% NOTE -- ONLY EDIT THE .Rnw FILE!!! The .tex file is
% likely to be overwritten.
\documentclass[dvipsnames,pdflatex,beamer]{beamer}
%\documentclass[a4paper,11pt,notitlepage]{article}\usepackage{beamerarticle}
\mode{\usepackage[text={6.2in,9in},centering]{geometry}}
\mode{\usetheme{Boadilla}\usecolortheme{seahorse}\usecolortheme{rose}}
\usepackage{SweaveSlides}
\title[Simple scalar models]{Mixed models in R using the lme4 package\\%
Part 1: Linear mixed models with simple, scalar random effects}
\SweaveOpts{engine=R,eps=FALSE,pdf=TRUE,width=10,height=6.5,strip.white=all}
\SweaveOpts{prefix=TRUE,prefix.string=figs/simple,include=TRUE}
\SweaveOpts{keep.source=TRUE}
\mode{\setkeys{Gin}{width=\textwidth}}
\mode{\setkeys{Gin}{width=0.8\textwidth}}
<>=
options(width=69,show.signif.stars=FALSE,str=strOptions(strict.width="cut"))
library(lattice)
library(Rcpp)
library(minqa)
library(Matrix)
library(MatrixModels)
library(lme4a)
lattice.options(default.theme = function() standard.theme())
#lattice.options(default.theme = function() standard.theme(color=FALSE))
if (file.exists("classroom.rda")) {
load("classroom.rda")
} else {
classroom <- within(read.csv("http://www-personal.umich.edu/~bwest/classroom.csv"),
{
classid <- factor(classid)
schoolid <- factor(schoolid)
sex <- factor(sex, labels = c("M","F"))
minority <- factor(minority, labels = c("N", "Y"))
})
save(classroom, file="classroom.rda")
}
@
\begin{document}
\mode{\maketitle\tableofcontents}
\mode{\frame{\titlepage}}
\mode{\frame{\frametitle{Outline}\tableofcontents[pausesections,hideallsubsections]}}
\section[Packages]{R packages and data in packages}
\begin{frame}[fragile]
\frametitle{R packages}
\begin{itemize}
\item Packages incorporate functions, data and documentation.
\item You can produce packages for private or in-house use or you
can contribute your package to the Comprehensive R Archive Network
(CRAN), \url{http://cran.R-project.org}
\item We will be using the \Emph{lme4a} package from R-forge.
Install it from the \Emph{Packages} menu item or with
<>=
install.packages("lme4a", repos="http://r-forge.r-project.org/")
@
\item You only need to install a package once. If a new version
becomes available you can update (see the menu item).
\item To use a package in an R session you attach it using
<>=
require(lme4a)
@
or
<>=
library(lme4a)
@
(Causing widespread confusion of the terms ``package'' and ``library''.)
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Accessing documentation}
\begin{itemize}
\item To be added to CRAN, a package must pass a series of quality
control checks. In particular, all functions and data sets must
be documented. Examples and tests can also be included.
\item The \code{data} function provides names and brief descriptions
of the data sets in a package.
\begin{Schunk}
\begin{Sinput}
> data(package = "lme4a")
\end{Sinput}
\begin{Soutput}
Data sets in package 'lme4a':
Dyestuff Yield of dyestuff by batch
Dyestuff2 Yield of dyestuff by batch
Pastes Paste strength by batch and cask
Penicillin Variation in penicillin testing
cake Breakage angle of chocolate cakes
cbpp Contagious bovine pleuropneumonia
sleepstudy Reaction times in a sleep deprivation study
\end{Soutput}
\end{Schunk}
\item Use \code{?} followed by the name of a function or data set to
view its documentation. If the documentation contains an example
section, you can execute it with the \code{example} function.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Effects - fixed and random}
\begin{itemize}
\item Mixed-effects models, like many statistical models, describe
the relationship between a \Emph{response} variable and one or
more \Emph{covariates} recorded with it.
\item The models we will discuss are based on a \Emph{linear
predictor} expression incorporating \Emph{coefficients} that are
estimated from the observed data.
\item Coefficients associated with the levels of a categorical
covariate are sometimes called the \Emph{effects} of the levels.
\item When the levels of a covariate are fixed and reproducible
(e.g. a covariate \code{sex} that has levels \code{male} and
\code{female}) we incorporate them as fixed-effects parameters.
\item When the levels of a covariate correspond to the particular
observational or experimental units in the experiment we
incorporate them as \code{random effects}.
\end{itemize}
\end{frame}
\section[Dyestuff]{The Dyestuff data and model}
\begin{frame}[fragile]
\frametitle{The Dyestuff data set}
\begin{itemize}
\item The \code{Dyestuff}, \code{Penicillin} and \code{Pastes} data
sets all come from the classic book \Emph{Statistical Methods in
Research and Production}, edited by O.L. Davies and first published
in 1947.
\item The \code{Dyestuff} data are a balanced one-way classification
of the \code{Yield} of dyestuff from samples produced from six
\code{Batch}es of an intermediate product. See \code{?Dyestuff}.
\end{itemize}
<>=
str(Dyestuff)
summary(Dyestuff)
@
\end{frame}
\begin{frame}
\frametitle{The effect of the batches}
\begin{itemize}
\item To emphasize that \code{Batch} is categorical, we use letters
instead of numbers to designate the levels.
\item Because there is no inherent ordering of the levels of
\code{Batch}, we will reorder the levels if, say, doing so can
make a plot more informative.
