University of Southern California Dissertations and Theses

Untangling the developmental relations between depression and externalizing behavior among maltreated adolescents

(USC Thesis Other)

Untangling the developmental relations between depression and externalizing behavior among maltreated adolescents

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

UNTANGLING THE DEVELOPMENTAL RELATIONS BETWEEN DEPRESSION
AND EXTERNALIZING BEHAVIOR AMONG MALTREATED ADOLESCENTS

by

Matthew Brensilver

_________________________________________________________________

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(SOCIAL WORK)

August 2010

Copyright 2010 Matthew Brensilver

ii

Acknowledgements
I would like to acknowledge Ferol Mennen, whose consistent support and guidance
leaves me with a feeling of deep gratitude. I am also grateful for the time and support
of Penny Trickett and for the opportunity to work on the Young Adolescent Project.
Jack McArdle graciously agreed to sit on both my Qualifying Examination
Committee and the Dissertation Committee and for this, I‟m indebted. My time at
USC has been made richer by several other professors, from whom I‟ve benefitted in
numerous ways: Bruce Jansson, John Brekke, Doni Whitsett, John Bola and Robert
Nishimoto. Malinda Sampson has been great and I‟m thankful for the support from
the School of Social Work and the Provost‟s Office. I‟ve been enriched through
discussion and enjoyed the company of many students and other friends, especially
Juye Ji, Kris Stevens, Michelle Yeh, Sonya Negriff, Shari Leskowitz, Jen Green,
Matt and Anna Dowling, Gloria Kamler, Cara Pohle, Bill Monro, Julia Hernandez
and Irving Argaez. I feel an abiding sense of gratitude for Noah Levine and Shinzen
Young, whose efforts have been transformational. And, of course, my family – Greg,
Becca, Joanne and Jeff. They have been wonderful and I consider myself very
fortunate. I hope that my education may be of some benefit for others.

iii

Table of Contents
Acknowledgments ii

List of Tables v

List of Figures viii

Abstract ix

Chapter 1: Background and Significance 1
Contextualizing Comorbidity 1
Aims of Current Study 4
Importance of Studying Phenomenon of Comorbidity 6
Depression, Externalizing Behavior and their Relation during

Adolescence 14
Deficiencies in Extant Literature : Moderating Variables of Gender and

Maltreatment Affecting Previous Findings & Differing Statistical

Methodologies 22
Present Study 27

Chapter 2: Research Design and Methods 29
Procedure 29
Participants 29
Maltreatment Experience Description 31
Measures 31
Analytic Plan 35

Chapter 3: Results 50
Preliminary Analyses 50
Determining Factor Structure for Child-Rated Depression and
Externalizing Behavior 58
Determining Factor Structure for Parent-Rated Depression and
Externalizing Behavior 71
Substantive Analyses

Specific Aim 1 81
Specific Aim 2 89
Specific Aim 3 108

iv

Chapter 4: Discussion 118

Bibliography 134

v

List of Tables

Table 1. Sample Characteristics for Maltreated & Comparison Subjects at Time 1,
Time 2 and Time 3 51

Table 2. Sample Characteristics for Girls & Boys at Time 1, Time 2 and Time 3 52

Table 3. Attrition Information 55

Table 4. Descriptive Statistics for Child Report Aggression, Delinquency and
Depression 56

Table 5. Descriptive Statistics for Parent Report Aggression, Delinquency and
Depression 57

Table 6. Correlations between Time 1, Time 2 and Time 3 Outcome Variables 60

Table 7. Means and Standard Deviations for Full Sample on Children's Depression
Inventory by Age 65

Table 8. Means and Standard Deviations for Boys on Children's Depression
Inventory by Age 65

Table 9. Means and Standard Deviations for Girls on Children's Depression
Inventory by Age 65

Table 10. Correlations between Mean Scores on Children's Depression Inventory
by Age 66

Table 11. Means and Standard Deviations for Full Sample on Child-Report
Externalizing Behavior by Age 69

Table 12. Means and Standard Deviations for Boys on Child-Report
Externalizing Behavior by Age 69

Table 13. Means and Standard Deviations for Girls on Child-Report
Externalizing Behavior by Age 69

vi

Table 14. Correlations between Mean Scores on Child-Report Externalizing
Behavior by Age 70

Table 15. Means and Standard Deviations for Full Sample on CBCL Depression by
Age 75

Table 16. Means and Standard Deviations for Boys on CBCL Depression by Age 75

Table 17. Means and Standard Deviations for Girls on CBCL Depression by Age 75

Table 18. Correlations between Mean Scores on Children's Depression Inventory
by Age 76

Table 19. Means and Standard Deviations for Full Sample on CBCL
Externalizing Behavior by Age 79

Table 20. Means and Standard Deviations for Boys on CBCL Externalizing
Behavior by Age 79

