13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.

sgm LaTeX2e(2002/01/18) P1: IBD

10.1146/annurev.publhealth.25.102802.124410

Annu. Rev. Public Health 2004. 25:99–117

doi: 10.1146/annurev.publhealth.25.102802.124410
Copyright °
c 2004 by Annual Reviews. All rights reserved

WHAT DO WE DO WITH MISSING DATA? SOME

OPTIONS FOR ANALYSIS OF INCOMPLETE DATA
Trivellore E. Raghunathan
Department of Biostatistics and Institute for Social Research,
University of Michigan, Ann Arbor, Michigan 48109; email: [email protected]
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

■ Abstract Missing data are a pervasive problem in many public health investiga-
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

tions. The standard approach is to restrict the analysis to subjects with complete data
on the variables involved in the analysis. Estimates from such analysis can be biased,
especially if the subjects who are included in the analysis are systematically different
from those who were excluded in terms of one or more key variables. Severity of
bias in the estimates is illustrated through a simulation study in a logistic regression
setting. This article reviews three approaches for analyzing incomplete data. The first
approach involves weighting subjects who are included in the analysis to compensate
for those who were excluded because of missing values. The second approach is based
on multiple imputation where missing values are replaced by two or more plausible
values. The final approach is based on constructing the likelihood based on the incom-
plete observed data. The same logistic regression example is used to illustrate the basic
concepts and methodology. Some software packages for analyzing incomplete data are
described.

Key Words available-case analysis, observed data likelihood, missing data

mechanism, multiple imputation, nonresponse bias, weighting

INTRODUCTION

Missing data are a ubiquitous problem in public health investigations involving

human populations. In a cross-sectional study relying on a survey, subjects may
refuse to participate entirely or may not answer all the questions in the question-
naire. The former type of missing data is called unit nonresponse, and the latter,
item nonresponse. In a longitudinal study, subjects may drop out, be unable, or
refuse to participate in subsequent waves of data collection. The missing data in
this context may be viewed as unit or item nonresponse. For instance, in a cross-
sectional analysis of data from a particular wave, drop-outs may be viewed as unit
nonrespondents, whereas in a longitudinal analysis involving data from all waves,
missing data due to drop-outs may be viewed as item nonresponse.
The standard approach, a default option in many statistical packages, is to restrict
the analysis to subjects with no missing values in the specific set of variables. This
0163-7525/04/0421-0099$14.00 99
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

100 RAGHUNATHAN

so-called available-case analysis can yield biased estimates. Sometimes, where

multiple analyses are involved, including descriptive analysis, the standard ap-
proach excludes subjects with any missing values in any of the variables used in
at least one analysis (the so-called complete-case or listwise deletion analysis).
Other ad hoc approaches such as treating the missing data as a separate category
also result in biased estimates (35).
To demonstrate the potential bias in the available-case estimates, consider a
population with a binary disease variable, D, a binary exposure variable, E, and a
continuous confounder, x. Suppose that this population adheres to the following
model assumptions:
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

x ∼ N (0, 1),
logit Pr(E = 1|x) = 0.25 + 0.75 × x, and
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

logit Pr(D = 1|x, E) = −0.5 + 0.5 × E + 0.5 × x.

This model and the parameters were partly motivated by the analysis of cohort
data. A sample size of 1000 is drawn from essentially an infinite population, and
a logistic regression model is fitted, with D as the outcome variable and x and E
as predictors (the last equation in the above model assumptions), resulting in the
estimates of −0.54, 0.57, and 0.53 for the intercept, regression coefficients for
E and x, respectively. These estimates are close to the true values in the logistic
regression model given above.
Now suppose that some values of x are deliberately set to be missing using the
following logistic model,

logit[Pr(x is missing)] = −1.11 − 1.09 × D − 1.85 × E + 2.31 × D × E.

That is, for each subject, generate a uniform random number between 0 and
1, and if this number is less than or equal to the probability computed from the
above equation, then set that value of x to missing. This model for deleting values
was designed to produce approximately 15% missing values, and the percentage
of observations with missing values in the 4 cells formed by D and E was 26% for
(D = 0, E = 0), 11% for (D = 1, E = 0), 5% for (D = 0, E = 1), and 14% for
(D = 0, E = 0). These parameters were chosen to match the missing values in
the cohort data that motivated this experiment. The same logistic regression model
was fitted to the data set, now restricted only to those subjects with no missing
values in x. The resulting estimates of intercept, regression coefficients for E and x
were −0.30, 0.28, and 0.52, respectively. The estimate of the regression coefficient
for E is remarkably different when compared to the “before-deletion” estimate as
well as the true value of 0.5. One might wonder whether the observed difference is
due to idiosyncrasies of this particular data set or has occurred purely by chance.
To investigate further, the experiment was replicated as follows:
1. Generate 2500 samples (we will call these before-deletion samples) each of
size 1000.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 101

2. Delete some values of x in each data set using the same logistic model
mechanism given above (we will call the corresponding data sets with values
of x set to missing as “after-deletion” data sets).
3. Fit logistic regression models to both 2500 before-deletion and the corre-
sponding after-deletion data sets.
4. The primary parameter of interest is the regression coefficient for E, the
log-odds ratio measuring the association between D and E adjusted for x.
Figure 1 provides the histogram of 2500 estimated regression coefficients
from before-deletion data sets, and Figure 2 provides the histogram of the
corresponding 2500 available-case estimates from after-deletion data sets.
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

Figure 1 is centered on the true value 0.5 and is normal in shape, given
the large sample size (n = 1000). Figure 2, on the other hand, is centered
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

approximately at 0.20, and in fact, the true value lies in the tail of the sam-
pling distribution. This basic simulation study demonstrates that the standard
practice of omitting subjects with any missing values can be invalid.