\item The particular batches observed are just a selection of the
possible batches and are entirely used up during the course of
the experiment.
\item It is not particularly important to estimate and compare
yields from these batches. Instead we wish to estimate the
variability in yields due to batch-to-batch variability.
\item The \code{Batch} factor will be used in \Emph{random-effects}
terms in models that we fit.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Dyestuff data plot}
\begin{center}
<>=
set.seed(1234543)
print(dotplot(reorder(Batch, Yield) ~ Yield, Dyestuff,
ylab = "Batch", jitter.y = TRUE, pch = 21, aspect = 0.32,
xlab = "Yield of dyestuff (grams of standard color)",
type = c("p", "a")))
@
\end{center}
\begin{itemize}
\item The line joins the mean yields of the six batches, which have
been reordered by increasing mean yield.
\item The vertical positions are jittered slightly to reduce
overplotting. The lowest yield for batch A was observed on two
distinct preparations from that batch.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{A mixed-effects model for the dyestuff yield}
<>=
fm1 <- lmer(Yield ~ 1 + (1|Batch), Dyestuff)
print(fm1)
@
\begin{itemize}
\item Fitted model \code{fm1} has one fixed-effect parameter, the mean
yield, and one random-effects term, generating a simple, scalar
random effect for each level of \code{Batch}.
\end{itemize}
\end{frame}
<>=
op <- options(digits=5)
@
\begin{frame}[fragile]
\frametitle{Extracting information from the fitted model}
\begin{itemize}
\item \code{fm1} is an object of class \code{"merMod"}
\item There are many \Emph{extractor} functions that can be applied
to such objects.
\end{itemize}
<>=
fixef(fm1)
ranef(fm1, drop = TRUE)
fitted(fm1)
@
\end{frame}
<>=
options(op)
@
\section[Mixed models]{Definition of mixed-effects models}
\begin{frame}[fragile]
\frametitle{Definition of mixed-effects models}
\begin{itemize}
\item Models with random effects are often written like
\begin{displaymath}
y_{ij}=\mu+b_i+\epsilon_{ij},\;
b_i\sim\mathcal{N}(0,\sigma_b^2),
\epsilon_{ij}\sim\mathcal{N}(0,\sigma^2),i=1,\dots,I,j=1,\dots,J_i
\end{displaymath}
\item This scalar notation quickly becomes unwieldy, degenerating
into ``subscript fests''. We will use a vector/matrix notation.
\item A mixed-effects model incorporates two vector-valued random
variables: the response vector, $\bc{Y}$, and the random effects
vector, $\bc{B}$. We observe the value, $\bm y$, of $\bc{Y}$. We
do not observe the value of $\bc{B}$.
\item In the models we will consider, the random effects are modeled
as a multivariate Gaussian (or ``normal'') random variable,
$\bc{B}\sim\mathcal{N}(\bm 0,\bm\Sigma(\bm\theta))$, where
$\bm\theta$ is a vector of \Emph{variance-component parameters}.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Linear mixed models}
\begin{itemize}
\item The conditional distribution, $(\bc Y|\bc B=\bm b)$, depends on
$\bm b$ only through its mean, $\bm\mu_{\bc Y|\bc B=\bm b}$.
\item The conditional mean, $\bm\mu_{\bc Y|\bc B=\bm b}$, depends on
$\bm b$ and on the fixed-effects parameter vector, $\bm\beta$, through a
\Emph{linear predictor} expression, $\bm Z\bm b+\bm X\bm\beta$.
The \Emph{model matrices} $\bm Z$ and $\bm X$ are determined from the
form of the model and the values of the covariates.
\item In a \Emph{linear mixed model} the conditional distribution is
a ``spherical'' multivariate Gaussian
\begin{displaymath}
(\bc{Y}|\bc{B}=\bm b)\sim\mathcal{N}(\bm Z\bm b+\bm X\bm\beta,
\sigma^2\bm I_n)
\end{displaymath}
\item The scalar $\sigma$ is the \Emph{common scale parameter}; the
dimension of $\bm y$ is $n$, $\bm b$ is $q$ and $\bm\beta$ is $p$,
hence $\bm Z$ is $n\times q$ and $\bm X$ is $n\times p$.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Simple, scalar random effects terms}
\begin{itemize}
\item A term like \code{(1|Batch)} in an \code{lmer} formula is
called a simple, scalar random-effects term.
\item The expression on the right of the \code{"|"} operator
(usually just the name of a variable) is evaluated as a factor,
called the \Emph{grouping factor} for the term.
\item Suppose we have $k$ such terms with $n_i,i=1,\dots,k$ levels
in the $i$th term's grouping factor. A scalar random-effects term
generates one random effect for each level of the grouping factor.
If all the random effects terms are scalar terms then
$q=\sum_{i=1}^kn_i$.
\item The model matrix $\bm Z$ is the horizontal concatenation of $k$
matrices. For a simple, scalar term, the $i$th vertical slice,
which has $n_i$ columns, is the indicator columns for the $n_i$
levels of the $i$th grouping factor.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Structure of the unconditional variance-covariance}
\begin{itemize}
\item Just as the matrix $\bm Z$ is the horizontal concatenation of
matrices generated by individual random-effects terms, the
(unconditional) variance-covariance matrix, $\bm\Sigma$, is
block-diagonal in $k$ blocks. In other words, the unconditional
distributions of random effects from different terms in the model
are independent. (However, the conditional distributions, given
the observed data, $(\bc B|\bc Y=\bm y)$, are not independent.)
