame
Data required to fit the model. \\
\code{model.matrix} &
Fixed-effects model matrix, $\bm X$.\\
\code{ngrps} (\ref{sec:summary}) &
Number of levels in each grouping factor. \\
\code{nobs} (\ref{sec:summary}) &
Number of observations.\\
\code{plot} &
Diagnostic plots for mixed-model fits. \\
\code{predict} (\ref{sec:predict}) &
Various types of predicted values. \\
\code{print} &
Basic printout of mixed-model objects. \\
\code{profile} (\ref{sec:profile}) &
Profiled likelihood over various model parameters. \\
\code{ranef} (\ref{sec:summary}) &
Conditional modes of the random effects. \\
\code{refit} (\ref{sec:pb}) &
A model (re)fitted to a new set of observations of the response variable. \\
\code{refitML} (\ref{sec:modComp}) &
A model (re)fitted by maximum likelihood. \\
\code{residuals} (\ref{sec:summary}) &
Various types of residual values. \\
\code{sigma} (\ref{sec:summary}) &
Residual standard deviation. \\
\code{simulate} (\ref{sec:predict}) &
Simulated data from a fitted mixed model. \\
\code{summary} (\ref{sec:summary}) &
Summary of a mixed model. \\
\code{terms} &
Terms representation of a mixed model. \\
\code{update} (\ref{sec:update}) &
An updated model using a revised formula or other arguments. \\
\code{VarCorr} (\ref{sec:summary}) &
Estimated random-effects variances, standard deviations, and correlations. \\
\code{vcov} (\ref{sec:summary}) &
Covariance matrix of the fixed-effect estimates. \\
\code{weights} &
Prior weights used in model fitting. \\
\hline
\end{tabular}
\caption{List of currently available methods for objects of the class `\code{merMod}'.}
\label{tab:methods}
\end{table}
The user interface for the output module largely consists of a set of
methods (Table~\ref{tab:methods}) for objects of class \code{merMod},
which is the class of objects returned by \code{lmer}. In addition to
these methods, the \code{getME} function can be used to extract
various objects from a fitted mixed model in \pkg{lme4}. Here we
illustrate the use of several of these methods.
\subsubsection{Updating fitted mixed models}
\label{sec:update}
To illustrate the \code{update} method for `\code{merMod}' objects we
construct a random intercept only model. This task could be done in
several ways, but we choose to first remove the random-effects term
\code{(Days | Subject)} and add a new term with a random intercept,
<<updateModel>>=
fm3 <- update(fm1, . ~ . - (Days | Subject) + (1 | Subject))
formula(fm3)
@ % |
\subsubsection{Model summary and associated accessors}
\label{sec:summary}
The \code{summary} method for `\code{merMod}' objects provides
information about the model fit. Here we consider the output of
\code{summary(fm1)} in four steps. The first few lines of output
indicate that the model was fitted by REML as well as the value of the
REML criterion (Equation~\ref{eq:REMLdeviance}) at convergence (which
may also be extracted using \code{REMLcrit(fm1)}). The beginning of
the summary also reproduces the model formula and the scaled Pearson
residuals,
<<summary1,echo=FALSE>>=
ss <- summary(fm1)
cc <- capture.output(print(ss))
reRow <- grep("^Random effects", cc)
cat(cc[1:(reRow - 2)], sep = "\n")
@
This information may also be obtained using standard accessor
functions,
<<summary1reproduced, eval=FALSE>>=
formula(fm1)
REMLcrit(fm1)
quantile(residuals(fm1, "pearson", scaled = TRUE))
@
The second piece of the \code{summary} output provides information
regarding the random-effects and residual variation,
<<summary2, echo=FALSE>>=
feRow <- grep("^Fixed effects", cc)
cat(cc[reRow:(feRow - 2)], sep = "\n")
@
which can also be accessed using,
<<summary2reproduced, eval=FALSE>>=
print(vc <- VarCorr(fm1), comp = c("Variance", "Std.Dev."))
nobs(fm1)
ngrps(fm1)
sigma(fm1)
@
The print method for objects returned by \code{VarCorr} hides the internal structure of these
`\code{VarCorr.merMod}' objects. The internal structure of an object of
this class is a list of matrices, one for each random-effects term.