Table 21. Means and Standard Deviations for Girls on CBCL Externalizing
Behavior by Age 79

Table 22. Correlations between Mean Scores on Children's Externalizing
Behavior by Age 80

Table 23. Model Fit Statistics for Multi-Group Cross-Lagged Regression Models 87

Table 24. Linear Latent Growth Curve Analyses of Child-Report Depression 94

Table 25. Effects of Predictors in Multi-Group Model of Child-Report
Depression 95

Table 26. Linear Latent Growth Curve Analyses of Child-Report Externalizing
Behavior 99

Table 27. Effects of Predictors in Multi-Group Model of Child-Report
Externalizing Behavior 100

Table 28. Effects of Predictors in Multi-Group Model of Parent-Rated
Externalizing Behavior 106
vii

Table 29. Linear Latent Growth Curve Analyses of Parent-Report Externalizing
Behavior 107

Table 30. Estimates from Univariate Latent Change Score Models for
Depression & Externalizing Behavior 109

Table 31. Estimates from Bivariate Latent Change Score Models for
Depression & Externalizing Behavior
114

Table 32. Fit for Bivariate Latent Change Score Models with and Without
Coupling Parameters 117

viii

List of Figures

Figure 1. Path Diagram of Latent Growth Model Organized by Age 43

Figure 2. Path Diagram of Univariate Latent Change Model of Depression 47

Figure 3. Path Diagram of Bivariate Latent Change Model 48

Figure 4. Parameter Estimates for Best-fitting Cross-Lagged Regression Model
with Full Sample 83

Figure 5. Parameter Estimates for Best-fitting Gender Multi-Group Cross-
Lagged Regression Model 86

Figure 6. Estimated Growth Trajectories for Child-Report Depression 93

Figure 7. Estimated Growth Trajectories for Child-Report Externalizing Behavior 98

Figure 8. Estimated Growth Trajectories for Parent-Report Externalizing
Behavior 104

Figure 9. Estimated Trajectories of Externalizing Behavior from Growth Mixture
Models 112

Figure 10. Results from Bivariate Latent Change Score Model 116

ix

Abstract
To study psychopathology is to study comorbidity. In both clinical and representative
samples, the co-occurrence of two or more disorders is the rule, rather than the exception.
While the virtual ubiquity of psychiatric comorbidity has been established, efforts have
been redirected to explaining the phenomenon. The current study investigated the
sources of covariation between depression and externalizing behavior among adolescent
boys and girls. Specifically, it examined the plausibility that one symptom cluster
exerted a causal effect on the other cluster of symptoms – a hypothesis known as
pathogenic comorbidity. This hypothesis has more support among girls, and for the
downstream effects of externalizing behavior on depression. A gender-balanced group of
454 young adolescents was sampled, with 303 of those having met specific criteria for
the recent experience of maltreatment. The youth and their caretakers were assessed
three times over the period of the next two and a half years. Pathogenic comorbidity was
examined utilizing different statistical methods, each with their own unique model of
change. Findings employing child and parent-report data were compared. The
preponderance of the evidence suggested that when controlling for risk factors associated
with both depression and externalizing behavior, no support was marshaled for the
existence of direct casual connections between the disorders in either gender. Robust
within-time associations were observed, suggesting a complex relationship. Bivariate
latent change score model findings were contrasted with a cross-lagged regression
analyses. While the traditional model lent some support for pathogenic comorbidity

x

among girls, the latent change score model suggested that finding was an artifact of the
failure to control for nonstationarity. Results are consistent with the hypotheses that
depression and externalizing behavior covary because the mechanisms of risk for one
disorder are shared for the other disorder or that the phenomenological differences
obscure a global psychopathological trait. Limitations of the present research and future
directions are discussed.

1

Chapter 1: Background and Significance
Contextualizing Comorbidity
In the assessment of psychopathology, the issue of diagnostic comorbidity has
figured prominently. Comorbidity, a term first introduced by Feinstein (1970) in a
medical context, refers to the presence of a clinically meaningful „ailment‟ that coexists
with a particular disease state. Its application to psychopathology has been controversial,
and the use of the term has not always adhered to Feinstein‟s original parameters
(Lilienfeld, 2003). The ambiguities involved in importing this construct into the field of
psychopathology have produced a construct that has been defined in several ways. The
broadest definition states that two different psychiatric diagnoses occurring at any point
during a person‟s life qualifies as comorbidity. The National Comorbidity Study
researchers have referred to this as „lifetime comorbidity‟ (Kessler, DuPont, Berglund &
Wittchen, 1999). Angold, Costello and Erkanli (1999) have distinguished lifetime
comorbidity from concurrent comorbidity. This is defined as the coterminous presence
of two or more disorders. This definition does not imply that the disorders had similar
times of onset or remission but merely co-exist at a single time point.
Simple principles of probability suggest that limited co-occurrence of mental
disorders should be witnessed. This phenomenon would be observed even if the
psychiatric nosology were perfectly valid, and there were no interrelationships between
disorders. For example, if the base rate of depression is 15%, and the base rate of panic
disorder is 2%, then we would expect that 0.3% of the population (.15 x .02) would
exhibit both depression and panic disorder. However, the rates of comorbidity have far
exceeded those expected by simple probability. In fact, the virtual ubiquity of
2