This article reviews three approaches for correctly analyzing incomplete data,
and these will be evaluated using the same simulated data sets used to illustrate the

Figure 1 Histogram of logistic regression coefficient from 2500 simulated

data sets before deleting any covariate values.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

102 RAGHUNATHAN
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Figure 2 Histogram of logistic regression coefficient from 2500

simulated data sets after deleting some covariate values and using
available-cases.

perils of using the available-case approach. The first approach involves attaching
weights to each subject included in the analysis to represent subjects who were
excluded. This is often used to compensate for unit nonresponse in surveys [see
(9) for a review]. The second approach is through multiple imputation (29, 30),
where the missing set of values is replaced by more than one plausible set of
values. Each plausible set of values in conjunction with the observed data results
in a completed data set. Each completed data set is analyzed separately using the
standard complete data software, and the resulting point estimates and standard
errors are combined using a simple formula described later.
The distinction between the observed and filled-in values must be incorpo-
rated in any subsequent analysis of data with imputed values. That is, the filled-in
values for any one subject with missing values should not be considered as micro-
data for that subject but rather as values that are statistically plausible given other
information on that subject. These filled-in or completed data sets are plausible
samples from the population under certain assumptions. Thus, the completed data
sets should result in inferences (point estimates and confidence intervals, for ex-
ample) that are within the realm of statistical plausibility of inferences that would
have been obtained had there been no missing data. In that respect, the multiple
imputation approach is a statistical approach of “rectangularizing” the observed
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 103

data to exploit the available complete data software to obtain valid inferences. The
inferential validity based on the multiple imputed data sets is the goal, and any
imputation procedure should not viewed as a method for recovering the missing
values for any given individual.
The third approach is based on the likelihood constructed from the observed in-
complete data. This approach has a long history: The earliest reference seems to be
McKendrick (21), who used an algorithm similar to the Expectation-Maximization
(EM) algorithm (5) to obtain estimates from a sample with missing values. The
EM algorithm is a popular approach for maximizing the observed data likelihood.
This paper emphasizes weighting and imputation approaches and briefly discusses
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

the maximum likelihood approach. The maximum likelihood approach is available

only for very few selected models such as multivariate normal linear regression,
contingency tables, and certain other multivariate analyses. A detailed discussion
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

of the maximum likelihood approach can be found in Reference 19 and a Bayesian

approach similar in spirit in Reference 34.
All these methods make certain assumptions about why data are missing. The
answer to this question is stated in terms of assumptions about the missing data
mechanism. In the next section, this critical concept is described. The missing
data mechanisms comprise probabilistic specifications of why data are missing.
These mechanisms were developed by Rubin (28) and later extended to longitu-
dinal data by Little (17). All three approaches described above are valid under the
so-called missing at random (MAR) mechanism, and, generally, the available-case
approach is valid under a stricter mechanism called missing completely at random
(MCAR). The third section describes the weighting approach; the fourth section,
the multiple imputation approach; and the fifth section describes the likelihood-
based methods. The final section concludes with discussion and limitations.

MISSING DATA MECHANISM

To understand the taxonomy of missing data mechanisms, consider a simple case
where the analysis of interest is concerned with a set of variables (U, V), the
missing values are in a single variable U, and variables in vector V have no missing
values. Our simulation study falls into this category. Suppose that R denotes an
indicator variable taking the value 1 if U is observed and 0 if U is not observed.
Thus, the observed data are R, (U, V), if R = 1 and V, if R = 0. The missing data
mechanism or, equivalently, the response mechanism is the conditional probability
distribution of R, Pr(R|U,V), given (U, V). That is, what probabilistic mechanism
governs the observation of U, specifically in relation to the variables of interest U
and V?
The missing data mechanism is called MCAR, if Pr(R = 1|U,V) = c, a con-
stant. That is, the probability that U is observed is independent on the underlying
values of U or V. Under this assumption, the available-cases (that is, with R =
1) constitute a random subsample of the original sample. In a more general case,
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

104 RAGHUNATHAN

where missing values can also occur in V, the MCAR assumption implies that
the missing values in any variable are independent of the underlying values of U
or V. The subjects included in the available-case analysis, therefore, constitute a
random subsample of the original sample. Thus, the analysis that includes only
those who have U and V observed is generally valid under this assumption because
the process of excluding the subjects with any missing values does not distort
the representativeness of the original sample. This assumption is clearly violated
in our simulation study where the percentages with missing values differ across
the four cells based on D and E. The MCAR is a rather strong assumption and
is rarely satisfied in practical applications. Sometimes the available-case analysis
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

may be valid under a weaker assumption [see, for example, (7, 12, 15)], but these
exceptions are few and idiosyncratic.
A weaker assumption is MAR where, again, resorting to the simple case first,
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Pr(R = 1|U,V) = f(V), a function that depends on V but not on U. The dele-
tion mechanism used in the simulation study falls under this category where the
missingness in x depends on D and E but not on x. In essence, this assumption
entails that for two individuals with the same value of V, one has U observed
and the other has U missing; the missing U is arising from the same distribution
(for a given V) as the observed one. That is, conditional on V, the missing U is
predictable from the observed distribution of U. In a more general situation where
the missing values can be in several variables, the exact analytical specification
of this assumption is difficult. Loosely speaking, suppose that di,obs denotes the
observed components of complete data, (Ui , Vi ), for subject i, and di,miss denotes
the missing components. For two individuals, i and j, when di,obs = d j,obs , the
missing components, di,miss and d j,miss , have the same distribution.
Given this conditional nature of the assumption, the stronger the correlates of
U in V, the weaker is the assumption about the missing data mechanism for U.
For example, if U were income, then having a rich set of variables to condition
on in V, such as age, gender, education, occupation, property values, monthly
expenditures, and neighborhood level information, makes the assumption about
missing data considerably weaker when compared to MCAR or when the list of
variables to be conditioned on is limited to, say, age and gender. The limitation
of the MAR assumption is due to lack of appropriate variables that can be con-
ditioned on in the analysis, and empirically it has been shown to be reasonable
in practical situations (4, 33). The three approaches discussed in this paper are
valid under different versions of this weaker assumption about the missing data
mechanism.
Finally, the missing data mechanism is said to be Not-Missing at Random
(NMAR), if Pr(R = 1|U, V) = f(U, V), a function that certainly depends on U but
may also depend on V. That is, even after conditioning on V, the distribution of U
for the respondents and nonrespondents are dissimilar. This function, however, is
not estimable from the observed data because whenever R = 0, U is unobserved.
Therefore, an explicit form of f has to be specified and the data cannot be used to
empirically verify the validity of this assumption.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 105