The standard deviations and correlation matrices for each term are
stored as attributes, \code{stddev} and \code{correlation},
respectively, of the variance-covariance matrix, and the residual
standard deviation is stored as attribute \code{sc}. For programming
use, these objects can be summarized differently using their
\code{as.data.frame} method,
<<VarCorr>>=
as.data.frame(VarCorr(fm1))
@
which contains one row for each variance or covariance parameter. The
\code{grp} column gives the grouping factor associated with this
parameter. The \code{var1} and \code{var2} columns give the names of
the variables associated with the parameter (\code{var2} is \code{<NA>}
unless it is a covariance parameter). The \code{vcov} column gives the
variances and covariances, and the \code{sdcor} column gives these
numbers on the standard deviation and correlation scales.
The next chunk of output gives the fixed-effect estimates,
<<summary3, echo=FALSE>>=
corRow <- grep("^Correlation", cc)
cat(cc[feRow:(corRow - 2)], sep = "\n")
@
Note that there are no $p$~values (see Section~\ref{sec:pvalues}).
The fixed-effect point estimates may be obtained separately via
\code{fixef(fm1)}.
Conditional modes of the random-effects coefficients can be obtained
with \code{ranef} (see Section~\ref{sec:condVar} for information on
the theory). Finally, we have the correlations between the fixed-effect estimates
<<summary4, echo=FALSE>>=
cat(cc[corRow:length(cc)], sep = "\n")
@
The full variance-covariance matrix of the fixed-effects estimates can
be obtained in the usual \proglang{R} way with the \code{vcov} method
(Section~\ref{sec:vcov}). Alternatively, \code{coef(summary(fm1))} will
return the full fixed-effects parameter table as shown in the summary.
\subsubsection{Diagnostic plots}
\label{sec:diagnostics}
\pkg{lme4} provides tools for generating most of the standard
graphical diagnostic plots (with the exception of those incorporating
influence measures, e.g., Cook's distance and leverage plots), in a
way modeled on the diagnostic graphics of \pkg{nlme}
\citep{R:Pinheiro+Bates:2000}. In particular, the familiar
\code{plot} method in base \proglang{R} for linear models (objects of
class `\code{lm}') can be used to create the standard fitted
vs. residual plots,
<<diagplot1,fig.keep="none">>=
plot(fm1, type = c("p", "smooth"))
@
scale-location plots,
<<diagplot2,fig.keep="none">>=
plot(fm1, sqrt(abs(resid(.))) ~ fitted(.),
type = c("p", "smooth"))
@
and quantile-quantile plots (from \pkg{lattice}),
<<diagplot3,fig.keep="none">>=
qqmath(fm1, id = 0.05)
@
In contrast to \code{plot.lm}, these scale-location and Q-Q plots
are based on raw rather than standardized residuals.
In addition to these standard diagnostic plots, which
examine the validity of various assumptions (linearity,
homoscedasticity, normality) at the level of the residuals,
one can also use the \code{dotplot} and \code{qqmath} methods
for the conditional modes (i.e., `\code{ranef.mer}'
objects generated by \code{ranef(fit)})
to check for interesting patterns and normality in the conditional
modes. \pkg{lme4} does not provide influence diagnostics, but
these (and other useful diagnostic procedures) are available
in the dependent packages \pkg{HLMdiag} and \pkg{influence.ME}
\citep{HLMdiag_pkg,influenceME_pkg}.
Finally, \emph{posterior predictive simulation}
\citep{gelman_data_2006} is a generally useful diagnostic tool,
adapted from Bayesian methods, for exploring model fit. Users pick
some summary statistic of interest. They then compute the summary
statistic for an ensemble of simulations (Section~\ref{sec:predict}),
and see where the observed data falls within the simulated
distribution; if the observed data is extreme, we might conclude that
the model is a poor representation of reality. For example, using the
sleep study fit and choosing the interquartile range
of the reaction times as the summary
statistic:
<<ppsim,results="hide">>=
iqrvec <- sapply(simulate(fm1, 1000), IQR)
obsval <- IQR(sleepstudy$Reaction)
post.pred.p <- mean(obsval >= c(obsval, iqrvec))
@
The (one-tailed) posterior predictive $p$~value of \Sexpr{round(post.pred.p,2)}
indicates that the model represents the data adequately, at least for
this summary statistic. In contrast to the full Bayesian case, the
procedure described here does not allow for the uncertainty in the
estimated parameters. However, it should be a reasonable approximation
when the residual variation is large.