\item If the $i$th term is a simple, scalar term then the $i$th
diagonal block is a multiple of the identity, $\sigma_i^2\bm I_{n_i}$.
\item This means that unconditional distributions of random effects
corresponding to different levels of the grouping factor are
independent.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Model matrices for model fm1}
\begin{itemize}
\item The formula for model \code{fm1} has a single fixed-effects
term, \code{1}, and one simple, scalar random-effects term,
\code{(1|Batch)}.
\item The model matrix, $\bm Z$, whose transpose is stored in a slot
called \code{Zt}, is the matrix of indicators for the six levels
of \code{Batch}.
\item The model matrix, $\bm X$, is $30\times 1$. All its elements
are unity.
\end{itemize}
<>=
str(as.matrix(model.matrix(fm1)))
fm1@re@Zt
@
\end{frame}
\begin{frame}
\frametitle{Conditional means of the random effects}
\begin{itemize}
\item Technically we do not provide ``estimates'' of the random
effects because they are not parameters.
\item One answer to the question, ``so what are those numbers
provided by \code{ranef} anyway?'' is that they are BLUPs (Best
Linear Unbiased Predictors) of the random effects. The acronym is
attractive but not very informative (what is a ``linear unbiased
predictor'' and in what sense are these the ``best''?). Also, the
concept does not generalize.
\item A better answer is that those values are the conditional
means, $\bm\mu_{\bc B|\bc Y =\bm y}$, evaluated at the estimated
parameter values. Regrettably, we can only evaluate the
conditional means for linear mixed models.
\item However, these values are also the conditional modes and that
concept does generalize to other types of mixed models.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Caterpillar plot for fm1}
\begin{itemize}
\item For linear mixed models the conditional distribution of the
random effects, given the data, written $(\bc B|\bc Y=\bm y)$, is
again a multivariate Gaussian distribution.
\item We can evaluate the means and standard deviations of the
individual conditional distributions, $(\bc B_j|\bc Y=\bm y), j =
1,\dots,q$. We show these in the form of a 95\% prediction
interval, with the levels of the grouping factor arranged in
increasing order of the conditional mean.
\item These are sometimes called ``caterpillar plots''.
\end{itemize}
\begin{center}
<>=
print(dotplot(ranef(fm1, postVar = TRUE), strip = FALSE)[[1]])
@
\end{center}
\end{frame}
\begin{frame}
\frametitle{REML estimates versus ML estimates}
\begin{itemize}
\item The default parameter estimation criterion for linear mixed
models is restricted (or ``residual'') maximum likelihood (REML).
\item Maximum likelihood (ML) estimates (sometimes called ``full
maximum likelihood'') can be requested by specifying \code{REML =
FALSE} in the call to \code{lmer}.
\item Generally REML estimates of variance components are
preferred. ML estimates are known to be biased. Although REML
estimates are not guaranteed to be unbiased, they are usually less
biased than ML estimates.
\item Roughly, the difference between REML and ML estimates of
variance components is comparable to estimating $\sigma^2$ in a
fixed-effects regression by $\mathit{SSR}/(n-p)$ versus
$\mathit{SSR}/n$, where $\mathit{SSR}$ is the residual sum of
squares.
\item For a balanced, one-way classification like the
\code{Dyestuff} data, the REML and ML estimates of the fixed-effects
are identical.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Re-fitting the model for ML estimates}
<>=
(fm1M <- update(fm1, REML = FALSE))
@
(The extra parentheses around the assignment cause the value to
be printed. Generally the results of assignments are not printed.)
\end{frame}
\begin{frame}
\frametitle{Verbose fitting}
\begin{itemize}
\item When fitting a large model or if the estimates of the variance
components seem peculiar, it is a good idea to monitor the
progress of the iterations in optimizing the deviance or the REML
criterion.
\item The optional argument \code{verbose = TRUE} causes \code{lmer}
to print iteration information during the optimzation of the
parameter estimates.
\item The quantity being minimized is the \Emph{profiled deviance}
or the \Emph{profiled REML criterion} of the model. The deviance
is negative twice the log-likelihood. It is profiled in the sense
that it is a function of $\bm\theta$ only --- $\bm\beta$ and
$\sigma$ are at their conditional estimates.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Obtain the verbose output for fitting fm1}
<>=
invisible(update(fm1, verbose = TRUE))
@
\begin{itemize}
\item The lines include a ``trust region'' radius, the cumulative
number of function evaluations, the profiled deviance or profiled
REML criterion (i.e. the quantity being minimized) and the
components of the parameter vector,
$\bm\theta$.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Estimates of variance components can be zero}
\begin{itemize}
\item We have been careful to state the variance of the random
effects is $\ge0$.
\item For some data sets the maximum likelihood or REML estimate,
$\widehat{\sigma_b^2}$ ends up as exactly zero. That is, the
optimal parameter value is on the boundary of the region of
allowable values.
\item Box and Tiao (1973) provide simulated data with a structure
like the \code{Dyestuff} data illustrating this.