The specification of Pr(R = 1|U, V), in conjunction with the substantive model
Pr(U, V), is used to construct inferences about the parameters of interest. This is
a selection model method and was first proposed by Heckman (8). The alternative
approach is to specify how different the distributions of (U, V) are for the respon-
dents and nonrespondents. That is, specify the population distribution as a mixture
of two components, Pr(U, V|R = 1) and Pr(U, V|R = 0), for respondents and non-
respondents, respectively. This mixture is used to construct inferences about the
population. Again, there is no data to specify the part of the mixture Pr(U, V|R =
0) because U is unobserved whenever R = 0. This approach was first proposed by
Rubin (27, 29) and later extended by Little (16). In any event, both these approaches
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

make empirically unverifiable assumptions, and their use is very limited to situ-
ations where some prior knowledge may exist to specify the mixture distribution
or the selection model. For more details see Chapter 15 in Reference 19.
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

WEIGHTING

The origins of weighting may be traced to sample survey practice where it is

used as a simple device to account for unequal probabilities of selection (11).
The survey weight for an individual is the inverse of his/her selection probability.
To be concrete, suppose that a national probability survey sampled 1000 subjects
with 500 Whites and 500 African Americans. That is, the African Americans were
oversampled with respect to their representation in the population. Suppose that
one were to ignore this fact and proceeded to compute a simple average of 1000
observations as an estimate of the population mean. If there are large differences
between Whites and African Americans in the survey variable of interest, the
sample mean will be a distorted representation of the population mean. Down-
weighting the observations on African Americans to their representation in the
population and up-weighting the observations on Whites to their representation
in the population would obtain the accurate picture of the population. Thus, the
weighted average (where the weights are inverse of their selection rate) is an
unbiased estimate of the population mean whereas the simple mean is not. However,
special software is needed to compute standard errors and confidence intervals
that use these weights. Currently, popular packages such as SAS, STATA, and
SUDAAN have built-in routines to take into account these survey weights.
Weighting to compensate for nonresponse is an extension of the same idea.
That is, excluding the subjects because of missing values is a distortion of the
representation in the original sample, and weights are attached to subjects included
in the analysis to restore the representation. To be concrete, consider the simulation
example. Suppose that n de is the number of subjects with D = d and E = e where
d, e = 0,1. Let the number of subjects with observed x in the corresponding cell
be rde . To restore the available-case analysis to its original sample representation,
we should weight each respondent in (d, e) cell by wde = n de /rde , inverse of the
response rate in that cell or the inverse of the selection probability into the data
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

106 RAGHUNATHAN
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Figure 3 Histogram of logistic regression coefficient from 2500 simulated

data sets after deleting some covariate values and using weights to compensate
for missing data.

analysis. A weighted logistic regression can be used to estimate the parameter of

interest (2).
Figure 3 gives the point estimates for the same 2500 data sets analyzed previ-
ously, but this time based on weighted logistic regression where the weights were
computed as the inverse of the response rates in the four cells. As can be seen, the
weighted estimates are unbiased and, understandably, slightly more variable than
the before-deletion estimates.
There are several approaches for constructing weights (9). Consider the example
given earlier where the analysis of interest is concerned with a set of variables (U,
V), the missing values are in a single variable U, and variables in a vector V have no
missing values. The adjustment-cell method involves constructing a contingency
table through cross-classification of sampled subjects based on V. The inverse of
the response rate in any cell is the weight attached to each respondent in that cell.
The underlying assumption is that, conditional on belonging to an adjustment cell,
the respondents and nonrespondents have similar distribution of variables with
missing values (that is, respondents and nonrespondents are exchangeable within
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 107

the adjustment cell). The simulation example described above used this method to
construct the weights based on D and E. The disadvantage of this approach is that
the continuous variables in V have to be categorized, and if V has a large number
of variables, the contingency table can be sparse, leading to unstable weights.
An alternative approach is to estimate the response propensity through a logistic
regression model, Pr(R = 1|V ) = [1 + exp(−βo − V t β1 )]−1 , where βo and β1 are
the unknown regression coefficients and the superscript t denotes matrix transpose.
Suppose that β̂o and β̂1 are the estimated regression coefficients; the weights for
a respondent j is then defined as w j = 1 + exp(−β̂o − V jt β̂1 ). One may include
the interaction terms between different variables or transformation in V. There is
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

no need to categorize continuous variables. This response propensity approach is

more practical when we have large numbers of variables. Sometimes categorization
of estimated response probabilities forms adjustment cells.
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