%% \bmb{it should be possible to put together something with
%% \code{arm::sim} to simulate beta values and \code{simulate} with
%% \code{newparams}, but (a) this would require some testing and
%% development (I don't think \code{newparams} can take a matrix) and
%% (b) it doesn't account for uncertainty in theta)}
\subsubsection{Sequential decomposition and model comparison}
\label{sec:modComp}
Objects of class `\code{merMod}' have an \code{anova} method which
returns $F$~statistics corresponding to the sequential decomposition
of the contributions of fixed-effects terms. In order to illustrate
this sequential ANOVA decomposition, we fit a model with orthogonal
linear and quadratic \code{Days} terms,
<<anovaQuadraticModel>>=
fm4 <- lmer(Reaction ~ polyDays[ , 1] + polyDays[ , 2] +
(polyDays[ , 1] + polyDays[ , 2] | Subject),
within(sleepstudy, polyDays <- poly(Days, 2)))
anova(fm4)
@
<<anovaSanityCheck, include=FALSE>>=
(getME(fm4, "RX")[2, ] %*% getME(fm4, "fixef"))^2
@
The relative magnitudes of the two sums of squares indicate that the
quadratic term explains much less variation than the linear term.
Furthermore, the magnitudes of the two $F$~statistics strongly
suggest significance of the linear term, but not the quadratic term.
Notice that this is only an informal assessment of significance as
there are no $p$~values associated with these $F$~statistics, an issue
discussed in more detail in the next subsection (``Computing $p$-values'',
p.~\pageref{sec:pvalues}).
To understand how these quantities are computed, let $\bm R_i$ contain
the rows of $\bm R_X$ (Equation~\ref{eq:blockCholeskyDecomp}) associated
with the $i$th fixed-effects term. Then the sum of squares for term
$i$ is,
\begin{equation}
\label{eq:SS}
\mathit{SS}_i = \widehat{\bm\beta}\trans\bm R_i\trans \bm R_i \widehat{\bm\beta}
\end{equation}
If $\mathit{DF}_i$ is the number of columns in $\bm R_i$, then the
$F$~statistic for term $i$ is,
\begin{equation}
\label{eq:Fstat}
F_i = \frac{\mathit{SS}_i}{\widehat{\sigma}^2 \mathit{DF}_i}
\end{equation}
For multiple arguments, the \code{anova} method returns model
comparison statistics.
<<anovaManyModels>>=
anova(fm1, fm2, fm3)
@
<<anovaRes,echo=FALSE>>=
fm3ML <- refitML(fm3)
fm2ML <- refitML(fm2)
fm1ML <- refitML(fm1)
ddiff <- deviance(fm3ML) - deviance(fm2ML)
dp <- pchisq(ddiff, 1, lower.tail = FALSE)
ddiff2 <- deviance(fm2ML) - deviance(fm1ML)
@
The output shows $\chi^2$ statistics representing the difference in
deviance between successive models, as well as $p$~values based on
likelihood ratio test comparisons. In this case, the sequential
comparison shows that adding a linear effect of time uncorrelated with
the intercept leads to an enormous and significant drop in deviance
($\Delta\mbox{deviance} \approx \Sexpr{round(ddiff)}$, $p \approx
\Sexpr{round(dp,10)}$), while the further addition of correlation
between the slope and intercept leads to a trivial and non-significant
change in deviance ($\Delta\mbox{deviance} \approx
\Sexpr{round(ddiff2,2)}$). For objects of class `\code{lmerMod}' the
default behavior is to refit the models with ML if fitted with
\code{REML = TRUE}, which is necessary in order to get sensible
answers when comparing models that differ in their fixed effects; this
can be controlled via the \code{refit} argument.
\subsubsection[Computing p values]{Computing $p$~values}
\label{sec:pvalues}
One of the more controversial design decisions of \pkg{lme4} has been
to omit the output of $p$~values associated with sequential ANOVA
decompositions of fixed effects. The absence of analytical results for null
distributions of parameter estimates in complex situations
(e.g., unbalanced or partially crossed designs) is a long-standing
problem in mixed-model inference. While the null distributions (and
the sampling distributions of non-null estimates) are asymptotically
normal, these distributions are not $t$~distributed for finite size
samples -- nor are the corresponding null distributions of
differences in scaled deviances $F$~distributed. Thus approximate methods
for computing the approximate degrees of freedom for
$t$~distributions, or the denominator degrees of freedom for
$F$~statistics \citep{Satterthwaite_1946,kenward_small_1997}, are at
best \emph{ad hoc} solutions.