\end{itemize}
<>=
str(Dyestuff2)
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Plot of the Dyestuff2 data}
\begin{center}
<>=
print(dotplot(reorder(Batch, Yield) ~ Yield, Dyestuff2,
ylab = "Batch", jitter.y = TRUE, pch = 21, aspect = 0.42,
xlab = "Simulated response (dimensionless)",
type = c("p", "a")))
@
\end{center}
\begin{itemize}
\item For these data the batch-to-batch variability is not large
compared to the within-batch variability.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Fitting the model to Dyestuff2}
<>=
(fm1A <- lmer(Yield ~ 1 + (1|Batch), Dyestuff2, REML=FALSE))
@
\end{frame}
\begin{frame}[fragile]
\frametitle{A trivial mixed-effects model is a fixed-effects model}
\begin{itemize}
\item The mixed model \code{fm1A} with an estimated variance
$\widehat{\sigma_b^2}=0$ is equivalent to a model with only
fixed-effects terms.
\end{itemize}
<>=
summary(lm1 <- lm(Yield ~ 1, Dyestuff2))
logLik(lm1)
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Recap of the Dyestuff model}
\begin{itemize}
\item The model is fit as
<>=
fm1@call
@
\item There is one random-effects term, \code{(1|Batch)}, in the model
formula. It is a simple, scalar term for the grouping factor
\code{Batch} with $n_1=6$ levels. Thus $q=6$.
\item The model matrix $\bm Z$ is the $30\times 6$ matrix of
indicators of the levels of \code{Batch}.
\item The variance-covariance matrix, $\bm\Sigma$, is a
nonnegative multiple of the $6\times 6$ identity matrix, $\bm I_6$.
\item The fixed-effects parameter vector, $\bm\beta$, is of length
$p=1$. All the elements of the $30\times 1$ model matrix $\bm X$
are unity.
\end{itemize}
\end{frame}
\section[Penicillin]{Crossed random-effects grouping: Penicillin}
\begin{frame}[fragile]
\frametitle{The Penicillin data (see also the ?Penicillin description)}
<>=
str(Penicillin)
xtabs(~ sample + plate, Penicillin)
@
\begin{itemize}
\item These are measurements of the potency (measured by the diameter
of a clear area on a Petri dish) of penicillin samples in a
balanced, unreplicated two-way crossed classification with the test
medium, \code{plate}.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Penicillin data plot}
\begin{center}
<>=
print(dotplot(reorder(plate, diameter) ~ diameter, Penicillin, groups = sample,
ylab = "Plate", xlab = "Diameter of growth inhibition zone (mm)",
type = c("p", "a"), auto.key = list(columns = 6, lines = TRUE)))
@
\end{center}
\end{frame}
\begin{frame}[fragile]
\frametitle{Model with crossed simple random effects for Penicillin}
<>=
(fm2 <- lmer(diameter ~ 1 + (1|plate) + (1|sample), Penicillin))
@
\end{frame}
<>=
op <- options(digits = 5)
@
\begin{frame}[fragile]
\frametitle{Random effects for fm2}
\begin{itemize}
\item The model for the $n=144$ observations has $p=1$ fixed-effects
parameter and $q=30$ random effects from $k=2$ random effects
terms in the formula.
\end{itemize}
<>=
ranef(fm2, drop = TRUE)
@
\end{frame}
<>=
options(op)
@
\begin{frame}[fragile]
\frametitle{Prediction intervals for random effects}
<>=
qrr2 <- dotplot(ranef(fm2, postVar = TRUE), strip = FALSE)
print(qrr2[[1]], pos = c(0,0,1,0.75), more = TRUE)
print(qrr2[[2]], pos = c(0,0.65,1,1))
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Model matrix Z for fm2}
\begin{itemize}
\item Because the model matrix $\bm Z$ is generated from $k=2$
simple, scalar random effects terms, it consists of two sets of
indicator columns.
\item The structure of $\bm Z\trans$ is shown below. (Generally we
will show the transpose of these model matrices - they fit better
on slides.)
\end{itemize}
\begin{center}
<>=
print(image(fm2@re@Zt, xlab = NULL, ylab = NULL, sub = "Z'"))
@
\end{center}
\end{frame}
\begin{frame}
\frametitle{Models with crossed random effects}
\begin{itemize}
\item Many people believe that mixed-effects models are equivalent
to hierarchical linear models (HLMs) or ``multilevel models''.
This is not true. The \code{plate} and \code{sample} factors in
\code{fm2} are crossed. They do not represent levels in a hierarchy.
\item There is no difficulty in defining and fitting models with
crossed random effects (meaning random-effects terms whose
grouping factors are crossed). However, fitting models with
crossed random effects can be somewhat slower.
\item The crucial calculation in each \code{lmer} iteration is
evaluation of a $q\times q$ sparse, lower triangular, Cholesky
factor, $\bm L(\bm\theta)$, derived from $\bm Z$ and
$\bm\Sigma(\bm\theta)$. Crossing of grouping factors increases
the number of nonzeros in $\bm L(\bm\theta)$ and causes some
``fill-in'' of $\bm L$ relative to $\bm Z\trans\bm Z$.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{All HLMs are mixed models but not vice-versa}
\begin{itemize}
\item Even though Raudenbush and Bryk (2002) do discuss models for
crossed factors in their HLM book, such models are not
hierarchical.
\item Experimental situations with crossed random factors, such as
``subject'' and ``stimulus'', are common. We can, and should, model
such data according to its structure.
\item In longitudinal studies of subjects in social contexts (e.g.
students in classrooms or in schools) we almost always have partial
crossing of the subject and the context factors, meaning that, over
the course of the study, a particular student may be observed in
more than one class but not all students are
observed in all classes. The student and class factors are
neither fully crossed nor strictly nested.
\item For longitudinal data, ``nested'' is only important if it means
``nested across time''. ``Nested at a particular time'' doesn't
count.