comorbidity has been well established. In large nationally representative samples, rates
of psychiatric comorbidity are consistently elevated above rates expected by chance, with
numerous disorders evidencing a lifetime co-occurrence rates greater than 50% (Kessler,
Berglund, Demler, Jin, Merikangas & Walters, 2005). Co-occurrence has been
documented within and between disorders on the internalizing and externalizing spectrum.
The internalizing spectrum broadly includes many of disorders that the Diagnostic and
Statistical Manual of Mental Disorders (DSM) groups into mood and anxiety disorders.
Depression, dysthymia, panic, phobias and generalized anxiety disorder evidence strong
interrelationships (Krueger, 1999; Kendler, Prescott, Myers & Neale, 2003). The
disorders comprising the externalizing spectrum – antisocial behavior, drug and alcohol
dependence – evidence similarly strong relationships (Krueger, Hicks, Patrick, Carslon,
Iacono & McGue, 2002). Although more moderate than the relations within the
internalizing spectrum and within the externalizing spectrum, substantial association is
found between internalizing and externalizing dimensions (Krueger & Markon, 2006).
The findings from representative samples have been replicated in clinical samples, where
rates of comorbidity are elevated beyond the levels seen in community samples (Lahey,
Loeber, Burke, Rathouz & McBurnet, 2002).
The phenomenon is not merely limited to one developmental phase but is
evidenced across the lifespan. In early childhood, relations between certain symptom
clusters can be so highly interrelated that they cannot be statistically distinguished. For
example, Cole, Peeke and Truglio (1997) found that among 3
rd
grade children, depression
and anxiety effectively represent a unified construct. Developmental theories suggest
that greater differentiation will emerge with age (Sroufe, 1979). With greater physical,
3

affective and cognitive capacities, we can expect greater specificity in the types of
symptoms exhibited. While Cole et al. (1997) identified some evidence for this
differentiation, comorbidity is clearly not limited to developmental phases when
cognitive and affective faculties are underdeveloped. In adolescent (Krueger, Caspi,
Moffitt & Silva, 1998) and adult samples (Krueger, 1999; Kessler et al. 2005),
comorbidity is the rule rather than the exception. A meta-analysis conducted by Angold,
Costello and Erkanli (1999) finds the existence of a single psychiatric diagnosis among
adolescents confers substantial risk for a comorbid condition with odds ratios ranging
from three to eleven, depending on the pair of diagnoses examined.
Among common disorders, the development of depression and externalizing
behavior figure prominently. After puberty, depression exerts a significant toll on the
adolescent population, especially on girls (Goodyer, 2008). The chronic nature of
depression is increasingly clear as substantial percentages of patients relapse or show
significant symptoms even following an adequate course of treatment (Hollon, DeRubeis,
Shelton, Amsterdam, Salomon, & O‟Reardon et al. 2005). Depression exerts a
staggering effect on health care costs and loss of productivity due to disability (Austrian,
2005). Externalizing behavior is broadly encompassed by aggression, delinquency,
oppositional behavior, risky sexual behavior and substance abuse. These behaviors
warrant immediate attention as they are both prevalent and feature high levels of
functional impairment (Loeber, Burke, Lahey, Winters & Zera, 2000). With prevalence
estimates ranging from 1 to 16%, the impact is notable as a substantial percentage of
conduct-disordered youth will progress to antisocial behavior as adults (Myers, Stewart &
Brown, 1998). Although not completely recalcitrant to intervention, conduct-disordered
4

individuals require comprehensive and costly intervention (Burke, Loeber & Birmaher,
2002).
For two decades, the co-occurrence of depression and externalizing behavior has
been noted. Several reviews and meta-analyses attest to the robust relation between these
two clusters of symptoms (Angold & Costello, 1993; Kennan & Loeber, 1994; Angold,
Costello & Erkanli, 1999). The robustness of this relation is virtually unquestioned and
research focus has shifted towards more nuanced developmental models of the dynamic
relationship between these two clusters of symptoms.