However, computing finite-size-corrected $p$~values is
sometimes necessary. Therefore, although the package does not provide
them (except via parametric bootstrapping,
Section~\ref{sec:pb}), we have provided a help page to guide users in
finding appropriate methods:
<<pvaluesHelp, eval=FALSE>>=
help("pvalues")
@
This help page provides pointers to other packages that provide
machinery for calculating $p$~values associated with `\code{merMod}'
objects. It also suggests framing the inference problem as a
likelihood ratio test (achieved by supplying multiple `\code{merMod}'
objects to the \code{anova} method (Section~\ref{sec:modComp}), as
well as alternatives to $p$~values such as confidence intervals
(Section~\ref{sec:CI}).
Previous versions of \pkg{lme4} provided the \code{mcmcsamp} function,
which generated a Markov chain Monte Carlo sample from the posterior
distribution of the parameters (assuming flat priors). Due to
difficulty in constructing a version of \code{mcmcsamp} that was
reliable even in cases where the estimated random-effect variances
were near zero, \code{mcmcsamp} has been withdrawn.
\subsubsection{Computing confidence intervals}
\label{sec:CI}
As described above (Section~\ref{sec:vcov}), \pkg{lme4} provides confidence intervals (using
\code{confint}) via Wald approximations (for fixed-effect parameters
only), likelihood profiling, or parametric bootstrapping (the
\code{boot.type} argument selects the bootstrap confidence interval type).
<<compareCI,echo=FALSE,cache=TRUE,message=FALSE,warning=FALSE>>=
ccw <- confint(fm1, method = "Wald")
ccp <- confint(fm1, method = "profile", oldNames = FALSE)
ccb <- confint(fm1, method = "boot")
@
<<CIqcomp,echo=FALSE,eval=TRUE>>=
rtol <- function(x,y) {
abs((x - y) / ((x + y) / 2))
}
bw <- apply(ccb, 1, diff)
pw <- apply(ccp, 1, diff)
mdiffpct <- round(100 * max(rtol(bw, pw)))
@
<<CIplot,echo=FALSE,eval=FALSE>>=
## obsolete
## ,fig.cap="Comparison of confidence intervals",fig.scap="CI comparison"
tf <- function(x, method) data.frame(method = method,
par = rownames(x),
setNames(as.data.frame(x), c("lwr", "upr")))
cc.all <- do.call(rbind, mapply(tf, list(ccw, ccp, ccb),
c("Wald", "profile", "boot"),
SIMPLIFY = FALSE))
ggplot(cc.all, aes(x = 1, ymin = lwr, ymax = upr, colour = method)) +
geom_linerange(position = position_dodge(width = 0.5)) +
facet_wrap( ~ par, scale = "free") +
theme(axis.text.x = element_blank()) +
labs(x = "")
@ %
As is typical for computationally intensive profile confidence
intervals in \proglang{R}, the intervals can be computed either directly from
fitted `\code{merMod}' objects (in which case profiling is done as an
interim step), or from a previously computed likelihood profile (of
class `\code{thpr}', for ``theta profile''). Parametric bootstrapping
confidence intervals use a thin wrapper around the \code{bootMer}
function that passes the results to \code{boot.ci} from the \pkg{boot}
package \citep{boot_pkg,DavisonHinkley1997} for confidence interval calculation.
In the running sleep study example, the 95\% confidence intervals
estimated by all three methods are quite similar. The largest
change is a \Sexpr{mdiffpct}\% difference in confidence
interval widths between profile and parametric bootstrap methods for
the correlation between the intercept and slope random effects
(\{\Sexpr{round(ccb[2,1],2)},\Sexpr{round(ccb[2,2],2)}\} vs.
\{\Sexpr{round(ccp[2,1],2)},\Sexpr{round(ccp[2,2],2)}\}).
%% \bmb{Mention that in our experience the profile CIs are
%% reasonably reliable for large and/or well-behaved data sets?
%% Is that too vague? Disclaimer that a simulation study would
%% be required to know for sure?}
\subsubsection{General profile zeta and related plots}
\label{sec:profzeta}
\pkg{lme4} provides several functions (built on \pkg{lattice}
graphics, \citealt{lattice}) for plotting the profile zeta functions
(Section~\ref{sec:profile}) and other related quantities.
\begin{itemize}
\item The \emph{profile zeta} plot
(Figure~\ref{fig:profile_zeta_plot}; \code{xyplot}) is simply a plot
of the profile zeta function for each model parameter; linearity of
this plot for a given parameter implies that the likelihood profile
is quadratic (and thus that Wald approximations would be reasonably
accurate).