MULTIPLE IMPUTATION

Weighting is a simple approach for making the subjects included in the available-
case analysis representative of the original sample and is effective in removing
nonresponse bias. However, by including only subjects with complete data, it
ignores partial information from subjects with incomplete data. For example, in a
multiple linear regression with some subjects missing one variable at the most, it
can be very inefficient to ignore information on the rest of the variables, especially
if they are good predictors of the variables with missing values.
An alternative approach is based on filling in or imputing the missing values in
the data set. Again, this approach can be traced back to survey practices adopted
by the U.S. Bureau of Census [See (6) for historical accounts]. This practice of
filling in the missing values (this is called single imputation method) in the survey
practice is attractive for several reasons. First, imputation adjusts for differences
between nonrespondents and respondents on variables observed for both and in-
cluded in the imputation process, as well as differences on variables not included
in the model that are predicted by the model; such an adjustment is generally not
made by available-case analysis. Second, the complete data software can be used
to process the data to obtain descriptive statistics and other statistical measures.
This is a significant advantage because complete-data software has kept closer
pace with the statistical methodological developments than the incomplete-data
software. Third, when a data set is being produced for analysis by the public or
multiple researchers, imputation by the data producer allows the incorporation of
specialized knowledge about the reasons for missing data in the imputation pro-
cedure, including confidential information that cannot be released to the public or
other variables in the imputation process that may not be used in substantive anal-
ysis by a particular researcher. Raghunathan & Siscovick (24) demonstrate that
using an auxiliary variable in the imputation process can improve the efficiency
considerably. Moreover, the nonresponse problem is solved in the same way for
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

108 RAGHUNATHAN

all users so that analyses will be consistent across users. The researcher using
the filled-in data can concentrate on addressing substantive questions of interest
and not be distracted by incomplete data. See Reference 31 for a detailed list of
applications of this approach.
Although single imputation, that is, imputing one value for each missing datum,
enjoys the positive attributes just mentioned, analysis of a singly imputed data set
using standard software fails to reflect the uncertainty due to the fact that the
imputed values are plausible replacements for the missing values but are not the
true values themselves. As a result, such analyses of singly imputed data tend to
produce estimated standard errors that are too small, confidence intervals that are
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

too narrow, and significance tests with p-values that are too small.
Multiple imputation (29, 30) is a technique that seeks to retain the advantages
of single imputation while also allowing the uncertainty due to imputation to be
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

incorporated into the analysis. The idea is to create more than one, say M, plausible
sets of replacements for the missing values, thereby generating M completed data
sets. The variation across the M completed data sets reflects the uncertainty due to
imputation. Typically, M is not larger than five.
The analysis of the M completed data sets resulting from multiple imputation
proceeds as follows:

1. Analyze each completed data set separately using a suitable software package
designed for complete data (for example, SAS, SPSS, or STATA).
2. Extract the point estimate and the estimated standard error from each
analysis.
3. Combine the point estimates and the estimated standard errors to arrive
at a single point estimate, its estimated standard error, and the associated
confidence interval or significance test.

Suppose el is the estimate and sl its standard error, based on the completed data
set l = 1, 2, · · · , M, whereM ≥ 2. The multiply imputed estimate is the average,

1 XM
ē M I = el ,
M l=1

and the standard error of the multiply imputed estimate is

r
M +1
s M I = ū M + bM ,
M
where

1 XM
ū M = s2 and
M l=1 l
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 109

1 X M
¡ ¢2
bM = el − ē M I
M − 1 l=1

The sampling variance (term inside the square root sign) has two parts: The first
part is the average sampling variance by treating the imputed values as though they
are real. This is called within-imputation component of variance. The second part
is the variability across the imputed values (the between-imputation component
of variance), which is not estimable unless more than one plausible set of values
are used as fill-in. Rubin & Schenker (32) and Rubin (30) derived the sampling
distribution as t-distribution with degrees of freedom, ν = (M − 1)(1 + r M )2 ,
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

where r M = ū M /[(1 + M −1 )b M ]. Several other methods have been developed to

construct intervals (1, 13, 14).
Most straightforward justification for generating more than one plausible set
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

of values is through draws from the predictive distribution of the missing val-
ues conditional on the observed values. Revisiting the simulation example, one
could generate several plausible values from the predictive distribution based on
a regression model,
x = βo + β1 D + β2 E + β3 D × E + ε,
where β = (βo , β1 , β2 , β3 ) is a vector of regression coefficients, and the residual
ε ∼ N (0, σ 2 ). A simple approach is to estimate the regression coefficients and
the residual variance using subjects with x, D, and E observed. Suppose x̂ is the
predicted value for an individual with missing x. Adding different noise variables
z ∼ N (0, σ̂ 2 ), where σ̂ 2 is the estimate of residual variance, to the predicted value
generates plausible values. This method is reasonable for large sample size but
still is not proper (30) because plausible values are generated without reflecting
uncertainty in the estimates of the regression coefficients and residual variances.
A proper approach reflects uncertainty in every estimate while generating plausi-
ble values. A proper approach for generating the plausible values is the Bayesian
approach where the missing values are drawn from the posterior predictive distri-
bution of the missing observations conditional on the observed data. The approach
is technical and the procedure for the simulation example is described in the ap-
pendix. More formal description of a proper method can be found in References
19, 30, and 34.
The fully Bayesian approach described in the appendix was implemented on
2500 simulated data sets with missing values described earlier. Five imputations
were created for each of 2500 data sets with missing values. Five completed data
sets were analyzed by fitting a logistic regression model to each. The multiple
imputation estimate and its standard error were computed using the formula given
above. Figure 4 gives the histogram of 2500 multiple imputation estimates, which
shows that the sampling distribution is centered on the true value 0.5.
The most straightforward approach for creating multiple imputations is model-
based using a Bayesian formulation. That is, draw values from the posterior
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

