Exploring Traumatic Influences
farm1980
O R I G I N A L C O N T R I B U T I O N
Adverse life events, area socioeconomic disadvantage, and psychopathology and resilience in young children: the importance of risk factors’ accumulation and protective factors’ specificity
Eirini Flouri • Nikos Tzavidis • Constantinos Kallis
Received: 6 March 2009 / Accepted: 21 September 2009 / Published online: 10 October 2009
� Springer-Verlag 2009
Abstract Few studies on resilience in young children
model risk appropriately and test theory-led hypotheses
about its moderation. This study addressed both issues. Our
hypothesis was that for preschool children’s emotional/
behavioral adjustment in the face of contextual risk pro-
tective factors should be located in the cognitive domain.
Data were from the first two sweeps of the UK’s Millen-
nium Cohort Study. The final study sample was 4,748
three-year-old children clustered in 1,549 Lower layer
Super Output Areas in nine strata. Contextual risk was
measured at both area (with the Index of Multiple Depri-
vation) and family (with proximal and distal adverse life
events experienced) level. Moderator variables were par-
enting, verbal and non-verbal ability, developmental
milestones, and temperament. Multivariate multilevel
models—that allowed for correlated residuals at both
individual and area level—and univariate multilevel mod-
els estimated risk effects on specific and broad psychopa-
thology. At baseline, proximal family risk, distal family
risk and area risk were all associated with broad psy-
chopathology, although the most parsimonious was the
proximal family risk model. The area risk/broad psycho-
pathology association remained significant even after
family risk was controlled but not after family level
socioeconomic disadvantage was controlled. The cumula-
tive family risk was more parsimonious than the specific
family risks model. Non-verbal ability moderated the effect
of proximal family risk on conduct and emotional prob-
lems, and developmental milestones moderated the effect
of proximal family risk on conduct problems. The findings
highlight the importance of modeling contextual risk
appropriately and of locating in the cognitive domain
factors that buffer its effect on young children’s
adjustment.
Keywords Hierarchical data � MCS � Multilevel models � Multivariate multilevel models � Psychopathology � Resilience � Strengths and Difficulties Questionnaire
Introduction
In the classic approach to cumulative contextual risk in
child psychiatry [1], organismic characteristics as well as
proximal and distal qualities of the environment are mod-
eled collectively. For each environmental construct, a
dichotomous classification of risk exposure is determined,
typically by a statistical cut off (e.g., greater than one
standard deviation above the mean, upper quartile, etc.) or
on the basis of a conceptual categorization (e.g., being
below the poverty line). Cumulative risk is then calculated
by a simple summation of the multiple risk categories.
These risk categories are usually not weighted. In this
approach, therefore, risk is viewed as an accumulation of
stressors, and the number of risks that children experience
carries more gravity than the experience of any particular
E. Flouri (&) Department of Psychology and Human Development,
Institute of Education, University of London,
25 Woburn Square, London WC1H 0AA, UK
e-mail: [email protected]
N. Tzavidis
Social Statistics and Centre for Census and Survey Research,
University of Manchester, Manchester, UK
e-mail: [email protected]
C. Kallis
London School of Hygiene and Tropical Medicine,
University of London, London, UK
e-mail: [email protected]
123
Eur Child Adolesc Psychiatry (2010) 19:535–546
DOI 10.1007/s00787-009-0068-x
risk. Indeed, cumulative risk indexes have been noted for
their potential to capture the natural covariation of risk
factors. For example, physical risk factors, such as poor
housing quality, noise and pollution are strongly interre-
lated as are psychosocial risk parameters such as family
turmoil and violence [2]. Furthermore, aggregate variables
of risk are more stable than any individual measure, and
there is increased power to detect effects because errors of
measurement decrease as scores are summed and degrees
of freedom are preserved [3]. Cumulative risk measures are
consistently found to explain more variance in children’s
outcomes than single factors [4–7].
Despite this, few studies [8] have examined the rela-
tionship between multiple risk exposure and young chil-
dren’s psychopathology using a cumulative risk approach.
In fact, even studies with school-aged children have yet to
deal with several issues with respect to the modeling of
cumulative risk. First, studies have yet to establish con-
vincingly the effect of the timing of cumulative contextual
risk on psychopathology [4, 9, 10]. Second, with few
exceptions [4, 11–14], studies do not examine the func-
tional form of cumulative contextual risk’s effect on psy-
chopathology. There is evidence for a linear effect whereby
increments in risk factors have a steady, additive impact on
mental health problems in school-aged children [5].
However, as few researchers actually report whether their
investigations included appropriate tests for nonlinear
patterns of cumulative risk, this ignores the possibility of a
nonlinear relationship that might manifest itself as an
acceleration [14] or a leveling-off [13] of problems at a
critical level of risk. Third, despite the renewed interest in
neighborhood effects on children’s development [15], only
few studies [16] compare family and neighborhood risk.
Fourth, although it is possible that the effect of one type or
one level of risk on psychopathology is conditional upon
the value of another, interactions between types and levels
of risk on child psychopathology are not usually studied.
Fifth and finally, with few exceptions [4, 8, 17, 18], most
studies do not examine factors that protect from negative
outcomes in children exposed to cumulative contextual
risk. The dearth of such research is unfortunate as various
protective factors have been identified as moderating the
impact of specific contextual risk. These factors are
grouped under two domains, namely, individual attributes
and connections to competent and caring adults in the
family and the community [19, 20].
Moderation of risk
To the best of our knowledge, the only study modeling the
moderation of cumulative contextual risk in preschool
children’s emotional and behavioral adjustment [8] found
that low vagal tone during tasks protected from the effect
of multiple risks, and concluded that this is because low
vagal tone during tasks may reflect regulatory capacities
that allow children to engage with learning opportunities.
This suggests that children’s advanced cognitive develop-
ment may be the factor that moderates the effect of con-
textual risk on emotional and behavioral outcomes.
Although previous research has certainly located in the
cognitive domain protective factors for school-aged chil-
dren’s emotional and behavioral adjustment in the face of
cumulative contextual risk [4, 18], studies have yet to
establish the importance of cognitive protective factors for
preschool children’s adjustment in the face of cumulative
contextual risk or test the hypothesis that for preschool
children’s adjustment in the face of cumulative contextual
risk factors promoting resilience should be located in the
cognitive rather than the biological or the social domain.
This study was undertaken to meet this aim, i.e., to test
that for preschool children’s adjustment in the face of
cumulative contextual risk factors promoting resilience
should be located in the cognitive domain. In doing so, it
also extended in several ways prior work on the role of
contextual risk in child psychopathology. First, it used
items from a well-validated measure of multiple family
risks. However, since only a selection of the original items
could be used, it also tested appropriately for the effect or
risk accumulation and risk specificity when modeling the
effect of family risk on child psychopathology. Second, it
measured both distal and proximal family risk to test for
the effect of the timing of risk on psychopathology. Third,
it measured area risk and compared area with family risk
effects. Fourth, it searched for an appropriate functional
form of both area and family risk’s effect on psychopa-
thology. Fifth and finally, it tested for the presence of
interaction effects between proximal, distal and area risk
on psychopathology.