\item \code{lme4} handles fully or partially crossed factors gracefully.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Recap of the Penicillin model}
\begin{itemize}
\item The model formula is
<>=
fm2@call[["formula"]]
@
\item There are two random-effects terms, \code{(1|plate)} and
\code{(1|sample)}. Both are simple, scalar random effects terms,
with $n_1=24$ and $n_2=6$ levels, respectively. Thus
$q=q_1n_1+q_2n_2=30$.
\item The model matrix $\bm Z$ is the $144\times 30$ matrix created
from two sets of indicator columns.
\item The relative variance-covariance matrix, $\bm\Sigma$, is block
diagonal in two blocks that are nonnegative multiples of identity
matrices.
\item The fixed-effects parameter vector, $\bm\beta$, is of length
$p=1$. All the elements of the $144\times 1$ model matrix $\bm X$
are unity.
\end{itemize}
\end{frame}
\section[Pastes]{Nested random-effects grouping: Pastes}
\begin{frame}[fragile]
\frametitle{The Pastes data (see also the ?Pastes description)}
<>=
str(Pastes)
xtabs(~ batch + sample, Pastes, sparse = TRUE)
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Structure of the Pastes data}
\begin{itemize}
\item The \code{sample} factor is nested within the \code{batch}
factor. Each sample is from one of three casks selected from a
particular batch.
\item Note that there are 30, not 3, distinct samples.
\item We can label the casks as `a', `b' and `c' but then the
\code{cask} factor by itself is meaningless (because cask `a' in
batch `A' is unrelated to cask `a'in batches `B', `C', $\dots$).
The \code{cask} factor is only meaningful within a \code{batch}.
\item Only the \code{batch} and \code{cask} factors, which are
apparently crossed, were present in the original data set.
\code{cask} may be described as being nested within \code{batch}
but that is not reflected in the data. It is \Emph{implicitly
nested}, not explicitly nested.
\item You can save yourself a lot of grief by immediately creating
the explicitly nested factor. The recipe is
\end{itemize}
<>=
Pastes <- within(Pastes, sample <- factor(batch:cask))
@
\end{frame}
\begin{frame}
\frametitle{Avoid implicitly nested representations}
\begin{itemize}
\item The \code{lme4} package allows for very general model
specifications. It does not require that factors associated with
random effects be hierarchical or ``multilevel'' factors in the
design.
\item The same model specification can be used for data with nested
or crossed or partially crossed factors. Nesting or crossing is
determined from the structure of the factors in the data, not the
model specification.
\item You can avoid confusion about nested and crossed factors by
following one simple rule: ensure that different levels of a
factor in the experiment correspond to different labels of the
factor in the data.
\item Samples were drawn from 30, not 3, distinct casks in this
experiment. We should specify models using the \code{sample}
factor with 30 levels, not the \code{cask} factor with 3 levels.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Pastes data plot}
<>=
Pastes <- within(Pastes, bb <- reorder(batch, strength))
Pastes <- within(Pastes, ss <- reorder(reorder(sample, strength),
as.numeric(batch)))
print(dotplot(ss ~ strength | bb, Pastes,
strip = FALSE, strip.left = TRUE, layout = c(1, 10),
scales = list(y = list(relation = "free")),
ylab = "Sample within batch", type = c("p", "a"),
xlab = "Paste strength", jitter.y = TRUE))
@
\end{frame}
\begin{frame}[fragile]
\frametitle{A model with nested random effects}
<>=
(fm3 <- lmer(strength ~ 1 + (1|batch) + (1|sample), Pastes))
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Random effects from model fm3}
\begin{center}
<>=
qrr3 <- dotplot(ranef(fm3, postVar = TRUE), strip = FALSE)
print(qrr3[[1]], pos = c(0,0,1,0.75), more = TRUE)
print(qrr3[[2]], pos = c(0,0.65,1,1))
@
\end{center}
Batch-to-batch variability is low compared to sample-to-sample.
\end{frame}
\begin{frame}[fragile]
\frametitle{Dimensions and relationships in fm3}
\begin{itemize}
\item There are $n=60$ observations, $p=1$ fixed-effects parameter,
$k=2$ simple, scalar random-effects terms ($q_1=q_2=1$) with
grouping factors having $n_1=30$ and $n_2=10$ levels.
\item Because both random-effects terms are simple, scalar terms,
$\bm\Sigma(\bm\theta)$ is block-diagonal in two diagonal blocks of
sizes $30$ and $10$, respectively. $\bm Z$ is generated from two
sets of indicators.
\end{itemize}
\begin{center}
<>=
print(image(fm3@re@Zt, xlab = NULL, ylab = NULL, sub = NULL))
@
\end{center}
\end{frame}
\begin{frame}
\frametitle{Eliminate the random-effects term for batch?}
\begin{itemize}
\item We have seen that there is little batch-to-batch variability
beyond that induced by the variability of samples within batches.
\item We can fit a reduced model without that term and compare it to
the original model.
\item Somewhat confusingly, model comparisons from likelihood ratio
tests are obtained by calling the \code{anova} function on the two
models. (Put the simpler model first in the call to \code{anova}.)