110 RAGHUNATHAN
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Figure 4 Histogram of logistic regression coefficient from 2500 simulated

data sets after deleting some covariate values and using multiple imputation
method.

predictive distribution, Pr(dmiss |dobs ), of the missing values, dmiss , conditional

on the observed values, dobs . Little & Raghunathan (18) argue that the impu-
tation should condition on as much observed information as possible to make
MAR plausible and imputations efficient. Schafer (34) has developed several rou-
tines for implementing the Bayesian method where one can achieve approximate
normality of continuous variables through transformation and a limited number
of categorical variables (see the website http://www.stat.psu.edu/∼jls). However,
this model-building task can be quite difficult in practical situations with hundreds
of variables, skip patterns in the questionnaire, different types of variables such
as continuous, categorical, count, and semicontinuous (basically, continuous but
with a spike at 0). Other common problems include structural dependencies. For
example, the question asking about years of smoking is applicable only to current
and former smokers, whereas years since quitting is applicable only to former
smokers. Also, years smoked cannot exceed the age of the person.
An alternative, though not a fully model-based Bayesian but fully conditional
approach, is the sequential regression approach (23), which builds on a version for
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 111

continuous variables originally proposed by Kennickel (10). A brief description of

SRMI is as follows: Let X denote the fully observed variables, and let Y1 , Y2 , . . . , Yk
denote k variables with missing values in any order. The imputation process for
Y1 , Y2 , . . . , Yk proceeds in c rounds. In the first round, Y1 is regressed on X, and
the missing values of Y1 are imputed (using a process analogous to that described
for the logistic regression example in the Appendix); then Y2 is regressed on X and
Y1 (including the imputed values of Y1 ), and the missing values of Y2 are imputed;
then Y3 is regressed on X, Y1 and Y2 , and the missing values of Y3 are imputed;
and so on, until Yk is regressed on X, Y1 ,Y2 . . . , Yk−1 , and the missing values of
Yk are imputed.
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

In Rounds 2 through c, the imputation process carried out in Round 1 is repeated,

except that now, in each regression, all variables except for the variable to be im-
puted are included as predictors. Thus, Y1 is regressed on (X, Y2 , Y3 , . . . , Yk ), and
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

the missing values of Y1 are imputed; then Y2 is regressed on (X, Y1 , Y3 , . . . , Yk ),

and the missing values of Y2 are imputed; and so on. After c rounds, the final
imputations of the missing values in (Y1 , Y2 , . . . , Yk ) are used.
An SAS-based software IVEware implementing this approach is available from
the website http://www.isr.umich.edu/src/smp/ive. The S-plus version of a similar
approach is available from http://www.multiple-imputation.com. For the regres-
sions in the SRMI procedure, IVEware allows the following models:
1. a normal linear regression model if the Y-variable is continuous;
2. a logistic regression model if the Y-variable is binary;
3. a polytomous or generalized logit regression model if the Y-variable is cat-
egorical with more than two categories;
4. a Poisson loglinear model if the Y-variable is a count;
5. a two-stage model if the Y-variable is mixed (i.e., semicontinuous). In the
first stage zero/nonzero status is imputed using a logistic regression model.
Conditional on being nonzero, a normal linear regression model is used to
imput a nonzero value.
Because SRMI requires only the specification of individual regression models
for each of the Y-variables, it does not necessarily imply a joint model for all of the
Y-variables conditional on X. This procedure compares well with the fully Bayesian
approach as demonstrated through the simulation study by Raghunathan et al.
(23). This is the most practical approach in many situations involving structural
dependencies and the large number of predictors of varying types.

MAXIMUM LIKELIHOOD
In the complete data statistical methodology, maximum likelihood for a given
model is a dominant inferential procedure, for example, the linear, logistic, Poisson,
log-linear, and random effects models. All use likelihood as a basis for constructing
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

112 RAGHUNATHAN

inferences. Extending the same notion, one possibility is to base our inferences
on likelihood function constructed from the actual observed data set. To motivate
this approach, consider an example based on a random sample of size n from a
bivariate normal distribution,
µ ¶ ·µ ¶ µ 2 ¶¸
X µX σX σX Y
∼N , .
Y µY σ X Y σY2
Suppose that p subjects provide both X and Y, q subjects provide only X and
not Y, and r subjects provide Y but not X. The objective is to estimate the five
unknown parameters (µ X , µY , σ X2 , σY2 , σ X Y ). Once these estimates are obtained,
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

any function of these parameters, such as correlation coefficient or regression

coefficients, can be computed. Wilks (36) addressed this estimation problem using
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

several approaches including the maximum likelihood method.

The observed data likelihood is a product of three components,

L obs = L p × L q × L r ,

where the first component is the contribution from p subjects who provided both
X and Y. This is a product of bivariate normal density function evaluated at the
observed values for those p subjects, and it involves all five parameters. The second
component is the contribution from q subjects who provided only X. This is a
product of univariate normal density functions involving only (µ X , σ X2 ). Finally, the
third component, based on r subjects who provided only Y, is a product of univariate
normal density functions involving only (µY , σY2 ). The observed data likelihood
is then maximized with respect to (µ X , µY , σ X2 , σY2 , σ X Y ) using some iterative
routines such as Newton-Raphson method or, more popular, the EM algorithm (5).
More generally, suppose that Yi is a complete data vector on subject i = 1, 2,
. . ., n; Yi,obs denotes the observed components and Yi,miss the missing components;
andYi = (Yi,obs , Yi,miss ). Suppose the complete data model is

f (Yi |θ ) = f (Yi,obs , Yi,miss |θ ),

where θ is the unknown parameter to be estimated. The observed data likelihood

is
Y
n n Z
Y
L obs (θ) = L(θ|Yi,obs ) ∝ f (Yi,obs , Yi,miss |θ ) dYi,miss .
i=1 i=1