\item The \emph{profile density plot}
(Figure~\ref{fig:profile_density_plot}; \code{densityplot}) displays
an approximation of the probability density function of the sampling
distribution for each parameter. These densities are derived by
setting the cumulative distribution function (c.d.f) to be
$\Phi(\zeta)$ where $\Phi$ is the c.d.f.{} of the standard normal
distribution. This is not quite the same as evaluating the
distribution of the estimator of the parameter, which for mixed
models can be very difficult, but it gives a reasonable
approximation. If the profile zeta plot is linear, then the profile
density plot will be Gaussian.
\item The \emph{profile pairs plot}
(Figure~\ref{fig:profile_pairs_plot}; \code{splom}) gives an
approximation of the two-dimensional profiles of pairs of
parameters, interpolated from the univariate profiles as described
in \citet[Chapter~6]{bateswatts88:_nonlin}. The profile pairs plot
shows two-dimensional 50\%, 80\%, 90\%, 95\% and 99\% marginal
confidence regions based on the likelihood ratio, as well as the
\emph{profile traces}, which indicate the conditional estimates of
each parameter for fixed values of the other parameters. While
panels above the diagonal show profiles with respect to the original
parameters (with random-effects parameters on the standard
deviation\slash{}correlation scale, as for all profile plots), the
panels below the diagonal show plots on the $(\zeta_i,\zeta_j)$
scale. The below-diagonal panels allow us to see distortions from
an elliptical shape due to nonlinearity of the traces, separately
from the one-dimensional distortions caused by a poor choice of
scale for the parameters. The $\zeta$ scales provide, in some sense,
the best possible set of single-parameter transformations for
assessing the contours. On the $\zeta$ scales the extent of a
contour on the horizontal axis is exactly the same as the extent on
the vertical axis and both are centered about zero.
\end{itemize}
<<profile_calc,echo=FALSE,cache=TRUE>>=
pf <- profile(fm1)
@
<<profile_zeta_plot,fig.cap="Profile zeta plot: \\code{xyplot(prof.obj)}",fig.scap="Profile zeta plot",echo=FALSE,fig.align='center',fig.pos='tb'>>=
xyplot(pf)
@
<<profile_density_plot,fig.cap="Profile density plot: \\code{densityplot(prof.obj)}",echo=FALSE,fig.align='center',fig.pos='tb'>>=
densityplot(pf)
@
<<profile_pairs_plot,fig.cap="Profile pairs plot: \\code{splom(prof.obj)}",echo=FALSE,fig.height=8,fig.width=8,fig.align='center',fig.pos='htb',out.height='5.5in'>>=
splom(pf)
@
For users who want to build their own graphical displays of the
profile, there is a method for \code{as.data.frame} that converts profile
(`\code{thpr}') objects to a more convenient format.
\subsubsection{Computing fitted and predicted values; simulating}
\label{sec:predict}
Because mixed models involve random coefficients, one must always
clarify whether predictions should be based on the marginal
distribution of the response variable or on the distribution that is
conditional on the modes of the random effects
(Equation~\ref{eq:distYgivenU}). The \code{fitted} method retrieves fitted
values that are conditional on all of the modes of the random effects;
the \code{predict} method returns the same values by default, but also
allows for predictions to be made conditional on different sets of
random effects. For example, if the \code{re.form} argument is set to
\code{NULL} (the default), then the predictions are conditional on all
random effects in the model; if \code{re.form} is \verb+~0+ or
\code{NA}, then the predictions are made at the population level (all
random-effect values set to zero). In models with multiple random
effects, the user can give \code{re.form} as a formula that specifies
which random effects are conditioned on.
\pkg{lme4} also provides a \code{simulate} method, which allows
similar flexibility in conditioning on random effects; in addition it
allows the user to choose (via the \code{use.u} argument) between
conditioning on the fitted conditional modes or choosing a new set of
simulated conditional modes (zero-mean Normal deviates chosen
according to the estimated random-effects variance-covariance
matrices). Finally, the \code{simulate} method has a method for
`\code{formula}' objects, which allows for \emph{de novo} simulation in
the absence of a fitted model. In this case, the user must specify
the random-effects ($\bm\theta$), fixed-effects ($\bm\beta$), and
residual standard deviation ($\sigma$) parameters via the
\code{newparams} argument. The standard simulation method (based on a
`\code{merMod}' object) is the basis of parametric bootstrapping
(Section~\ref{sec:pb}) and posterior predictive simulation
(Section~\ref{sec:diagnostics}); \emph{de novo} simulation based on a
formula provides a flexible framework for power analysis.