Aims of Current Study
Although some progress has been made in charting the developmental relations of
depression and externalizing behavior, methodologically rigorous studies are uncommon.
While boys and girls likely feature distinctive trajectories of co-occurring depression and
externalizing behavior, boys have been disproportionately featured in longitudinal studies
of conduct problems. Explicit inclusion of childhood maltreatment – a risk factor for
both depression and externalizing behavior and thus a potential confounding variable in
comorbidity studies – has been notably lacking. This dissertation aimed to build and
extend previous research in mapping and explaining the frequent co-occurrence of these
two forms of psychopathology. Specifically, it examined the plausibility of pathogenic
comorbidity – one disorder catalyzing another – separately, for boys and girls. Further,
the study contrasted the implications from different statistical methodologies that have
been utilized to assess pathogenic comorbidity. The study has three specific aims.

5

Specific Aim #1: Cross-Lagged Regression Models of Depression and Externalizing
Behavior Contrasted with Latent Change Score Models

1A. Using traditional cross-lagged analytic strategies, determine the
longitudinal relationship between depression and externalizing behavior for
the full sample of girls and boys

1B. Using multi-group analyses, determine if gender moderates the
relationship between depression and externalizing behavior

Specific Aim #2: Predicting the Trajectory of Depression and Externalizing Behavior

2A. Characterize the trajectories of depression and externalizing behavior
separately for boys and girls with unconditional latent growth curve analyses

2B. Determine if baseline depression predicts level or growth of externalizing
behavior, before and after accounting for common risk factors (peer
delinquency, child maltreatment, parental depression, harsh parenting)

2C. Determine if the baseline externalizing behavior predicts level or growth
of externalizing behavior, before and after accounting for common risk
factors (peer delinquency, child maltreatment, parental depression, harsh
parenting)

2D. Assess the moderating role of gender in the longitudinal association of
depression and externalizing behavior

Specific Aim #3: Dynamic Modeling of Latent Change Scores

3A. Characterize the course of depression and externalizing behavior
separately with univariate latent change modeling

3B. Determine the dynamic relationships between symptom change and
previous psychological status with bivariate latent change score models

3C. Contrast these model findings with those of the cross-lagged analysis

6

Importance of Studying Phenomenon of Comorbidity
Psychiatric Nosology & Comorbidity
The psychiatric classification system – codified in the Diagnostic and Statistical
Manual of Mental Disorders (DSM) – has exerted profound impact on the public‟s
understanding of mental disorder, the program of behavioral research, and treatment and
prevention efforts. Undoubtedly, the formulation and refinement of the nosology has
offered critical advantages. Adherence to this classification system permits
generalizability of research findings and offers the possibility of synthesizing research on
psychopathology and the principles through which we can intervene. Nevertheless,
evidence challenging the empirical integrity of the DSM has emerged (Krueger, 1999;
Helzer, Kraemer & Krueger, 2006; Clark, Watson & Reynolds, 1995). This evidence has
taken different forms. Some evidence suggests that dimensional, rather than categorical
distinctions, better fit the data, at least for some disorders (Flett, Vredenburg, & Krames,
1997; Markon & Krueger, 2005). The DSM has also been challenged on the grounds that
diagnostic definitions permit extreme heterogeneity within categories. The polythetic
format of the criteria – diagnosis is met, for example, when 5 of 9 criteria are met –
concedes a substantial degree of diagnostic heterogeneity (Skodol, Gudnerson, Pfohl,
Widiger, Livesley, & Siever, 2002). The pervasive evidence of co-occurrence of two (or
more) apparently distinct diagnoses – the subject of the current research – has also
challenged us to rethink the current classification system. In evaluating the implications
of comorbidity for our thinking about psychopathology, it is first important acknowledge
the profound influences of our psychiatric nosology on the field of mental health.
7

Studies of etiological pathways for particular disorders have been of critical
interest in the field of psychopathology. This is logical, as the identification of causal
factors facilitates the development of interventions that effectively interrupt problematic
developmental processes. Studies on the etiology of mental disorder have typically been
guided by traditional psychiatric categories of disorder. Research efforts have commonly
sought links between particular environmental or biological vulnerability factors and
specific mental disorders. Retrospective designs commonly begin with a clinical sample
defined by the presence of a DSM diagnosis and link particular characteristics of the
individual‟s history with their current status. Prospective designs assess risk
longitudinally, and then, similarly seek to differentiate the individuals who develop a
psychiatric condition from those who do not. In both designs, psychiatric nosology
exerts a profound effect. The findings of linkages between risk factors and the outcome
are effectively constrained by the validity of the classification system. If the validity of
the DSM is compromised, we would expect conflicting or null findings in the effort to
link risks with particular psychopathologies.
Mental health intervention research has also been shaped by adherence to the
traditional psychiatric nosology. In recent years, clinical researchers have sought to
identify, develop and disseminate empirically supported treatments. Emerging largely
from the evidence-based medicine paradigm (Sackett, Richardson, Rosenberg & Haynes,
1997), „evidence-based practice‟ within the mental health field suggests that client
outcomes can be maximized by basing clinical decisions on the best available evidence.
Accordingly, the research program has sought to identify the particular therapeutic
techniques and other factors, such as the therapeutic alliance, that predict outcomes for a
8