The justification that this is the correct likelihood to be maximized when data
are MAR is given in Reference 28.
The approximate sampling variances are typically estimated by inverting the
negative second derivative of the logarithm of the observed data likelihood. Though
the estimates based on this approach are perhaps the most efficient and enjoy
all the nice properties accorded to maximum likelihood estimates, implementa-
tion is quite difficult even in the simple logistic regression case considered ear-
lier. Typically special software needs to be developed for a particular problem.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 113

Software for multivariate normal model is available in SPSS 10 and as an add-on

package in GAUSS. The general location model, which can be used to model
normal continuous and discrete variables simultaneously, was considered by Little
& Schluchter (20), Raghunathan & Grizzle (22) and Schafer (34). Software for
fitting such models using a Bayesian approach was developed by Schafer (34).
Due to these technical difficulties, this method is not practical in many real world
applications.

DISCUSSION AND LIMITATIONS

Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

Analysis of data with some missing values is an important problem, and the stan-
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

dard strategy of including only those on whom a particular analysis can be carried
out can lead to biased estimates. Three approaches have been discussed with in-
creasing levels of statistical sophistication. Weighting is the simplest approach;
multiple imputation is at the second level but is still a more general approach. The
maximum likelihood approach is the most difficult, often requiring user-developed
software for implementation. All these approaches are valid under a general class
of mechanisms called MAR, whereas the available-case analysis is generally valid
under MCAR, though there are some exceptions (7, 12). These exceptions are few
and idiosyncratic, and are often difficult to verify in a practical setting. Even if
the data are MCAR, the available-case method is less efficient owing to discard-
ing subjects with partial information. It is not uncommon for substantial numbers
of subjects to be excluded in a regression analysis, even though each subject is
missing only a few variables.
Weighting is a simple device to correct for bias, but it suffers from the same
disadvantages as the available-case method in terms of efficiency. It still discards
partial information from subjects with missing values. Thus, if the bias correction
is the motivating factor, then weighting should be used to compensate for missing
data. Either the adjustment-cell method or the response propensity method can be
used to derive the weights, though the response propensity method relies on the
logistic regression model. A compromise might be to use the response propensities
to form adjustment cells to rely less on the correctness of the response propensity
model specification.
Perhaps the most practical approach is based on multiple imputation. This ap-
proach involves an upfront investment in multiply imputing the missing values in
the database. Once multiply imputed, any complete data software can be used to
repeatedly analyze the completed data sets, extract the point estimates and their
standard errors, and combine them using the formula given in the third section of
this review. The last step can be carried out using a spreadsheet program such as
Excel. There are user-developed routines now available in STATA (3) and IVE-
WARE (23), and in commercial software such as SAS version 8.2 and SOLAS.
Though this method requires additional storage and extra steps of repeated analysis
and combining estimates, in the grand scheme of public health investigations, it is
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

114 RAGHUNATHAN

a minor step, especially owing to the availability of software for creating multiple
imputations.
Though we emphasize multiple imputations, it is possible to correct the standard
errors using the single imputation method. For a limited set of variables such as
means and proportions and the correlation coefficients, Rao & Shao (26) and Rao
(25) have proposed the Jackknife method for computing correct standard errors
from singly imputed data sets. Nevertheless, the multiple imputation approach
seems to be the most practical approach in a setting involving a large data set with
multiple researchers using different portions of the same data set, as well as for a
single researcher analyzing a particular data set with missing values, provided an
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

upfront investment is made to develop multiple imputations.

Perhaps the gold standard is the maximum likelihood method, working directly
with the likelihood based on observed incomplete data. This option is preferable
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

for a limited set of models for which the software is available. In fact, the multi-
ple imputation based on Bayesian formulation can be viewed as an approximate
maximum likelihood method. Specifically, suppose that L(θ |Dobs , Dmiss ) is the
likelihood that would have been constructed had there been no missing data. The
observed data likelihood is,
R
L(θ|Dobs ) = L(θ|Dobs , Dmiss ) Pr(Dmiss |Dobs ) d Dmiss .

(l)
In the event that the M imputations, (Dmiss , l = 1, 2, . . . , M), are draws from
the posterior predictive distribution, Pr(Dmiss |Dobs ), the observed data likelihood
can be approximated by the average,

1 X ¡ (l) ¢
L(θ|Dobs ) ≈ L θ |Dobs , Dmiss ,
M l

of the completed-data likelihoods. That is, the multiple imputation analysis that
combines the likelihood-based analysis from each completed data set is approxi-
mately equivalent to the analysis based on the observed data likelihood. This dis-
cussion is another justification for Bayesian imputation or something very close
to it.
Clearly, several possible options exist for a public health researcher to perform
a correct analysis with incomplete data. Both user-driven software and commercial
software are becoming available to implement these methods. Though most meth-
ods rely on a MAR assumption, its lack of applicability is related to the lack of
variables that can be used to predict the missing values. Because the missing data
are inevitable, a prudent step, from the design perspective, is to investigate potential
predictors of variables with missing data and include them in the data-collection
process. Such auxiliary variables can include administrative data, neighborhood-
level observations, and interviewer observations. These additional variables can be
used in the multiple imputation process. It is important that missing data be con-
sidered not solely a data analysis problem, but also a design and analysis problem.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 115

APPENDIX

Suppose in a sample of n, r subjects are missing values in x. Let UR denote the

design matrix with r rows and four columns representing intercept, D, E, and
D × E. Similarly, let UM denote the design matrix with n-r rows and four columns
for the nonrespondents. Let β̂ = (U Rt U R )−1 U Rt X R be the least square estimate
of the regression of XR on UR, where XR is a vector with r rows containing ob-
served values of the covariate x. Let s = (X R − U R β̂)t (X R − U R β̂) denote the
residual sum of squares. The following algorithm then represents a draw from the
posterior predictive distribution of missing covariates conditional on the observed
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

data.
1. Draw a chi-square random variable, c, with r-4 of freedom and define,σ∗2 =
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

s/c.
2. Draw r independent standard normal deviates and arrange them as a
vector z.
3. Define β∗ = β̂ + σ∗ TR z, where TR is the square root (Cholesky decomposi-
tion) of the matrix (U Rt U R )−1 .
4. Draw n-r independent random normal deviates and arrange them as a
vector v.
5. Define x∗ = U M β∗ + σ∗ v as imputed values.
6. Repeat steps 1–3 independently to generate multiple imputation.

