-
This article was downloaded by: [University of Hawaii at
Manoa]On: 28 February 2015, At: 13:34Publisher: RoutledgeInforma
Ltd Registered in England and Wales Registered Number: 1072954
Registered office: Mortimer House, 37-41 MortimerStreet, London W1T
3JH, UK
Educational PsychologistPublication details, including
instructions for authors and subscription
information:http://www.tandfonline.com/loi/hedp20
Assesing the effects of context in studies of child and
youthdevelopmentGreg J. Duncan & Stephen W. RaudenbushPublished
online: 08 Jun 2010.
To cite this article: Greg J. Duncan & Stephen W. Raudenbush
(1999) Assesing the effects of context in studies of child and
youth development,Educational Psychologist, 34:1, 29-41
To link to this article:
http://dx.doi.org/10.1207/s15326985ep3401_3
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy
of all the information (the Content) contained in the publications
onour platform. However, Taylor & Francis, our agents, and our
licensors make no representations or warranties whatsoever as to
theaccuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are
theopinions and views of the authors, and are not the views of or
endorsed by Taylor & Francis. The accuracy of the Content
should notbe relied upon and should be independently verified with
primary sources of information. Taylor and Francis shall not be
liable forany losses, actions, claims, proceedings, demands, costs,
expenses, damages, and other liabilities whatsoever or howsoever
causedarising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private
study purposes. Any substantial or systematic
reproduction,redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is
expressly forbidden. Terms &Conditions of access and use can be
found at http://www.tandfonline.com/page/terms-and-conditions
http://www.tandfonline.com/loi/hedp20http://dx.doi.org/10.1207/s15326985ep3401_3http://www.tandfonline.com/page/terms-and-conditions
-
EDUCATIONAL PSYCHOLOGIST, 34(1), 29-41 Copyright O 1999,
Lawrence Erlbaum Associates, Inc.
Assessing the Effects of Context in Studies of Child and Youth
Development
Greg J. Duncan Northwestern University
Stephen W. Raudenbush University of Michigan
Children develop in a multitude of social environments
(Bronfenbrenner, 1979). Indeed, human development is un- thinkable
without social settings created to protect, feed, and nurture
children and to teach them to speak and interact with others; or
without inputs from a broader community of family friends,
relatives, and institutions (Haveman & Wolfe, 1994).
Because social settings such as the family and larger com-
munity-including neighborhoods, schools, and peers-are essential to
making the child fully human, it may seem odd that social science
research is far from definitive about whether "context matters."
This is due in large part to the varying senses in which context
might be said to matter within the logic of social science. For
many social scientists, including sociologists who study status
attainment and psy- chologists who focus on individual differences,
context can be said to matter if differences among social contexts
are found to be important in explaining individual differences in
achieving ends most of us value-mental health, literacy, in-
tellectual growth, educational attainment, occupational sta- tus,
and the like. Program evaluators wony primarily about effect sizes,
rather than explained variance, and economists extend that wony to
include benefits associated with effect sizes relative to cost.
In considering the methodological issues facing the design- ers
of studies of contextual effects, we generally refer to such
effects in the narrow sense used by social scientists who study
individual differences. Thus, we consider past research on the
ability of variations in contexts to account for variation in child
outcomes and to identify how designs of future studies might use
such variation to identify important contextual influences.
This research is relevant to social policy aimed at improv- ing
settings such as neighborhoods and schools; for if certain
Requests for reprints should be sent to Greg J. Duncan,
Institute for Policy Research, Northwestern University, 2040
Sheridan Road, Evanston, IL
settings are found to be especially helpful in promoting de-
sired child and youth outcomes, policy might aim to recreate those
settings on a broader scale. And this research is also rel- evant
for the design of new studies of child development that would rely
on "naturally occurring" variation in social set- tings to gauge
contextual effects. However, it is important to realize that
effects may turn out to be small because the de- gree of natural
variation is small, rather than because the set- ting is
irrelevant. Correlational research based on naturally occurring
variation can identify plausible consequences only if that
variation currently exists.
The purpose of this article is to formulate research de- signs
for studies of child and youth development that would "do context
right." We choose neighborhood contexts to il- lustrate our points,
but much of what we say also carries over to contexts such as
schools. We begin with theoretical sto- ries about neighborhood
effects to indicate the kinds of theo- ries that one might want to
collect data to test. We then present some evidence regarding where
the contextual "ac- tion" might be.
We next present an overview of practical design and statis-
tical issues for modeling contextual effects, emphasizing that
although there is a need for theoretically appropriate contex- tual
information from record-based sources, it is generally unavailable.
For example, with respect to the neighborhood literature, we may
believe that the degree of informal social control is a key
theoretical construct (Sampson & Groves, 1989), but we may be
constrained to neighborhood measures that appear on the decennial
census form. Special data collec- tion efforts can be undertaken to
measure constructs more precisely, but they are often very
expensive and become even more so in longitudinal studies, because
geographic mobility increases the number of neighborhoods or
schools exponen- tially as a sample is followed year by year.
In addition to measurement problems, a host of statistical
difficulties is caused by the nonrandom selection of parents
60208. E-mail: greg-duncan@ nwu.edu and children into
neighborhood, school, and other contexts.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
30 DUNCAN AND RAUDENBUSH
Most important, apparent effects of neighborhood character-
istics may merely reflect unobserved parental characteristics such
as concern for their children's development, parental mental
health, or permanent family income. Residential mo- bility caused
by natural or randomized experiments is rare. Researchers relying
on nonexperimental data must find model-based solutions to these
problems.
We conclude with a discussion of implications for study designs.
One possible approach to the measurement problem is to
geographically cluster the initial sample selection so that
measurements taken from individual respondents can be ag- gregated
across all respondents in a given geographic cluster to provide the
contextual measures. As appealing as this strat- egy may seem, we
argue that it is rarely satisfactory. For ex- ample, if measures of
neighborhood quality are obtained by aggregating reports of survey
participants and used as predic- tors of the mental health outcomes
of those same participants, a same-source bias is likely. On the
other hand, aggregating from independent samples of informants to
predict outcomes of others can be a very attractive, if expensive,
method for obtain- ing contextual information. Less expensive and
also attractive is intensive interviewer-based observation of the
environments. Except in cases of true or quasiexperimental designs,
definitive solutions to the problem of possible biases caused by
nonrandom parental selection of context are more elusive.
THEORIES AND EVIDENCE ABOUT NEIGHBORHOOD EFFECTS
Why might extrafamilial contexts such as a neighborhood af- fect
a child's development? The literature is filled with pro- posed
answers to this question, some, but not all, of which ar- gue that
environments of higher socioeconomic status (SES) are better for
children. We provide in this section a brief and selective review
of theories and evidence of neighborhood ef- fects.
Why Neighborhood Conditions Might Matter
Neighborhoods have been conceptualized in various ways,
encompassing geographic areas that range from a few blocks to
entire community areas and defined by both objective and subjective
means (Gephart, 1997). Jencks and Mayer (1990) developed a taxonomy
of theoretical ways in which neighbor- hoods may affect child
development. They distinguished (a) "epidemic" theories, based
primarily on the power of peer in- fluences to spread problem
behavior; (b) theories of "collec- tive socialization," in which
neighborhood role models and monitoring are important ingredients
in a child's socializa- tion; (c) "institutional" models, in which
the neighborhood's institutions (e.g., schools, police protection)
rather than neighbors per se make the difference; (d) "competition"
mod- els, in which neighbors (including classmates) compete for
scarce neighborhood resources; and (e) models of "relative
deprivation," in which individuals evaluate their situation or
relative standing vis-h-vis their neighbors (or classmates).