\section{Conclusion}
Mixed modeling is an extremely useful but computationally intensive
technique. Computational limitations are especially important because
mixed models are commonly applied to moderately large data sets
($10^4$--$10^6$ observations). By developing an efficient, general,
and (now) well-documented platform for fitted mixed models in
\proglang{R}, we hope to provide both a practical tool for end users
interested in analyzing data and a reusable, modular framework for
downstream developers interested in extending the class of models that
can be easily and efficiently analyzed in \proglang{R}.
%% Results of some quick Google scholar searches:
%% "mixed models" +
%% stata: 6130
%% "PROC MIXED": 12200
%% CRAN: 4020
%% lme4: 5110
%% "nlme": 2960
%% SAS: 49500
%% NLMIXED: 1960
%% GLIMMIX: 4920
%% winbugs: 2250
%% gllamm: 1140
We have learned much about linear mixed models in the process of
developing \pkg{lme4}, both from our own attempts to develop robust
and reliable procedures and from the broad community of \pkg{lme4}
users; we have attempted to describe many of these lessons here. In
moving forward, our main priorities are (1) to maintain the reference
implementation of \pkg{lme4} on the Comprehensive \proglang{R} Archive Network (CRAN), developing relatively few new
features; (2) to improve the flexibility, efficiency and scalability
of mixed-model analysis across multiple compatible implementations,
including both the \pkg{MixedModels} package for \proglang{Julia} and
the experimental \pkg{flexLambda} branch of \pkg{lme4}.
\section*{Acknowledgments}
Rune Haubo Christensen, Henrik Singmann, Fabian Scheipl,
Vincent Dorie, and Bin Dai contributed ideas
and code to the current version of \pkg{lme4}; the
large community of \pkg{lme4} users has exercised the software,
discovered bugs, and made many useful suggestions.
S{\o}ren H{\o}jsgaard participated in useful discussions
and Xi Xia and Christian Zingg exercised the code
and reported problems.
We would like to thank the Banff International Research Station
for hosting a working group on \pkg{lme4}, and
an NSERC Discovery grant and NSERC postdoctoral fellowship for funding.
\clearpage
\bibliography{lmer}
\clearpage
% \bigskip \par \pagebreak[3]
\begin{appendix}
\section{Modularization examples}
\label{sec:modularExamples}
The modularized functions in \pkg{lme4} allow for finer control of the
various steps in fitting mixed models (Table~\ref{tab:modular}). The
main advantage of these functions is the ability to extend the class
of models that \code{lmer} is capable of fitting. In general there
are three steps to this process.
\begin{itemize}
\item Construct a mixed-model formula that is similar to the intended
model.
\item Using the modular functions, hook into and modify the fitting
process so that the desired model is estimated.
\item Use the estimated parameter values to make inferences.
\end{itemize}
Note that the many tools of the output module become unavailable for
sufficiently extensive departures from the class of models fitted by
\code{lmer}. Therefore, this last inference step may be more involved
in the modular approach than with inference from models fitted by
\code{lmer}. Here we provide two examples that illustrate this
process.
\subsection{Zero slope-slope covariance models}
\label{sec:zeroSlopeSlope}
Consider the following mixed-model formula,
<<modularSimulationFormula>>=
form <- respVar ~ 1 + (explVar1 + explVar2 | groupFac)
@ %
in which we have a fixed intercept and two random-effects terms, each
with a random slope and intercept. The \code{lmer} function would fit
this formula to data by estimating an unstructured $3 \times 3$
covariance matrix giving the variances and covariances among the
intercept and the two slopes. It is not possible to use the current
version of \code{lmer} to fit a model with the following two
properties: (1) the slopes are uncorrelated and (2) the intercept is
correlated with both of the slopes. We illustrate the use of the
modular functions in \pkg{lme4} to fit such a model.
\subsubsection{Simulate from a zero slope-slope covariance model}
We simulate data from such a model using a balanced design with a
single grouping factor. We use \pkg{lme4}'s utilities for
constructing a template from the model formula, with which to conduct
these simulations. We simulate a balanced design with 500 levels to
the grouping factor and 20 observations per group,
<<dataTemplate>>=
set.seed(1)
dat <- mkDataTemplate(form,
nGrps = 500,
nPerGrp = 20,
rfunc = rnorm)
head(dat)
@ %
The initial parameters for this model may be obtained using
\code{mkParsTemplate}.