Method
Participants and procedure
Data were obtained from the first two sweeps of the
Millennium Cohort Study (MCS), a longitudinal survey
drawing its sample from all live births in the UK over a
period of 12 months, beginning on 1 September 2000 in
England and Wales, and 3 months later in Scotland and
Northern Ireland. The sample was drawn slightly later in
Scotland and Northern Ireland so as not to coincide with
other surveys being carried out on families with babies in
these areas at the same time. Sweep 1 took place when the
children were aged 9 months, and Sweep 2 took place when
the children were around 3 years. The sample design
536 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
allowed for disproportionate representation of families liv-
ing in areas of child poverty in Northern Ireland, Scotland
and Wales and in areas with high ethnic minority
populations in England. In all there were nine strata: Eng-
land-advantaged, England-disadvantaged, England-ethnic,
Wales-advantaged, Wales-disadvantaged, Scotland-advan-
taged, Scotland-disadvantaged, Northern Ireland-advan-
taged, and Northern Ireland-disadvantaged. Data on child
psychopathology were collected at Sweep 2 from the main
respondent. Although the MCS is a study of 19,244 children,
there were complete data on broad psychopathology for a
total of 9,736 children clustered in Lower layer Super Output
Areas (LSOAs). LSOAs are built from groups of Output
Areas (typically 4–6), and are constrained by the boundaries
of the Standard Table wards used for 2001 Census outputs.
They have, on average, 1,500 residents. For example, there
are 175,434 Output Areas and 34,378 LSOAs in England
and Wales. In this study, LSOAs, rather than the electoral
wards on which the MCS survey design was built, were used
to get smaller ‘neighborhoods’ and more up-to-date eco-
logical correlates. For one child, there was no information on
the LSOA at Sweep 2 and, therefore, this case was removed
from our analysis.
Measures
Area contextual risk was measured with the Index of
Multiple Deprivation (IMD), a weighted area level aggre-
gation of specific dimensions of deprivation. The dimen-
sions of deprivation used to construct the English IMD
2004, for instance, were (1) income deprivation, (2)
employment deprivation, (3) health deprivation and dis-
ability, (4) education, skills and training deprivation, (5)
barriers to housing and services, (6) living environment
deprivation and (7) crime [21]. As the various UK coun-
tries’ indices of multiple deprivation are not comparable,
IMD ranks were used, which for the purposes of this study
were standardized. As discussed above, the geography we
worked with was at LSOA level and the IMD score was
measured at LSOA level. Many families changed LSOAs
between Sweeps 1 and 2 (of the 19,244 cohort members,
52.4% did not change LSOA between Sweeps 1 and 2, 44%
did, and for 3.6% there was missing information about
LSOA at Sweep 1), but most remained within the same
LSOA. The IMD scores at Sweeps 1 and 2 were highly
correlated (the correlation between the IMD ranks at
Sweeps 1 and 2 was 0.911), and we, therefore, used IMD at
Sweep 1 in all our analyses.
Distal (birth to Sweep 1) and proximal (Sweep 1 to
Sweep 2) family contextual risk was measured with a
subset of Tiet et al.’s [18] Adverse Life Events Scale. This
scale is composed of 25 possible events for which children
had little or no control over and is a modification of the
Life Events Checklist (LEC) [22–24], a psychometrically
valid [22] measure of exposure to potentially traumatic
events developed at the National Center for Posttraumatic
Stress Disorder (PTSD) to facilitate the diagnosis of
PTSD. The self-report version of the Adverse Life Events
Scale has been used with children as young as 9 years
[25], and measures exposure to adverse life events at both
family (e.g., ‘negative change in parents’ financial situa-
tion’), and child (e.g., ‘got seriously ill or injured’, ‘lost a
close friend’) levels. In view of this study’s research aims
the version of the scale used comprised items which (a)
measured only family level risk, (b) were developmentally
appropriate, (c) could be reconstructed from the MCS
data, and (d) measured both distal and proximal risk. In
all, information about eight adverse life events was
available both between birth and Sweep 1 (‘distal family
risk’) and between Sweeps 1 and 2 (‘proximal family
risk’) in MCS. The eight events corresponded to the fol-
lowing Adverse Life Events Scale items: ‘family member
died’, ‘family member was seriously injured’, ‘negative
change in parents’ financial situation’, ‘family member
had mental/emotional problem’, ‘family moved’, ‘got a
new brother or sister’, ‘one of the parents went to jail’, and
‘parents separated’. In contrast to the high correlation in
IMD scores at Sweeps 1 and 2, the correlation between
family risk at Sweep 1 and family risk at Sweep 2 was
0.201.
Broad and specific psychopathology was assessed with
the main caregiver’s report of the Strengths and Difficulties
Questionnaire (SDQ), a 25-item 3-point (ranging from 0 to
2) scale measuring four difficulties (hyperactivity, emo-
tional symptoms, conduct problems, and peer problems), as
well as prosocial behavior [26, 27]. Each subscale has five
items. A total difficulties (broad psychopathology) scale is
calculated by summing the scores for hyperactivity, emo-
tional symptoms, conduct problems and peer problems
(http://www.sdqinfo.com). The SDQ has both good test–
retest reliability [28] and excellent concurrent and dis-
criminant validity [29, 30].
The child level covariates were age and sex, and the
family level covariates, all measured at Sweep 1, were
family structure, maternal psychological distress, and, to
index family’s socioeconomic status, maternal social class
and qualifications. Maternal psychological distress was
measured with nine items from the Malaise Inventory [31],
a psychometrically valid [32] measure of depressed mood.
Maternal social class, measured with the National Statistics
Socio-economic Classification (NS-SEC), ranged from 1
(‘high managerial’/‘professional’) to 7 (‘routine’). Moth-
ers’ highest qualifications were grouped into six major
categories, ranging from 1 to 6 (‘level 5’, i.e., first/higher
degree), roughly equivalent to National Vocational Quali-
fication (NVQ) levels.