Social disorganization theory suggests other neighbor- hood
factors likely to influence child and adolescent develop- ment. For
instance, following Shaw and McKay (1942), Sampson argued that a
high degree of ethnic heterogeneity and residential instability
leads to an erosion of adult friend- ship networks and of a values
consensus in the neighborhood (Sampson & Lauritsen 1994).
Wilson's (1987) explanation of inner-city poverty in Chicago relied
on a more complicated model in which massive changes in the
economic structure, when combined with residential mobility among
more advan- taged Blacks, leave behind homogeneously impoverished
neighborhoods that provide neither resources nor positive role
models for children and adolescents growing up in them.
Furstenberg (1993) argued for the importance of under- standing
the role of family process in assessing neighborhood effects.
Basing his work on ethnographic studies, he pointed out that
families formulate different strategies for raising children in
high-risk neighborhoods, ranging from extreme protection and
insulation to an active role in developing com- munity-based
"social capital" networks that can help children at key points in
their academic or labor-market careers.
Because adolescents typically spend a good deal of time away
from their homes, explanations of neighborhood influ- ences based
on peers, role models, schools, and other neigh- borhood-based
resources would appear to be more relevant for them than for
younger children. However, it is possible that neighborhood
influences begin long before adolescence. A substantial minority of
3- and Cyear-olds are enrolled in center-based day care or
preschool (Hofferth & Chaplin, 1994). Physically dangerous
neighborhoods may force moth- ers to be isolated in their homes and
thus restrict opportunities for their children's interactions with
peers and adults. Parks, libraries, and children's programs provide
more enriching opportunities in relatively affluent neighborhoods
than are available in resource-poor neighborhoods. Parents of high
SES may be observed to resort less frequently to corporal
punishment and to engage more frequently in learn- ing-related
play. Thus, there are many ways in which neigh- borhood conditions
might affect both children and adolescents (Chase-Lansdale, Gordon,
Brooks-Gunn, & Klebanov, 1997).
Distinguishing empirically among these competing theo- ries is
not an easy task. As Jencks and Mayer (1990) pointed out, relative
deprivation and competition models of context generally predict
negative effects of higher-SES neighbors on youth achievement and
behavior, whereas all other models predict the opposite. Epidemic
models focus on the presence of "problematic" peers and have often
been implemented with measures of neighborhood poverty or adult
unemploy- ment.
In contrast, social control and institutional models focus more
on the presence of higher SES neighbors than the pres-
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
EFFECTS OF CONTEXT 31
ence or absence of low-SES neighbors. This distinction is
subtle, but easily conceived if SES is thought to have at least
three strata-say, low, medium, and high level of SES. The diversity
of U.S. neighborhoods produces different combina- tions of these
three strata, which enables researchers to distin- guish
empirically among their effects on developmental outcomes
(Brooks-Gunn, Duncan, Klebanov, & Sealand, 1993).
Social disorganization and family process models are the hardest
to implement empirically, because the required mea- sures are not
readily available from sources such as the decen- nial census.
Thus, the empirical literature displays an uneasy tension between
more representative studies using readily available but
theoretically flawed neighborhood measures and smaller, more
specialized studies with better context measures.
Empirical Studies of Neighborhood Effects
Every 10 years, the U.S. Census Bureau provides information that
can be used to construct neighborhood-based measures, such as the
fraction of individuals who are poor, the fraction of adults with a
college degree, and the fraction of adult men without jobs. Such
data are available for U.S. Census tracts (geographic areas
encompassing 4,000 to 6,000 individuals with boundaries drawn to
approximate neighborhood areas), zip codes, cities, counties,
metropolitan areas, and other use- ful geographically defined
areas.
As an example of a study of neighborhood effects using
cross-sectional census data, Crane (1991) used data from a special
linked family-tract file from the 1970 U.S. Cen- sus-based Public
Use Microdata Sample (PUMS) file. He used these data to relate
tract conditions to out-of-wedlock birth rates and high school
dropout rates of adolescents in the tracts. He found highly
nonlinear effects of neighborhood quality on adolescent outcomes,
effects that are consistent with the so-called epidemic models of
adolescent behavior. Dropping out of high school was very likely to
occur among individuals, both Black and White, living in
neighborhoods where fewer than 5% of workers in the neighborhood
held professional or managerial jobs. Apart from neighborhoods in
this extreme category, however, there was little evidence that
neighborhood characteristics mattered.
The power of the epidemic model to describe patterns of
neighborhood effects is called into question by Clark's (1992)
failure to replicate Crane's results using similar data from the
1980 census. Although Clark found that several measures of
neighborhood resources predict the high school dropout status of
male adolescents, she failed to find substan- tial evidence of
nonlinear effects such as those represented by "tipping points"
beyond which neighborhood effects become visible.
As pointed out by Manski (1993), analyses such as these that are
based on cross-sectional data may suffer from the "re-
flection problem." This occurs when the association between
neighborhood- and family-level characteristics at the time the
census is taken reflects the fact that neighborhood-level char-
acteristics are nothing more than the aggregation of family- and
individual-level characteristics. The neighborhood or peer-group
crime rate may indeed correlate with the chance of observing
criminal activity on the part of an adolescent liv- ing in that
neighborhood or having those peers, but to what extent is this
association truly causal?
Although certainly no panacea, longitudinal data provide some
statistical leverage to help solve this problem, because the
measurement of neighborhood characteristics can pre- cede the
outcome variables of interest. Brooks-Gunn et al. (1993) used
longitudinal data from the Infant Health and De- velopment Program
(IHDP) and Panel Study of Income Dy- namics (PSID) to examine the
impact of census-based neighborhood data-singly and in concert with
family-level variables--on early-childhood IQ and behavior problems
(in the IHDP) and adolescent school-leaving and out-of-wedlock
childbearing (in the PSID). They found that the absence of af-
fluent neighbors is much more important than the presence of
low-income neighbors-findings that support models of ben- eficial
institutions and collective socialization.
Analysts contributing chapters to Brooks-Gunn, Duncan, and Aber
(1997) matched a number of developmental data sets to census-based
neighborhood data and subjected them to parallel analyses. They
found that (a) although there is some evidence of neighborhood
effects in the preschool years, the most consistent evidence shows
up among school-age chil- dren; (b) cognitive and achievement
measures appear some- what more sensitive to neighborhood
influences than do behavioral and mental-health measures; (c) among
the five neighborhood factors used (low SES, high SES, ethnic
diver- sity, male joblessness, and the concentration of families in
the neighborhood), the high-SES factor had the most consistently
powerful effects; (d) Blacks were somewhat less affected by the
neighborhood measures than Whites; and (e) important from a
methodological point of view, multicolinearity prob- lems arose in
attempts to estimate separate effects of the five neighborhood
factors using data from single cities or from fairly homogeneous
neighborhoods across a small number of cities, but not in using
data from national (PSID, National Longitudinal Survey of Youth) or
heterogeneous, multisite data (IHDP).