range of defined problems. This research has largely modeled itself around the
biomedical model which posits that discrete disease entities are to be addressed with
specific courses of treatment. Accordingly, the nation‟s program of psychotherapy
research has relied primarily on DSM diagnoses to define the treatment population
(Westen, Novotny, & Thompson-Brenner, 2004). The population for a study sample has
typically been defined as the presence of DSM diagnosis. Consequently, evidence-based
practice has been built on the foundation of the DSM – treatment manuals typically
feature a particular disorder as the target. The dissemination of treatment manuals based
on a focal diagnostic category is a central feature of the evidence-based practice
movement. Dissemination of evidence-based practice has exerted broad impact on
graduate education and training programs, professional licensing standards, and
requirements for reimbursable mental health care. If the validity of the DSM is
compromised, current treatment guidelines may not be optimal.
Given its profound role in shaping psychopathology research, and the DSM‟s
influence on a network of institutions involved in training and delivering mental health
services, tracing the implications of comorbidity becomes critical. Of course, the
existence of comorbidity does not necessarily imply that the validity of the DSM is
compromised. Recall that the DSM itself implicitly acknowledges the clustering
tendency of psychopathological symptoms. The organization into mood, anxiety, and
psychotic disorders (among twelve others) suggests some higher-order relationships
between disorders exist within the same category. This framework implies that some
comorbidity should be expected, as disorders within categories are related through their
membership in the higher-order category. Nevertheless, the current nosology implies that
9

the most important distinctions are made at the level of the disorder, not at the level of the
higher-order category.
Thus, studies of comorbidity are important as they can suggest a more empirically
viable and parsimonious nosology. Given the DSM‟s global impact on mental health
treatment, professional training and policy funding, such studies represent a worthy
research goal.
Different Explanations of Comorbidity Have Different Implications for Intervention
While the implications for the psychiatric nosology are important, understanding
comorbidity has direct implications for understanding etiology, the course of a syndrome,
prevention and treatment (Biederman, Faraone, Mick & Lelon, 1995). Temporarily
placing aside the importance of a valid diagnostic system, it‟s clear that different
mechanisms explaining comorbidity carry distinct implications for intervention and
research programs. In their seminal paper, Klein and Riso (1993) trace eleven possible
explanations of comorbidity. Among the bivariate models, five suggest that comorbidity
is merely an artifact of sampling design or diagnostic fuzziness. These possibilities are
unsupported by the available evidence (Angold et al., 1999). Other models include the
possibility of shared or associated liabilities. In this case, risk factors for one disorder
might be correlated with the risk factors for another disorder. Alternatively, a single risk
factor – maternal depression, for example – might act as a causal factor for a range of
disorders.
Multiformity models of comorbidity suggest that phenomenologically diverse
symptoms may share a common underlying structure. Substantial evidence has emerged
for this model. In a series of papers, Krueger and colleagues have argued that the
10

pervasive nature of comorbidity is neither artifactual nor a nuisance (Krueger, 1999;
Krueger & Markon, 2006). Instead, it reflects the natural consequence of the latent
structure of psychopathology. From this perspective, what unifies disorders may be more
fundamental than their phenomenological differences. This hypothesis draws from both
the multiformity and shared liability models. Through factor analytic studies of
diagnostic data on a range of anxiety, mood and externalizing disorders, several groups
have established the relevance of higher-order factors in the explanation of comorbidity
of DSM disorders (Krueger, 1999; Vollebergh Iedema, Bijl, Graaf, Smit & Ormel, 2001;
Slade & Watson, 2006). Kreuger (1999) utilized data from the National Comorbidity
Survey (Kessler et al., 1994) to model the structure of common mental disorders and
identified three underlying dimensions that best accounted for the variation of major
depression, dysthymia, generalized anxiety disorder, panic disorder, social phobia,
specific phobia, agoraphobia, alcohol dependence, drug dependence, and antisocial
personality disorder. All mood and anxiety disorders loaded on one of two correlated
latent factors identified as “anxious-misery” and “fear.” The drug and alcohol diagnoses
and the antisocial personality disorder loaded on the externalizing factor, which was
moderately correlated with the internalizing factor. Subsequently, Vollebergh et al.
(2001) sought to replicate Kreuger‟s (1999) findings in a large representative sample of
the Dutch population. Results were strikingly similar to those reached in the American
sample. Vollebergh et al. (2001) reached the same factor structure, and extended the
previous analysis by demonstrating high longitudinal stability of this structure over the
period of one year. Slade and Watson (2006) incorporated the International
11