Eur Child Adolesc Psychiatry (2010) 19:535–546 537
123
The moderator variables examined were developmental
milestones, temperament and parenting, all assessed in
Sweep 1 by the main respondent, and verbal and non-
verbal ability, measured in Sweep 2. Developmental mile-
stones are a set of functional skills or age-specific tasks that
most children can do at a certain age range. In the MCS
developmental milestones assessing communicative ges-
tures as well as fine and gross motor coordination typical
for a 9-month-old child were measured with 12 3-point
scales from the Denver Developmental Screening Test
[33], the most popular tool to screen for potential devel-
opmental problems. A high score on the test indicated
developmental delay. Temperament or individual differ-
ences in reactivity and self-regulation that are assumed to
have a relatively enduring, biological basis, was measured
in the MCS with items selected from the Carey Infant
Temperament Scale [34, 35], used also in the Avon Lon-
gitudinal Study of Parents and Children (ALSPAC) and the
National Longitudinal Survey of Youth. The items inclu-
ded in the MCS measured three dimensions of tempera-
ment, namely mood (measured with five items such as ‘is
pleasant’), adaptability (measured with five items such as
‘is rarely or almost never wary of strangers’) and regularity
(measured with four items such as ‘gets sleepy at about the
same time’). Parenting was measured with four 5-point
scales, originally derived from the European Longitudinal
Study of Pregnancy and Childhood, and used in other UK
longitudinal studies (such as the ALSPAC), assessing to
what extent the main respondent agreed that ‘it is important
to develop a regular pattern of feeding and sleeping with a
baby’, ‘babies need to be stimulated if they are to develop
well’, ‘talking, even to a young baby, is important’, and
‘cuddling a baby is very important’. The four-item scale for
the whole MCS at Sweep 1 yielded a Cronbach’s alpha of
0.67, with a high scale score indicating negligent parenting.
Finally, verbal and non-verbal ability were measured with
the naming vocabulary subtest of the British Ability Scales
(BAS) and with the School Readiness Composite (SRC) of
the Revised Bracken Basic Concept Scale, respectively.
The BAS Naming Vocabulary subtest measures children’s
expressive language skills, and the SRC measures ‘readi-
ness’ for formal education. Both assessments were indi-
vidually administered in Sweep 2 by survey team members
in computer-assisted interviews. The BAS Naming
Vocabulary subtest is part of a cognitive test battery
designed for children aged 3–17 [36]. Children are asked to
name a series of pictures of everyday items. There are 36
items in total, and the assessment is terminated if five
successive items are answered incorrectly. In the MCS, the
test was not administered to children who did not speak
English. The SRC comprises six subtests of the Revised
Bracken Basic Concept Scale measuring children’s
knowledge of those ‘readiness’ concepts that parents and
teachers traditionally teach children in preparation for
formal education [37]. The test has been designed for
children in the age range of 2.5–7 years and 11 months.
The six subtests of the SRC comprise the assessment of
children’s basic concepts of colors, letters, numbers/
counting, sizes, comparisons and shapes.
Models
The structure of the MSC data with children clustered
within LSOAs dictates the use of statistical models that
appropriately account for the nested structure. One class of
statistical models that provide valid inferences in the case
of hierarchical data is multilevel random effects models
[38]. To be consistent with the hierarchical structure of the
MCS data, we describe multilevel models using a two-level
structure under which children are clustered within LSOAs
at Sweep 2. The simplest two-level model for total diffi-
culties is a random intercepts model described by
yij ¼ xTij b þ uj þ eij; i ¼ 1; . . .; n; j ¼ 1; . . .; d ð1Þ
where yij is the response variable (total difficulties) for
child i in LSOA j, xij denotes a set of explanatory variables
that can be defined at child, family and LSOA level, uj denotes a vector of random effects associated with LSOA j
and eijk denotes the level 1 (child level) residual term. As information for each child was available for each of
the five SDQ subscales, allowing for a different model for
each SDQ subscale may provide a deeper insight. How-
ever, it is possible that responses to the different SDQ
subscales are correlated. To account for the existence of
correlation in each child’s responses, we fitted multivariate
response multilevel models that allowed for the error terms
of the different models to be correlated. Let us denote by i
the subscript referring to the child, by j the subscript
referring to the LSOA, and by l the subscript denoting a
specific SDQ subscale. The multivariate response multi-
level model is then defined as
yijl ¼ xTijlbl þ ujl þ eijl; i ¼ 1; . . .; n; j ¼ 1; . . .; d; l ¼ 1; . . .; 5
ð2Þ
where eijl * N(0, Re) with Re denoting the variance covariance matrix between the level 1 error terms and
ujl * N(0, Ru) with Ru denoting the variance covariance matrix between the level 2 error terms. Modeling specific
psychopathologies in a multivariate way offers a more
flexible modeling framework as it can accommodate dif-
ferent covariates for the different SDQ subscales as well as
allow for the correlation between unobserved factors
affecting the scores on the different SDQ subscales at
different hierarchical levels. When compared with a single
multilevel model for broad psychopathology, a multivariate
538 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
multilevel model offers an additional advantage with
respect to the handling of missing data. In particular, when
modeling broad psychopathology, a missing value in one of
the items comprising total difficulties on the SDQ results in
a missing total difficulties value. This is not the case with
multivariate models, since missing values in some of the
SDQ subscales do not result in the omission of the corre-
sponding unit from the analysis. In this way, the estimates
obtained from the multivariate multilevel model are
efficient.
The complex survey design of the MCS was accounted
for by conditioning on the design variables, i.e., the vari-
ables used in designing the MCS survey. Therefore, in the
models described below, stratification was accounted for
by including a set of dummy variables representing the
nine strata of the MCS, whereas clustering was accounted
for by running multilevel models. The models assumed that
data were missing at random (MAR) [38].
Results
We used STATA version 10 and MLwiN to run random
intercepts models (see Table 1 for a model summary),
described by (1), allowing the constant to vary at level 2,
i.e., LSOA at Sweep 2. We first estimated an empty
model for total difficulties in the 9,735 children with valid
data on total difficulties. This model included only the
design variables to adjust for the effects of stratification.
The two-level empty model showed that the average total
difficulties score as reflected in the intercept was 8.219
(SE = 0.102). The child level variance component (level
1 variance) was 23.367 (SE = 0.353), and the variance
due to differences in LSOAs was 0.850 (SE = 0.166).
This suggests a significant between LSOA variation.
Using the two variance components to partition the var-
iance across levels we found that the intra-cluster corre-
lation coefficient was 0.035. This suggests that controlling
only for stratification 3.5% of the variance in total diffi-
culties scores was attributable to the area level. Therefore,
although small the intra-cluster correlation justified the
use of hierarchical modeling. We first investigated the
contextual risk/total difficulties association by running
three baseline models, one with proximal family risk, one
with distal family risk and one with area risk. Area risk
(b = -0.938, SE = 0.083), distal family risk (b = 0.852,
SE = 0.060), and proximal family risk (b = 0.715,
SE = 0.043) were all related to total difficulties. As
expected, the amount of variance due to differences in
areas explained in this model when area risk was added to
Table 1 Model summary (broad psychopathology)
a The design variables used
were the MCS strata defined by
country (England, Scotland,
Wales and Northern Ireland)
and disadvantage status
according to the Child Poverty
Index (CPI). For England, there
was an additional stratum, i.e.