Garner and Raudenbush (199 1) focused on neighborhood social
deprivation as a predictor of overall educational attain- ment in
Scotland. Key to their work was that any neighbor- hood effect
prior to secondary schooling was effectively controlled by
including measures of Primary 7 achievement (verbal IQ and reading
proficiency at Primary Grade 7). Thus, the test of the neighborhood
effect was a stringent one. Addi- tional control variables included
SES (parental education and occupation, unemployment, family size)
and school attended. The model accounted for essentially all of the
variation be- tween neighborhoods (enumeration districts, which are
simi-
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
32 DUNCAN AND RAUDENBUSH
lar to U.S. Census tracts) and between schools, as well as for
over half of the variation within schools. Neighborhood so- cial
deprivation (a composite index from the British census) was
strongly negatively related to overall attainment, after
controlling for the above factors.
The work of Sampson, Raudenbush, and Earls (1997) is a
noteworthy exception to the rule that studies of extrafamilial
context are hampered by measures of context that fail to cor-
respond closely to theoretical constructs. In a study of the de-
linquent behavior of youth in a sample of Chicago neighborhoods,
they measured the "collective efficacy" of neighborhoods by
conducting a survey of adult residents in sampled neighborhoods
rather than by relying exclusively on some collection of decennnial
census measures. For reasons detailed later, it is important to
note that their neighborhood measures were derived from an
independent survey of neigh- borhood residents and not an
aggregation of the characteris- tics of the youth whose possible
delinquent behavior was being studied.
Collective efficacy combines social cohesion (the extent to
which neighbors trust each other and share common val- ues) with
informal social control (the extent to whlch neigh- bors can count
on each other to monitor and supervise youth and protect public
order). It is thus a capacity for collective action shared by
neighbors. Sampson et al. (1997) found that collective efficacy so
defined relates strongly to neighbor- hood levels of violence,
personal victimization, and homicide in Chicago, after controlling
for social composition (as indi- cated by census variables) and for
prior crime.
Sampson et al. (1997) also found that collective efficacy
substantially mediates associations of concentrated disad- vantage,
residential instability, and immigrant concentration with violence
and crime. Key, then, is not so much the criminogenic character of
neighborhoods but rather the ca- pacity of adults to informally
regulate social behavior, partic- ularly that of young people.
Thus, collective efficacy exists relative to a particular task (in
this case, protecting public or- der), and its consequences ought
to be specific to the outcome of interest (curbing antisocial
behavior, especially of young people).
All of the previously cited studies relied on nonexperimental
data, and none fully accounted for the possi- ble biases caused by
the unmeasured characteristics of parents that lead them to choose
to live in one neighborhood over an- other (Duncan, Connell, &
Klebanov, 1997). A more complete discussion of nonexperimental
approaches to the bias problem is provided later. Here we note that
Rosenbaum (1991) was able to circumvent these problems by using
data from an un- usual quasiexperiment involving low-income black
families from public-housing projects in Chicago. As part of the
Gautreaux court case, nearly 4,000 families volunteered to par-
ticipate in a subsidized program that arranged for private hous-
ing, much of it in predominantly White Chicago suburbs, but some of
it in predominantly White sections of the city of Chi- cago itself.
Because participants were assigned to the first
available housing and were not allowed to choose between city
and suburban locations, their assignment to locations ap- proached
the experimental ideal of randomized assignment.
Rosenbaum (1991) reported an impressive series of positive
differences, both in the employment outcomes for adults and in
developmental outcomes for their children, for the families as-
signed to the suburban as opposed to the city locations. A crucial
question for reconciling the large effects found by Rosenbaum with
the more modest ones found in the nonexperimental litera- ture is
whether these effects were produced because of the
quasiexperimental nature of his data, because large neighbor- hood
effects exist for underclass Blacks but not for other popula- tion
groups, or because the volunteer nature of hls sample produced
larger effects than would be the case for a more gen- eral sample
of low-income, inner-city Blacks.
An alternative approach to assessing the strength of con-
textual effects relies on correlations between children who are
neighbors or classmates or on the explained variance of
neighborhoods, schools, or classmates to provide an upper bound on
the possible effect of these contexts. Many studies have used
sibling correlations to estimate the importance of shared family
and other environmental experiences. For ex- ample, sibling
correlations for years of completed schooling are quite high-around
.%-indicating that there are impor- tant elements of the families
(including genetic influences), neighborhoods, schools, and other
aspects of the shared envi- ronments of siblings that make the
siblings much more alike in terms of completed schooling than two
individuals drawn at random from the population.
An analogous correlation for children growing up in the same
neighborhood but not in the same family indicates how much of what
is important in the shared environments of sib- lings lies outside
the immediate family. A high com- pleted-schooling correlation for
unrelated neighbor children, for example, is consistent with a
strong neighborhood effect and would imply that shared neighborhood
conditions are an important component of the sibling correlations.
(An alterna- tive interpretation is that the extrafamilial
correlations are driven by the often-similar family backgrounds of
children in neighboring families.) Neighbor correlations close to
zero suggest that family effects are driving the sibling
correlations and that the scope for pure (i.e., extrafamilial)
neighborhood effects is quite small.
National surveys such as the PSID and the National Longi-
tudinal Survey of Youth draw their samples from a set of tightly
clustered neighborhood areas. Thus, these clusters ap- proximate
neighborhood areas and, using an anonymous cluster identification,
it is possible to calculate both sibling and neighbor correlations
for various outcomes of interest. Solon, Page, and Duncan (1997)
calculated such sibling and neighbor correlations with a
representative PSID sample con- sisting of individuals between the
ages of 8 and 16 years in 1968. Neighborhoods were defined by
either sampling clus- ters (if available) or census tracts. For
their outcome mea- sure---years of completed schooling-the sibling
correlation
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
34 DUNCAN AND RAUDENBUSH
must be taken to account for the clustered nature of the data;
statistical methods for doing so are now widely available.3
The second type of analysis uses repeated measurements of the
same outcome variable to estimate trajectories of indi- vidual
growth or change. For example, Huttenlocher, Haight, Bryk, and
Seltzer (1991) asked: How does maternal speech affect vocabulary
growth during the 2nd year of achild's life? To answer this
question, vocabulary was assessed monthly and maternal speech was
hypothesized to predict acceleration in vocabulary. We might
similarly be interested in the chang- ing propensity to commit
crimes during early, middle, and late adolescence, or rates of
change in externalizing behaviors during the transition to middle
school. Because they must in- corporate the within-subject serial
correlation in the errors, analysis methods for such data differ
quite dramatically from those used for cross-sectional models.
Modeling growth versus status. Although most stud- ies of
neighborhood and school effects are static in the sense that they
relate the individual's status at a particular point to his or her
environment, there is growing evidence that a static fo- cus is
misguided. The status of a chlld or adolescent at a given time
reflects the cumulative effects of all past contexts-in- cluding
home learning and learning and other experiences oc- curring in
past neighborhoods and schools-and may reflect only slightly the
contribution of the current neighborhood or school. This concern is
amplified by evidence of substantial mobility across neighborhood
and school contexts.
In contrast, a neighborhood or school can more legiti- mately be
held accountable for a change in a child's achieve- ment or
behavior during the time that child lives in that neighborhood or
attends that school. Bryk and Raudenbush (1988) and Bryk,
Raudenbush, and Congdon (1996) found in two data sets that more
than 50% of the variation in rates of math learning can be
attributed to differences between ele- mentary schools. In the same
data sets, less than 20% of the variation in status can be
attributed to differences between schools. The implications are
that a longitudinal design is re- quired to understand the nature
of school effects and that
here are two widely used approaches for ensuring that
statistical infer- ences appropriately reflect the clustered
character of the data. First, one may explicitly model the
variability at each level in a hierarchical model (cf., Bryk &
Raudenbush. 1992), also known as a multilevel model (Goldstein,
1995) or a random coefficients model (Longford, 1993). This
approach en- ables study of variation and covariation at each level
and ensures that stan- dard errors reflect this variation and
covariation. The approach can be imple mented by a variety of
statistical packages, including HLM, MLN, Mixor, and SAS Proc
Mixed. The second approach uses standard least squares esti- mation
of regression coefficients, but computes "robust" standarderrors
(cf., Liang & Zeger, 1986). This approach is useful when the
regression coeffi- cients are of sole interest, for it gives no
information on the variability at each level. However, a benefit of
the approach is that inferences do not depend on distributional
assumptions. The approach can be implemented in several software
packages, including STATA. Cheong, Raudenbush, and Fotiu (1998)
compare the approaches and discuss how they can be used
together.
school effects are best conceptualized as effects on growth
rates rather than on status.
Similar results may hold for neighborhood effects. Be- cause
people move across social settings throughout life, their status on
an outcome at any time represents the cumula- tive effect of all
past settings. However, their rate of change while in a setting is
more directly influenced by that setting. Moreover, repeated
measures data, when analyzed effi- ciently, provide dense
information about develop- ment-more than can be provided by a
snapshot at one time-with likely increases in statistical
precision.
Social settings and mobility. If participants stayed within a
single school or neighborhood during the course of a study, the
study of contextual effects on development would be simpler than it
is. One could compare contexts by compar- ing the rates of growth
of participants in those contexts. How- ever, mobility is
exceedingly common during child- hood-five in six children move at
least once by age 1 5 . ~ Roughly one third of all children move
more than three times, and one sixth move more than five times.
Geographic mobil- ity is especially common in early childhood, with
more than half of all children moving at least once between birth
and age 3, nearly half moving between age 4 and 6, and between 25%
and 41% moving at least once in the 3-year periods in middle
childhood and early adolescence.
Thus, in contrast to participants in a cross-sectional analy-
sis, people in longitudinal studies will generally not be purely
nested within contexts. Rather, the data will have a
cross-classified structure, with time series data cross-classified
by individuals and settings. Raudenbush (1993) proposed and
estimated a cross-classified ran- dom-effects model for studying
the growth of children as they move across social contexts.
Sample Sizes and Resource Allocation
Number of time points per person. The number of time points per
person is determined by the frequency of ob- servation and the
length of the study. More time points are needed when trajectories
are more complex. For example, vo- cabulary over the life course
likely follows an exponential growth and decay function, or "S
curve." Multiple assess- ments are needed to fit such a curve.
However, one sees only a piece of that curve during a small range
of ages. Thus, since we see upward acceleration during the 2nd year
of life, we need a quadratic function to represent growth; growth
may look nearly linear for several years thereafter. Thus, more
fre-
- ---
4 ~ h e s e mobility calculations are based on unpublished data
from the na- tionally representative PSID sample of children.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
EFFECTS OF CONTEXT 35
quent observations are required to study vocabulary growth
around age 2 than around age 10.
The association between age and antisocial behavior looks like a
bell-shaped curve between ages 1 1 and 21 (an accelerat- ing
propensity during early adolescence reaching a peak at around age
17 and then declining during early adulthood). A nonlinear function
with at least three parameters is required to fit this curve. Apart
from the dramatically different learning rates during the summer
and the academic year, growth in reading and math look nearly
linear during the early school years.
The total number of participants. Of course, adding time points
costs money and thus reduces the overall sample size, given fixed
resources for the study. The trade-offs need not be one for one,
however. Duncan, Juster, and Morgan (1984) estimated that
interviewer costs associated with con- tact and persuasion make a
second observation on the same family cost only about two thirds as
much as the first observa- tion on that or another family. Overall,
the sample size must be very large if it is believed essential to
describe mean trajec- tories accurately for many subgroups defined,
for example, by gender, ethnicity, and social status. The precision
of esti- mation of the coefficients associated with time-invariant
covariates will depend largely on the total number of partici-
pants.
The number of contexts. Sampling many contexts is generally
expensive because travel costs for interviewing grow with the
geographic dispersion of the participants. In the case of schools,
there is also often a large cost of obtaining en- try to that
context. On the other hand, the precision of estima- tion of
contextual effects (coefficients associated with con- texts)
depends strongly on sampling a large number of contexts.
Variabiiity in contextual characteristics. The rela- tively
limited variability in neighborhood conditions found in sections of
cities or even in entire geographic areas of cities poses difficult
trade-offs for study design. To illustrate the scope of the
problem, we drew tract-based data from the 1980 decennial census.
We formed subsets of tracts to approximate typical study designs:
(a) all tracts in the United States (to ap- proximate national
samples); (b) all tracts in the city of Chi- cago (to approximate a
large study in a single but diverse city); (c) all Chicago tracts
with a 30% or greater poverty rate (to approximate an "underclass"
study in a large city); (d) all tracts in the city of Atlanta, GA
(to approximate a large study in a less diverse city); (e) all
Atlanta tracts with a 30% or greater rate (to approximate an
"underclass" study in a less di- verse large city); and (f) all
tracts in the city of Rochester, NY (to approximate a study in a
medium-sized city).
We drew from the census files seven tract-level demo- graphic
measures often used in neighborhood-based re-
search: race-the percentage of individuals in the tract who are
Black; female headship-the percentage of households headed by
women; welfare-the percentage of households receiving public
assistance; poverty-the percentage of nonelderly individuals with
below-poverty household in- comes; high education-the percentage of
adults with college degrees; neighborhood stability-the percentage
of house- holds who lived in the same dwelling 5 years before; and
job- lessness-the percentage of adult men who worked less than 26
weeks in 1979.
Descriptive statistics for these measures differed dramati-
cally across the subsets. Although the standard deviations of the
seven measures were as great in Chicago, Atlanta, and Rochester as
in the entire set of U.S. Census tracts, limiting tracts to areas
of cities with a relatively high poverty rate re- duced the
standard deviations substantially, especially for the schooling and
residential stability measures.
An important analytic concern is the extent to which sam- pled
neighborhoods enable analysts to estimate the distinct effects of
neighborhood characteristics. For example, there are theoretical
reasons to suspect that concentrations of low- and high-SES
neighbors have distinct effects on child out- comes. But if
measures of low and high SES are too closely correlated in the
chosen sample of tracts, then it will be im- possible to
distinguish their separate effects.
To assess potential multicolinearity problems, we took our
collection of tracts and regressed each of the neighborhood
measures on the remaining six neighborhood measures.5 A high
R-squared indicates a great deal of multicolinearity; modest
R-squared suggest the potential for estimating distinct effects.
Table 2 presents the R-squared from the 42 (7 mea- sures by 6
geographic areas) regressions.
In almost all cases, multicolinearity is considerably higher in
the city-specific samples than for the national set of tracts. This
was particularly hue for the high-SES, stability, and job- lessness
indicators. For example, only 29% of the variation in the fraction
of college-graduate adults could be accounted for by the other six
measures in the national sets of tracts. In the city-specific
samples, the squared multiple correlations ranged from .31 to -75.
Overall, the average extent to which the city-specific squared
correlations exceeded those for all U.S. tracts ranged from .07 to
.25.
Endogeneity of Contextual Effects
The contexts in which children develop are not allocated by a
random process. This is most clearly seen in the case of selec-
tion of preschool child-care arrangements, in which the deci- sion
is almost always made by the parent and is affected by parental
preferences, financial constraints, and local supply
-
'For example, the .53 entry in the first row and column of Table
2 comes from a regression, using all U.S. Census tracts, of race on
the six other tract-based measures of SES. The R-squared from that
regression was .53.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
36 DUNCAN AND RAUDENBUSH
TABLE 2 Fraction of Variance in a Given Census-Based
Neighborhood Measure Explained by Six Other Census-Based
Measures,
by Geographic Area
U.S." chicagob Chicago 30+%' ~ t l a n t a ~ Atlanta 30+%'
~ochesref
Race Black
Female headship Households headed by women
Welfare Households receiving public assistance
Poverty Nonelderly individuals with below-poverty
household incomes High education
Adults with college degrees Neighborhood stability
Households living in the same dwelling 5 years before
Joblessness Adult men working less than 26 weeks in
1979 Average difference from all U.S. tracts
Note. All data are in fractions. "All U.S. tracts. b ~ l l
Chicago tracts. 'All Chicago tracts with 30+% poverty rates
Rochester. NY, tracts.
conditions. A child's immediate neighborhood and, to a somewhat
smaller extent, schools also have an element of pa- rental choice.
The propensity of individuals to choose higher or lower quality
child care or to move to better or worse neighborhoods or schools
depends on background character- istics and current circumstances.
Apart from the rare case of pseudoexperiments such as Gautreaux or
genuine experi- ments such as the one in Tennessee with class size
(Mosteller, 1995), substantial effort is required to model these
propensi- ties as a precondition to drawing conclusions regarding
the causal nature of context influences.
The possibility of bias in estimates from nonexperimental data
arising from nonrandom parental selection of context is more
certain than its likely direction. Suppose parents choose between
(a) holding two jobs and using the extra income to buy a better
neighborhood and (b) having a single earner and living in a poorer
neighborhood. Suppose further that those who live in poorer
neighborhoods or send their children to worse schools, or both,
make up for the deficiencies of the neighborhood or school through
the additional time that mothers spend with their children.
Neighborhood or school conditions matter in this scenario, but an
empirical analysis will show this to be the case only if it adjusts
for differences in parental time use. Failure to adjust for
parental employment will cause conventional regression-based
approaches to un- derstate neighborhood or school effects.
Another scenario, also leading to an understatement of
neighborhood or school effects, is one in which parents
well-equipped to resist the effects of bad neighborhoods
. d ~ l l Atlanta, GA, tracts."All Atlanta, GA, tracts with 30+%
poverty rates. 'AH
choose to live in them to take advantage of cheaper housing or
perhaps shorter commuting times. Unless measures of paren- tal
competence are included in the model, the estimated ef- fects of
bad neighborhoods or schools on child development will be smaller
than if parents were randomly allocated across neighborhoods.
It is perhaps more likely that parents especially ill-equipped
to handle bad neighborhoods or schools are most likely to live in
them, because these parents lack the (partly unmeasured)
wherewithal to move to better neighborhoods. In this case, the
coincidence of a poor neighborhood or school and the poor
developmental outcomes of their children results from their
inability to avoid either, thus leading to an overesti- mation of
the effects of current neighborhood conditions. Conversely, parents
who are effective in promoting the de- velopmental success of their
children may find their neigh- borhood choices dominated by
considerations of developmental consequences. If this capacity is
not captured in measured parental characteristics, then the
coincidence of positive developmental outcomes for their children
and liv- ing in a better neighborhood would be misattributed to
cur- rent neighborhood conditions and also lead to an
overestimation of neighborhood effects. In terms of a re- gression
model relating some child outcome to family and neighborhood
characteristics, the omitted factors amount to unobserved
characteristics of the parents (e.g., concern for their children's
development) that affect developmental outcomes. A key problem with
most existing studies is that they estimate regressions without
controlling for all relevant
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
EFFECTS OF CONTEXT 37
parenting variables, thus biasing estimates of contextual ef-
fects.
There are three approaches for addressing this problem. The best
situation would be one in which families are ran- domly assigned to
child care settings, neighborhoods, or schools. The Department of
Housing and Urban Develop- ment's Moving to Opportunity experiment
contains such ex- perimental data on neighborhood context. With
funding for 10 years, Moving to Opprotunity is randomly assigning
hous- ing-project residents in five of the nation's largest cities
to (a) a group that is moved to a low-poverty area; (b) a control
group receiving conventional Section 8 housing assistance; (c) a
second control group receiving no special assistance. Second-best
solutions to the nonrandom neighborhood selec- tion problem are
quasiexperimental data such as Gautreaux.
A nonexperimental approach to the selection bias problem is to
locate data that measure the crucial omitted variables. Some
child-development data sets contain fairly sophisti- cated measures
of parenting characteristics, including, for ex- ample, an
assessment of the home learning environment provided by parents.
Controls for such measures in regres- sion-based analyses can help
reduce the omitted-variable bias.
Another nonexperimental approach to the bias problem is to
replace the contextual measures in the regression analysis of
interest with an instrumental variable for those variables. This
instrumental variable is purged of the measures' spuri- ous
correlation with unobserved parenting measures. Instru- mental
variables estimation is a two-step procedure. The first consists of
predicting the contextual variables themselves, ideally using at
least some independent variables that do not belong in the child
outcome equation. In the second, the child outcome equation is
estimated using the predicted value of context obtained in the
first stage.
A study by Evans, Oates, and Schwab (1992) adopted this strategy
to investigate selection bias in school-based effects, although it
relied on dubious instrumental variables. Their dependent variables
of interest are high school completion and out-of-wedlock teen
childbearing. Their contextual vari- able was the SES of the
student body. When they ignored se- lection issues and regressed
their outcomes on student-body SES and family-level controls, they
found highly significant, beneficial effects of high student-body
SES. However, when they estimated a two-equation model, with the
first equation regressing student-body SES on characteristics of
the metro- politan area in which the student resided and the second
re- gressing the developmental outcomes on predicted student-body
SES and family-level controls, the effects of student-body SES
disappeared.
Yet another approach to the bias problem is to eliminate the
biasing influence of omitted persistent, unmeasured pa- rental
characteristics by differencing them out using sib- ling-based
fixed-effects models. In fixed-effects models, each sibling's
scores on the dependent and independent vari- ables are subtracted
from the average values of all siblings in
the family. The influence of persistent family characteristics
that affect residential choices, whether measurable or not, are
differenced out of the model. However, as Griliches (1979) pointed
out, differencing between siblings reduces but does not eliminate
endogenous variation in neighborhood regressors and, at the same
time, filters out much of the exog- enous variation as well.
Aaronson (in press) demonstrated the feasibility of this ap-
proach on PSID adolescents. He used family residential changes as a
source of neighborhood background variation within families to
estimate sibling-based neighborhood ef- fects that are
substantially free of family-specific heterogene- ity biases
associated with neighborhood selection. Using a sample of
multiple-child PSID families in which the adoles- cent siblings are
separated in age by at least 3 years, he esti- mated family
fixed-effect equations of children's educational outcomes and found
that the impact of neighborhoods exists even when family-specific
unobservables are controlled for. In fact, family fixed-effect
regressions that use the neighbor- hood poverty rate as the measure
of neighborhood conditions show even larger neighborhood effects on
high-school gradu- ation and grades completed than the models
without fixed ef- fects.
Measuring Contextual Characteristics
Alternative approaches to assessment of contextual charac-
teristics have strong implications for design. First, neighbor-
hood data from the decennial census offer a number of mea- sures of
the economic and demographic composition of individuals and
families in the census tracts in which sample children live. The
data provided by the decennial census about these neighborhood
areas come from the census forms the population is asked to fill
out on April 1 of the 1st year of every decade. Abundant
information about the economic and demographic characteristics of
the population is provided by the completed census forms. As
illustrated by the data pre- sented in Table 2, this enables one to
characterize neighbor- hoods according to a number of key
dimensions, such as the extent of neighborhood poverty, female
headship, public as- sistance receipt, and male joblessness.
Regrettably absent from the census forms are measures of crime,
drugs, gang activity, neighborhood collective effi- cacy, churches,
community centers, and school quality. Other national databases on
neighborhood conditions provide such data, but on either county or
zip code areas. These are large geographic areas that contain
substantial internal variation in neighborhood conditions. Other
administrative databases can be used for measuring certain physical
characteristics of neighborhoods and schools as well as certain
ecological risk factors (such as crime and infant mortality rates
of neighbor- hoods). However, none of these data sources are
appropriate for assessing the social organizational dynamics that
have more proximal theoretical linkages to outcomes.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
38 DUNCAN AND RAUDENBUSH
Second, the use of participants or parents of participants as
informants about the qualities of their neighborhoods or schools is
generally not to be recommended. One reason for this is that
measurement errors in these assessments are likely to be correlated
with the measurement errors of other predic- tors and many
outcomes. For example, parents' mental health may lead their
reports of neighborhood context to be spuri- ously correlated with
their reports of their children's social behavior. And participant
reports of peer attitudes will be spuriously correlated with
participant self-reports of attitudes and behavior. We do not
dispute the utility of aggregating de- mographic characteristics
such as ethnicity, sex, or social class to construct segregation
indexes or other measures of social composition. For example, Lee
and Bryk (1989) con- structed measures of the social and ethnic
composition of U.S. high schools from student survey data and used
those measures to predict the same students' academic achieve-
ment. There is a small risk that a student's report of demo-
graphic background is influenced by his or her achievement. In
contrast, an aggregated measure of perceived instructional quality
would quite plausibly reflect the achievement of the reporters, and
it would therefore be inadvisable to use such a measure of
instructional quality as a predictor of their achievement.
A much more satisfying, if expensive, strategy is to obtain an
independent sample of capable informants about a context and to
pool their reports to create context-level measures. Ttus approach
has been successfully used in national data on school climate (cf.,
Raudenbush, Rowan, & Kang, 1991), with multiple teachers
surveyed about their degree of control, collaboration, and
supportive administrative leadership; and in data assessing the
social cohesion, informal social control, and collective efficacy
of neighbors in Chicago (Sampson et al., 1997). In both cases, 15
to 30 informants per context were required to obtain reliable
contextual-level measures. Clearly, the expense of this measurement
strategy grows rap- idly with the number of contexts sampled, and
it will increase during the course of a longitudinal study as
mobility creates greater dispersion of participants across contexts
and, hence, produces more contexts to be assessed.
Systematic social observation (SSO; Reiss, 1971) pro- vides an
alternative source of contextual information inde- pendent of the
sample of participants. Using this approach, trained observers can
fairly quickly assess aspects of a neigh- borhood, such as its
degree of social and physical disorder. In- terviewers dispatched
to conduct interviews can also be used to conduct such observations
at an expense that is far less than that of conducting an
independent survey of residents.
The Project on Human Development in Chicago Neigh- borhoods
implemented SSO by having a van drive 5 miles an hour down every
street within 80 target neighborhood clus- ters. Videotape
recorders on both sides of the van captured physical
characteristics of the streets and buildings on each side of the
street as well as visible aspects of social interac- tion. Trained
observers then coded the videotapes, noting the
status of buildings (residential versus commercial, detached
homes or apartments, whether vacant or burned out, their general
condition, presence of security precautions such as bars or grates,
etc.), presence of garbage, litter, graffiti, drug paraphernalia,
broken bottles, abandoned cars, and other as- pects of the physical
environment.
The driver and a second rider in the van, trained to observe
social interactions, also recorded their observations via au-
diotape. Social interactions included, for example, adults drinking
in public, drug sales, children playing in the street, and apparent
gang activity. Scales tapping social and physical disorder, housing
conditions, and other aspects of the neigh- borhood environment
showed high internal consistency across face blocks within
neighborhood clusters and reason- ably high construct validity as
indicated by correlations with theoretically linked constructs
measured by an independent community survey, by the census, and by
official crime data. Analyses now underway are estimating the value
added by the videotapes, above the information gleaned from the au-
diotapes. Generally, the videotaped data are far more expen- sive
than the audiotaped data. It is feasible to use the audiotape
strategy even when samples are not highly clus- tered because data
collection per block face is comparatively cheap.
SSO has substantial promise for efficient collection of data on
the social organization of neighborhoods-data not available from
administrative records. However, some of the constructs that can be
captured through interviews, such as "collective efficacy" in
Sampson et al. (1997), are not acces- sible via observational
methods. Given the expense of inter- viewing residents in
unclustered samples, researchers interested in neighborhood effects
face difficult trade-offs, which are discussed later. Similar
trade-offs face school re- searchers, who might opt for
observational measures (cf., Mortimore, Sammons, Stoll, Lewis,
& Ecob, 1988) as an al- ternative to survey methods designed to
capture school orga- nization and climate.
BOTTOM-LINE RECOMMENDATIONS
The multitude of possible designs for developmental studies
makes it difficult to present a succinct set of recommenda- tions.
We organize our summary discussion with recommen- dations relevant
for any developmental study, followed by recommendations for
studies that focus on specific periods of childhood.
Universal Recommendations
1. The diversity of individual developmental trajectories argues
for a longitudinal design in which outcomes and ex- planatory
factors of interest are measured on at least several occasions.
Capturing the dynamics of achievement or behav-
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
EFFECTS OF CONTEXT 39
ior surrounding transitions (e.g., from elementary to middle
school) or conditions (e.g., summer vs. school months) of in-
terest requires outcome measurement before and after the
transitions or conditions. Measurement at equal intervals across
the study period is less important than measurement surrounding key
transitions and conditions of interest.
2. All else equal, nonexperimental studies of contextual effects
are best conducted using data collected from partici- pants living
in very diverse contexts. In the case of neighbor- hood variation,
this argues for representative samples drawn from many diverse
neighborhoods. In the case of schools, this argues for samples
drawn from many schools in varied set- tings.
3. If the only goal of a study is to estimate effect sizes via
regression coefficients, and if the cost of data collection were
irrelevant, clustering samples by context (e.g., sampling mul-
tiple families within neighborhoods or multiple students within
classrooms or schools) is undesirable. Such clustering creates a
(statistically) inefficient dependence across obser- vations. In
the absence of cost savings in collecting interview or contextual
information, the optimal design would sample one participant per
context. Although such a design would be optimal for estimating
regression coefficients, it would pro- vide no information about
variation within and between con- texts. To the extent it is
important to gauge the magnitude of unmeasured sources of variation
within and between con- texts, the unclustered sample design is
problematic, even when costs are ignored.
4. In most instances, interviewing costs associated with
additional participants per context are substantially lower than
costs associated with participants drawn from different contexts.
The trade-off between interviewing costs and statis- tical
efficiency has long been a concern of sampling statisti- cians and
typically leads to designs with relatively modest cluster
sizes.
5. Per-participant costs of contextual information vary widely.
The costs of administrative data, such as census-tract demographic
conditions, school expenditures, or la- bor-market employment
conditions, are typically small and largely independent of the
number of contexts in which par- ticipants are found. Thus,
exclusive use of administra- tive-based contextual information
provides no rationale for clustering samples within context.
6. The utility of administrative data makes it crucial to
identify the location of participants on all measurement occa-
sions. In the case of schools, this means identifying the school
and school district. In the case of neighborhoods, this means the
identification of census tract and zip code. Dwell- ing-based
sample surveys almost always identify tract as part of the data
used to draw their samples, making the identifica- tion of Wave- 1
neighborhood location exceedingly simple. Residential mobility
makes it necessary to invest some re- sources in identification of
census tract in subsequent waves. Geographic Information Systems,
which facilitate geocoding of address data, pinpoint the census
tract or block group in
which interviewing takes place. If addresses are tracked for
purposes of respondent payment or other reasons, then com- mercial
services are available to convert addresses into tract
identifiers.
7. Contextual data drawn from nonadministrative
sources-teachers, independent representative samples of in-
dividuals in the context, or time-intensive systematic obser-
vation of the context-give rise to cost functions in which study
costs increase almost linearly with the number of sam- pled
contexts. These instances may argue for more heavily clustered
samples.
8. Contextual information reported by individuals them- selves
or obtained by aggregating information drawn from in- dividuals who
are part of the context is typically problematic. Data on ethnicity
and SES may be confidently aggregated from participants to
characterize those aspects of neighbor- hoods or schools (assuming
reasonably large samples of par- ticipants per neighborhood or
school). However, such data are typically available from records,
and because the records are based on larger samples (or even a
census), the record data will typically be more reliable.
Participant reports of the orga- nizational health or climate of
neighborhoods and schools should, in general, be avoided due to
reliability and bias prob- lems.
9. Contextual information drawn from independent sam- ples is
very expensive in unclustered samples and becomes more so as
mobility increases the number of contexts in which participants
reside. Study objectives may require the collec- tion of such
information, in which case, sample sizes in the 15 to 25 range
appear sufficient. If not, gathering such informa- tion is probably
too costly to be warranted.
10. We view systematic observational methods of gather- ing
contextual information as an underutilized but very prom- ising
compromise strategy for gathering needed contextual information.
Trained interviewer ratings of the learning envi- ronment of
child-care settings and SSO of neighborhood or school settings are
examples.
Additional Recommendations for Contextual Studies of Preschool
Children
1. Child-care settings are perhaps the most important
extrafamilial contexts for preschool children. Gauging the ef-
fects of child-care settings on children is exceedingly difficult
given the likely biases associated with the high degree of choice
in parental selection of child-care arrangements for their
children. Studies focused on the effects of child-care
characteristics should consider extraordinary measures (e.g.,
experimental or quasiexperimental design) to solve the prob- lem of
likely bias associated with nonrandom selection.
2. In the case of intensive studies of language and perhaps
other aspects of the cognitive development of preschool chil- dren,
the likely nonlinear trajectories argue for numerous
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
40 DUNCAN AND RAUDENBUSH
measurements of outcomes over likely periods of rapid
change.
Additional Recommendations for Contextual Studies of School-Age
Children
1. School and neighborhood settings are the most impor- tant
extrafamilial contexts for school-age children. Gauging the effects
of these contexts on children is complicated by the degree of
choice in parental selection of neighborhoods and schools for their
children. Studies focused on the effects of neighborhoods or
schools should consider carefully how their proposed designs might
solve the problem of likely bias asso- ciated with nonrandom
selection.
2. In the case of achievement during the school years, dif-
ferential rates of learning between summer and during the school
year argue for at least two measurements per year.
In summary, problems associated with obtaining reliable and
valid measures of the social organization of neighbor- hoods or
schools and unbiased estimates of the effects of these contexts on
child and adolescent outcomes present the researcher with sobering
design choices. If nationally repre- sentative descriptions of
child development are essential, it is hard to imagine a feasible
research strategy that would also provide exceptionally
high-quality measures of context. For example, the decennial census
provides high-quality and na- tionally uniform measures of
neighborhood context, but only for the demographic measures sought
in the census enumera- tion forms. The connection between these
demographic mea- sures and crucial theoretical dimensions of
neighborhood conditions is often remote. Similar problems exist for
school-based administrative record data and theoretically de-
sirable measures of school context.
A tempting alternative for broad-based studies is to mea- sure
context using the aggregated responses to theoretically appropriate
questions from clusters of neighboring children, classmates or
parents. We argue against this strategy because measurement errors
in these assessments are likely to be cor- related with the
measurement errors of other predictors and many outcomes. Moreover,
individual reports of context are likely to be quite
unreliable.
Smaller-scale studies are better suited for preferred mea-
surement strategies such as SSO. The problem with such studies is
that they may restrict unduly the variability of con- textual
conditions and not support the estimation of the rich,
multidimensional contextual models implied by theory.
A final, and perhaps most difficult problem, is that of
nonrandom selection of parents and children into their natu- rally
occumng contexts. Pursuit of sturdy causal inferences in light of
the high degree of self-selection in parents' choice of context
pushes the designer in the direction of randomized experiments or
carefully controlled quasiexperiments. Thus, it appears that there
is no single study design and no single
study for best understanding how neighborhood and school
contexts affect child development. Required, then, is a sensi- ble
research agenda drawing evidence from nationally repre- sentative
data, geographically concentrated data, experimental data, and
quasiexperimental data. Durable knowledge will cumulate as the
research community synthe- sizes evidence from these multiple
sources.
ACKNOWLEDGMENTS
Portions of this article were presented at the conference "Re-
search Ideas and Data Needs for Studying the Well-Being of Children
and Families," October 2 1-23, 1997, Washington, DC. It has
benefitted from discussions with Gary Solon, Jens Ludwig, and
fellow members of the MacArthur Foundation Methodology Working
Group-Robert Sampson, Tom Cook, Helena Kraemer, Ron Kessler, and
John Nesselroade. We are grateful to the Family and Child
Well-Being Research Net- work of the National Institute of Child
Health and Human De- velopment for supporting this research and to
Eric Petersen for research assistance.
REFERENCES
Aaronson, D. (in press). Sibling estimates of neighborhood
effect. Journal of' Human Resources.
Bronfenbrenner, U. (1979). The ecology of human development:
Experi- ments by nature and design. Cambridge, MA: Harvard
University Press.
Brooks-Gunn, J., Duncan, G., & Aber, J . L. (1997).
Neighborhoodpoverty: Context and consequences for children. New
York: Russell Sage Foun- dation.
Brooks-Gunn, J.,Duncan, G., Klebanov, P., & Sealand, N.
(1993). Do neigh- borhoods affect child and adolescent development?
American Journal of Sociology, 99, 353-395.
Bryk, A., & Raudenbush, S. (1988). Toward a more appropriate
conceptual- ization of research on school effects: A three-level
hierarchical linear model. American Journal of Education,
97,65-108.
Bryk, A,, & Raudenbush, S. W. (1992). Hierarchical linear
models for social and behavioral research: Applications and data
analysis methods. Newbury Park, CA: Sage.
Bryk, A. S., Raudenbush, S. W., & Congdon, R. T. (1996).
HLM: Hierarchi- cal linear and nonlinear modeling with the HLMRLl
and HLMI3Lpro- grams. Chicago: Scientific Software
International.
Cain, G., &Watts, H. (1972). Problems in making policy
inferences from the Coleman report. American Sociological Review,
35. 228-252.
Chase-Lansdale, L., Gordon, R., Brooks-Gunn, J., & Klebanov,
P. (1997). Neighborhood and family influences on the intellectual
and behavioral competence of preschool and early school-age
children. In J . Brooks-Gunn, G. Duncan, & L. Aber (Eds.),
Neighborhood poverty: Context and consequences for children (pp.
119-145). New York: Rus- sell Sage Foundation.
Cheong, Y. F., Fotiu. R. P., & Raudenbush, S. W. (1998).
Analytic alterna- tives in the analysis of data from the National
Assessment of Educa- tional Progress (Longitudinal and Multilevel
Methods Project). Ann Arbor: University of Michigan.
Clark, R. (1992). Neighborhood effects on dropping out of school
among teenage boys. Unpublished manuscript.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015
-
Crane, J. (1991). The epidemictheory of ghettosand neighborhood
effects on dropping out and tccnage childbearing. American Journal
of Sociology, %, 1126-1159.
Duncan, G. J., Connell. J., & Klcbanov, P. (1997).
Conceptual and method- ological issues in estimating causal effects
of neighborhoods and family conditions on individual development.
In J. Brooks-Gunn, G. Duncan, & L. Aber (Eds.), Neighborhood
poverty: Context and consequences for children @p. 219-250). New
York: Russell Sage Foundation.
Duncan, G., Juster, F. T., &Morgan. J. (1984). The role of
panel studies in a world of scarce resources. In S. Sudman & M.
A. Spaeth (Eds.), The col- lection and analysis of economic and
consumer behavior data @p. 301-328). Champaign, IL: Bureau of
Economic and Business Re- search.
Evans, W. N., Oates, W. E., & Schwab, R. M. (1992).
Measuring peer group effects: A study of teenage behavior. Journal
of Political Economy, 100,966-991.
Furstenberg, F. F. (1993). How families manage risk and
oppormnity in dan- gerous neighborhoods. In W. J. Wilson (Ed.).
Sociology and the public agenda (pp. 231-258). Newbury Park, CA:
Sage.
Garner, C., & Raudenbush, S. (1991). Neighborhood effects on
educational attainment: A multi-level analysis of the influence of
pupil ability, fam- ily, school, and neighborhood. Sociology of
Education, 64,251-262.
Gephart, M. (1997). Neighborhoods and communities as contexts
for devel- opment. In J. Brooks-Gunn, G. Duncan, & L. Aber
(Eds.), Neighbor- hood poverty: Contcxr and consequences for
children (pp. 1-43). New York: Russell Sage Foundation.
Goldstein, H. (1995). Multilevel statistical models (2nd 4.).
New York: Wiley.
Griliches, Z. (1979). Sibling m&ls and data in economics:
Beginnings of a survey. Journal of Political Economy, 87,
S37-S64.
Haveman, R., & Wolfe, B. (1994). Succeedinggencrations: On
the effects of investments in children. New York: Russell Sage
Foundation.
Hoffenh, S., & Chaplin, D. (1994). Caring for young children
whik parents work: Public policies andprivate strategies.
Washington. DC: The Ur- ban Institute.
Huttenlocher, J., Haight, W., Bryk, A., & Seltzer, M.
(1991). Early vocabu- lary growth: Relation to language input and
gender. Developmenfa1 Psychology, 27,236-249.
Jencks, C., & Mayer, S. (1990). The social consequences of
growing up in a poor neighborhood. In L. Lynn & M. McGeary
(Eds.), Inner-citypov- erty in the United States (pp. 11 1-186).
Washington, DC: National Academy Press.
Lee, V., & Bryk, A. (1989). A multilevel model of the social
distribution of educational achievement. Sociology of Education,
62, 172-192.
Liang, L., & Zeger, S. (1986). Longitudinal data analysis
using generalized linear models. Biometrika, 73, 13-22.
Longford, N. (1993). Random coeflcient models. Oxford, England:
Claren- don.
Manski, C. (1993). Identification of endogenous social effects:
The refle- tion problem. Review of Economic Studies,
60,531-542.
Mortimore, P., Sammons, P., Stoll. L., Lewis, D., &Ecob, R.
(1988). School matters. Los Angeles: University of California
Press.
Mosteller, F. (1995). The Tennessee study of class size in the
early school grades. The Future of Children: Critical Issues for
Children and Youth, 5, 113-125.
Raudenbush, S. (1993). A crossed randomeffects model for
unbalanced data with applications in cross-sectional and
longitudinal research. Journal of Educational Statistics, 18(4), 32
1-349.
Raudenbush. S., Rowan, B., & Kang, S. (1991). A multilevel,
multivariate model for studying school climate in secondary schools
with estima- tion via the EM algorithm. Journal of Educational
Statistics, 14, 295-330.
Reiss. A. J., Jr. (1971). Systematicobse~ations of natural
social phenomena. In H. Costner (Ed.), Sociological methodology
(pp. 3-33). San Fran- cisco: Jossey-Bass.
Rosenbaum, J. (1991). Black pioneers-Do their moves to the
suburbs in- crease economic opportunity for mothers and children?
Housing Policy Debate, 2, 1179-1213.
Rosenthal, R., & Rubin, D. (1982). Comparing effect sizes of
independent studies. Psychological Bulletin, 92, 500-504.
Sampson, R., &Groves, W. (1989). Community structure and
crime: Testing social-disorganization theory. American Journal of
Sociology, 94, 774-802.
Sampson, R., & Lauritsen, J. L. (1994). Violent
victimization and offend- ing: Individual, situational and
community-level risk factors. In A. J. Reiss & J. Roth (Eds.),
Understanding andpreventing violence: So- cial influences, Vol. 3
(pp. 1-1 14). Washington DC: National Acad- emy Press.
Sampson, R., Raudenbush, S., & Earls, F. (1997).
Neighborhoods and vio- lent crime: A multilevel study of collective
efficacy. Science, 277, 918-924.
Shaw. C., & McKay, H. (1942). Juveniledelinquency and urban
areas. Chi- cago: University of Chicago Press.
Solon. G., Page, M., &Duncan, G. (1997). Correlations
between neighbor- ing children in their socioeconomic status as
adults. Unpublished manuscript, University of Michigan, AM
Arbor.
Wilson, W. J. (1987). The truly disadvantaged: The inner city,
the underclass andpublic policy. Chicago: University of Chicago
Press.
Dow
nloa
ded
by [
Uni
vers
ity o
f H
awai
i at M
anoa
] at
13:
34 2
8 Fe
brua
ry 2
015