\item Sometimes likelihood ratio tests can be evaluated using the REML
criterion and sometimes they can't. Instead of learning the rules
of when you can and when you can't, it is easiest always to refit the
models with \code{REML = FALSE} before comparing.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Comparing ML fits of the full and reduced models}
<>=
fm3M <- update(fm3, REML = FALSE)
fm4M <- lmer(strength ~ 1 + (1|sample),
Pastes, REML = FALSE)
anova(fm4M, fm3M)
@
\end{frame}
\begin{frame}
\frametitle{p-values of LR tests on variance components}
\begin{itemize}
\item The likelihood ratio is a reasonable criterion for comparing
these two models. However, the theory behind using a $\chi^2$
distribution with 1 degree of freedom as a reference distribution
for this test statistic does not apply in this case. The null
hypothesis is on the boundary of the parameter space.
\item Even at the best of times, the p-values for such tests are only
approximate because they are based on the asymptotic behavior of the
test statistic. To carry the argument further, all results in
statistics are based on models and, as George Box famously said,
``All models are wrong; some models are useful.''
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{LR tests on variance components (cont'd)}
\begin{itemize}
\item In this case the problem with the boundary condition results in
a p-value that is larger than it would be if, say, you compared this
likelihood ratio to values obtained for data simulated from the null
hypothesis model. We say these results are ``conservative''.
\item As a rule of thumb, the p-value for the $\chi^2$ test on a
simple, scalar term is roughly twice as large as it should be.
\item In this case, dividing the p-value in half would not affect our
conclusion.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Updated model, REML estimates}
<>=
(fm4 <- update(fm4M, REML = TRUE))
@
\end{frame}
\begin{frame}
\frametitle{Recap of the analysis of the Pastes data}
\begin{itemize}
\item The data consist of $n=60$ observations on $n_1=30$ samples
nested within $n_2=10$ batches.
\item The data are labelled with a \code{cask} factor with $3$
levels but that is an implicitly nested factor. Create the
explicit factor \code{sample} and ignore \code{cask} from then
on.
\item Specification of a model for nested factors is exactly the
same as specification of a model with crossed or partially crossed
factors --- provided that you avoid using implicitly nested factors.
\item In this case the \code{batch} factor was inert --- it did not
``explain'' substantial variability in addition to that attributed
to the \code{sample} factor. We therefore prefer the simpler model.
\item At the risk of ``beating a dead horse'', notice that, if we had
used the \code{cask} factor in some way, we would still need to
create a factor like \code{sample} to be able to reduce the
model. The \code{cask} factor is only meaningful within \code{batch}.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{This is all very nice, but $\dots$}
\begin{itemize}
\item These methods are interesting but the results are not really
new. Similar results are quoted in \Emph{Statistical Methods in
Research and Production}, which is a very old book.
\item The approach described in that book is actually quite
sophisticated, especially when you consider that the methods
described there, based on observed and expected mean squares, are
for hand calculation --- in pre-calculator days!
\item Why go to all the trouble of working with sparse matrices and
all that if you could get the same results with paper and pencil?
The one-word answer is \Emph{balance}.
\item Those methods depend on the data being balanced. The design
must be completely balanced and the resulting data must also be
completely balanced.
\item Balance is fragile. Even if the design is balanced, a single
missing or questionable observation destroys the balance.
Observational studies (as opposed to, say, laboratory experiments)
cannot be expected to yield balanced data sets.
\item Also, the models involve only simple, scalar random effects
and do not incorporate covariates.
\end{itemize}
\end{frame}
\section[Fixed-effects]{Incorporating fixed-effects terms: classroom}
\begin{frame}[fragile]
\frametitle{Structure of the classroom data}
\begin{itemize}
\item The \code{classroom} data are a cross-section of students
within classes within schools. The \code{mathgain} variable is
the difference in mathematics achievement scores in grade 1 and
kindergarten.
\item These data are quite unbalanced. The distribution of the
number of students observed per classroom is
<>=
xtabs( ~ xtabs(~ classid, classroom))
@
\item Similarly, the distribution of the number of classes observed
per school is
<>=
table(xtabs(~ schoolid,
unique(subset(classroom, select = c(classid, schoolid)))))
@
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Twelve schools, each with 5 classrooms}
<>=
refactor <- function(x) if(is.factor(x)) factor(x) else x
sch12 <- do.call(data.frame,
lapply(subset(classroom,
schoolid %in% c(12,15, 17, 33,46, 57,
68, 70, 71, 76, 85, 99)),
refactor))
sch12 <- within(sch12, ss <- reorder(schoolid, mathgain))
sch12 <- within(sch12, cc <- reorder(reorder(classid, mathgain),
as.numeric(schoolid)))
print(dotplot(cc ~ mathgain | ss , sch12,
strip = FALSE, strip.left = TRUE, layout = c(1, 12),
scales = list(y = list(relation = "free")), pch = 21,
ylab = "Class within school", type = c("p", "a"),
xlab = "Mathematics gain from kindergarten to grade 1",
jitter.y = TRUE))
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Simple, ``unconditional'' model for the classroom data}
<>=
(fm5 <- lmer(mathgain ~ 1 + (1|classid) + (1|schoolid),
classroom))
@
\end{frame}
\begin{frame}
\frametitle{Some comments on the ``unconditional'' model}
\begin{itemize}
\item In the multilevel modeling literature a model such as
\code{fm5} that does not incorporate fixed-effects terms for
demographic characteristics of the student, class or school, is
called an ``unconditional'' model.
\item Notice that the dominant level of variability is the residual
variability. It is unlikely that random effects for both classes
and schools are needed when modeling these data.
\item We have seen in Exercises 2 that there seem to be trends with
respect to the \code{minority} factor and the \code{mathkind}
score but no overall trends with respect to \code{sex}.