Classification of Diseases-10 diagnoses into the same analytic strategy and reached a
very similar solution in an Australian sample.
In contrast, causal explanations of comorbidity suggest that one disorder acts as a
direct risk factor for the development of a second disorder. More fully developed later,
causal explanations posit that comorbidity arises as a function of the cascading effects of
impairments created by one disorder.
These brief descriptions of explanatory models of comorbidity suggest that
psychosocial intervention strategies would vary dramatically depending on the model of
comorbidity assumed. Psychological interventions are designed to target causally
relevant processes of psychopathology (Rosen & Davison, 2003). Diagnostic
formulations serve to name the nature of the dysfunction that will be targeted by
treatment. However, different models of comorbidity imply very different
understandings of the problem itself. For example, Krueger‟s model suggests that core
psychopathological processes underlie a diverse range of disorders. Researchers have not
fully traced the implications of this model for treatment, but it is clear that traditional
approaches might be modified. While some disorder-specific characteristics of the
individual could be addressed, the main therapeutic thrust would address the core
psychopathological vulnerability. Kreuger (1999), following Watson and Clark (1992)
hypothesizes that the internalizing cluster of disorders might be unified by negative
emotionality. Additional support for Kreuger‟s model might encourage the development
of treatments that address this core vulnerability, rather than focusing exclusively on the
symptomatic differences. Thus treatments, for example, of depression and generalized
anxiety disorder, might begin to converge.
12

In contrast, causal models suggest different avenues for treatment development.
In this case, one disorder compromises functioning and catalyzes development of a
second disorder. Identifying the processes that amplify vulnerabilities and lead to new
domains of dysfunction becomes critical. For example, if it were determined that
conduct disorder exerts a causal impact on depression, the strategic target for intervention
would be clear. Interrupting this developmental cascade would prove valuable in both
addressing conduct problems and decreasing the likelihood that depressive symptoms
would emerge. Additional research is needed in order to assess the plausibility of
different comorbidity models and trace the implications for intervention.
Multiple Explanatory Mechanisms of Comorbidity
Although rarely acknowledged, it is possible, if not likely, that several
explanatory factors operate to explain comorbidity. Investigations have traditionally
sought to indentify singular explanations of the phenomenon. However, the models of
comorbidity are not mutually exclusive. Evidence from developmental psychopathology
suggests that the emergence of maladaptive behavior arises from complex networks of
interacting systems (Cicchetti and Rogosch, 2002). The complexity of etiology is
acknowledged for the emergence of single disorders. Depression, for example, is no
longer explained by reference to a simple main effect of stress (Cole, Nolen-Hoeksema,
Girgus & Paul, 2006) or a straightforward genetic effect (Kendler, 2005). However, the
explanation of comorbidity has tended to focus on a single explanatory mechanism. It is
important, therefore, to evaluate the possible roles of various processes which may lead
to comorbidity. Rather than seeking the definitive explanation, research should seek to
understand the plausibility of multiple mechanisms of comorbidity. Such a framework
13

makes some of the conflicting findings interpretable. Causal mechanisms, for example,
can co-exist with explanations based on associated liabilities. Thus, research can proceed
without the agenda of establishing a singular explanation for comorbidity but instead
highlight diverse processes that may collectively explain the phenomenon.
Categorical vs. Dimensional Perspectives on ‘Comorbidity’
Any discussion of co-occurring symptoms requires explanation of dimensional
and categorical assessments of psychopathology. Most investigations of the co-
occurrence of diagnoses have employed categorical models of psychopathology where
only one distinction is made: the presence or absence of the disorder. Although the
construct of comorbidity might appear to provide a conceptual framework to understand
co-occurrence of disorders, the validity of the construct in psychopathology research has
been questioned. Lilienfeld, Waldman and Israel (1994) argue that the limited
understanding of the etiology and pathology of mental disorders compromises the valid
use of the term. The construct implicitly presupposes valid disease entities underlying
symptom patterns despite the fact that the investigation of the taxonicity of mental
disorder is ongoing and offers substantial evidence for dimensional perspectives (Watson,
2005). Dimensional perspectives on psychopathology, which consider
psychopathological traits as continuously distributed in the population, conceptualize
„comorbidity‟ differently. From this perspective, co-occurring symptom clusters are said
to represent symptom or diagnostic or symptom covariation. While Neale and Kendler
(1995) have pioneered modeling strategies for categorical models of risk, they
acknowledge that dimensional models can be profitably employed and have suggested
examining covariation between latent factors representing psychopathological traits. This
14

approach may afford significant advantages. Consequently, the analytic strategy
proposed will adhere to the construct of diagnostic covariation. In the discussion of
background material, the choice to employ “comorbidity” or “diagnostic covariation”
will be made on the basis of the literature being discussed. Although each term implies a
different model of psychopathology, the phenomenon depicted by both terms is
conceptually similar.