ethnic minority indicating wards
which, in the 1991 Census of
Population, had an ethnic
minority indicator of at least
30%. In other words, at least
30% of their total population
fell into the categories ‘Black’
or ‘Asian’
Models Specification
Model 1 Design variables a
Model 2 Model 1 ? area random effect
Model 3.1 Model 2 ? proximal family risk
Model 3.2 Model 2 ? distal family risk
Model 3.3 Model 2 ? area risk
Model 4 Model 2 ? area risk ? measures of family risk (proximal and distal family risk)
Model 5 Model 4 ? family level fixed effects (family structure, mother’s qualifications,
mother’s social class and mother’s psychological distress)
Model 6 Model 5 ? child level fixed effects (gender, age) ? moderator variables (mood,
regularity, adaptability, non-verbal ability, verbal ability, developmental delay,
negligent parenting)
Model 7.1 Model 6 ? quadratic effect for area risk
Model 7.2 Model 6 ? quadratic effect for distal family risk
Model 7.3 Model 6 ? quadratic effect for proximal family risk
Model 8.1 Model 6 ? interaction between proximal family risk and distal family risk
Model 8.2 Model 6 ? interaction between proximal family risk and area risk
Model 8.3 Model 6 ? interaction between distal family risk and area risk
Model 9.1 Model 6 ? interaction between proximal family risk and mood
Model 9.2 Model 6 ? interaction between proximal family risk and regularity
Model 9.3 Model 6 ? interaction between proximal family risk and adaptability
Model 9.4 Model 6 ? interaction between proximal family risk and non-verbal ability
Model 9.5 Model 6 ? interaction between proximal family risk and verbal ability
Model 9.6 Model 6 ? interaction between proximal family risk and developmental delay
Model 9.7 Model 6 ? interaction between proximal family risk and parenting
Eur Child Adolesc Psychiatry (2010) 19:535–546 539
123
the empty model was reduced (0.697, SE = 0.152). To
compare the goodness of fit of these three baseline
models, we used an appropriate statistic that takes into
account that the models are not nested. Using the
Bayesian Information Criterion (BIC), a function of the
likelihood, the number of observations and the number of
free parameters of each model [39], we found that the
most parsimonious model was the proximal family risk
model. The association of area risk, distal family risk and
proximal family risk with total difficulties remained sig-
nificant even when all three risk variables were entered in
the same model. The amount of variance due to differ-
ences in areas explained in this model was further
reduced (0.606, SE = 0.140), but was still significant.
This suggests that although child psychopathology scores
differed by area partly because of differences in area level
socioeconomic disadvantage, even with area socioeco-
nomic disadvantage adjusted there was still area level
variance that remained unexplained.
In the next step, the family level variables of family
structure, and mother’s qualifications, social class and
psychological distress were added. With these variables in
the model, the effect of area risk on total difficulties
became nonsignificant. The amount of variance in total
psychopathology scores due to differences in area
explained in the model became nonsignificant (0.090,
SE = 0.093). Taken together, these two findings suggest
that both the area differences in total difficulties and the
effect of area risk on total difficulties operated via family
characteristics. The effect of distal and proximal family
risk, on the other hand, remained significant (b = 0.238,
SE = 0.065, and b = 0.376, SE = 0.044, respectively)
suggesting that family contextual risk predicts child psy-
chopathology directly and independently of family struc-
ture, socioeconomic status or maternal psychopathology.
Next, the full model (Model 6) was introduced. This
added the child level control variables and the proposed
moderators. The final study sample obtained after omitting
cases with missing values on both broad psychopathology
and all its predictors was 4,748 children clustered in 1,549
LSOAs. The number of children per LSOA ranged from 1
to 41, with an average of 3.1 children per LSOA. The
distribution of the 4,748 children in the nine MCS strata in
Sweep 2 was as follows: England-advantaged: 1,847;
England-disadvantaged: 884; England-ethnic: 139; Wales-
advantaged: 319; Wales-disadvantaged: 421; Scotland-
advantaged: 434; Scotland-disadvantaged: 264; Northern
Ireland-advantaged: 220, and Northern Ireland-disadvan-
taged: 220. As can be seen in Table 2, although negligent
parenting was marginally significant, all the child level
moderator variables were statistically significant in pre-
dicting broad psychopathology. In this final model, the
child level variance component (level 1 variance) was
16.409 (SE = 0.359), and the variance due to differences
in LSOAs was 0.203 (SE = 0.133).
As explained above, the initial MCS sample is 19,244
children. Of these, 9,736 had information about broad
psychopathology, our main response variable, and of these,
4,748 children were the final study sample. As expected,
these 4,748 children differed from the 14,496 children not
included in the final study sample in most of the study
variables. In particular, with regards to differences in
Sweep 1 variables, compared with their counterparts those
in the final study sample tended to live in less disadvan-
taged areas, in intact families, and with mothers who were
less negligent, less distressed, more educated and of higher
social class. With regards to differences in Sweep 2 vari-
ables, compared with their counterparts those in the study
sample tended to present with less broad psychopathology,
be female, have easier temperament, higher verbal and
non-verbal cognitive ability, and have less developmental
delay (all differences were statistically significant at
P \ 0.001; results available from the authors). However, there was no difference between the two groups in age or in
number of either proximal or distal family adverse life
events experienced.
Comparing cumulative and specific family risk
specifications
We could not compare the proximal cumulative risk
specification with the proximal specific risk specification
as one item of the proximal Adverse Life Events Scale
(‘one of the parents went to jail’) was dropped due to
collinearity. However, the comparison of the distal
cumulative with the distal specific risk specification
showed that, after applying a Bonferroni correction
(alpha of 0.05/8 adverse events = 0.00625), none of the
eight specific distal risks were significantly associated
with broad psychopathology. The only distal risk item
approaching significance was ‘negative change in parents’
financial situation’ (b = 0.421, SE = 0.158). As the BIC
for the cumulative risk model (27,176.89) was lower than
that for the specific risk model (27,216.41), we concluded
that the cumulative risk model specification should be
preferred, and therefore this risk specification was used for
the remaining analysis.
Investigating the appropriate functional
form of contextual risk’s effect and testing
for moderator effects
We further investigated the appropriate functional form of
contextual risk’s effect on broad psychopathology by
introducing quadratic terms for contextual risk separately
in the full model (Model 6). None of the three quadratic
540 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
terms for area, distal family, and proximal family risk were
significant. We also tested if the effect of one type of risk
was conditional upon the value of another. We entered
these two-way interactions between distal, proximal and
area risk variables separately in the full model, but neither
interaction term was significant. This suggests that the
effect of proximal family risk on broad psychopathology
did not depend on the level of distal risk, and that area risk
did not moderate the effect of either proximal or distal risk
on broad psychopathology.