The Annual Review of Public Health is online at

http://publhealth.annualreviews.org

LITERATURE CITED
1. Barnard J, Rubin DB. 1999. Small-sample 1977. Maximum likelihood from incom-
degrees of freedom with multiple imputa- plete data via the EM algorithm (with
tion. Biometrika 86:949–55 discussion). J. Roy. Statist. Soc. 39:1–
2. Binder DA. 1983. On the variances of 38
asymptotically normal estimators from 6. Ford BN. 1983. An overview of hot deck
complex survey data. Int. Statist. Rev. procedures. In Incomplete Data in Sam-
51:279–92 ple Surveys, Vol II: Theory and Anno-
3. Carlin JB, Li N, Greenwood P, Coffey C. tated Bibliography, ed. WG Meadow, I
2002. Tools for analyzing multiple imputed Olkin, DB Rubin, pp. 185–206. New York:
data sets. Tech. Rep., Univ. Melbourne, Academic
Australia 7. Glynn RJ, Laird NM. 1986. Regression es-
4. David MH, Little RJA, Samuhel ME, Tri- timates and missing data: complete case
est RK. 1986. Alternative methods for CPS analysis. Tech. Rep., Harvard School of
income imputation. J. Am. Statist. Assoc. Public Health, Dep. Biostatistics
81:29–41 8. Heckman JI. 1976. The common structure
5. Dempster AP, Laird NM, Rubin DB. of statistical models of truncation, sample
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

116 RAGHUNATHAN

selection and limited dependent variables, 21. McKendrick AG. 1926. Applications of
and a simple estimator for such models. mathematics to medical problems. Proc.
Ann. Econ. Soc. Meas. 5:475–92 Edinburgh Math. Soc. 44:98–130
9. Holt D, Elliot D. 1991. Methods of weight- 22. Raghunathan TE, Grizzle JE. 1995. A split-
ing for unit nonresponse (correction: v41, questionnaire survey design. J. Am. Statist.
p. 599). Statistician 40:333–42 Assoc. 90:55–63
10. Kennickell AB. 1991. Imputation of the 23. Raghunathan TE, Lepkowski JM, van
1989 Survey of Consumer Finances: Hoewyk M, Solenberger PW. 2001. A
stochastic relaxation and multiple imputa- multivariate technique for multiply imput-
tion. Proc. Sec. Surv. Res. Meth. Am. Statist. ing missing values using a sequence of re-
Assoc. pp. 1–10 gression models. Survey Methodol. 27:85–
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

11. Kish L. 1965. Survey Sampling. New York: 95. For associated IVEware software, see
Wiley http://www.isr.umich.edu/src/smp/ive
12. Kleinbaum DG, Morgenstern H, Kupper 24. Raghunathan TE, Siscovick DS. 1996. A
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

LL. 1981. Selection bias in epidemio- multiple imputation analysis of a case-

logical studies. Am. J. Epidem. 113:452– control study of the risk of primary
63 cardiac arrest among pharmacologically
13. Li KH, Meng XL, Raghunathan TE, Ru- treated hypertensive. Appl. Statist. 45:335–
bin DB. 1991. Significance levels from 52
repeated p-values with multiply imputed 25. Rao JNK. 1996. On variance estimation
data. Statist. Sinica 1:65–92 with imputed survey data. J. Am. Statist.
14. Li KH, Raghunathan TE, Rubin DB. Assoc. 91:499–506
1991. Large sample significance levels 26. Rao JNK, Shao J. 1992. Jackknife variance
from multiply-imputed data using moment- estimation with survey data under hot deck
based sta tistics and an F-reference dis- imputation. Biometrika 79:811–22
tribution. J. Am. Statist. Assoc. 86:1065– 27. Rubin DB. 1974. Characterizing the esti-
73 mation of parameters in incomplete data
15. Little RJA. 1992. Regression with miss- problems. J. Am. Statist. Assoc. 69:467–
ing X’s: a review. J. Am. Statist. Assoc. 74
87:1227–37 28. Rubin DB. 1976. Inference and missing
16. Little RJA. 1993. Pattern-mixture models data (with discussion). Biometrika 63:581–
for multivariate incomplete data. J. Am. 92
Statist. Assoc. 88:125–34 29. Rubin DB. 1977. Formalizing subjective
17. Little RJA. 1995. Modeling the drop-out notions about the effect of nonrespondents
mechanism in longitudinal studies. J. Am. in sample surveys. J. Am. Statist. Assoc.
Statist. Assoc. 90:1112–21 72:538–43
18. Little RJA, Raghunathan TE. 1997. Should 30. Rubin DB. 1987. Multiple Imputation
imputation of missing data condition on for Nonresponse in Surveys. New York:
all observed variables? Proc. Sec. Surv. Wiley
Res. Meth. Am. Statist. Assoc. pp. 617– 31. Rubin DB. 1996. Multiple imputation after
22 18+ years (with discussion). J. Am. Statist.
19. Little RJA, Rubin DB. 2002. Statistical Assoc. 91:473–89
Analysis with Missing Data. New York: 32. Rubin DB, Scehnker N. 1986. Multiple im-
Wiley putation for interval estimation from sim-
20. Little RJA, Schluchter MD. 1985. Maxi- ple random samples with ignorable non-
mum likelihood estimation for mixed con- response. J. Am. Statist. Assoc. 81:366–
tinuous and categorical data with missing 74
values. Biometrika 72:497–512 33. Rubin DB, Stern H, Vehovar V. 1995.
 13 Feb 2004 14:30 AR AR209-PU25-05.tex AR209-PU25-05.sgm LaTeX2e(2002/01/18) P1: IBD