Depression, Externalizing Behavior and their Relation during Adolescence

Adolescence as a critical period for study of psychopathology
Developmentalists have characterized adolescence as a period of massive physical,
cognitive and psychological reorganization (Allen & Land, 1999). This period is notable
as it marks a phase of personality consolidation (Bosma & Gerlsma, 2003). Adolescence
may serve as a „critical period‟ where underlying vulnerabilities intersect with
environmental stressors to yield psychopathological outcomes in adulthood (Allen &
Land, 1999). Research investigating the etiology of psychological disorders has found
that adolescence represents a sensitive period for the development of addiction
(Chambers, Taylor & Potenza, 2003), behavioral disorders (Hofstra, van der Ende &
Verhulst, 2002) personality disorders (Reich & Zanarini, 2001) mood disorders and
suicidality (Levy & Deykin, 1989). Adolescence is commonly the period of onset for
numerous disorders including eating disorders, mood disorders, anxiety disorders and
schizophrenia (American Psychiatric Association, 2000). Given these data, adolescence
serves as an important phase for studying developmental mechanisms that can point
towards more efficient targets of intervention.
15

Depression and Externalizing Behavior: Extent of their Relationship
As previously noted, depression and externalizing conditions are prevalent,
feature high levels of functional impairment and deleterious developmental consequences.
At a superficial level, we might expect the marked symptom differences of depression
and externalizing behavior to yield especially low levels of diagnostic comorbidity
(Wolff & Ollendick, 2006). Whereas the diagnostic descriptions between depression and
anxiety, for example, are similar, depression and externalizing behavior apparently
represent different poles on a spectrum of disinhibition. Depression is characterized by
failures of behavioral activation while externalizing disorders are characterized by the
failure to inhibit behavior. Nevertheless, rates of comorbidity are consistently elevated
above chance (Angold et al. 1999; Wolff & Ollendick, 2006).
In one of the first papers devoted to depressive comorbidity, Angold and Costello
(1993) surveyed the available epidemiological studies. In each study reviewed, they
found significant associations between depression and externalizing behavior. Among
children and adolescents with depression, the rate of comorbidity with conduct disorder
or oppositional defiant disorder ranged from 22 to 83%. Of those with oppositional
defiant disorder or conduct disorder, between 9 and 45% also met criteria for depression.
In a clinical sample of adolescent boys, Lahey et al. (2002) found an average correlation
of .35 between symptoms of the two disorders across six waves of data. Greene,
Biederman, Zerwas, Monuteaux, Goring and Faraone (2002) suggested that between one
third and one half of a clinical sample with oppositional defiant disorder met criteria for
„severe‟ major depression – a major depressive episode with marked impairment. A
meta-analysis of general population studies conducted by Angold, Costello and Erkanli
16

(1999) found that the existence of depression or conduct disorder in youth confers
substantial risk for the other disorder with a median odds ratios of 6.6. Ninety-five
percent confidence intervals ranged from 4.4 to 11. Kennan and Loeber (1994) reviewed
a range of studies and found that comorbidity is consistently elevated in both clinical and
community samples (Zoccolillo and Rogers, 1991; Fergusson, Homwood, & Lynskey,
1993; Bird, Canino, Rubio-Stipec, Gould, Ribera, Sesman, et al., 1988; Feehan, McGee,
Raja, &Williams, 1994). Robust relations between depression and externalizing behavior
are virtually unquestioned and research focus has shifted from documenting the
phenomenon to its explanation.
Explaining the Relationship between Depression and Externalizing Behavior: Causal
Explanations

Most broadly, the relationship between depression and externalizing behavior can
be characterized as causal or non-causal. A causal explanation would require that (1) one
set of symptoms precedes the other, (2) that there is an association between the symptom
clusters and (3) that this association is non-spurious. The causal explanation for the
phenomenon has been termed pathogenic comorbidity (Mineka, Watson, & Clark, 1998).
This hypothesis suggests that depression, or externalizing behavior, acts directly as a
causal factor for the other disorder. Of course, the relationship may be meditated by a
complex circuit of causal mechanisms, but the most stringent version of the pathogenic
comorbidity hypothesis implies that the second disorder would not emerge in the absence
of the first. There is some theoretical precedent to suggest such a scenario, particularly
for the causal effect of externalizing behavior on depression. Externalizing behavior is
associated with various functional impairments including academic underachievement
17