Table 2 Broad psychopathology
Predictors Model 2 Model 4 Model 6
Coeff. SE Coeff. SE Coeff. SE
Constant 8.219 0.102 7.250 0.174 13.749 1.371
Stratum (Ref: England-advantaged)
England-disadvantaged 1.927 0.170 0.849 0.186 0.315 0.208
England-ethnic 2.732 0.289 1.983 0.294 0.630 0.405
Wales-advantaged -0.155 0.257 -1.609 0.289 -0.528 0.321
Wales-disadvantaged 1.276 0.224 -0.406 0.266 0.021 0.302
Scotland-advantaged -0.357 0.234 21.595 0.257 -0.178 0.281
Scotland-disadvantaged 1.115 0.268 -0.403 0.293 0.188 0.334
Northern Ireland-advantaged 20.715 0.286 22.142 0.318 20.844 0.360
Northern Ireland-disadvantaged 1.194 0.260 -0.435 0.298 20.854 0.360
Standardized IMD rank 20.776 0.081 -0.039 0.095
Proximal adverse life events 0.581 0.044 0.394 0.054
Distal adverse life events 0.666 0.061 0.308 0.079
Family structure (Ref: two natural parents)
Natural mother only 0.260 0.208
Other -0.959 1.828
Mother’s NS-SEC (Ref: high managerial/professional)
Low managerial/professional -0.047 0.229
Intermediate 0.098 0.259
Small emp and self-employed 20.824 0.361
Low supervisory and technical 0.517 0.336
Semi routine 0.728 0.275
Routine 0.553 0.310
Mother’s highest qualifications (Ref: level 5)
Level 4 -0.426 0.306
Level 3 0.372 0.330
A/AS Level 0.345 0.333
Level 2 0.920 0.314
Level 1 1.623 0.364
Mother’s psychological distress 0.472 0.040
Girl 20.755 0.121
Age 0.669 0.327
Mood 20.153 0.019
Regularity 20.134 0.023
Adaptability 20.123 0.018
Non-verbal ability 20.038 0.006
Verbal ability 20.072 0.017
Developmental delay 0.049 0.025
Negligent parenting 0.073 0.042
Between area variability (ru 2 ) 0.849 0.155 0.606 0.133 0.171 0.135
Within area variability (re 2 ) 23.367 0.351 22.379 0.335 16.317 0.357
Eur Child Adolesc Psychiatry (2010) 19:535–546 541
123
We finally explored if parenting, verbal and non-verbal
ability, developmental milestones, and mood, regularity
and adaptability moderate the association between con-
textual risk and young children’s broad psychopathology.
As proximal risk was the type of risk more strongly asso-
ciated with child psychopathology both at baseline and in
the full model, interaction terms with proximal family risk
and each of the proposed moderators were calculated and
entered separately in the full model (Model 6). None of the
interaction terms were significant, however, suggesting that
the effect of proximal risk on broad psychopathology was
not buffered either by child’s easy temperament or
advanced development, or by involved parenting. Table 2
presents the results of Model 6 and of the intermediate
models (Models 2 and 4) that were of particular substantive
significance.
Multivariate response models
We fitted multivariate response multilevel models descri-
bed by [8] that allowed the error terms of the different
models to be correlated and that included random area
(LSOA) effects. These multivariate response two level
models were effectively treated in MLwiN as three level
models, i.e., with the responses as the additional lower
level.
First, we ran an empty multivariate response multilevel
model (Model 10). This model included only the design
variables to adjust for the effects of stratification. The
variance partition coefficients obtained showed that 1% of
the total variation in prosocial behavior, 2% of the total
variation in emotional symptoms, 2% of the total variation
in conduct problems, 2% of the total variation in hyper-
activity, and 2% of the total variation in peer problems
were due to between LSOA variation.
We started our analysis by fitting three baseline models.
The first model examined the effect of proximal family
risk, which was significant in predicting scores in all five
SDQ subscales. The second model examined the effect of
distal risk, which was also significant in predicting scores
in all five SDQ subscales. The third model examined the
effect of area risk, which also had a significant, albeit
weaker, effect on the four difficulties and a nonsignificant
effect on prosocial behavior. The effect of proximal, distal
and area risk on all four difficulties remained significant
even when all risk variables were entered simultaneously
(Model 13). However, the effect of distal risk on prosocial
behavior became nonsignificant after adjusting for proxi-
mal and area risk suggesting that the effect of distal risk on
prosocial behavior was via its impact on proximal risk.
Subsequently, we introduced the full model (Model 14).
This added to the model including the design variables and
the three risk variables all the family- and child level fixed
effects, and all the moderator variables. The sample size in
this full model was 28,672. 1
In this model, the effect of
proximal risk on prosocial behavior became nonsignificant,
as did the effect of area risk on all four difficulties.
Although the effect of proximal risk remained significant
on all four difficulties, the effect of distal risk was signif-
icant only on hyperactivity and conduct problems.
Inspection of the between area and between children var-
iance covariance matrices (Table 3) from the various
models showed the covariances between the SDQ sub-
scales at area level became less significant as we controlled
for background characteristics. The covariances between
the SDQ subscales at child level, however, remained sig-
nificant even after controlling for a range of background
characteristics in this final model (Model 14), which jus-
tifies the use of a multivariate model.
Next, we tested if the effect of one type of risk was
conditional upon the value of another. We calculated two-
way interactions between distal, proximal and area risk
variables and entered these separately in the full model,
but none of these interaction terms was significant. This
suggests that the effect of proximal family risk on specific
psychopathology did not depend on the level of distal
risk, and that area risk did not moderate the effect of
either proximal or distal risk on any specific child
psychopathology.
Finally, we explored if parenting, verbal and non-verbal
ability, developmental milestones, and mood, regularity
and adaptability moderate the association between con-
textual risk and young children’s specific psychopathology.
As proximal risk was the type of risk more strongly asso-
ciated with child psychopathology both at baseline and in
the full model, interaction terms with proximal family risk
and each of the proposed moderators were calculated
and entered separately in the full model. The interaction
between proximal family risk and developmental
milestones predicted conduct problems (b = -0.017,
SE = 0.008), and the interaction between proximal family
risk and non-verbal ability predicted both conduct prob-
lems (b = -0.002, SE = 0.001) and emotional symptoms
(b = -0.002, SE = 0.001), but no other interaction effects
were significant. These effects remained significant even
when both interaction terms were entered in the same
model (Model 17). These findings suggest that delayed
development buffered the effect of proximal risk on con-
duct problems, and non-verbal ability buffered the effect
of proximal risk on both emotional symptoms and con-
duct problems. Although developmental delay was not
1 This was the total number of observations in the multivariate
(specific psychopathology) model. The number of observations for
each SDQ outcome was as follows: Prosocial behavior: 5,719;
Emotional symptoms: 5,966; Conduct problems: 5,954; Hyperactiv-
ity: 5,689, and Peer problems: 5,344.
542 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
significantly associated with any externalizing psychopa-
thology (although it was positively associated with emo-
tional symptoms and peer problems and negatively
associated with prosocial behavior) in the full model
(Model 14), it buffered the effect of proximal adverse life
events on conduct problems. Similarly, although non-ver-
bal ability was not significantly associated with any
internalizing psychopathology (although it was positively
related to prosocial behavior and negatively related to both
conduct problems and hyperactivity) in the full model
(Model 14), it buffered the effect of proximal risk on
conduct problems but also on emotional symptoms.