\item A coefficient for a continuous covariate, such as
\code{mathkind}, or for fixed, reproducible levels of a factor
like \code{sex} or \code{minority} is incorporated in the
fixed-effects terms.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Model-building approach}
\begin{itemize}
\item Note that these unbalanced data have, for the most part,
very few classes per school (sometimes as few as 1) and very few
students per class (also sometimes as few as 1). Under these
circumstances, it is optimistic to expect to be able to partition
the variability across students, classes and schools.
\item We should consider adding fixed-effects terms and perhaps
removing one of the random-effects terms.
\item We will start by incorporating fixed-effects terms then
revisit the need for both random-effects terms.
\item We will begin with the fixed-effects terms adopted as a final
model in chapter 4 of West, Welch and Ga\l{}ecki (2007).
\item For brevity, we only display the output of model fits as this
contains enough information to reconstruct the call to \code{lmer}.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{A model with fixed-effects terms}
<>=
print(fm6 <- lmer(mathgain ~ 1 + mathkind + minority + sex + ses + housepov
+ (1|classid) + (1|schoolid), classroom), corr = FALSE)
@
\end{frame}
\begin{frame}
\frametitle{Where are the p-values?!!}
\begin{itemize}
\item The first thing that most users notice is that there are no
p-values for the fixed-effects coefficients! Calculating a p-value
for $H_0:\beta_j=0$ versus $H_a:\beta_j\ne0$ is not as
straightforward as it may seem. The ratio called a ``t value'' in
the output does not have a Student's T distribution under the null
hypothesis.
\item For simple models fit to small, balanced data sets one can
calculate a p-value. Not so for unbalanced data. When the number
of groups and observations are large, approximations don't matter
--- you can consider the ratio as having a standard normal
distribution.
\item The only time that you can calculate an ``exact'' p-value
and the difference between this and the standard normal dist'n is
important is for small, balanced data sets, which are
exactly the cases that appear in text books. People get very,
very upset if the values calculated by the software don't agree
perfectly with the text book answers.
\item Here, just say a coefficient is ``significant'' if $|t|> 2$.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Removing the insignificant term for sex}
<>=
print(fm7 <- lmer(mathgain ~ 1 + mathkind + minority + ses + housepov
+ (1|classid) + (1|schoolid), classroom), corr = FALSE)
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Removing the insignificant term for housepov}
<>=
print(fm8 <- lmer(mathgain ~ mathkind + minority + ses
+ (1|classid) + (1|schoolid), classroom), corr = FALSE)
@
\end{frame}
\begin{frame}
\frametitle{Prediction intervals on random effects for class}
\begin{center}
<>=
print(dotplot(ranef(fm8, post = "TRUE"), strip = FALSE,
scales = list(y = list(draw = FALSE)))$classid)
@
\end{center}
\end{frame}
\begin{frame}[fragile]
\frametitle{Normal probability plot of random effects for class}
With many levels of the grouping factor, use a normal probability
plot of the prediction intervals for the random effects.
<>=
qqmath(ranef(fm8, post = TRUE))$classid
@
\begin{center}
<>=
print(qqmath(ranef(fm8, post = TRUE),strip=FALSE)$classid)
@
\end{center}
\end{frame}
\begin{frame}[fragile]
\mode{\frametitle{Normal probability plot of random effects for school}}
\begin{figure}[htb]
\centering
<>=
print(qqmath(ranef(fm8, post = TRUE),strip=FALSE)$schoolid)
@
\mode{\caption{Normal probability plot of random effects for school}}
\label{fig:Schoolpred}
\end{figure}
\end{frame}
\begin{frame}[fragile]
\frametitle{Refit without random effects for class}
<>=
print(fm9M <- lmer(mathgain ~ mathkind + minority + ses
+ (1|schoolid), classroom, REML = FALSE), corr = FALSE)
@
\end{frame}
\begin{frame}[fragile]
\frametitle{Check if random effects for class are significant}
<>=
fm8M <- update(fm8, REML = FALSE)
anova(fm9M, fm8M)
@
\begin{itemize}
\item Contrary to what we saw in the plots, the random-effects term
for \code{classid} is significant even in the presence of the
\code{schoolid} term
\item Part of the reason for this inconsistency is our incorporating
312 random effects at a ``cost'' of 1 parameter. In some way we are
undercounting the number of degrees of freedom added to the model
with this term.
\end{itemize}
\end{frame}
\begin{frame}[fragile]\frametitle{A large observational data set}
\begin{itemize}
\item A large U.S. university (not mine) provided data on the grade
point score (\code{gr.pt}) by student (\code{id}), instructor
(\code{instr}) and department (\code{dept}) from a 10 year period.
I regret that I cannot make these data available to others.
\item These factors are unbalanced and partially crossed.
\end{itemize}
\begin{Schunk}
\begin{Sinput}
> str(anon.grades.df)
\end{Sinput}
\begin{Soutput}
'data.frame': 1721024 obs. of 9 variables:
$ instr : Factor w/ 7964 levels "10000","10001",..: 1 1 1 1 1 1 1 1 1 1 ...
$ dept : Factor w/ 106 levels "AERO","AFAM",..: 43 43 43 43 43 43 43 43 43 43 ...
$ id : Factor w/ 54711 levels "900000001","900000002",..: 12152 1405 23882 18875 18294 20922 4150 13540 5499 6425 ...
$ nclass : num 40 29 33 13 47 49 37 14 21 20 ...
$ vgpa : num NA NA NA NA NA NA NA NA NA NA ...