<<parsTemplate>>=
(pars <- mkParsTemplate(form, dat))
@ %
We wish to simulate random effects from the following covariance
matrix,
<<simCorr>>=
vc <- matrix(c(1.0, 0.5, 0.5,
0.5, 1.0, 0.0,
0.5, 0.0, 1.0), 3, 3)
@ %
Here the intercept is correlated with each of the two slopes, but the
slopes are uncorrelated. The \code{theta} vector that corresponds to
this covariance matrix can be computed using \pkg{lme4}'s facilities
for converting between various representations of (co-)variance
structures (see \code{?vcconv}).
<<vcTheta>>=
pars$theta[] <- Vv_to_Cv(mlist2vec(vc))
@ %
With these new parameters installed, we simulate a response variable.
<<varCorrStructure>>=
dat$respVar <- simulate(form,
newdata = dat,
newparams = pars,
family = "gaussian")[[1]]
@ %
Indeed this data set appears to contain intercept-slope correlations,
but not slope-slope correlations, as we see using the \code{lmList}
function.
<<graphSims>>=
formLm <- form
formLm[[3]] <- findbars(form)[[1]]
print(formLm)
cor(t(sapply(lmList(formLm, dat), coef)))
@ %
This \code{lmList} function computes ordinary linear models for each
level of the grouping factor. Next we show how to fit a proper mixed
model to these data.
\subsubsection{Fitting a zero slope-slope covariance model}
The main challenge with fitting a zero slope-slope model is the need
to figure out which Cholesky decompositions correspond to
variance-covariance matrices with structural zeros between the slopes.
There are two general approaches to solving this problem. The first
is to reorder the rows and columns of the covariance matrix so that
the intercept column comes last. In this case, setting the second
component of the covariance parameter vector, $\bm\theta$, to zero will
achieve the desired covariance structure. However, we consider a
second approach, which retains the ordering produced by the
\code{lFormula} function: (1) intercept, (2) first slope, and (3)
second slope. The covariance structure for the random effects in this
model is given by the following covariance matrix and its associated
Cholesky decomposition.
\begin{equation}
\label{eq:1}
\begin{bmatrix}
\sigma_1^2 & \sigma_{12} & \sigma_{13} \\
\sigma_{21} & \sigma_2^2 & 0 \\
\sigma_{31} & 0 & \sigma_3^2 \\
\end{bmatrix} =
\begin{bmatrix}
\theta_1 & 0 & 0 \\
\theta_2 & \theta_4 & 0 \\
\theta_3 & \theta_5 & \theta_6 \\
\end{bmatrix}
\begin{bmatrix}
\theta_1 & \theta_2 & \theta_3 \\
0 & \theta_4 & \theta_5 \\
0 & 0 & \theta_6 \\
\end{bmatrix}
\end{equation}
where the $\sigma$ parameters give the standard deviations and
covariances, and the $\theta$ parameters are the elements of the
Cholesky decomposition. The key feature of this covariance matrix is
the structural zero indicating no covariance between the two
slopes. Therefore, in order to ensure the structural zero, we must
have,
\begin{equation}
\label{eq:2}
\theta_5 = -\frac{\theta_2\theta_3}{\theta_4}.
\end{equation}
To represent this restriction in \proglang{R} we introduce a length
five vector, \code{phi}, which we convert into an appropriate
\code{theta} vector using the following function.
<<phiToTheta, cache = FALSE>>=
phiToTheta <- function(phi) {
theta5 <- -(phi[2]*phi[3])/phi[4]
c(phi[1:4], theta5, phi[5])
}
@ %
We may therefore optimize the structured model by optimizing over
\code{phi} instead of \code{theta}. We begin to construct an
appropriate objective function for such an optimization by taking the
first two modular steps, resulting in a deviance function of
\code{theta}.
<<compute deviance function modular, cache = FALSE>>=
lf <- lFormula(formula = form, data = dat, REML = TRUE)
devf <- do.call(mkLmerDevfun, lf)
@ %
This deviance function is a function of the unconstrained \code{theta}
vector, but we require a deviance function of the constrained
\code{phi} vector. We therefore construct a wrapper around the
deviance function as follows.
<<wrapper modular, cache = FALSE>>=
devfWrap <- function(phi) devf(phiToTheta(phi))
@ %
Finally, we optimize the deviance function,
<<opt modular, cache = FALSE>>=
opt <- bobyqa(par = lf$reTrms$theta[-5],
fn = devfWrap,
lower = lf$reTrms$lower[-5])
@ %
To extract and label the estimated covariance matrix of the random
effects, we may use the conversion functions (\code{?vcconv}) and the
column names of the random-effects model matrix (\code{cnms}). We
construct this estimated matrix and compare it with the true matrix
used to simulate the data.