Table 4 shows the model summary for specific psychopa-
thology, and Table 5 shows the results of Model 17.
Table 3 Child level and area level covariance matrices between the five SDQ subscales estimated by fitting multivariate models (Models 10, 13, and 14)
Prosocial Emotional Conduct Hyperactivity Peer
Model 10
Between area variance covariance matrix
Prosocial 0.021 (0.011)
Emotional -0.011 (0.007) 0.042 (0.008)
Conduct -0.018 (0.010) 0.050 (0.009) 0.088 (0.017)
Hyperactivity -0.018 (0.011) 0.048 (0.010) 0.080 (0.016) 0.090 (0.022)
Peer -0.009 (0.007) 0.039 (0.007) 0.051 (0.010) 0.047 (0.011) 0.043 (0.010)
Between children variance covariance matrix
Prosocial 3.302 (0.043)
Emotional -0.209 (0.024) 1.938 (0.025)
Conduct -1.246 (0.036) 0.757 (0.026) 3.937 (0.050)
Hyperactivity -1.249 (0.041) 0.715 (0.031) 2.173 (0.047) 5.202 (0.069)
Peer -0.759 (0.027) 0.611 (0.020) 0.719 (0.029) 0.745 (0.033) 2.105 (0.029)
Model 13
Between area variance covariance matrix
Prosocial 0.022 (0.011)
Emotional -0.013 (0.007) 0.032 (0.008)
Conduct -0.023 (0.009) 0.031 (0.008) 0.054 (0.015)
Hyperactivity -0.023 (0.011) 0.033 (0.009) 0.053 (0.014) 0.068 (0.002)
Peer -0.011 (0.007) 0.028 (0.006) 0.029 (0.008) 0.029 (0.010) 0.031 (0.009)
Between children variance covariance matrix
Prosocial 3.299 (0.043)
Emotional -0.200 (0.024) 1.915 (0.024)
Conduct -1.223 (0.035) 0.700 (0.026) 3.799 (0.048)
Hyperactivity -1.228 (0.041) 0.663 (0.030) 2.043 (0.045) 5.079 (0.067)
Peer -0.750 (0.027) 0.587 (0.020) 0.665 (0.028) 0.697 (0.033) 2.081 (0.028)
Model 14
Between area variance covariance matrix
Prosocial 0.005 (0.017)
Emotional -0.005 (0.009) 0.008 (0.009)
Conduct -0.006 (0.014) 0.015 (0.01) 0.029 (0.02)
Hyperactivity -0.004 (0.016) 0.029 (0.012) 0.011 (0.018) 0.024 (0.027)
Peer -0.011 (0.011) 0.007 (0.008) 0.003 (0.011) -0.003 (0.013) 0.023 (0.013)
Between children variance covariance matrix
Prosocial 2.848 (0.056)
Emotional -0.093 (0.029) 1.468 (0.028)
Conduct -0.949 (0.043) 0.375 (0.029) 2.992 (0.058)
Hyperactivity -0.871 (0.051) 0.358 (0.036) 1.478 (0.054) 4.304 (0.085)
Peer -0.500 (0.033) 0.384 (0.023) 0.361 (0.033) 0.387 (0.04) 1.668 (0.034)
Eur Child Adolesc Psychiatry (2010) 19:535–546 543
123
Discussion
This study was carried out to test the hypothesis that for
preschool children’s behavioral and emotional adjustment
in the face of contextual risk, factors promoting resilience
should be located in the cognitive domain. Indeed, this
study showed that neither parenting nor temperament
buffered the effect of contextual risk on young children’s
emotional and behavioral adjustment. However, non-verbal
ability moderated the effect of cumulative proximal risk on
both emotional symptoms and conduct problems.
In testing this hypothesis, we measured and modeled
contextual risk appropriately and we, therefore, also
extended in several ways prior work on contextual risk
effects on children’s behavioral and emotional adjustment.
First, assessing with well-validated measures family con-
textual risk, we showed that, although the number of family
adversities experienced in the child’s first year (distal
family risk) did predict broad and externalizing psycho-
pathology, the number of proximal (i.e., in the second and
third year) family adversities (proximal family risk) pre-
dicted both broad and externalizing but also internalizing
psychopathology. Second, we showed that the most parsi-
monious model was the model that included cumulative
rather than specific family risk. Taken together, these
findings highlight the importance of proximal family risk in
predicting both broad and specific psychopathology, and
the importance of considering family risk accumulation
rather than specificity in predicting psychopathology in
young children. Third, we showed that although the effect
of area and the effect of area level contextual risk were
significant on broad psychopathology even after distal and
proximal family risk were controlled for, they became
nonsignificant after adjusting for maternal socioeconomic
status and mental health. This finding suggests that the
effect of area and the effect of area risk on child psycho-
pathology operate via family characteristics. The effect of
distal and proximal family risk, on the other hand,
remained significant after adjusting for these maternal
characteristics suggesting that family contextual risk pre-
dicts child psychopathology directly and independently,
and not because it is associated with parental psychopa-
thology or social class. In other words, the effect of family
contextual risk on child psychopathology transcends social
origins and genetic predispositions. Fourth and finally, by
testing for the functional form of the effect of both area and
family risk and for the interaction between the various
types and levels of risk, we joined the few other researchers
[4, 11–14] that have tested for ‘threshold’ models of mul-
tiple risk. Ours, however, was the first study to test such
effects on young children’s psychopathology.
Of course, the study’s limitations should also be
acknowledged. For example, the children of the final study
sample were clearly more privileged than those not inclu-
ded in the final study sample. However, in our modelling
framework data were treated under the MAR assumption.