$ rawai : num 2.88 -1.15 -0.08 -1.94 3.00 ...
$ gr.pt : num 4 1.7 2 0 3.7 1.7 2 4 2 2.7 ...
$ section : Factor w/ 70366 levels "19959 AERO011A001",..: 18417 18417 18417 18417 9428 18417 18417 9428 9428 9428 ...
$ semester: num 19989 19989 19989 19989 19972 ...
\end{Soutput}
\end{Schunk}%$
\end{frame}
\begin{frame}[fragile]\frametitle{A preliminary model}
\begin{Schunk}
\begin{Soutput}
Linear mixed model fit by REML
Formula: gr.pt ~ (1 | id) + (1 | instr) + (1 | dept)
Data: anon.grades.df
AIC BIC logLik deviance REMLdev
3447389 3447451 -1723690 3447374 3447379
Random effects:
Groups Name Variance Std.Dev.
id (Intercept) 0.3085 0.555
instr (Intercept) 0.0795 0.282
dept (Intercept) 0.0909 0.301
Residual 0.4037 0.635
Number of obs: 1685394, groups: id, 54711; instr, 7915; dept, 102
Fixed effects:
Estimate Std. Error t value
(Intercept) 3.1996 0.0314 102
\end{Soutput}
\end{Schunk}
\end{frame}
\begin{frame}
\frametitle{Comments on the model fit}
\begin{itemize}
\item $n=1685394$, $p=1$, $k=3$, $n_1=54711$, $n_2=7915$, $n_3=102$,
$q_1=q_2=q_3=1$, $q=62728$
\item This model is sometimes called the ``unconditional'' model in
that it does not incorporate covariates beyond the grouping factors.
\item It takes less than an hour to fit an "unconditional" model
with random effects for student (\code{id}), instructor
(\code{inst}) and department (\code{dept}) to these data.
\item Naturally, this is just the first step. We want to look at
possible time trends and the possible influences of the
covariates.
\item This is an example of what ``large'' and ``unbalanced'' mean
today. The size of the data sets and the complexity of the models
in mixed modeling can be formidable.
\end{itemize}
\end{frame}
\section[Large]{A model fit to a large data set}
\begin{frame}[fragile]
\frametitle{A model fit to a large data set (by today's standards)}
\begin{itemize}
\item Harold Doran recently fit a linear mixed model to the annual
achievement test results for the last 4 years in one of the United
States. There were $n=5212017$ observations on a total of
$n_1=1876788$ students and $n_2=47480$ teachers.
\item The models had simple, scalar random effects for student and
for teacher resulting in $q=1924268$ (i.e. nearly 2 million!)
\item There were a total of $p=29$ fixed-effects parameters.
\item At present Harold needed to fit the model to a subset and only
evaluate the conditional means for all the students and teachers
but we should be able to get around that limitation and actually
fit the model to all these responses and random effects.
\item I don't know of other software that can be used to fit a
model this large.
\end{itemize}
\end{frame}
\begin{frame}[fragile]
\frametitle{Size of the decomposition for this large model}
\begin{itemize}
\item The limiting factor on the memory size in such a model is the
Cholesky factor $\bm L(\bm\theta)$.
\item In this case the \code{x} slot is itself over 1GB in size and the \code{i} slot is over 0.5 GB.
\item These are close to an inherent limit on atomic R objects (the
range of an index into an atomic object cannot exceed $2^{31}$.
\end{itemize}
\begin{Schunk}
\begin{Sinput}
> str(L)
\end{Sinput}
\begin{Soutput}
Formal class 'dCHMsimpl' [package "Matrix"] with 10 slots
..@ x : num [1:174396181] 1.71 2.16 1.4 1.32 2.29 ...
..@ p : int [1:1924269] 0 2 4 5 7 9 10 12 14 15 ...
..@ i : int [1:174396181] 0 2 1 2 2 3 5 4 5 5 ...
..@ nz : int [1:1924268] 2 2 1 2 2 1 2 2 1 2 ...
..@ nxt : int [1:1924270] 1 2 3 4 5 6 7 8 9 10 ...
..@ prv : int [1:1924270] 1924269 0 1 2 3 4 5 6 7 8 ...
..@ colcount: int [1:1924268] 2 2 1 2 2 1 2 2 1 2 ...
..@ perm : int [1:1924268] 1922843 1886519 134451 1921046 1893309 183471 1912388 1888309 196670 1922626 ...
..@ type : int [1:4] 2 1 0 1
..@ Dim : int [1:2] 1924268 192426
\end{Soutput}
\end{Schunk}
\end{frame}
\begin{frame}
\frametitle{Recap of simple, scalar random-effects terms}
\begin{itemize}
\item For \code{lmer} a simple, scalar random effects term is of the
form \code{(1|F)}.
\item The number of random effects generated by the $i$th such
term is the number of levels, $n_i$, of \code{F} (after dropping
``unused'' levels --- those that do not occur in the data. The idea
of having such levels is not as peculiar as it may seem if, say,
you are fitting a model to a subset of the original data.)
\item Such a term contributes $n_i$ columns to $\bm Z$. These
columns are the indicator columns of the grouping factor.
\item Such a term contributes a diagonal block $\sigma^2_i\bm
I_{n_i}$ to $\bm\Sigma$. The multipliers $\sigma_i^2$ can be
different for different terms. The term contributes exactly one
element (which happens to be $\sigma_i/\sigma$) to $\bm\theta$.
\end{itemize}
\end{frame}
\end{document}