ANALYSIS OF INCOMPLETE DATA 117

Handling “don’t know” survey responses: 35. Vach W. 1994. Logistic Regression with
the case of Slovenian plebiscite. J. Am. Missing Values in Covariates. New York:
Statist. Assoc. 90:822–28 Springer-Verlag
34. Schafer JL. 1997. Analysis of Incomplete 36. Wilks SS. 1932. Moment and distribu-
Multivariate Data. New York: CRC Press. tion of estimates of population parame-
For associated software, see http://www. ters from fragmentary samples. Ann. Math.
stat.psu.edu/∼jls Stat. 3:163–95
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org
Access provided by 190.45.4.164 on 03/27/18. For personal use only.
P1: FRK
February 20, 2004 14:36 Annual Reviews AR209-FM

Annual Review of Public Health

Volume 25, 2004

CONTENTS
PERSPECTIVES
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

The Third Revolution in Health, Lester Breslow xiii

EPIDEMIOLOGY AND BIOSTATISTICS
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Can Dementia Be Prevented? Brain Aging in a Population-Based Context,

Mary N. Haan and Robert Wallace 1
Public Health Surveillance of Low-Frequency Populations,
Elena M. Andresen, Paula H. Diehr, and Douglas A. Luke 25
Statistical and Substantive Inferences in Public Health: Issues in the
Application of Multilevel Models, Jeffrey B. Bingenheimer
and Stephen W. Raudenbush 53
Trends in the Health of the Elderly, Eileen M. Crimmins 79
What Do We Do with Missing Data? Some Options for Analysis
of Incomplete Data, Trivellore E. Raghunathan 99
ENVIRONMENTAL AND OCCUPATIONAL HEALTH
Emission Trading and Public Health, Alexander E. Farrell
and Lester B. Lave 119
Genetic Testing in the Workplace: Ethical, Legal, and Social Implications,
Paul W. Brandt-Rauf and Sherry I. Brandt-Rauf 139
Health Effects of Chronic Pesticide Exposure: Cancer and Neurotoxicity,
Michael C.R. Alavanja, Jane A. Hoppin, and Freya Kamel 155
Implications of the Precautionary Principle for Primary Prevention
and Research, Philippe Grandjean 199
Issues of Agricultural Safety and Health, Arthur L. Frank,
Robert McKnight, Steven R. Kirkhorn, and Paul Gunderson 225
Time-Series Studies of Particulate Matter, Michelle L. Bell,
Jonathan Samet, and Francesca Dominici 247
PUBLIC HEALTH PRACTICE
Developing and Using the Guide to Community Preventive Services:
Lessons Learned About Evidence-Based Public Health, Peter A. Briss,
Ross C. Brownson, Jonathan E. Fielding, and Stephanie Zaza 281

vii
P1: FRK
February 24, 2004 20:39 Annual Reviews AR209-FM

viii CONTENTS

Modeling Infection Transmission, Jim Koopman 303

The Current State of Public Health in China, Liming Lee 327
The Public Health Workforce, Hugh Tilson and Kristine M. Gebbie 341
Lessons Learned from Public Health Mass Media Campaigns: Marketing
Health in a Crowded Media World, Whitney Randolph and K. Viswanath 419
SOCIAL ENVIRONMENT AND BEHAVIOR
Assuring the Health of Immigrants: What the Leading Health Indicators
Tell Us, Namratha R. Kandula, Margaret Kersey, and Nicole Lurie 357
Harm Reduction Approaches to Reducing Tobacco-Related Mortality,
Annu. Rev. Public Health 2004.25:99-117. Downloaded from www.annualreviews.org

Dorothy K. Hatsukami, Jack E. Henningfield, and Michael Kotlyar 377

Housing and Public Health, Mary Shaw 397
Access provided by 190.45.4.164 on 03/27/18. For personal use only.

Lessons Learned from Public Health Mass Media Campaigns: Marketing

Health in a Crowded Media World, Whitney Randolph and K. Viswanath 419
The Role of Culture in Health Communication, Matthew W. Kreuter
and Stephanie M. McClure 439
HEALTH SERVICES
Economic Implications of Increased Longevity in the United States,
Dorothy P. Rice and Norman Fineman 457
International Differences in Drug Prices, Judith L. Wagner
and Elizabeth McCarthy 475
Physician Gender and Patient-Centered Communication: A Critical
Review of Empirical Research, Debra L. Roter and Judith A. Hall 497
The Direct Care Worker: The Third Rail of Home Care Policy, Robyn Stone 521
Developing and Using the Guide to Community Preventive Services:
Lessons Learned About Evidence-Based Public Health, Peter A. Briss,
Ross C. Brownson, Jonathan E. Fielding, and Stephanie Zaza 281
INDEXES
Subject Index 539
Cumulative Index of Contributing Authors, Volumes 16–25 577
Cumulative Index of Chapter Titles, Volumes 16–25 581
ERRATA
An online log of corrections to Annual Review of Public Health
chapters may be found at http://publhealth.annualreviews.org/