Table 4 Model summary (specific psychopathology)
Models Specification
Model 10 Design variables
Model 11 Model 10 ? area random effect
Model 12.1 Model 11 ? proximal family risk
Model 12.2 Model 11 ? distal family risk
Model 12.3 Model 11 ? area risk
Model 13 Model 11 ? area risk ? measures of family risk (proximal and distal family risk)
Model 14 Model 13 ? family level fixed effects (family structure, mother’s qualifications,
mother’s social class, mother’s psychological distress) ? child level fixed effects
(age, gender) ? moderator variables (mood, regularity, adaptability, non-verbal
ability, verbal ability, developmental delay, negligent parenting)
Model 15.1 Model 14 ? interaction between proximal family risk and distal family risk
Model 15.2 Model 14 ? interaction between proximal family risk and area risk
Model 15.3 Model 14 ? interaction between distal family risk and area risk
Model 16.1 Model 14 ? interaction between proximal family risk and mood
Model 16.2 Model 14 ? interaction between proximal family risk and regularity
Model 16.3 Model 14 ? interaction between proximal family risk and adaptability
Model 16.4 Model 14 ? interaction between proximal family risk and non-verbal ability
Model 16.5 Model 14 ? interaction between proximal family risk and verbal ability
Model 16.6 Model 14 ? interaction between proximal family risk and developmental delay
Model 16.7 Model 14 ? interaction between proximal family risk and parenting
Model 17 Model 14 ? interaction between proximal family risk and developmental
delay ? interaction between proximal family risk and non-verbal ability
544 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
Our study findings have important implications for the
study of both contextual risk and resilience as they suggest
that, even among young children, contextual factors that
impede adjustment are proximal rather than distal, and
should be modeled cumulatively and located in the family
rather than the area, and, importantly for the study of
resilience, that factors that promote adjustment in the face
of such contextual risk should be located among individual
Table 5 Specific psychopathology (Model 17)
Predictors Prosocial
behavior
Coeff. (SE)
Emotional
symptoms
Coeff. (SE)
Conduct
problems
Coeff. (SE)
Hyperactivity
Coeff. (SE)
Peer problems
Coeff. (SE)
Constant 6.633 (0.558) 2.306 (0.401) 3.789 (0.573) 3.989 (0.698) 2.870 (0.445)
Stratum (Ref: England-advantaged)
England-disadvantaged -0.050 (0.076) -0.032 (0.054) 0.017 (0.078) 0.078 (0.094) 0.181 (0.063)
England-ethnic 0.375 (0.139) 0.128 (0.099) 0.038 (0.146) 0.076 (0.176) 0.259 (0.117)
Wales-advantaged 0.033 (0.118) -0.046 (0.084) -0.213 (0.122) 0.118 (0.146) -0.134 (0.097)
Wales-disadvantaged 0.177 (0.109) -0.044 (0.079) 0.154 (0.114) 0.237 (0.136) -0.082 (0.091)
Scotland-advantaged -0.037 (0.105) -0.058 (0.074) -0.042 (0.108) -0.042 (0.129) -0.067 (0.086)
Scotland-disadvantaged -0.194 (0.123) -0.079 (0.088) 0.141 (0.128) 0.139 (0.153) 0.062 (0.101)
Northern Ireland-advantaged 0.024 (0.134) -0.036 (0.094) -0.232 (0.136) -0.197 (0.164) -0.187 (0.108)
Northern Ireland-disadvantaged 0.154 (0.133) -0.085 (0.094) -0.134 (0.137) -0.408 (0.164) -0.082 (0.108)
Standardized IMD rank -0.037 (0.035) -0.012 (0.025) -0.028 (0.036) 0.077 (0.043) -0.033 (0.029)
Proximal adverse life events -0.071 (0.147) 0.171 (0.105) 0.511 (0.149) 0.134 (0.181) 0.073 (0.118)
Distal adverse life events 0.025 (0.030) 0.031 (0.021) 0.096 (0.030) 0.132 (0.037) 0.035 (0.024)
Family structure (Ref: two natural parents)
Natural mother only 0.079 (0.077) 0.079 (0.054) 0.170 (0.078) 0.110 (0.095) 0.082 (0.061)
Other -0.092 (0.060) -0.131 (0.459) -0.745 (0.617) -1.340 (0.833) 0.425 (0.525)
Mother’s NS-SEC (Ref: high managerial/professional)
Low managerial/professional 0.002 (0.088) -0.065 (0.062) -0.064 (0.089) 0.120 (0.108) 0.039 (0.070)
Intermediate 0.015 (0.099) -0.136 (0.069) 0.069 (0.099) 0.280 (0.121) -0.082 (0.078)
Small emp and self-employed 0.109 (0.138) -0.159 (0.097) -0.243 (0.138) -0.173 (0.168) -0.035 (0.109)
Low supervisory and technical 0.123 (0.126) -0.123 (0.089) 0.056 (0.127) 0.445 (0.156) 0.174 (0.100)
Semi routine -0.107 (0.104) -0.037 (0.073) 0.347 (0.104) 0.332 (0.128) 0.082 (0.083)
Routine 0.017 (0.116) 0.001 (0.082) 0.173 (0.117) 0.276 (0.142) 0.173 (0.092)
Mother’s highest qualifications (Ref: level 5)
Level 4 -0.100 (0.115) -0.050 (0.081) -0.096 (0.117) -0.035 (0.140) -0.120 (0.094)
Level 3 -0.048 (0.124) -0.004 (0.088) 0.010 (0.126) 0.345 (0.152) 0.067 (0.101)
A/AS Level -0.108 (0.125) 0.022 (0.088) -0.058 (0.127) 0.272 (0.153) 0.056 (0.102)
Level 2 -0.111 (0.117) 0.096 (0.083) 0.170 (0.119) 0.495 (0.143) 0.141 (0.096)
Level 1 -0.092 (0.134) 0.311 (0.095) 0.400 (0.136) 0.722 (0.165) 0.229 (0.110)
Mother’s psychological distress -0.032 (0.014) 0.085 (0.010) 0.158 (0.015) 0.159 (0.018) 0.064 (0.012)
Girl 0.306 (0.045) -0.009 (0.032) -0.143 (0.046) -0.505 (0.056) -0.152 (0.036)
Age 0.051 (0.121) 0.109 (0.088) 0.013 (0.124) 0.580 (0.152) -0.066 (0.097)
Mood 0.064 (0.007) -0.025 (0.005) -0.053 (0.007) -0.051 (0.009) -0.017 (0.006)
Regularity 0.003 (0.008) -0.019 (0.006) -0.031 (0.009) -0.043 (0.010) -0.035 (0.007)
Adaptability 0.017 (0.007) -0.050 (0.005) -0.020 (0.007) -0.015 (0.008) -0.039 (0.005)
Non-verbal ability 0.009 (0.003) -0.000 (0.002) -0.011 (0.003) -0.016 (0.004) -0.002 (0.003)
Verbal ability 0.021 (0.007) -0.016 (0.005) -0.014 (0.007) -0.033 (0.008) -0.020 (0.005)
Developmental delay -0.076 (0.015) 0.023 (0.011) 0.029 (0.015) 0.009 (0.019) 0.026 (0.012)
Negligent parenting -0.095 (0.015) 0.005 (0.011) 0.015 (0.015) 0.035 (0.019) 0.028 (0.012)
Proximal adverse life events by developmental delay 0.004 (0.008) -0.004 (0.005) -0.017 (0.008) 0.004 (0.010) -0.002 (0.006)
Proximal adverse life events by non-verbal ability -0.001 (0.001) -0.002 (0.001) -0.002 (0.001) -0.002 (0.002) 0.001 (0.001)
Eur Child Adolesc Psychiatry (2010) 19:535–546 545
123
attributes pertaining specifically to cognitive strengths.
Future studies should test this as well as the generalise-
ability of these findings in older child samples as the
adverse characteristics of a neighborhood might impact
differently upon the psychopathology of older children.
Acknowledgments This program of research was supported by a grant from the British Academy to the first two authors. The authors
are grateful to Jon Johnson, Rachel Rosenberg, and Tina Roberts for
their help with the construction of the dataset.
References
1. Rutter M (1979) Protective factors in children’s responses to
stress and disadvantage. In: Kent MW, Rolf JE (eds) Primary
prevention of psychopathology: III. Promoting social competence
and coping in children. University Press of New England,
Hanover, pp 49–74
2. Evans GW (2003) A multimethodological analysis of cumulative
risk and allostatic load among rural children. Dev Psychol
39:924–933
3. Burchinal M, Roberts J, Hooper S, Zeisel S (2000) Cumulative
risk and early cognitive development: a comparison of statistical
risk models. Dev Psychol 36:793–807
4. Flouri E, Kallis C (2007) Adverse life events and psychopa-
thology and prosocial behavior in late adolescence: testing the
timing, specificity, accumulation, gradient, and moderation of
contextual risk. J Am Acad Child Adolesc Psychiatry 46:1651–
1659
5. Deater-Deckard K, Dodge KA, Bates JE, Pettit GS (1998) Mul-
tiple risk factors in the development of externalizing behavior
problems: group and individual differences. Dev Psychopathol
10:469–493
6. Atzaba-Poria N, Pike A, Deater-Deckard K (2004) Do risk factors
for problem behaviour act in a cumulative manner? An exami-
nation of ethnic minority and majority children through an eco-
logical perspective. J Child Psychol Psychiatry 45:707–718
7. Sameroff AJ, Seifer R, Baldwin A, Baldwin C (1993) Stability of
intelligence from preschool to adolescence: the influence of
social and family risk factors. Child Dev 64:80–97
8. Sheinkopf SJ, Lagasse LL, Lester BM et al (2007) Vagal tone as
a resilience factor in children with prenatal cocaine exposure.
Dev Psychopathol 19:649–673
9. Burchinal MR, Roberts JE, Zeisel SA, Rowley SJ (2008) Social
risk and protective factors for African American children’s
academic achievement and adjustment during the transition to
middle school. Dev Psychol 44:286–292
10. Ackerman BP, Brown ED, Izard CE (2004) The relations
between persistent poverty and contextual risk and children’s
behavior in elementary school. Dev Psychol 40:367–377
11. Flaherty EG, Thompson R, Litrownik AJ et al (2006) Effect of
early childhood adversity on child health. Arch Pediatr Adolesc
Med 160:1232–1238
12. Gerard JM, Buehler C (2004) Cumulative environmental risk and
youth problem behavior. J Marriage Fam 66:702–720
13. Morales JR, Guerra NG (2006) Effects of multiple context and
cumulative stress on urban children’s adjustment in elementary
school. Child Dev 77:907–923
14. Simmons RG, Burgeson R, Carlton-Ford S, Blyth DA (1987) The
impact of cumulative change in early adolescence. Child Dev
58:1220–1234
15. Leventhal T, Brooks-Gunn J (2000) The neighborhoods they live
in: the effects of neighborhood residence on children and ado-
lescent outcomes. Psychol Bull 126:309–337
16. McCulloch A, Joshi HE (2001) Neighborhood and family influ-
ences on the cognitive ability of children in the British National
Child Development Study. Soc Sci Med 53:59–591
17. Ackerman BP, Izard CE, Schoff K, Youngstrom EA, Kogos J
(1999) Contextual risk, caregiver emotionality, and the problem
behaviors of six- and seven-year-old children from economically
disadvantaged families. Child Dev 70:1415–1427
18. Tiet QQ, Bird HR, Davies M et al (1998) Adverse life events and
resilience. J Am Acad Child Adolesc Psychiatry 37:1191–1200
19. Masten AS (2001) Ordinary magic: resilience processes in
development. Am Psychol 56:227–238
20. Masten AS, Burt KB, Roisman GI, Obradovic J, Long JD, Tellegen
A (2004) Resources and resilience in the transition to adulthood:
continuity and change. Dev Psychopathol 16:1071–1094
21. Noble M, Wright G, Smith G, Dibben C (2006) Measuring
multiple deprivation at the small area level. Environ Plann A
38:169–185
22. Brand AH, Johnson JH (1982) Note on reliability of the Life
Events Checklist. Psychol Rep 50:1274
23. Coddington RD (1972) The significance of life events as etiologic
factors in the diseases of children: I. A survey of professional
workers. J Psychosom Res 16:7–18
24. Coddington RD (1972) The significance of life events as etiologic
factors in the diseases of children: II. A study of a normal pop-
ulation. J Psychosom Res 16:205–213
25. Tiet QQ, Bird HR, Hoven CW et al (2001) Relationship between
specific adverse life events and psychiatric disorders. J Abnorm
Child Psychol 29:153–164
26. Goodman R (1994) A modified version of the Rutter Parent
Questionnaire including extra items on children’s strengths: a
research note. J Child Psychol Psychiatry 35:1483–1494
27. Goodman R (1997) The Strengths and Difficulties Questionnaire:
a research note. J Child Psychol Psychiatry 38:581–586
28. Goodman R (2001) Psychometric properties of the Strengths and
Difficulties Questionnaire. J Am Acad Child Adolesc Psychiatry
40:1337–1345
29. Goodman R, Scott S (1999) Comparing the Strengths and Diffi-
culties Questionnaire and the Child Behavior Checklist: is small
beautiful? J Abnorm Child Psychol 27:17–24
30. Goodman R, Meltzer H, Bailey V (1998) The Strengths and
Difficulties Questionnaire: a pilot study on the validity of the self-
report version. Eur Child Adolesc Psychiatry 7:125–130
31. Rutter MJ, Tizard J, Whitmore K (1970) Education, health and
behaviour. Longman, London
32. Rodgers B, Pickles A, Power C, Collishaw S, Maughan B (1999)
Validity of the Malaise Inventory in general population samples.
Soc Psychiatry Psychiatr Epidemiol 34:333–341
33. Frankenburg WK, Dodds JB (1967) Denver Developmental
Screening Test. J Paediatr 71:181–191
34. Carey W, McDevitt S (1978) Revision of the Infant Temperament
Questionnaire. Pediatrics 61:735–739
35. Carey W, McDevitt S (1995) Coping with children’s tempera-
ment: a guide for professionals. Basic Books, New York
36. Elliot CD (1983) British Ability Scales. NFER-Nelson, Windsor
37. Bracken BA (1998) Bracken Basic Concept Scale-revised.
The Psychological Corporation, Harcourt Brace and Company,
San Antonio
38. Goldstein H (2003) Multilevel statistical models, 3rd edn.
Arnold, London
39. Schwarz G (1978) Estimating the dimension of a model. Ann Stat
6:461–464
546 Eur Child Adolesc Psychiatry (2010) 19:535–546
123
Copyright of European Child & Adolescent Psychiatry is the property of Springer Science & Business Media
B.V. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright
holder's express written permission. However, users may print, download, or email articles for individual use.