<<varCorr modular, cache = FALSE>>=
vcEst <- vec2mlist(Cv_to_Vv(phiToTheta(opt$par)))[[1]]
dimnames(vcEst) <- rep(lf$reTrms$cnms, 2)
round(vcEst, 2)
vc
@ %
We see that the estimated covariance matrix provides a good estimate
of the true matrix.
\subsection{Additive models}
\label{sec:additive}
It is possible to interpret additive models as a particular class of
mixed models \citep{gamm4}. The main benefit of this approach is
that it bypasses the need to select a smoothness parameter using
cross-validation or generalized cross-validation. However, it is
inconvenient to specify additive models using the \pkg{lme4} formula
interface. The \pkg{gamm4} package wraps the modularized functions of
\pkg{lme4} within a more convenient interface \citep{gamm4}. In
particular, the strategy involves the following steps,
\begin{enumerate}
\item Convert a \pkg{gamm4} formula into an \pkg{lme4} formula that
approximates the intended model.
\item Parse this formula using \code{lFormula}.
\item Modify the resulting transposed random-effects model matrix, $\bm
Z\trans$, so that the intended additive model results.
\item Fit the resulting model using the remaining modularized
functions (Table~\ref{tab:modular}).
\end{enumerate}
Here we illustrate this general strategy using a simple simulated data
set,
<<simulateSplineData,message=FALSE>>=
library("gamm4")
library("splines")
set.seed(1)
n <- 100
pSimulation <- 3
pStatistical <- 8
x <- rnorm(n)
Bsimulation <- ns(x, pSimulation)
Bstatistical <- ns(x, pStatistical)
beta <- rnorm(pSimulation)
y <- as.numeric(Bsimulation %*% beta + rnorm(n, sd = 0.3))
@ %
where \code{pSimulation} is the number of columns in the simulation
model matrix and \code{pStatistical} is the number of columns in the
statistical model matrix.
We plot the resulting data along with the predictions from the
generating model, %% BMB: stopped here
<<splineExampleDataPlot, fig.width=4, fig.height=3,fig.align="center">>=
par(mar = c(4, 4, 1, 1), las = 1, bty = "l")
plot(x, y, las = 1)
lines(x[order(x)], (Bsimulation %*% beta)[order(x)])
@ %
We now set up an approximate \pkg{lme4} model formula, and parse it
using \code{lFormula},
<<splineExampleApproximateFormula>>=
pseudoGroups <- as.factor(rep(1:pStatistical, length = n))
parsedFormula <- lFormula(y ~ x + (1 | pseudoGroups))
@ %
We now insert a spline basis into the \code{parsedFormula} object,
<<splineExampleModifyZt>>=
parsedFormula$reTrms <- within(parsedFormula$reTrms, {
Bt <- t(as.matrix(Bstatistical))[]
cnms$pseudoGroups <- "spline"
Zt <- as(Bt, class(Zt))
})
@ %
Finally we continue with the remaining modular steps,
<<splineExampleRemainingModularSteps>>=
devianceFunction <- do.call(mkLmerDevfun, parsedFormula)
optimizerOutput <- optimizeLmer(devianceFunction)
mSpline <- mkMerMod( rho = environment(devianceFunction),
opt = optimizerOutput,
reTrms = parsedFormula$reTrms,
fr = parsedFormula$fr)
mSpline
@ %
Computing the fitted values of this additive model requires some
custom code, and illustrates the general principle that methods for
\code{merMod} objects constructed from modular fits should only be
used if the user knows what she is doing,
<<splineExampleFittedModelPlot, fig.width=4, fig.height=3, fig.align="center">>=
xNew <- seq(min(x), max(x), length = 100)
newBstatistical <- predict(Bstatistical, xNew)
yHat <- cbind(1, xNew) %*% getME(mSpline, "fixef") +
newBstatistical %*% getME(mSpline, "u")
par(mar = c(4, 4, 1, 1), las = 1, bty = "l")
plot(x, y)
lines(xNew, yHat)
lines(x[order(x)], (Bsimulation %*% beta)[order(x)],lty = 2)
legend("topright", bty = "n", c("fitted", "generating"), lty = 1:2,col = 1)
@
\end{appendix}
\end{document}
%
%%% Local Variables:
%%% ess-swv-processor: knitr
%%% TeX-master: t
%%% End: