Some Ways in Which Neighborhoods, Nuclear Families, Friendship Groups, and Schools Jointly Affect Changes in Early Adolescent Development

by Thomas D. Cook, Melissa R. Herman, Meredith Phillips, Richard A. Settersten, Jr.
Some Ways in Which Neighborhoods, Nuclear Families, Friendship Groups, and Schools Jointly Affect Changes in Early Adolescent Development
Thomas D. Cook, Melissa R. Herman, Meredith Phillips, Richard A. Settersten, Jr.
Child Development
Start Page: 
End Page: 
Select license: 
Select License

Child Development, July / August 2002, Volume 73, Number 4, Pages 1283-1309
Some Ways in Which Neighborhoods, Nuclear Families, Friendship Groups,
and Schools Jointly Affect Changes in Early Adolescent Development
Thomas D. Cook, Melissa R. Herman, Meredith Phillips, and Richard A. 5ettersten, Jr.
This study assessed some ways in which schools, neighborhoods, nuclear families, and friendship groups
jointly contribute to positive change during early adolescence. For each context, existing theory was used to
develop a multiattribute index that should promote successful development. Descriptive analyses showed that
the four resulting context indices were only modestly intercorrelated at the individual student level (N =
12,398), but clustered more tightly at the school and neighborhood levels (N = 23 and 151 respectively). Only
for aggregated units did knowing the developmental capacity of anyone context strongly predict the corresponding
capacity of the other contexts. Analyses also revealed that each context facilitated individual change
in a success index that tapped into student academic performance, mental health, and social behavior. However,
individual context effects were only modest in size over the 19 months studied and did not vary much by
context. The joint influence of all four contexts was cumulatively large, however, and because it was generally
additive in form, no constellation of contexts was identified whose total effect reliably surpassed the sum of its
individual context main effects. These results suggest that achieving significant population changes in multidimensional
student growth during early adolescence most likely requires both theory and interventions that are
exp licitly pan-contextual.
Social contents that promote healthier development.
Early adolescence can be a trying period for young
people, and social contexts contribute to making it so
(Bronfenbrenner, 1986). In the United States, early adolescents
must negotiate the transition from elementary
to middle school and to high school. Performance
standards increase along the way, making it
more difficult to obtain good grades (Eccles, Midgley,
Wigfield, Buchanan, & MacIver, 1993) and to participate
in extracurricular activities (Brown, 1998). At
home, conflicts with parents arise, especially over autonomy
and self-discipline. Parents expect more selfregulation
and initiative at the very time that their
early teenagers are beginning to experiment with all
kinds of risky behaviors (Barber, 2002). Peers are central
to this experimentation, whether it be around sex,
drugs, or novel forms of self-expression (Brown, 1990).
The power of peer approval is also reflected in the
growing social significance of who one's friends are
and how successfully one relates to the opposite sex
(Brown, 1990). Relations within the neighborhood
also change. Community-based organizations seem
increasingly less relevant as adolescents age, making
it ever more difficult to keep them attached to
the agencies trying to serve them (Handler et al.,
1995); and some local streets seem both more dangerous
and more exciting than earlier (Cook & Murphy,
Even developmental changes that seem to be aeontextual
are not so. Consider, for example, the biological
change of puberty. Both its onset and consequences
are mediated by family, peer, and school
influences (Simmons & Blyth, 1987). In the same vein
are psychological needs. However universal adolescent
autonomy and identity consolidation may seem
to be, these needs are still expressed in different ways
in different cultures. Social contexts are always causally
involved in human development, although some
links are more direct than others. This article deals with
four of the most direct links- those involving the school,
the neighborhood, the nuclear family, and the friendship
group-and the joint role these four contexts play
in fostering positive development between the early
seventh and late eighth grades (see also Barber &
Olsen, 1997).
It is easy to measure a context's structural features,
be it a school's size, a family's social class, a neighborhood's
racial composition, or the age distribution of a
friendship group. Such variables do not specify the
processes most directly responsible for individual
change, however; for example, how a child responds
to his or her parent's actions and how the parent then
responds to the child (Bronfenbrenner, 1988). Understanding
such processes is all the more important because
they are usually only modestly correlated with
structure (Cronbach, 1982). For example, many chil-
© 2002 by the Society for Research in Child Development, Inc.
All rights reserved. 0009-3920/2002/7304-0020
1284 Child Development
dren from poor or single-parent homes do well in
later life, perhaps because many of the transactions
that occurred in their homes, schools, neighborhoods,
and peer groups were not very different from those
that occur with greater frequency in more economically
advantaged circumstances. Because social structures
are not deterministic, it is invidious to designate
a setting as developmentally superior solely because
of, for example, the race or class characteristics of the
persons in that setting. To describe a developmentally
superior context, existing theory is used that specifies
the microlevel processes presumed to be responsible
for individual change in that context, assuming its
structural attributes to be causally prior to its process
What constitutes a developmentally superior family,
school, neighborhood, or friendship group? In the
family setting, sociologists tend to emphasize structural
features such as size, household income, and the
number of biological parents in the home. To describe
high-quality family processes, psychologists (e.g.,
Steinberg et al., 1992) tend to draw on Baumrind's
(1975) theory of authoritative parenting. This posits
that healthy development results from a composite of
parental warmth, autonomy promotion, and monitoring.
Other attributes of a good family can also be
invoked, however, such as the quality of communication
between parent and child (Baumrind, 1971), exposure
to high culture (Bourdieu, 1977; Furstenberg,
Cook, Eccles, Elder, & Sameroff, 1999), and exposure
to stressful events such as death or divorce (Sameroff,
1983). Unfortunately, no theories specify in detail
how these particular attributes are related to each
other, to positive development, or to family structure.
Indeed, few empirically grounded theories have detailed
interrelationships among any set of family attributes,
although Bradley and Caldwell (1995) discussed
some interrelationships among attributes of
Lists of the attributes of good schools are also
available, but they, too, outnumber theories that relate
these attributes to each other and to individual
development (Maugham, 1994; Talbert & McLaughlin,
1999). Structural factors typically examined include
school size and location, the race and class mix
of the student body, whether a school is public or private,
and what kind of choice among schools a district
policy allows. With regard to within-school processes,
most theories emphasize the school as a community
dedicated to enhancing cognitive performance and
social development (e.g., Bryk, Lee, & Holland, 1993;
Comer, 1995; Slavin et al., 1996). Important causal
roles are assigned to (1) high academic standards for
all students, (2) a core curriculum of basics with much
time spent on task, (3) parents who are involved in
the school, (4) a simple organizational structure that
allows staff a voice in governance, and (5) teachers
who are well trained, know their students, and understand
human development. However, little of this
knowledge of positive school attributes comes from
experiments or strong quasiexperiments (Cook &
Payne, 2002), making it particularly tentative.
Researchers interested in neighborhood structure
typically focus on the percentage of residents who are
poor or unemployed (Jargowsky, 1997; Wilson, 1987)
or on the percentage of adults working in managerial
and professional occupations (Brooks-Cunn, Duncan,
Klebanov, & Sealand, 1993). Process researchers focus
on social cohesion and social control (Sampson, Raudenbush,
& Earls, 1997), social disorganization (Wilson,
1987), neighborhood satisfaction (Wilson, 1987),
adult involvement in local organizations (Elliott,
1989), and the availability of local institutions that
support families and youth (Furstenberg et al., 1999;
McLaughlin, Irby, & Langman, 1994). In the best theories,
structure and process are interconnected, as with
Wilson's (1987) claim that local unemployment and
poverty drive the need for a local drug trade that, in
turn, undermines neighborhood social consensus and
social control. This then encourages middle-class residents
to leave the community, making it harder to
raise those children who remain. Brooks-Gunn et al.
(1993) suggest that neighborhoods with dense concentrations
of white-collar workers provide young
people with more adult models of conventional behavior
and thus reproduce the same behavior in the
next generation.
Most theories of quality friendship groups stress
the extent to which members identify with social
goals that most adults interpret as conventional, with
an emphasis on participating in organized neighborhood
youth groups and avoiding delinquent or otherwise
oppositional peers (Brown, 1990; Elliott, 1989;
Fordham & Ogbu, 1986). However, behavioral risk is
also associated with not having any close friends,
changing friends often, and having friends who are
older by more than a year (Magnusson, Stattin, &
Allen, 1986; Savin-Williams & Berndt, 1990). Also, a
balance of advice from both friends and adults is important,
because early adolescents who turn solely to
friends for advice about school, friends, moods,
drugs, and sex may be especially vulnerable to negative
peer pressure.
Configurational theories that specify how a context's
various process attributes relate to each other to
influence adolescent development are rare. Thus, in
the present study for each context the attributes listed
above were measured, weighted equally, and then
combined additively. This resulted in a unit-weighted,
linearly scaled index of each context's presumed ability
to promote positive development. Future research
may reveal limitations in the scaling and variables selected.
However, results of the present study showed
that each of the four context quality indices so created
was related to changes in adolescent outcomes, thus
providing some empirical support for how these context
attributes were selected and scaled.
The research questions. Scaling is necessary to examine
two questions about interdependencies among
social contexts. Scholars have long surmised that human
lives are embedded in, and shaped by, multiple
contexts that are causally interdependent in complex
ways (Bronfenbrenner & Ceci, 1994; Brown, 1999). For
instance, school differences in early adolescent outcomes
may depend on the distribution of social cliques
in a school (e.g., Rutter, Maugham, Mortimore, &
Ouston, 1979). Neighborhood influences on individual
behavior may depend on the behavioral norms set
by local families (e.g., Massey & Denton, 1993). Effective
parents influence the selection of a child's school,
who the child's friends are, and how their child participates
in local youth programs (e.g., Furstenberg et
al., 1999; Silbereisen, 1995). Moreover, each school receives
students from many neighborhoods that presumably
vary in quality. Each neighborhood contains
children from different schools that can also vary in
quality. Students can have some school friends who
are from their neighborhood and others who are not,
and they can have neighborhood friends who do and
do not attend their own school. It is also important to
remember that both schools and neighborhoods
contain multiple peer groups that can differ widely
in how conventional they are (Kinney, 1993), and
many students have multiple friends whose guardians'
parenting skills may vary considerably. It is interdependencies
such as these that led to our first
research question.
QUESTION 1: How tightly are social contexts coupled
(Weick, 1976) once they are scaled for their
capacity to promote successful development?
To what degree are quality schools associated with
quality neighborhoods, quality friends, and quality
nuclear families? Two substantive literatures suggest
that contexts will be tightly coupled. On the one
hand, studies of the urban underclass portray poor
minority neighborhoods as "reservoirs of dysfunction"
(Kelley, 1997) characterized by troubled schools,
poorly functioning families, dangerous streets, and
peer groups that actively resist conventional goals
(e.g., Jencks & Peterson, 1991; Wilson, 1987). On the
other hand, studies of residential mobility indicate
Cook et al. 1285
that middle-class couples select new places to live
under the assumption that higher housing prices entail
superior neighborhoods, better schools, and both
families and students who are more conventional in
social orientation (e.g., Brett, Stroh, & Reilly, 1993;
Hormuth, 1990).
If tight contextual coupling were discovered it
would have several important implications. The first
is that it would be difficult to trust most past tests of
how a given social context affects young people
(Duncan & Raudenbush, 2001), because school influences,
for example, would then be so co-linear with
neighborhood or parenting influences that it would
be impossible to identify the independent contribution
of each. Tight contextual coupling also has implications
with regard to theories of risk and protection
(Garmezy, 1987; Sameroff, 1983). For example, if most
students lived in consistently worse or better contexts,
then it would not be possible to explore whether better
contexts protect students against worse ones. Few
students would experience the mix of context qualities
that such an analysis would require.
However, social contexts may not cluster tightly
when scaled for their capacity to promote healthy development.
Ethnographic accounts portray many peer
crowds within a school, and these crowds vary in
terms of how conventional they are (Brown, 1990;
Kinney, 1993). Even in the poorest of neighborhoods
there are some families with mostly conventional values
and relationships (Clark, 1983), and some peer groups
are quite conventional in orientation (MacLeod, 1995).
Within peer groups, members differ in many ways
(Sullivan, 1989). MacLeod (1995) described one adolescent
group in which most members attended ineffective
local public schools but one member had,
until recently, attended an exclusive private school.
Finally, siblings from within the same family systematically
differ in many important ways (Plomin,
1994). Such within-context variability suggests that
a given student may attend a better school and live
in a better neighborhood, but still hang out with a
deviant crowd or live in a disorganized family. Likewise,
a student might attend a poor school and live
in a disorganized neighborhood, but still come from
a well-organized family and hang out with conventional
The loose contextual coupling suggested by ethnographic
accounts is at odds with the tight clustering
predicted from assumptions made in work on the urban
underclass and on middle-class residential mobility.
Therefore, one goal of the present study was to
provide empirical estimates of the extent of contextual
clustering and thus to determine whether it is
best characterized as loose or tight.
1286 Child Development
QUESTION 2: How do multiple contexts jointly influence
developmental changes?
If the four contexts (school, family, neighborhood,
and friendship group) are not tightly coupled, it is
possible to examine their singular and, especially,
their joint influence on healthy development. The
simplest model of joint influence is additive. In this
model, each context positively affects development,
but the effects of anyone context do not depend on
the effects of another. This assumption is frequently
made in the literature on risk and protective factors
(Meyer-Probst, Rossler & Teichmann, 1983; Sameroff,
Seifer, Baldwin, & Baldwin, 1993),and has several important
implications. First, a given context can operate
as a risk factor at levels that are negative for positive
development and also as a protective factor at
higher levels. Second, the combined contextual effect
should obviously be larger than any single context effect.
Third, strong empirical support for an additive
model suggests that it is futile to seek to identify those
combinations of contexts in which the joint effect surpasses
the sum of individual context effects. Such
non-additive influences direct research and policy attention
to those specific configurations of contexts
that are especially important for development because
they add an extra causal force to the individual
context effects involved in the statistical interaction.
An additive model can also provide theoretical and
policy guidance, especially when the main effect differences
between contexts are convincing and large in
size. This helps to identify which particular contexts
are responsible for larger or smaller effects between
the age range studied.
Multiplicative models suggest more complex associations
with regard to how interdependencies among
social contexts affect development. For instance,
Brown and Huang (1995) showed that family quality
was linearly related to social development for young
people whose peer groups were average on the quality
scale used; however, this relationship was not linear
for adolescents whose peers were of different quality.
Assuming that this interaction was not due to
chance implies that combining these particular family
and peer factors at the levels specified creates a
unique force over and above any causal forces created
by simply adding the peer and family effects.
Given the four contexts examined in the present
study, it was possible to specify many setting combinations
that might confer more protection than the
sum of the main effects comprising them. Does attending
a better school offer special protection against
poor parenting? Does the combination of a good family
and good peer group inoculate against troubled
neighborhoods and schools? Do good parenting and
schools inoculate against troubled neighborhoods
and friends? The potential combinations are numerous.
However, well-specified claims about interactive
contextual effects are rare. Thus, our intention was
to explore all possible interactions among developmentally
effective schools, neighborhoods, friendship
groups, and nuclear families. The hope was to identify
combinations of contexts that especially protect
young people by more than additively countervailing
against any negative contextual forces in their lives.
In the perhaps rarer case of young people for whom
all social contexts tend toward the positive, the intent
was to learn whether the pattern of individual change
that resulted was positively accelerated and thus
more than merely additive.
Although contextual influences are more important
when they apply across all kinds of young
people, it is necessary to guard against premature overgeneralization.
Theory is not yet developed enough
to support grounded hypotheses about the kinds of
young adolescents who are most and least influenced
by constellations of social contexts. However, the
large sample involved in this study and the heterogeneous
nature of the county from which the sample
was drawn provided that opportunity to explore for
interactions between student attributes and social
contexts taken singly and especially jointly. In particular
we probed for variations by race, gender, family
SES, household composition, and elementary school
achievement test scores.
Setting. This study took place in Prince George's
County, a geographically large county that wraps
around the south and east of Washington, DC. Areas
of the county that are closest to the city are the most
densely populated and feel quite urban; most other
areas are distinctly suburban except for a large but
sparsely populated and rural part in the southeast. In
1990, the year before this research began, 729,268
people resided in the county. Fifty-one percent were
Black, 43% were White, 4% were Asian, and 4% were
Latino. The median household income was $41,265
for Blacks and $46,822 for Whites. Because the corresponding
national incomes were $18,676 and $31,231,
the county was more affluent and racially equal than
the nation at large. Since 1970, the county has experienced
considerable White flight. By 1990, 150,000
Whites had moved out of the county and were replaced
by an approximately similar number of Blacks.
Internal variation in social class was considerable, although
inevitably more limited than in the nation at
Cook et al. 1287
Table 1 Selected 1990 Census Characteristics, Prince George's County versus the United States
Prince George's
County United States
Median M
Household income ($)
Poverty status (%)
Female headed
Female headed and child <18 years
Female headed and child <5 years
Education (%)
At least high school diploma
At least college degree
School enrollment (%)
Public schools
Private schools
Labor force status (%)
In labor force
Men in labor force
Women in labor force
Industry (%)
Health, education, and other
Public administration
Retail trade
Class or work
Private wage and salary
Government workers
State and local government
Federal government
Median M
43,127 48,606
46,822 52,769
41,265 45,197
40,795 46,446
36,195 41,530
Note: N.A. = not available.
large. However, although many communities bordering
on Washington are "inner city" in nature, there are
no Census tracts with poverty rates over 30% and none
are as wealthy as can be found in adjacent Maryland
and Virginia (see Table 1 for relevant county details).
Population andsample selection. The county has 25
middle schools, 23 of which participated in this study.
One nonparticipating school taught children with
special educational needs, and the other was undergoing
extensive multiyear restoration. The population
of the 23 participating schools consisted of 22,314
seventh graders from entering cohorts in the 1990,
1991, and 1992 school years. Of these, 17,276 (77%)
had parents who gave consent to be in the study; and
12,702 of these children then entered the longitudinal
sample, so named because each student had outcome
data at the beginning of seventh and the end of eighth
grade and still attended the same school. All 304 Latino
students were subsequently dropped because
50% of their parents did not provide consent to be in
the study. The remaining sample was believed to be
too small and perhaps too biased for reliable population-
specific results. That left a longitudinal sample of
12,398 White, Black, and Asian students who constitute
the study sample.
To examine bias, the district supplied selected data
on all students in the district during the study years
(N == 25,627), including students from the two nonsampled
schools, those who joined the school system
partway through the year, and those for whom we
did not receive parental consent. The absence of consent
meant that the district would not supply us with
the individual identifiers needed to equate our study
population with the one for which it provided data.
1288 Child Development
Nonetheless, analyses not reported here compared
the percentage of seventh graders of various types in
the district-provided population and in our longitudinal
sample (e.g., the percentage Asian, Black, White,
Latino, male, reduced or free lunch, and even Black
males with reduced or free lunch). Results showed
that the longitudinal sample had a small underrepresentation
of males, students receiving subsidized school
lunches, and Black males who were receiving subsidized
lunches; but for no variable was the distribution
truncated. To illustrate this, the grade point average
(CPA) of the population was compared to that of
the sample because grades are anchored at zero and,
of all the measures examined, CPA showed the most
sample bias in standard deviation units. Participants
in the longitudinal sample were indeed somewhat
better students and were underrepresented at the
lower end of the CPA distribution. However, the sample
and population distributions overlapped considerably
and the sample included many cases at the very
lowest CPA levels. Thus, there was some bias in the
sample relative to the county population, but it did not
lead to nonrepresented values at the extreme of ranges.
Datacollection. Between 1990 and 1995, students
from each cohort provided information about their
beliefs and behaviors on the"Attitude Questionnaire"
completed in early seventh and late eighth grade. Another
questionnaire, the "Climate Questionnaire," dealt
mainly with perceptions of the school's climate and
was administered at the end of seventh grade. Students
were assured of confidentiality and only their
ID numbers appeared on questionnaires. A single
make-up occurred about a week after the original
data collection.
Each spring between 1991 and 1994, approximately
1,500 teachers and school staff provided information
on the climate and management of their
school. Telephone interviews were also conducted with
1,046 parents of Cohort 2 students-54% of those invited
from a random sample within each school. About
1,500 parents of students in the same cohort volunteered
for a face-to-face interview that dealt with their
child, parenting practices, and perceptions of the
neighborhood and school. Parents in this last substudy
were slightly more affluent and educated than
the longitudinal sample average, and their children
had slightly higher CPAs (Nadherny, Marcantonio, &
Cook, 1993).
Annual data were also collected from district records.
At the school level, they dealt with each school's
size, percentage of students in the school lunch program,
the racial composition of students, test score
averages, absenteeism rates, teacher racial composition,
teacher training, and teacher longevity at the
school. At the student level, the information was about
gender; age; race; schools attended; school lunch program
status; absenteeism; quarterly grades in math,
English, social studies, and science; third- and fifthgrade
California Achievement Test (CAT) scores; early
seventh-grade math scores on the Maryland Functional
Math and Reading Test (MRT); and the early
ninth-grade MRTscores in both reading and math.
1990 Census reports were used to describe the tract
in which each student lived along many dimensions,
including its percentage of poor, percentage of Whites,
percentage of adults in managerial and professional
jobs, and percentage of homes with a female head.
Outcome constructs. Nine student outcome constructs
were examined-five from self-reports and
four from school records. Three constructs tapped
into school performance (CPA, test scores, and algebra
placement), three tapped into social behavior
(drug use, acting out, and engagement in conventional
activities outside of class), and two tapped into
mental health (positive self and negative self). It was
anticipated that the ninth student outcome constructschool
attendance-would fit into the mental health
domain, but it was more highly correlated with the
school performance variables.
Table 2 presents the relevant constructs, their seventh-
and eighth-grade means, standard deviations,
and, where appropriate, a representative item. Reliabilities
are also provided. For multi-item tests, reliabilities
are reported as the average of coefficients across
time points within cohorts as well as across cohorts.
All constructs created from multiple items had ex
values of at least .65, most of them notably higher. For
the state achievement examination, the reliability was
the value reported in the manual. Misclassification of
algebra placement was judged to be very rare in
school archives, and therefore was assigned a reliability
of .99. Misclassification of the total number of
archived absences was judged to be more prevalent
but still minor, and thus was assigned a reliability of
.96. All nine constructs also showed clear evidence
of discriminant validity. That is, the individual level
reliability-corrected correlations between constructs
never exceeded .60, whatever the cohort. Nonetheless,
there was distinct patterning among the outcome
correlations, and they generally fell into the three
domains noted above-school performance, mental
health, and social behavior.
Dimensionality was less clear longitudinally. Raw
changes were moderately intercorrelated among positive
and negative mental health, acting out, drug
use, and CPA. However, there was scant evidence of
discriminant validity among these five, or evidence
that they were related to the other four outcomes, or
Cook et al. 1289
Table 2 Outcome Constructs and Measures
Measure Items Reliability- M SD Descriptions and Sample Questions"
Overall success at seventh grade
First-quarter seventh-grade 4 .83 3.69 .85 Average of first quarter math, science, English, and social science
grade point average grades; data from transcripts.
Attendance in grade 7 1 .95 -.82 .42 Attendance measured from school records, logged.
Conventional activities 11 .80 .54 .23 Hours spent doing homework, organized recreation, reading,
volunteer work, lessons, clubs, or youth groups.
(Lack of) negative self 14 .67 3.66 .74 Feelings of (nonclinical) depression, anger, and negative coping
(Avoidance of) soft drugs 3 .82 .89 .21 "In the last 6 months have you drunk alcohol, smoked cigarettes, or
used marijuana?"
Positive self 19 .70 3.89 .56 Self-efficacy in school, satisfaction with self, likelihood of success in
future life.
Taking pre-algebra in grade 7 1 .99 .86 .35 Math placement in pre-algebra.
(Lack of) misbehavior 11 .81 7.53 2.37 "In the past year have you hit anyone, damaged property, lied to your
parents, cheated on exams, cut classes, done risky things just for the
kick of it, been sent to the principal's office, or stolen property?"
Standardized math test score 1 .93 .00 1.00 Maryland State Math Readiness Test.
Total success 9 .92 .71 .20 Number of above domains in which adolescent is "getting by."
Overall success at eighth grade
Fourth quarter eighth-grade 4 .82 3.48 .90 Average of fourth-quarter math, science, English, and social science
grade point average grades; data from transcripts.
Attendance in grade 8 1 .95 -.88 .43 Attendance measured from school records, logged.
Conventional activities 11 .75 .44 .20 Hours spent doing homework, organized recreation, reading,
volunteer work, lessons, clubs, or youth groups.
(Lack of) negative self 14 .66 3.68 .74 Feelings of (nonclinical) depression, anger, and negative coping
(Avoidance of) soft drugs 3 .75 .81 .28 "In the last 6 months have you drunk alcohol, smoked cigarettes, or
used marijuana?"
Positive self 19 .71 3.82 .61 Self-efficacy in school, satisfaction with self, likelihood of success in
future life.
Taking algebra in grade 8 1 .99 .53 .50 Math placement in pre-algebra.
(Lack of) misbehavior 11 .87 6.18 2.81 "In the past year have you hit anyone, damaged property, lied to your
parents, cheated on exams, cut classes, done risky things just for the
kick of it, been sent to the principal's office, or stolen property?"
Standardized math test score 1 .92 .00 .90 Maryland State Math Readiness Test.
Total success 9 .93 .61 .22 Number of above domains in which adolescent is "getting by."
Individual demographic variables
Male 1 .99 .48 .50 Student report of gender.
California Achievement 2 .98 -.01 .91 Average of third- and fifth-grade achievement test scores.
Test scores
SES 3 .90 2.18 .79 Free or reduced-price lunch ability and parental educational attainment.
Two-parent family 1 .90 .63 .48 Two (biological) parent families contrasted with all other family structures.
White 1 .95 .26 .44 Student report of race.
Asian 1 .95 .04 .20 Student report of race.
a For single-item measures, reliabilities were estimated.
b For multiple-item measures, questions were paraphrased and run together for ease of presentation.
that these last outcomes were related to each other.
Although there was clearly a group of young people
in this county that was simultaneously changing toward
lower grades, more negative social behavior,
and worse mental health, none of the change correlations
exceeded .30. Change was not strongly yoked
across the nine outcomes studied.
Due to space limitations, the present article does
not consider all nine outcome constructs separately
and in detail, but rather focuses on a single success index.
For each outcome, we first developed a threshold
that corresponded to our interpretation of what it
takes to "get by" in American culture. We then added
the number of constructs on which this threshold was
surpassed to create a count from 0 to 9. This was the
foundation of the index. To be included in the analysis,
students also had to have complete data on at
least seven of the nine outcomes. This led to the loss
1290 Child Development
of 3% of the students and required that the success index
be rescaled as the proportion of outcomes on
which a student was getting by. State authorities created
some cutoffs, and we created the rest so that about
two thirds of students would "get by," thus approximating
success levels across constructs. The actual
cutoffs were (1) grades: a quarterly GPAequivalent to
C+; (2) academic achievement: official passing score
on the state competency test; (3) math track: seventhgrade
enrollment in pre-algebra or higher; eighthgrade
enrollment in algebra or higher; (4) school attendance:
fewer than 10 days absent per school year;
(5) conventional activities: engagement in conventional
activities for at least half the free time noted; (6)
positive self: scores greater than 3 (just over the midpoint)
for academic self-efficacy, satisfaction with self,
and positive feelings about the future; (7) negative
self: reverse scored so that scores above 3 indicated a
student who rarely felt depressed and was angry less
than once a month; (8) soft drugs: no previous use of
any drugs, given that this was a very early age for
drug onset; and (9) misbehavior: reverse scored to reflect
never having participated in 8 or more of 11 misbehaviors,
most of which are illegal.
This success index makes several important assumptions.
One is that it is meaningful to conceptualize
success as "consistently getting by." We believe
that low cutoffs are consonant with the democratic
notion that productive and happy citizens need not
excel in anything. They are also consonant with the
cumulative risk assumption that undesirable adult
consequences follow from early adolescent troubles
with some combination of failing school performance,
breaking the law, using drugs early, and being
depressed. The more such negatives there are in individual
lives, the greater is the threat to autonomy as
an adult. Another assumption is that the chosen cutoffs
are not so arbitrary that varying them would entail
different results. At eighth grade, the "getting by"
success index correlated .71 with a composite without
cutoffs that was formed by simply standardizing and
summing all nine scores. It also correlated .74 with an
"excellence" index that depended on high rather than
low cutoff points. Such correlations suggest that the
results would not be radically different if other success
indices had been used. A third assumption is that
lumping the nine outcomes into a single index has
advantages that outweigh the costs. Our later analyses
are essentially of changes in success over 19
months and, as noted above, such changes are largely
independent across the various outcomes. By measuring
the number of areas in which a young person
is getting by, we adopt more of a configurational
than variable-centered conception of young people. It
is a conception that seeks to capture the reality that
young people are judged (and perhaps judge themselves)
multidimensionally rather than unidimensionally.
To study just one domain-or just one variable
within that domain-would be to lose the person.
Still, to avoid overgeneralization, outcome-specific
analyses of the most central research questions are
also reported.
The success index was normally distributed at
each time point and, in this county, students were
generally doing well on it. At the beginning of seventh
grade, the average student was getting by on
over 70% of the outcomes. Nineteen months later, the
average had dropped to just over 60%, primarily because
of higher percentages of students failing the
state competency test, getting lower course grades,
entering lower math tracks, being more involved
with drugs, acting out more, and engaging in fewer
conventional activities outside of class. The same secular
trend was hardly discernible, however, for absenteeism
from school and for positive and negative
self, although females tended to score increasingly
worse on the latter. This general pattern of temporal
decrease meant that it was not necessary to worry
about ceiling effects. However, it also meant that the
analyses would classify some young people as becoming
more successful than others, even though
their absolute scores were declining.
Individual demographic variables. Some of the demographic
or background variables included in our
models require commentary. The average of students'
third- and fifth-grade CAT scores, which were taken
from school records, was used to indicate prior achievement.
However, as many as 32% of the middle-school
students had no elementary-school CAT data (by far
the most for any construct), and therefore these missing
scores were estimated using a maximum likelihood
algorithm (Little & Rubin, 1987). After significant
pilot work, family composition was coded as to
whether children lived with both biological parents
or not. To describe SES, school lunch eligibility alone
was not used because its distribution is so curtailed.
Instead, school lunch information was combined with
student reports of parent education, using the highest
attainment of any residential parent. The high-SES
category included families without a lunch subsidy
that had at least one college graduate parent (N ==
5,056). The moderate-SES category included families
that either (1) received a lunch subsidy and had a college
graduate parent, or (2) did not receive a lunch
subsidy but had a parent who had graduated high
school but not college (total N == 4/045). The low-SES
category included families that either received a free
lunch subsidy or had no parent who had completed
Cook et al. 1291
Table 3 Family Context: Constructs and Measures
Measure Items
Family quality index 6
Family orientation 7
Culture promotion 4
Collaborative decision making 2
Parental control/monitoring 3
Parent / child trust and acceptance
Sum of family stressful events 5
Reliabilitya M SD Descriptions and Sample Questions"
.80 .00 .49
.74 3.50 2.08 "Do you discuss the following with your parents: drugs,
sex, problems with friends, or problems with teachers?"
.61 2.93 .83 "Do your parents encourage you to read in your spare
time, go to the library, museum, concerts, or plays?"
.80 1.69 .63 "Who decides where and how you spend your time after
school and your curfew?"
.57 3.95 .83 Parental response to school problems, rule breaking, and
monitoring of student's homework.
.63 3.39 .66 "My parents trust me" (strongly agree to strongly
.80 1.16 1.01 "In the last year have you moved to a new house, or have
your parents lost their jobs, gotten divorced,
been remarried, or gotten seriously ill?"
aFor single-item measures, reliabilities were estimated.
b For multiple-item measures, questions were paraphrased and run together for ease of presentation.
high school (N == 2,918). Table 2 reports the reliabilities
for the individual difference measures. For single
items they were estimated, with gender and race
being cross-validated across school records and four
self-ratings over 2 years.
Family process measures. Table 3'provides data about
the six family process measures that students rated in
early seventh and late eighth grade. The family orientation
measure speaks to the quality of communication
between parent and child. The culture promotion
measure has to do with visits to museums, concerts,
other parts of the country and abroad, and the like.
Also measured was collaborative decision making,
parental trust and acceptance, and parental control
and monitoring. Within a risk and protective factors
framework it is common to measure family stress
(Sameroff, 1983), although it is less common to exclude
from the measure all items that refer either to
schools, neighborhoods, or peers or to family demography.
However, in the present study, these items
were excluded to obtain a purer family measure. The
six family measures were standardized and summed
to create the family quality composite. Most family
variables were measured at the beginning of seventh
grade, but two components of parental control were
assessed at the end of seventh grade, as was parental
acceptance. No family variables were assessed uniquely
at the end of eighth grade.
Friend process measures. Table 4 presents details on
the eight friend measures, all from self-reports completed
in seventh grade and eighth grade. The friend
orientation measure counted the number of domains
(out of seven) for which a student turned to friends
rather than to parents, school adults, or other adults
for advice. Four measures of negative friends dealt
with those who acted out, used drugs, were sexually
active at this early age, or were older by at least a year.
Also measured was the presence of friends who regularly
participated in extracurricular activities that
most adults consider desirable. Finally, the availability
of close and of continuing friends was assessed,
thus addressing the emotional quality of friendships.
These eight measures were then standardized and
summed to form a composite measure of peer context,
which had more to do with friendship circles than
school crowds or peer groups, and focused more on
deviant friends than positive ones.
Neighborhood process measures. The present study defined
neighborhoods as Census tracts. The county
from which this study's participants were drawn had
173 tracts, 151 of which were used in the analyses.
Of the 22 omitted, 1 tract had no residents, 3 had no
middle schoolers in the study, 13 were primarily
served by the two middle schools not in the original
school sample, 3 straddled a county line but most students
attended the neighboring county's schools, and
2 had far fewer students than Census specifics led us
to expect-perhaps because they disproportionately
contained the 14% of county children who attended
private schools. The loss of 12% of the tracts resulted
in a loss of only 1%of the students. Another 7%of students
could not be assigned to a tract, mostly because
they lived in new housing developments for which
tract boundaries had not been determined in time for
the 1990 Census. These students tended to be more
successful than others by about one fifth of a domain
in both seventh and eighth grade, were from higher
SES families, and were 2% more likely to be Black
.85 .47
2.82 2.00
.67 .39
3.48 .92
2.92 1.17
3.82 1.06
2.73 .98
3.56 1.03
3.17 1.31
1292 Child Development
Table 4 Friendship Context: Constructs and Measures
Measure Items Reliability"
Peer quality index 8 .92
Friend orientation 7 .67
(Lack of) negative friends 3 .79
Friends disapprove of drugs 3 .83
Friends (not) having sex 1 .80
Older friends 1 .80
Positive friends 3 .64
Close friends 1 .80
Continuing friends 1 .80
M SD Items and Sample Questions"
Extent to which you turn to friends (instead of others) with
concerns about school, drugs, sex, and relationships.
"Have any of your friends suggested that you break the law,
damage property, or steal something over $50?"
"How do your friends feel about drinking alcohol, smoking
cigarettes, and using marijuana?"
"How many friends of your own gender have had sexual
"How many of your friends are older than you by at least a
year?" (reverse coded)
"Do your friends participate in religious activities, do
community service, and plan to attend college?"
"Of the kids you know, how many do you consider to be your
close friends?"
"How many of your close friends now were also your friends in
your last grade?"
a For single-item measures, reliabilities were estimated.
b For multiple-item measures, questions were paraphrased and run together for ease of presentation.
than White. Such bias presumably works in the opposite
direction to the attrition bias. Of the 151 tracts in
the study, 135 were pure Census tracts, and 16 were
synthetic tracts created by merging each tract that had
one or two parent respondents with its contiguous
and demographically most similar tract that had
three or more respondents (this was done because
pilot work showed that a minimum of three parent
respondents was required to achieve stable tractlevel
correlations between neighborhood process attributes).
The average number of parents per tract
was 11.
Table 5 describes the seven neighborhood process
attributes. All came from household interviews with
the approximately 1,500 volunteer parents of Cohort
2 students. Respondents were free to interpret their
neighborhood boundaries as they wanted. The data
from reporting parents were aggregated to create
tract-level descriptions of seven constructs: the availability
of local resources for youth, the degree of cohesion,
the degree of social control, the level of social
problems, the participation level of adults in local
organizations, the stability of the neighborhood and
satisfaction with it, plus judgments about how conducive
the neighborhood was to raising successful
children. Factor analyses with tracts as the unit of
analysis revealed that a single factor dominated. Therefore,
an additive composite was formed from the
seven standardized variables.
The 1990 Census was also used to provide information
on structural characteristics of each tract, especially
those related to race and SES (see the lower
part of Table 5). Factor analyses of these variables
suggested a two-factor structure. The first contained
poverty and SES attributes as well as the percentage
of Whites and the percentage of Blacks. (The correlation
between the percentage of Whites and average
SES was .58.) The second tapped into immigration
and residential stability and included the percentage
of Latinos and the percentage of Asians. Discriminating
between neighborhood structural and process
characteristics is difficult because they are so highly
correlated (Cook, Shagle, & Degirmencioglu, 1997);
therefore, we proceeded under the assumption that
neighborhood processes are causally more proximal
to individual change than are neighborhood demographics.
As a result, when multicollinearity at the
tract level made it impossible to include both process
and structural attributes in the same analysis, only
the former was used.
School process measures. The upper panel of Table 6
describes the six school process measures: the developmental
climate of the school, parent involvement
in the school, setting high standards for all students,
teachers having high expectations for students, the
school's specialist to staff ratio, and the average teacher
training level. The six constructs were standardized
and summed to create a composite of school quality.
The analytic models also included some schoollevel
demographic variables: the percentage of Black
students, the percentage of Black teachers, and the average
SES of the student body. These were all modestly
correlated with each other and with some school
process attributes. Specifically, better trained teachers
Cook et al. 1293
Table 5 Neighbor Context: Constructs and Measures
Measure Items Reliability- M SD Descriptions and Sample Questions"
Neighborhood quality index 7 .95 .00 .75
Avail ability of resources 6 .74 3.75 .90 Whether these resources exist: health-care center,
day-care, youth clubs, summer recreation
Social cohesion 4 .77 3.38 .34 "Is yours a close-knit neighborhood where people
share views and information on opportunities for
Social control 3 .87 3.28 .27 "Would people in your neighborhood do something
if there was a fight in front of their house and
someone was being beaten, someone was trying
to sell drugs to their children in plain sight, kids
were getting into trouble?"
Neighborhood problems 12 .95 2.44 .27 Whether some of these problems exist: vandalism,
sexual assault, organized crime, drug use, and
Adult participation in neighborhood 9 .64 3.65 .78 "Have adults in your household participated in:
organizations tenant's council, religious services, CEO
program, volunteer programs, local political
activities, civil rights activities, recreation
programs, etc.?"
Stability and satisfaction with 2 .63 -.01 .73 "Are you satisfied with the neighborhood and do
neighborhood you think you'll live here in 5 years?"
Probability of success for neighborhood 3 .86 3.58 .48 "Chances that neighborhood children will graduate
children high school, complete college, and find a wellpaying
Neighborhood demographic measures
%Black 1 .99 .64 .31 Census variable
%Latino 1 .99 .03 .06 Census variable
%Asian 1 .99 .04 .07 Census variable
%White 1 .99 .28 .29 Census variable
Poverty 3 .84 .00 .87 Percentage of families in poverty, percentage of
female-headed households, and percentage
receiving public assistance.
SES 3 .85 .00 .87 Percentage of college graduates, percentage of
professional managerial, and median household
a For single-item measures, reliabilities were estimated.
b For multiple-item measures, questions were paraphrased and run together for ease of presentation.
who held higher expectations for their students were
overrepresented in the higher SES schools with more
White students. Thus, the demographically most advantaged
schools appeared to get the better teachers.
Even so, demography and process were not as closely
linked for schools as for neighborhoods, which made it
possible to examine statistical models of the association
between school processes and student success that
took into account some potential school-level selection
confounds as well as many individual-level ones.
It is important to note several other things about
the school measures. First, no eighth-grade student
reports were included as school descriptors, because
they could have artifactually inflated relations with
student outcomes measured at approximately the
same time. Second, some school measures could not
be used because of reliable but extremely low intraclass
correlation (ICC) values-for example, school
safety and order. Third, with only 23 schools, other
measures could not be used because of confounding.
For instance, teacher preparation was especially highly
correlated with a student-based measure of academic
climate, making it necessary to drop the student academic
climate measure in favor of the archival measure
based on the years of preparation the average
teacher had had. Also dropped were both student
mobility rates and average CAT scores because of
their extremely high correlations with average SES
and the percentage of Black students.
Statistical analysis. Important theoretical issues were
involved in estimating the degree of clustering among
contexts after they had been scaled for their ability to
1294 Child Development
Table 6 School Context: Constructs and Measures
Measure Items Reliability- M SD Descriptions and Sample Questions"
School Quality Index 6 .90 .00 .48
Social developmental climate 7 .85 .02 .81
Democratic governance 5 .94 3.45 .39 "00 teachers think the principal delegates responsibility, shares
power, supports decisions made by the teams?"
Shared leadership 3 .51 -.02 .13 "How much input does staff have into school activities, how much do
teacher teams collaborate, and how much do teachers seek help
from each other?"
Respect for all voices 4 .71 .00 .16 "00 staff listen to student concerns, is the school sensitive to ethnic
diversity, do teachers use culturally sensitive learning materials?"
Staff morale 4 .72 -.01 .21 "How attached are teachers to the school, how much pride do they
take in working at the school, how effective do they feel?"
Developmental sensitivity A 3 .82 .00 .79 Teacher's use of problem-solving strategies, giving students choices
and responsibilities.
Developmental sensitivity B 3 .76 .00 .91 Teacher's consideration of child or adolescent development when
designing programs and activities.
Perceptions of improvement 4 .82 -.02 .26 "00 teachers perceive an improvement in school climate with respect
to student thinking, staff morale, parent involvement, and school
Parent involvement 2 .98 -.08 .93
Parent invited to school 10 .71 5.02 2.53 Total number of invitations to parent-teacher conferences, plays,
events concerts, sports events, social events, open houses, etc.
Parent attends school events 10 .76 1.50 .38 Actual attendance at parent-teacher conferences, plays, concerts,
sports events, social events, open houses, etc.
High standards for all 4 .82 .01 .79
Reaching difficult students 3 .75 3.39 .20 "How much can you do to reach students who are disinterested,
behaviorally difficult, or don't do homework?"
Instilling the value of 2 .71 3.41 .34 "How often do you link educational performance to gaining respect
education from others and graduating high school?"
Teacher motivation 4 .70 3.10 .27 "00 you motivate students with group assignments, choice of
strategies projects, tailored assignments, or varied instructional strategies?"
Teacher response to skipping 5 .77 3.29 .28 "00 you respond by talking to the student, punishing, contacting the
parents, keeping the student after class, or doing nothing?"
Teacher expectations for students 2 .87 .05 .30
Expectations for attainment 2 .79 .02 .81 "What percentage of your students do you expect will graduate high
school and go on to college?"
Expectations for performance 2 .81 .01 .78 "What proportion of your students are willing to learn, try hard, do
extra-credit work, plan to graduate, and plan to attend college?"
Specialists-to-staff ratio 1 .83 .11 .04 Percentage of staff who work as guidance counselors, media
specialists, or reading specialists.
Teacher training level 1 .98 2.58 .31 District archival data on teacher education ranging from bachelor's to
School demographic measures
Percentage of Black teachers 1 .99 .36 .12 From district archival data.
Percentage of Black students 1 .99 .67 .17 From published district reports.
MeanSES 1 .90 4.46 .32 Aggregated from individual-level SES data.
a For single-item measures, reliabilities were estimated.
b For multiple-item measures, questions were paraphrased and run together for ease of presentation.
promote positive development. But the task is essentially
descriptive and statistical concerns were minor
once corrections for unreliability were made. Little difference
resulted from these adjustments, however,
because three context composites had reliabilities of
at least .92, while the reliability of the school composite
was .80 (Nunnally and Bernstein, 1994, p. 268
was used for estimating reliabilities for composites
created out of multiple constructs with their own
Statistical challenges are numerous when it comes
to learning how multiple contexts jointly influence individual
change. Although cause is implied in the question
formulation, one cannot randomly assign young
people to their families, friendship groups, schools,
and neighborhoods! Statistical adjustments are needed
to deal with selection into settings. Such adjustments
are necessarily imperfect even in the best of circumstances;
unfortunately, the circumstances of the present
research were far from optimal. No instrumental variable
approach seemed plausible (Angrist, Imbrens, &
Rubin, 1996);and a propensity score approach (Shadish,
Cook, & Campbell, 2002) requires a dichotomous
independent variable, whereas having four continuous
context indices constitutes one of the major
strengths of this research design.
Our central assumption is that contexts affect
early adolescent development. However, young persons'
behavior and feelings can also influence which
social contexts they are in, which specific processes
occur in these settings, and whether other people
choose or modify contexts based on their understanding
of how a young person is developing. Such causal
reciprocity suggests that it may be a Sisyphean task to
try to draw unidirectional causal conclusions on the
basis of statistical adjustments alone. Endogeneity
was not a predicament with the neighborhood and
school contexts, however, and early seventh-grade
standing on the success index was not reliably correlated
with either middle school or neighborhood
quality after controlling for individual student background
characteristics. In contrast, endogeneity was
an issue with friend and family quality. Each was measured
from student reports rather than from process
reports aggregated across all family members or from
self-reports aggregated across all those individuals
who sample members named as their friends. Thus,
the family and friend ratings may, to some extent,
have been products rather than causes of a young
person/s development.
Attempts were made to control for this statistically.
One part of the control strategy was to analyze only
19-month changes in the success index, ignoring final
performance levels for anything but descriptive purposes.
But even analyzing change was tricky. Debates
continue about when gain scores or individual trajectories
should be used rather than later scores that control
for earlier ones, the latter being equivalent to analyzing
gain scores while controlling for initial level
(Campbell & Kenny, 1999; Cronbach & Furby, 1970;
Willett, 1989). Because the present study included
only two time points, it was not possible to analyze
trajectories; therefore, we chose to analyze late eighthgrade
scores while statistically controlling for early
seventh-grade scores that had been adjusted for unreliability
by dividing the obtained score by the square
root of the reliability. This analytic decision reflected
our conceptual interest in a counterfactual assessment
of how students in different contexts would
have fared had they started out at similar seventh-
Cook et al. 1295
grade success levels. This was judged to be a more
general question than describing how students gain
when their starting points are already correlated with
the quality of social contexts. The latter is descriptively
meaningful because it makes change contingent
on where individuals start, and in the real world individuals
start at different places. However, the concern
here was analytical: What would happen to young
people who live in different kinds of social contexts if
they were to start out from the same position?
Another part of the statistical control strategy
involved supplementing the early seventh-grade
reliability-adjusted outcomes with individual background
variables-namel~family SES, family composition,
elementary school CAT score, individual race,
gender, and cohort, with reliability adjustments for
the first three of these. Additionally, in explorations of
model sensitivity, the interaction of individual race
and class and of individual race and gender were
sometimes added to the model, as were the school- and
tract-level demographic variables. However, none of
these accounted for much additional variance, given
the individual-level variables already in the model.
Three other technical details are worth noting.
First, only the seventh-grade friend and family quality
indices were used when testing the causal models,
because they did not entail any irrelevant variance
shared with the outcome, as would happen if eighthgrade
measures of these two contexts were related to
a success index that contained the five constructs assessed
at the same time as family and friend quality.
However, in answering descriptive questions about
context clustering in which temporal priority was not
an issue, the seventh- and eighth-grade family and
friend measures were combined to achieve more
valid assessment.
Second, data for the school and neighborhood process
composites were achieved from archives or by aggregating
information from adult respondents within
the context in question. Thus, sample sizes reflected
the number of tracts and schools, not individuals. To
take this design effect into account, hierarchical linear
modeling (HLM) was used when analyzing the school
and neighborhood contexts by themselves. When all
four contexts were analyzed together, the algorithm
built into the computer program STATA was used for
correcting design effects. Because STATA only allows
for clustering on one variable at a timet all analyses
were replicated, taking into account both tract- and
school-level clustering. This rarely led to different
substantive conclusions, and the standard errors used
in significance testing were from models corrected for
clustering within schools.
Finally, the individual-level demographic variables
1296 Child Development
Table 7 Reliability-Adjusted Correlations among Context Qualities
at Three Levels of Aggregation
How tightly were thecontexts coupled? Table 7 reports
the reliability-adjusted correlations among all four
contexts at the individual, tract, and school levels. At
the individual level, all were positive and statistically
significant, given the large sample of students. Thus,
contexts clustered once they had been scaled for their
presumed capacity to promote successful development.
However, the correlations were variable and
generally modest in size. The highest was between
family and friend quality (.43), but this was probably
inflated by irrelevancies of shared measurement because
each was assessed from the same questionnaire.
The next highest correlations were between neighborcould
be construed as causally antecedent to a particular
context and to seventh-grade success. Therefore,
a fully saturated LISREL model was also tested that
made this particular temporal assumption while also
letting the social contexts correlate with seventhgrade
success. The model then postulated direct
causal paths from both context quality and seventhgrade
success to eighth-grade success, as well as indirect
paths from each of the demographic variables via
context quality to eighth-grade success. (The model is
described in visual form as Figures 2, 3, and 4 in the Results
section.) The intent was to have the causal modeling
and regression analyses replicate each other under
slightly different assumptions, mostly with regard to
demographic variables preceding early seventh-grade
success and context quality.
hood and school quality (.18) and between neighborhood
and friend quality (.14). Here, bias from
correlated measurement was not an issue. However,
the magnitudes, although not trivial, provided little
predictive power. The remaining correlations were
even lower. Thus, although social contexts were positively
related in individual lives, they were generally
only loosely related. Knowing the quality of anyone
context did not help to reliably predict the quality of
others, except for the friend-family link.
Correlations between these same context indices
were noticeably higher at the tract and school levels.
Knowing the quality of a school or neighborhood permitted
good prediction of the aggregated peer and
family contexts that could be found within a school or
neighborhood, and knowing the quality of any two
contexts permitted quite accurate prediction of the
quality of the others. For abstractions such as the body
of students in a school or neighborhood, social contexts
clustered quite tightly.
One interpretation of these correlation differences
by level of analysis is methodological. All correlations
increase as measures became more reliable, and
studying neighborhood or school averages controls
for individual differences between students in how
the developmental quality of the four social contexts
relate to each other. Another interpretation is more
substantive. It suggests that the other contexts relate
to each other quite heterogeneously within neighborhoods
or schools. Intraclass correlations estimate how
much of the total variation lies within and between
aggregates such as schools or neighborhoods. Table 8
reports reliability-adjusted ICC values for many of
the different variables. Irrespective of the variable or
context examined, it is clear that most variation lies
within contexts and not between them.
Still, the ICC values did vary. For instance, neither
neighborhoods nor schools differed much in terms of
the cognitive skills of the students within them, nor in
their students' mental health or social behavior. However,
they differed much more with regard to socioeconomic
characteristics, presumably because of classbased
local housing patterns. Equally striking were
context differences in ICC values. The betweenneighborhood
and between-school variation in friend
and family quality never exceeded 3%, suggesting
that these larger contexts were very heterogeneous in
the quality of the families and peer groups located
within them. However, about 27% of the variation in
school quality lay between neighborhoods. The average
neighborhood contained students who attended
seven public schools, and the ICC values point to
schools that were somewhat similar to each other in
quality and were clearly different from what could be
Peer Family School Neighborhood
At the individual level
At the tract level
(N = 151)
At the school level
(N= 23)
Cook et al. 1297
Table 8 Reliability-Adjusted Intraclass Correlations (ICC) by School and Census Tract
ICC by School
ICC by
Census Tract
Grade 7 Grade 8 Grade 7 Grade 8
Test scores
Conventional behavior
(Lack of) misbehavior
Positive self
(Lack of) negative self
(Avoidance of) soft drugs
Taking algebra
Overall success
Contexts qualities
Family quality
Peer quality
Neighborhood quality
School quality
Specific dimensions of school quality
Social developmental climate
Democratic governance
Shared leadership
Respect for all voices
Perceptions of improvement
Developmental sensitivity A
Developmental Sensitivity B
High standards for all
Reaching difficult students
Teacher motivation strategies
Instilling value of education
Teacher response to skipping
Parent involvement
Parent attends school events
Parent invited to school events
Teacher training level
Teacher expectations for students
Specialists-to-staff ratio
Neighborhood quality
Social cohesion
Availability of resources
Adult participation in neighborhood organizations
Social control
Neighborhood problems
Probability of success for neighborhood children
Stability and satisfaction with neighborhood
found in other neighborhoods. Of the variation in
neighborhood quality, as much as 51% lay between
schools. This meant that, although schools drew from
a mean of 39 neighborhoods, the mix of neighborhood
qualities was relatively similar within schools
and was strikingly different from one school to the
next. Otherwise, however, ICC values were not large,
indicating considerable differences within neighborhoods
and schools in the mix of context qualities and
student attributes found there.
How did the four contexts jointly influence the positive
development of individuals? The neighborhood and
school quality composites and the early seventh-grade
measures of family and peer group quality were used
Table 9 Changes in Success Index as a Function of All Four
Contexts and Their Interactions
curvilinearity that would be expected if the most positive
context values promoted atypically large positive
None of the context interactions in Table 9 were reliable.
Accepting the null hypothesis about interaction
hypotheses is tricky. Because statistical power is
highest when cases are sampled from the extremes of
distributions (McClelland & Judd, 1993), the preceding
analysis was replicated using just the top and bottom
quartiles for each context. This also failed to result
in reliable interactions, suggesting that the contexts
jointly influenced early adolescent change in additive
fashion. Each good context promoted healthier development,
and thus may have buffered against bad contexts,
but no combination of contexts implied a special
degree of protection.
Although Table 9 illustrates the modesty of the
context effects taken individually, it also illustrates
how summing their separate effects resulted in a
large total effect. One way to estimate the magnitude
of this joint effect takes advantage of the large sample
size, the absence of reliable interactions among contexts,
and the only modest difference in individual
context coefficients. Those factors justified creating
for each respondent a sum of the four standardized
context qualities. The total quality measure thus
1298 Child Development
to examine this issue. The predictors of greatest interest
were the four (standardized) context composites
and all their possible interactions. However, the model
also included the seventh-grade success index along
with all the individual difference background variables.
In some versions of the general model, the
school and neighborhood demographic descriptors
were also added. All context and individual difference
variables that could be reliability adjusted were,
and the outcome was the index of developmental success
at the end of eighth grade.
Before doing this, each context was analyzed separately
using the same basic model. Wediscovered that
each context composite was related to changes in the
success index, whether analyzed in HLM or LISREL.
A minor but interesting complication occurred with
school quality because the better middle schools
graded increasingly more strictly across the middle
school years, resulting in a negative association between
school quality and changes in CPA. Thus, the
analyses that follow were replicated with and without
CPA in the success index. Being only one of nine
index attributes, however, its deletion made little
practical difference.
Analyses restricted to a single context cannot test
whether each context continues to be effective when
the others are in the model, nor can the functional
form of the relation between multiple social contexts
and changes in success be identified. Table 9 tackles
these issues, and reveals simple subgroup main effects.
Asians were changing at a faster positive rate than
were Blacks; and Blacks were changing at a faster rate
than were Whites (although further analyses showed
that Whites did better over time on academic performance
outcomes and Blacks did better on social behavioral
and mental health outcomes). Students living
with both biological parents changed more positively
than did other students. Females did better over time
than did males, as did students from higher SEShomes
and those with higher elementary school CAT scores.
More importantly, Table 9 shows that each context
continued to predict changes in the success composite
even with the other contexts in the model. However,
the neighborhood effect was smallest and was now
not quite statistically significant-perhaps because of
its correlation with school quality. In any event, it is
unlikely that any of the main effects from analyses of
single contexts was totally due to some other context
operating as an artifact or a mediating variable. In
this sense, "pure" context effects were obtained (Duncan
& Raudenbush, 2001).
When each context effect was separately plotted,
or when coefficients for curvilinear context terms were
examined in the data, no evidence emerged of the
Global success at grade 7
Context quality measures
Family quality (grade 7)
Friend quality (grade 7)
Neighborhood quality
School quality
Family X Friend
Family X Neighborhood
Family X School
Friend X Neighborhood
Friend X School
Neighborhood X School
Family X Friend X Neighborhood
Family X Friend X School
Family X Neighborhood X School
Friend X Neighborhood X School
Family X Friend X Neighborhood X School
Family structure
California Achievement Test score
Note: All coefficients are standardized.
*p < .05.
resulting was strongly related to changes in the success
index (after controlling for the usual reliabilityadjusted
background variables). The resulting standardized
coefficient was .13,which was about the same
size as the elementary school CAT coefficient (see Table
9) that was often used as an indicator of "human
capital" and thus as a potent promoter of adult success
(Hermstein & Murray; 1994;Jencks & Phillips, 1999).
Another way to describe the joint context effect entailed
dividing the total quality measure into 19
equally spaced intervals, which averaged about 500
students per interval. However, as Figure 1 shows,
the variation was from 30 students at one extreme of
the distribution to 1,174 at its central tendency. (Constructing
the intervals with equal sample sizes would
have restricted the range on the total context quality
scale and thus reduced the chances of observing nonlinearities).
More importantly, Figure 1 also shows
how total context quality was related to adjusted
eighth-grade success after controlling for initial success
and other individual differences. Success scores rose
linearly and systematically across the total context
quality scale, and even very small context quality differences
were generally associated with positive
change. A nonlinear component was apparent on close
visual inspection. However, it was not reliable even
when analyses were used that weighted the extreme
Cook et aI. 1299
context scores more than their otherwise small sample
sizes warranted. When the three highest and lowest
total context quality levels were separately averaged
to add stability to the extremes, the resulting
difference between them in unstandardized units was
about 1.5 outcomes on the nine-outcome success index.
Because the average temporal decrease was about
.9 units, being in consistently positive contexts negated
most of this expected developmental decrement
whereas being in consistently negative contexts almost
doubled it.
This same analysis was reproduced for each of the
nine separate outcomes, with its own reliabilityadjusted
seventh-grade value being used as a control
together with all the background variables. A reliable
linear trend from context quality to adjusted outcomes
was noted in all cases except for the test score
measure, which had the right sign but was only reliable
at the .10 level, partly because of a ceiling effect
in the state competency exam that was designed to
discriminate at the low cutoff score for passing the
exam rather than at the extremes of the scale. Thus,
the additive effect of multiple contexts seemed to be
general across the nine outcomes studied.
Probing some contingency variables. When relevant
first-order interactions were added to regression
models such as those just presented, the effects of
.75 -,-----------------------------
~ sr:
'"'d .65 m~ cI 0:5 ...c: co
~ .6
-lo0oi u:
1 52
2 66
3 114
4 215
5 339
6 491
7 731
8 823
9 1,002
10 1,174
11 1,114
12 1,027
13 827
14 627
15 412
16 241
17 112
18 53
19 30
o 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Total Context Quality
Figure 1 Adjusted eighth-grade success by total context quality, with 19 equally spaced context intervals.
1300 Child Development
Table 10 Changes in Overall Success as a Function of the
Context Summary and Demographics, With and Without the Context
x Race Interactions
Note: Family and friend measures were from early grade 7; success
was from late grade 8.
*p < .01.
total context quality did not reliably vary by family
composition, initial CAT score, gender, or family SESe
But joint context effects did seem to vary by race. Table
Global Success at Grade 8
10 presents the results of an analysis of the 19-point
total quality index with the usual predictors, including
the interactions of total context quality with the
Asian and White categories, Blacks being the reference
group. Context quality was associated with most
change for Blacks, with (nonreliably) less change for
Whites and with (reliably) least change for Asians.
These race differences could be seen more clearly
in LISREL analyses. The regressions that have been
presented are vulnerable to the criticism that a student's
individual characteristics may have caused the early
seventh-grade success standing and total context
quality. Thus, for each racial group a saturated LISREL
model was tested. It postulated that student individual
differences could cause both social contexts and
seventh-grade success. It also allowed seventh-grade
success and total context quality to be correlated, and
allowed for direct paths to eighth-grade success from
(1) the context sum, (2) seventh-grade standing, and
(3) each student background variable. All measures
that could be were adjusted for unreliability.
Figure 2 displays the results for the Black students.
Joint context quality was higher when students came
from intact homes and higher SES families. Seventh-
Variable Coefficient
Global success at grade 7
Context quality summary
Family structure
California Achievement
Test score
Context X Asian
Context X White
.12 .36
Figure 2 Path model of the determinants and consequences of total context quality for Blacks (N = 6,381). Values represent standardized
coefficients reliably differing from zero.
grade success was related to being female, having
higher CAT scores, and coming from a higher SES
family. Most important, there was a direct effect of .10
from the joint contexts to late eighth-grade success
adjusted for early seventh-grade success. The indirect
paths from SESand family structure totaled about .03.
Figure 3 displays the results for the White students.
Total context quality was related to CAT scores,
family SES,and family structure. Early seventh-grade
success depended on these same variables plus being
female. The direct path from context quality to adolescent
late eighth-grade success was reliable and
about .05, with the indirect paths totaling about .025
standard deviation units. These results suggest that
social contexts might count for less change in White
early adolescents than in Black adolescents, although
this race difference was never statistically reliable.
The total context coefficient for Asians reliably differed
from that for Blacks, even though the Asian
sample numbered only about 4% of the total. The
Asian coefficients in Figure 4 reveal the usual determinants
of both context quality and early seventhgrade
success, but there was no relation at all between
context quality and changes in the success index. In-
Cook et al. 1301
deed, the coefficient was trivially negative. At this age
period, Asian American students did not seem to be
responsive to the quality of the social environments
they were in. They marched to a different drummer.
LISREL analyses were also conducted for each of
the four social contexts separately, combining the
three race samples. In the school model, early success
and school quality were not related because students
had only just begun middle school when the first outcomes
were measured. Neighborhood quality was
also unrelated to seventh-grade success. However,
both friend and family quality were related to seventh-
grade success, implying that these two contexts
might be either products or causes of a young person's
standing. In any event, the direct effects of
schools and families were both about .06; of friends,
.04; and of neighborhoods, .02. This last coefficient
only reliably differed from zero when no account was
taken of clustering in the data. Thus, it seems that, as
with the joint context effect, individual context effects
were also generally replicated when a model was
used that involved temporal assumptions different
from those in the regression analyses.
However, the results might still have been contin-
Figure 3 Path model of the determinants and consequences of total context quality for Whites (N =2,665). Values represent standardized
coefficients reliably differing from zero.
1302 Child Development
Figure 4 Path model of the determinants and consequences of total context quality for Asians (N = 404). Values represent standardized
coefficients reliably differing from zero (except for the path from Context to Success Grade 8).
gent on using the global success index rather than individual
outcomes. Therefore, Table 11presents results
for each outcome separately. The model included early
seventh-grade measure of each outcome, plus all the
other early seventh-grade outcomes, all the context
main effects and interactions, plus the individual difference
variables. Where possible, all variables were
reliability adjusted. This model was designed to protect
against selection biases, but ran the risk of leaving
less variance for the contexts to predict than might
otherwise have been the case (e.g., if the other early
seventh-grade scores were omitted). Also, with so
many coefficients there was the problem of capitalizing
on chance.
Nonetheless, the results showed only as many interactions
among the four context quality indices as
would be expected by chance. Thus, use of the success
index did not seem to have obscured valid nonlinear
relations among the contexts when they were
scaled as was done in this instance. The context main
effects indicate that the family tended to have its
greatest influence on changes in the social behavioral
and mental health outcomes. The friend index most
consistently related to less acting out and drug use
and fewer mental health problems; the neighborhood
index was most often associated with school attendance
and participating in conventional activities;
and the school index was related to positive changes
in attendance, rates of entry into eighth-grade algebra,
and participating in conventional activities. Nonsignificant
trends were also observed in which test
scores increased at a steeper rate in better schools,
and grading became increasingly tougher.
Thus, the success index did indeed obscure differential
patterns of change associated with each context
having stronger or unique links to some outcomes.
However, our guiding assumption was that
individuals are multidimensional, with strengths in
some domains that can compensate for weaknesses
in others. To be a poor-performing student is not a
disaster if one has a cheerful disposition and friends
who are conventional in orientation. Thus, we hold
the index to be just as important as the outcomes
taken singly. The clear implication, however, is that
anyone wanting to effect change in specific outcomes
would do better to look to specific contexts
rather than to the total set of contexts examined in
the present study.
Cook et al. 1303
Table 11 Nine Individual Outcomes as a Function of All Context Main Effects and Interactions
Test Conventional Lack of Lack of Positive Mental
Algebra Grades scores Attendance Activities Misbehavior Drug Use Self Health
Variable (3 (3 (3 (3 (3 (3 (3 (3 (3
Pre-algebra .241* .045 .073* .006 -.012 .017 -.008 -.008 -.001
Grade point average
grade 7 .342* .575* .201* .085* .030 .108* .095* .011* .011
Attendance .019* .083* -.003 .603* .015 .000 .022 .015 .014
Conventional activities .017 .077* .019 .039* .364* .082* .076* -.008 .012
Lack of misbehavior -.023 .082* -.023 .071* .118* .532* .149* .012 -.033
Positive self -.047 -.042 -.049 .021 .058* -.045 -.067 .713* -.086*
Mental health .014 .004 .080* -.064* -.091* -.014 .028 -.032* .709*
Test score grade 7 .227* .089* .390* .035 .043 .014 .040 .062 .004
Lack of drug use .029 .001 .001 .015 .030* .034* .294* .022 .012
Family .042 .018 -.032* .013 .102* .061* .046 .014* .047
Friend .000 -.002 .007 .014 -.004 .026* .051* -.014 .020*
Neighborhood -.024 -.016 .019 .044* .038* .001 -.019 -.005 -.017
School .146* -.032 .029 .051* .039* .027 .001 -.019 .010
Family X Friend -.003 -.004 .005 .009 .017 .002 -.005 -.009 -.019
Family X Neighborhood -.007 .015 .012 .000 -.005 .008 -.007 .001 .007
Family X School -.019 .000 -.002 -.011 .000 -.023 .018 -.004 -.001
Friend X Neighborhood -.008 -.001 -.011 .002 .014 -.007 .003 .009 -.004
Friend X School -.018 -.006 .010 .007 .024 .000 -.002 -.006 -.007
Neighborhood X School -.006 .011 .000 .001 -.008 .006 .011 -.008 .006
Family X Friend X
Neighborhood -.010 .019 .002 -.013 -.020 .012 .014 .004 .011
Family X Friend X School -.017 -.005 -.003 -.011 -.004 -.004 .001 -.002 .008
Family X Neighborhood .000 .000 .008 -.019 -.006 -.014 -.002 .013 .021
Family X Friend -.014 .012 -.019 .011 .006 .015 .016 .005 -.019
Family X Friend X
Neighborhood .000 -.018 -.006 -.013 .010 -.009 -.008 .008 -.005
Asians .007 .050* .007 .003 .000 -.003 -.035 -.008 -.013
Whites .004 .041* .033* -.083* -.003 -.011 -.139* -.050* -.059*
Family structure -.003 .033* .013 .039* .013 .010 .013 .002 .006
Gender .030 -.073* .025 .066* .115* -.046* .028 -.010 -.006
SES .024 .015 .022* .033* .043* -.010 -.012 .022 -.009
California Achievement
Test score .097* .019 .250* .011 .007 -.050* -.076* -.023 -.017
Note: Family and friend quality measures were taken at grade 7.
*p < .01.
DISCUSSION Even so, some context links were stronger than
others. At the individual level, the strongest link was
Were the four social contexts tightly or loosely cou- between family and friend quality, although the estipled
when each was scaled for its capacity to promote mate was doubtless inflated by shared measurement
positive development? The answer to this simple variance. Nonetheless, early adolescents from homes
question is complex. For individual adolescents in with authoritative parenting and extensive parentthis
county, the contexts were generally only weakly child communication seemed to have friends who were
coupled. Some early adolescents could and did come more stable over time, less deviant, and more convenfrom
disorganized families and yet attended a good tional. Better estimated were the school/neighborschool.
Others could and did have deviant friends, hood quality links. These are important because theyet
lived in an organized neighborhood. Students ories of residential mobility stress that when families
could and did come from a well-organized family, with children choose a new neighborhood, they do so
had friends who were highly conventional, and yet based in part on their perception of the quality of 10-
lived in a troubled neighborhood. All combinations cal schools. The results obtained here are consistent
were obtained. Social contexts were internally hetero- with this theory. However, the link between schools
geneous with respect to the developmental quality of and neighborhoods was probably lower than most
the other contexts nested within them. commentators would probably have expected. Do
1304 Child Development
parents assess school quality using easily observed
attributes such as student race and social class composition
or published test scores? Directly observing
what actually takes place in the school building is
more difficult for parents, yet this is essentially what
our school quality index tapped into and what proximally
promoted student gains.
At the individual level, most of the correlations between
contexts were very low; yet at higher levels of
aggregation, such as the collectivity of young people
in a school or neighborhood, social contexts were
much more tightly coupled. Disciplines such as sociology
and education tend to focus more on groups
than individuals, with theorists seeking to explain relations
among abstractions such as neighborhoods,
schools, or classrooms. The current results indicate
that it is quite legitimate in these disciplines to conclude
that context coupling is tight and that, for the
collectivity of young students or neighborhood residents,
the developmental quality of anyone context
constrains the quality of others. Having said this,
however, scholars in these disciplines have to avoid
the ecological fallacy trap-assuming that knowledge
about aggregates generalizes to individuals.
In this connection, consider Wilson's (1987) theory,
which predicts that inner-city minority neighborhoods
entail troubled schools, disorganized parenting,
and dangerous peer groups. Although the county
in the present study had no "urban underclass" neighborhoods,
the findings suggest that a theory like Wilson's
offers a valid description of how contextual relations
cohere when neighborhoods are the unit of
analysis. However, the description may fit very few
individual lives in these neighborhoods. Some early
adolescents who live in a disorganized setting may
attend better schools, some (and they need not be the
same ones) may have effective parents, and some
(again not necessarily the same ones) may hang out
with quite conventional friends. Conversely, even in
affluent neighborhoods, there are some youngsters
who attend less effective schools, grow up in disorganized
families, and hang out with dangerous friends.
In psychology and economics, the individual is the
theoretical centerpiece, and in these disciplines it is
appropriate to conclude that context coupling is generally
loose. However, scholars from these disciplines
should not overlook the reality that context coupling
is more systematic once individual differences are accounted
for. Then, the social world is ordered in ways
that generally favor young persons from a better school
or neighborhood and that generally compromise the
welfare of young persons from a worse neighborhood
or school. In the present study, the large variation discovered
within schools and neighborhoods did not
negate the fact that the worse schools tended to have
youngsters from the worst neighborhoods, homes,
and friendship groups; nor did it negate the fact that
the worse neighborhoods tended to be inhabited by
young people from the worst schools and homes
whose friends were also the most socially imperiled.
Such relationships, however, only held on the average.
They have limited relevance for understanding
how context qualities cohered in individual lives.
The fact that more of the total variation lay within
schools and neighborhoods than between them could
be interpreted to imply that the study of individuals
deserves higher priority than the study of contexts. To
understand why this conclusion is not warranted,
consider applying the same criteria and standards to
context differences that ICC analyses apply to individual
differences. With ICCs, the role of a single setting
is contrasted to the role of all sources of individual
difference. To level the playing field entails cumulating
the between-context variation across all possible
settings, and thus dramatically increasing the total
between-context component. Conversely, imagine contrasting
the variation associated with a single context
(e.g.,schools) with the variation associated with a single
individual difference (e.g., gender). Only then could
we begin to decide whether research based on individuals
is more important than research based on
contexts. Even so, the framing seems naive. The need
is to examine how specific context and individual difference
factors jointly affect development.
Our second research question was, How do families,
friendship groups, neighborhoods, and schools
jointly influence change in early adolescence? The answer
to this question can be separated into three
parts: the assumptions made, the findings about individual
contexts, and the findings about joint contexts.
Assumptions. All the findings presented depended
on a single county that, despite its considerable internal
heterogeneity, is inevitably more curtailed on all
variables than the nation at large. The answers also
depend on the particular success and context measures
used, and hence on the substantive theories that
generated them. Moreover, family, friends, schools,
and neighborhoods were treated as simultaneous
causes of changes in the success index. However, it is
possible to argue for many different theories of causal
precedence. Do neighborhoods exercise their influence
through their effects on schools, family life, peer groups,
or all three? Do peer groups exercise their influence
through the different neighborhoods and even families
into which one's peers take one? Do families exercise
their influence through the neighborhoods or schools
they select or through the peer groups toward which
they steer their children? Are all four contexts interrelated
reciprocally, being affected by each other as well
as by how well a young person is doing, thus rendering
notions of unidirectional cause descriptively incomplete?
These and many other assumptions about
causal precedence deserve to be made and tested with
data covering longer time periods than those of the
present research. Our assumption about the simultaneity
of multiple causal relations is surely an oversimplification
of reality, and is probably more relevant to
shorter than longer time periods.
Given the impossibility of simultaneous random
assignment to famil)', friend, school, and neighborhood
quality, the findings we presented also inevitably depend
on the assumptions built into the substantive
models tested. These models examined late eighthgrade
student outcomes as a function of (1) the developmental
quality of four social contexts; (2) reliabilityadjusted
early seventh-grade scores on whatever
outcome was under study; and (3) many reliabilityadjusted
individual differences, sometimes modeled
as causally prior even to early seventh-grade standing.
Whereas the control variables adjust for much selection,
there is no way to know if the adjustment was
perfect. A shortfall is particularly likely for the family
and friend contexts, where some residual influence
from prior outcomes to measured contexts might still
remain unaccounted for. We strove to minimize bias,
realizing that no suitable instrumental variable was
available and that propensity scores require a dichotomous
independent variable and so entail losing the
many advantages associated with our continuous measures
of context quality.
Single-context effects. Constrained by these assumptions,
some important things were discovered about
individual contexts. First, each caused positive change
during the early adolescent years. This bolstered the
belief that the various composites included some socialprocess
attributes that promote healthier development.
Second, the effects attributed to anyone context
were not due to its correlation with other contexts.
This suggests that these were truly independent context
effects within the limits set by our particular theories
and measures of what constituted a developmentally
superior nuclear family, friendship group,
neighborhood, and school.
Third, no social context was curvilinearly related
to individual change, although the large sample size
and the absence of a ceiling effect made an unusually
powerful test of nonlinearity possible. It was not possible
to identify any extreme position on any context
scale that was associated with especially large shifts
in success and thus deserved special status. The better
the context quality, the better a student did whatever
the level of context quality under consideration. Of
Cook et al. 1305
course, we cannot be sure of linearity in other settings,
especially those in which both urban underclass
areas and enclaves of the super-rich are present.
Nonetheless, the county studied in the present research
provided an unusually heterogeneous suburban
setting and failed to reveal any meaningful
discontinuities in how anyone context related to
changes in early adolescent success.
Fourth, none of the single context effects were of a
magnitude that can be called large. To improve effect
sizes, various configurational approaches to combining
a context's attributes were tried; however, none
did noticeably better than a linear model that simply
summed context features. This, plus the high variation
demonstrated within schools and neighborhoods,
led us to conclude that anyone context probably had
only modest two-year effects at this age. There was no
single context that could be considered as a silver bullet,
just as there was no extreme value on any context
that could be considered in the same way. Of course,
if similar rates of change had regularly taken place
since infancy, or if these early adolescent rates were to
continue into adulthood, then the resulting multiyear
single context effects would be considerable, as would
some of the developmental differences among the
four contexts. Clearly what is needed is to learn about
longer time periods and earlier beginnings than those
examined in the present study.
Fifth, neighborhood coefficients were regularly
smaller than the other context coefficients and were
not even systematically reliable in models that included
the other contexts. To argue from this that
neighborhood effects were truly smaller presumes
that each context was measured equally validly. However,
neighborhoods could have been measured least
well. For instance, many Census tracts had geographic
boundaries by 1990 that were far from isomorphic
with their original social boundaries, even though the
social boundaries are plausibly more causally proximal.
In addition, in some tracts the number of parents
reporting on neighborhood processes was quite modest.
Thus, we prefer the general conclusion that each
context independently but modestly improved a
young person's life chances.
Finally, we discovered that each setting influenced
a different set of outcomes. Families were more potent
in the mental health domain; peers influenced
negative social behavior; schools impacted academic
performance; and neighborhoods influenced school
attendance and participation in conversational social
activities. Many scholars and practitioners are interested
in specific outcome domains, and these findings
suggest that they are probably correct to concentrate
on particular social contexts if the interest is in spe1306
Child Development
cific outcome domains. However, for any scholar who
adheres to a "whole child" approach to development,
the specificity of the links between domains and outcomes
challenges the utility of concentrating on a
single context.
Joint-context effects. We found that combining all
four contexts resulted in a large effect size, even over
a period as short as 19 months. As early adolescents
moved through the research county's public middle
schools, they came to fail in almost one more domain
on the average. However, teenagers living in four
consistently better contexts did not really experience
this normative decline, whereas adolescents living in
four consistently worse contexts did worse over time
by almost two outcomes. (Remember that these were
failures by very modest "getting by" performance
standards rather than "doing well.") Another perspective
on the joint effect size emphasized that, in
the regression analyses, the coefficient for the four
contexts combined was very similar to the direct coefficient
for the CAT scores, a measure that many economists,
sociologists, and behavioral geneticists treat
as an index of "human capital." Thus, the conclusion
seems warranted that, at this period in history and in
this age range, these four contexts count in young
people's lives about as much as human capital. Contexts
really matter in determining how young people's
lives develop, but they do so cumulatively more than
No evidence of reliable statistical interactions was
found among the four context qualities. If the null hypothesis
is accepted, it suggests that cumulative context
effects were probably additive rather than multiplicative.
However, tests of statistical interactions are
never easy. They are stronger if a particular form of
nonlinearity is specified in advance and if the appropriate
contrast is then tested. Unfortunately, substantive
theory on contextual interdependencies is not yet
developed enough to permit this. Therefore, the tests
conducted were of overall context-related differences
in slope. Because strong interaction tests also require
oversampling from the tails of distributions, subanalyses
were conducted that omitted the two middle
quartiles for each context. These also failed to produce
reliable interactions. Accepting the null hypothesis
also depends on scaling because interactions can
result from unequal intervals at different scale points
and any measure can be arithmetically rescaled to
make an interaction appear or disappear. Thus, accepting
the null hypothesis requires justifying the
scales used. The success index was basically a count
of the number of domains in which a student was getting
by. To transform this would have no substantive
meaning even if it did arithmetically result in an interaction.
Also, because it is a count, it is presumptively
an interval scale. However, no convincing substantive
justification can be offered for how the
context attributes were scaled; and the only evidence
about equal intervals for the context measures was indirect,
based on each context and the total context index
having normally distributed scores. Uncertainties with
regard to the scaling of the context measures meant
that our interaction tests were bound to be imperfect,
however good the scaling of the success outcome.
Nonetheless, the absence of reliable interaction results
was judged to be adequate for supporting the
following complex conclusions: (1) if there were any
undetected interactions among contexts in these data,
they were probably small in size; (2) if there were any
larger undetected interaction effects, they probably
involved context and outcome features more micro
than the multiattribute indices we used; and (3)if social
contexts more extreme than in this county had been
sampled, then larger interactions might have emerged,
even though such extremes are rare, as with dysfunctional
urban ghettos and enclaves of mansions.
This article has gone into unusual detail about testing
statistical interactions because those involving the
four contexts have important substantive implications.
One conclusion suggested by their absence
(and by the absence of nonlinearity within individual
contexts) is that there are no early adolescent "silver
bullets" that can radically transform young lives for
the better. Many theorists and program developers
have made strong claims about the especially positive
consequences of extreme values of a given context, or
about magical combinations of two or more contexts,
or about one social context being more important
than others at a particular age. However, this study
found little convincing evidence for any of these
claims with the measures used. Developmentally sensitive
contexts matter, and more of them matter more;
but in early adolescence they matter additively, not
nonlinearly or even perhaps differentially.
A second implication of the additive relation among
contexts is more theoretical. Social contexts are regularly
included in lists of risk and protective factors, albeit
in ways that confound different contexts, or that
weight some contexts more than others without explicit
justification, or that confound a context's internal
processes with its social demography. In the present
study, all context measures were kept process related,
specific to a single context, and independent of structural
attributes. When purified this way, the evidence
showed that it was legitimate to consider social contexts
as risk and protective factors. Individual social
contexts could and did individually affect how young
people develop. When a context was positive, it protected
early adolescents; and when it was negative, it
added to their risk. Moreover, positive benefits accrued
robustly as the number of developmentally
protective contexts increased, and this happened independently
of which specific contexts were being
combined. Similarly, dangers accrued robustly as the
number of risky contexts increased, irrespective of
which types of context were involved. Individual
context effects were largely substitutable and did not
depend on the developmental value of other contexts.
However, it must be noted that they were only substitutable
when using the global success index. When
individual outcomes were analyzed, different contexts
seemed to affect different outcomes. For the algebra
promotions that put a student on track for a
four-year college, neighborhoods were definitely not
substitutable for schools!
The protection that developmentally superior
contexts afforded was largely unconditional. That is,
the same relation between joint social contexts and
changes in the success index generally held across
variation in gender, SES,family composition, elementary
school CAT score, and even across Black and
White students, although Blacks may have been more
responsive to contextual quality than were Whites.
The main exception at this stage in their development
was for Asian students who were not affected
by the four contexts as measured in this study. Reexamining
the measures makes it easy to see how the
family composite may have inadvertently failed to
capture conceptions of parenting based on honor
and/ or shame that are relevant to many of the groups
within this county's ethnically diverse Asian population.
However, it is less clear how the neighborhood,
school, or even friend composites were so culturally
limited. In any event, factors other than the social
contexts assessed in this study will be needed to explain
individual variation in success among Asian
students. This one exception aside, in the county
studied in the present research, the relation between
joint context quality and changes in success seemed
to be broadly additive.
There appear to be no quick fixes during early adolescence.
Improving the developmental quality of
anyone context will help; but it will not dramatically
alter the multidimensional welfare of many young
people. Tomake anyone context superlative will help
more, but it will not lead to an unusually steep rise at
some point on the function relating context quality
to outcome changes; nor will combining any two contexts
lead to "emergent properties" that promote
development more than the sum of their individual
effects. However, positive change is significantly enhanced
when more of the contexts are developmen-
Cook et al. 1307
tally promotive in ways that the current literature
speaks to-the more of them the better. For anyone
interested in attaining the clearly positive developmental
consequences of pan-contextual improvement,
it will be a gigantic task to simultaneously improve
multiple contexts in the lives of many young
people. However, one thing seems certain from the
present research: It will not be easy to identify "silver
bullets"-details from just one or two social contexts
that, when changed, will significantly improve early
adolescents' general welfare.
Work on this article was facilitated by research support
from the MacArthur Foundation's Network on
Successful Adolescence in High-Risk Settings and by
fellowships to the first author from the Center for Advanced
Studies in the Behavioral Sciences, Stanford,
CA and from the Max Planck Institute for Human Development,
Berlin, Germany.
Corresponding author: Thomas D. Cook, Department
of Sociology and Institute for Policy Research, Northwestern
University, 2040 Sheridan Road, Evanston,
IL 60208-4100; e-mail:
Melissa R. Herman is also at Northwestern University;
Meredith Phillips is at the University of California,
Los Angeles; and Richard A. Settersten, [r., is at Case
Western Reserve University, Cleveland, OH.
Angrist, J. D., Imbrens, G. W., & Rubin, D. B. (1996). Identification
of causal effects using instrumental variables.
Journal of the American Statistical Association, 91, 444462.
Barber, B. (Ed.). (2002). Intrusive parenting: How psychological
control affects children and adolescents. Washington, DC:
American Psychological Association.
Barber, B., & Olsen, J. A. (1997). Socialization in context:
Connection, regulation, and autonomy in the family,
school, and neighborhood, and with peers. Journal ofAdolescent
Research, 12, 287-315.
Baumrind, D. (1971). Current patterns of parental authority.
Developmental Psychology, 4, 1-103.
Baumrind, D. (1975). The contributions of the family to the
development of competence in children. Schizophrenia
Bulletin, 14, 12-37.
Bourdieu, P. (1977). Cultural reproduction and social reproduction.
In J. Karabel & A. H. Halsey (Eds.), Power and
ideology in education (pp. 487-510). New York: Oxford
University Press.
1308 Child Development
Bradley, R. H., & Caldwell, B. M. (1995). The acts and conditions
of the care giving environment. Developmental Review,
15, 92-96.
Brett, J. M., Stroh, L. K., & Reilly, A. H. (1993). Pulling up
roots in the 1990s: Who's willing to relocate? Journal of
Organizational Behavior, 14, 49-60.
Bronfenbrenner, U. (1986). Ecology of the family as a context
for human development: Research perspectives. Developmental
Psychology, 22, 723-742.
Bronfenbrenner, U. (1988). Interacting systems in human
development: Research paradigms, present and future.
In N. Bolger, A. Caspi, G. Downey, & M. Moorehouse
(Eds.), Persons in context: Developmental processes (pp. 2549).
New York: Cambridge University Press.
Bronfenbrenner, U., & Ceci, S. J. (1994). Nature/nurture reconceptualized
in developmental perspective: A bioecological
model. Psychological Review, 101, 568-586.
Brooks-Gunn, J., Duncan, G. J., Klebanov, P. K., & Sealand,
N. (1993). Do neighborhoods influence child and adolescent
behavior? American Journal of Sociology, 99, 353395.
Brown, B. B. (1990). Peer groups and peer cultures. In S. S.
Feldman & G. R. Elliott (Eds.), At the threshold: The developingadolescent
(pp. 171-198). Cambridge, MA: Harvard
University Press.
Brown, B. B. (1998). Learning contexts beyond the classroom:
Extracurricular activities, community organizations,
and peer groups. In K. Borman & B. Schneider
(Eds.), The adolescent years: Social influences and educationalchallenges
(pp. 109-141). Chicago: The National Society
for the Study of Education.
Brown, B. B. (1999). Measuring the peer environment of
American adolescents. In S. L. Friedman & T. D. Wachs
(Eds.), Measuring environment across the life span: Emerging
methods concepts (pp. 59-90). Washington, DC: American
Psychological Association.
Brown, B. B., & Huang, B. H. (1995). Examining parenting
practices in different peer contexts: Implications for adolescent
trajectories. In L. J. Crockett & A. C. Crouter
(Eds.), Pathways through adolescence: Individual development
in relation to social contexts (pp. 119-138). Mahwah,
NJ: Erlbaum.
Bryk, A. S., Lee, V.E., & Holland, P.B. (1993). Catholic schools
and the common good. Cambridge, MA: Harvard University
Campbell, D. T.,& Kenny, D. A. (1999).A primeron regression
artifacts. New York: Guilford Press.
Clark, R. (1983). Family life and school achievement: Why poor
Black children succeed orfail. Chicago: University of Chicago
Comer, J. P. (1995). School power. New York: Free Press.
Cook, T. D., & Murphy, R. F. (1999). How inner-city children
seetheirfamily, school, peers, and neighborhood: Developmental
change during the transition to adulthood. (Working Paper).
Evanston, IL: Institute for Policy Research, Northwestern
Cook, T. D., & Payne, M. R. (2002). Objecting to the objections
to using random assignment in educational research.
In F.Mosteller & R. F.Boruch (Eds.), Evidence matters:
Randomized trials in education research. Washington,
DC: Brookings Institute Press.
Cook, T. D., Shagle, S. C., & Degirmencioglu, S. M. (1997).
Capturing social process for testing mediational models
of neighborhood effects. In J. Brooks-Gunn, G. J. Duncan,
& J. L. Aber (Eds.), Neighborhood poverty: Context and
consequences for children (Vol. 2). New York: Russell Sage
Foundation Press.
Cronbach, L. J. (1982). Designing evaluations of educational
and social programs. San Francisco: Jossey-Bass Publishers.
Cronbach, L. J., & Furby, L. (1970).How we should measure
"change"-or should we? Psychological Bulletin, 74, 6880.
Duncan, G. J., & Raudenbush, S. W. (2001). Neighborhoods
and adolescent development: How can we determine
the links? In A. Booth & A. C. Crouter (Eds.), Does it take
a village?: Community effects on children, adolescents, and
families (pp. 105-136). State College: Pennsylvania State
University Press.
Eccles, J. S., Midgley, C., Wigfield, A., Buchanan, A., & MacIver,
D. (1993). Development during adolescence: The
impact of stage environment fit on young adolescents'
experiences in schools and in families. AmericanPsychologist,
48, 90-101.
Elliott, D. S. (1989). Multiple problem youth. New York:
Fordham, S., & Ogbu, J. U. (1986). Black students' school
success: Coping with the "burden of I acting white'." UrbanReview,
18, 176-206.
Furstenberg, F. F., Cook, T. D., Eccles, J., Elder, G. H., &
Sameroff, A. (1999). Managing to makeit: Urban families in
high risk neighborhoods. Chicago: University of Chicago
Garmezy, N. (1987). Stress, competence, and development:
Continuities in the study of schizophrenic adults, children
vulnerable to psychopathology, and the search for
stress-resistant children. AmericanJournal of Orthopsychiatry,
57, 159-174.
Handler, J. F., Berlin, G. L., Cook, T. D., Crim, A. A., Dornbusch,
S. M., Dryfoos, J. G., Fernandez, R. M., Freeman,
R. M., Hagan, J., Irwin, C. E., Jr.,lessor, R.,Johnson-Powell,
G., Shirley, A., Starfield, B.,& Street, L. (1995). Losinggenerations:
Adolescents in high-risksettings. Washington, DC:
National Academy of Science Press.
Hermstein, R.J.,& Murray, C. (1994). Thebell curve: Intelligence
andclass structure in American life. New York:Free Press.
Hormuth, S. E. (1990). The ecology of the self Relocation and
self-concept change. Cambridge, U.K.: Cambridge University
Jargowsky, P.A. (1997). Povertyand place. New York: Russell
Sage Foundation.
Jencks, C., & Peterson, P. (Eds.). (1991). Theurbanunderclass.
Washington, DC: Brookings Institution.
Jencks, C., & Phillips, M. (Eds.). (1998). The Black-White
test score gap. Washington, DC: Brookings Institution
Kelley, R. G. (1997). Yo' mama's disFUNKtional!: Fighting the
culture wars in urbanAmerica. Boston: Beacon Press.
Kinney, D. A. (1993). From nerds to normals: The recovery
of identity among adolescents from middle school to
high school. Sociology of Education, 66, 21-40.
Little, R. J., & Rubin, D. B. (1987). Statistical analysis with
missing data. New York: Wiley.
MacLeod, J. (1995). Ain't no making it. Boulder, CO.: Westview
Magnusson, D., Stattin, H., & Allen, V. (1986). Differential
maturation among girls and its relation to social adjustment
in a longitudinal perspective. In P. B. Baltes, D.
Featherman, & R. Lerner (Eds.), Lifespandevelopment and
behavior (Vol. 7, pp. 143-158). Hillsdale, NJ: Erlbaum.
Massey, D. S., & Denton, N. A. (1993). American apartheid:
Segregation and the making of the underclass. Cambridge,
MA: Harvard University Press.
Maugham, B. (1994). School influences. In M. Rutter & D. F.
Hay (Eds.), Development through life: A handbook for clinicians.
Boston: Blackwell Scientific.
McClelland, G. H., & Judd, C. M. (1993). Statistical difficulties
of detecting interactions and moderator effects. Psychological
Bulletin, 114, 376-390.
McLaughlin, M. W., Irby, M. A., & Langman, J. (1994). Urban
sanctuaries: Neighborhood organizations in the lives and futures
of inner-city youth. San Francisco: Jossey-Bass.
Meyer-Probst, B., RosIer, H.-D., & Teichmann, H. (1983). Biological
and psychological risk factors and development
during childhood. In D. Magnusson & V. L. Allen (Eds.),
Human development: An interactional perspective (pp. 243259).
New York: Academic Press.
Nadherny, K., Marcantonio, R. J., & Cook, T. D. (1993). Identifying
biases in the cohort two sample interviewed at home.
(Working Paper). Evanston, IL: Institute for Poverty Research,
Northwestern University.
Nunnally, J. C., & Bernstein, 1.H. (1994). Psychometric theory
(3rd ed.). New York: McGraw-Hill.
Plomin, R. (1994). Genetics and experience: The interplay between
natureand nurture. Thousand Oaks, CA: Sage.
Rutter, M., Maugham, B., Mortimore, P.,& Ouston, J. (1979).
Fifteen thousand hours: Secondary schools and theireffects on
children. Cambridge, MA: Harvard University Press.
Sameroff, A. J. (1983). Developmental systems: Contexts
and evolution. In W. Kessen (Ed.), Handbook of Child
Psychology (Vol. I, 4th ed.). New York: Wiley.
Sameroff, A. J., Seifer, R., Baldwin, A., & Baldwin, C. (1993).
Stability of intelligence from preschool to adolescence:
The influence of social and family risk factors. Child Development,
64, 80-97.
Cook et al. 1309
Sampson, R. J., Raudenbush, S. W., & Earls, F. (1997). Neighborhoods
and violent crime: A multilevel study of collective
efficacy. Science, 277, 918-924.
Savin-Williams, R., & Berndt, T. J. (1990). Friendship and
peer relations. In S. S. Feldman & G. R. Elliott (Eds.), At
the threshold: The developing adolescent (pp, 277-307).
Cambridge, MA: Harvard University Press.
Shadish, W. J., Cook, T. D., & Campbell, D. T. (2002). Experimentalandquasi-
experimental designs for generalized causal
inference. Boston: Houghton Mifflin.
Silbereisen, R. K. (1995). How parenting styles and crowd
contexts interact in actualizing potentials for development:
Commentary. In L. J. Crockett & A. C. Crouter
(Eds.), Pathways through adolescence: Individual development
in relation to social contexts (pp. 197-207). Mahwah,
NJ: Erlbaum.
Simmons, R. G., & Blyth, D. A. (1987). Moving into adolescence:
Theimpactof pubertal change andschool context.New
York: Aldine de Gruyter.
Slavin, R. E., Madden, N. A., Dolan, L. J., Wasik, B. A., Ross,
S., Smith, L., & Dianda, M. (1996). Success for all: A summary
of research. Journal of Education for Students Placed
at Risk, 1, 41-76.
Steinberg, L., Lamborn, S. D., Dornbusch" S. M., & Darling,
N. (1992). Impact of parenting practices on adolescent
achievement: Authoritative parenting, school involvement
and encouragement to succeed. Child Development,
63, 1266-1281.
Sullivan, M. L. (1989). Getting paid. Ithaca, NY: Cornell University
Talbert, J. E., & McLaughlin, M. W. (1999). Assessing the
school environment: Embedded contexts and bottom-up
research strategies. In S. L. Friedman & T. D. Wachs
(Eds.), Measuring environment across the life span: Emerging
methods and concepts (pp. 197-227). Baltimore, MD:
United Book Press.
Weick, K. (1976). Educational organizations as loosely
coupled systems. Administrative Science Quarterly, 21"
Willett, J. B. (1989). Some results on reliability for the longitudinal
measurement of change: Implications for the design
of studies of individual growth. Educational andPsychological
Measurement, 49"587-602.
Wilson, W. J. (1987). The truly disadvantaged: The inner city,
the underclass and public policy. Chicago: University of
Chicago Press.

  • Recommend Us