(Hinshaw, 1992; Masten, Roisman, Long, Burt, Obradovic, Riley, Boelcke-Stennes &
Tellegen, 2003) and lower social competence (Renouf, Kovacs, & Mukerji, 1997) and
peer rejection (Dodge, 1983). These developmental impairments are, in turn, associated
as predictors of depression. These observations led Capaldi (1991, 1992) to formulate
the „failure model,‟ which posits that conduct problems compromise the development of
competencies, and this in turn, generates depressive symptoms.
Several studies have documented the existence of externalizing behavior
preceding the emergence of depression. Rhode, Lewinsohn, and Seely (1991) found that
for individuals with co-occurring depression and a disruptive behavioral disorder, the
disruptive behavior was more likely to emerge first. This finding was echoed by
Biederman, Faraone, Mick and Leon (1995) in their clinical sample. In a recent review
of these issues, Wolff and Ollendick (2006) conclude that depression most commonly
follows the emergence of conduct problems, rather than the reverse.
Temporal precedence alone does not establish causality. Longitudinal
associations between the two disorders have been probed in a number of studies. The
preponderance of the evidence suggests that if there is a causal relationship, it is conduct
problems exerting the impact, but as we will see, ambiguities exist in the data. In an
important early contribution to the comorbidity literature, Capaldi (1991, 1992) suggested
that conduct problems were prospectively related to depressive symptoms. In this model,
poor peer and parental relationships mediate the relation between conduct problems and
depression. Lahey et al. (2002), in an implicit replication of Capaldi‟s findings with
adolescent boys, assessed the prognostic power of conduct disorder over the course of
seven annual assessments of psychopathology. Analyses revealed that initial conduct
18

problems predicted depression at the six subsequent assessments, when controlling for
baseline depression. The reverse was not the case. However, these prospective relations
were not specific to depression – conduct disorder presaged anxiety and attention-deficit
symptoms. Further, the preponderance of the evidence suggested concurrent escalations
of symptoms rather than the failure model suggested by Capaldi (1991). In a subsequent
analysis with the same sample but more waves of data, Burke, Loeber, Lahey and
Rathouz (2005) explicitly tested the failure model. They first identified prospective
relations between conduct problems and depression. However, when including
psychosocial impairment as a covariate in the general estimating equation regressions,
conduct disorder was no longer predictive of depression at the subsequent time wave.
Little and Garber (2005) assessed mediational models linking externalizing
behavior and depression. They hypothesized that externalizing behavior, defined by
Achenbach‟s (1991) broadband cluster of symptoms, generates dependent social stressors.
Dependent social stress refers to stressful life events that are catalyzed by the individuals
own behavior (i.e. interpersonal conflicts). In a gender-balanced sample, they found that
the prospective relationship between externalizing behavior and depression was partially
mediated by dependent stressors. Independent stressors – those entirely unrelated to the
behavior of the individual – did not mediate the relationship. In a similar analysis, Kim,
Conger, Elder and Lorenz (2003) tested reciprocal influence models whereby
internalizing and externalizing behavior are exacerbated by stressful life events but also
increase exposure to stressful events. Although the modeling strategy did not assess the
relationship between the two symptom domains, results are consistent with the work of
Little and Garber (2005). Masten et al. (2003) assessed a sample at 7, 10, and 20 years.
19

They tested a model, consistent with the failure model, where externalizing problems in
childhood undermined academic success during adolescence, and in turn predicted
internalizing symptoms (operationalized as depression, anxiety and somatization) in
early adulthood. Data provided evidence for this „developmental cascade‟, particularly
among girls. Interestingly, there was no evidence for a similar causal effect of
internalizing behavior. Internalizing behavior was either unrelated or inversely related to
subsequent externalizing behavior. In an important contribution to the literature on girls,
Measelle, Stice and Hogansen (2006) assessed developmental trajectories of several
symptom domains including depression and antisocial behavior in a sample of 493
female adolescents. They found that initial antisocial symptoms were associated with
both the initial depression level, but also the escalation of depression over a period of five
years. Wiesner (2003) assessed depression and delinquency using a four-wave latent-
variable cross-lagged design in a mixed-gender sample of 15 and 16 year old adolescents.
Delinquency, operationalized as violence, theft and property destruction, was predictive
of subsequent depression in one of three crossed paths for boys and two of three paths for
girls.
Although the preponderance of the evidence suggests externalizing behavior as a
predictor of subsequent depression, the data are not consistent. Wolff and Ollendick
(2006) acknowledge that depression may exert a direct effect on conduct problems –
although the theoretical basis is less clear. It is conceivable that the hopelessness
associated with depression might diminish the deterrent effect of punishments associated
with delinquency. Further, the acting out behavior could serve a self-regulatory
mechanism, whereby the excitement of externalizing behavior counteracts the vegetative
20

aspects of depression. Several recent studies have marshaled evidence of this
relationship. In a male sample of middle adolescents, Beyers and Loeber (2003) assessed
concurrent and prospective relations of delinquency variety and depression. A valuable
feature of these analyses is the inclusion of risk factors in the models that are common to
both depression and delinquency. They found that even after controlling for the effects
of poor parent-adolescent communication, family SES, peer delinquency, low academic
achievement and aggression, concurrent and longitudinal effects of depression on
delinquency were found. Similar effects were found for the effects of delinquency on
depression, although the results were marginally significant. Wiesner (2003) found
bidirectional influence among girls, though not boys. Similarly, Measelle et al. (2006)
identified the same bidirectional effects in a female sample. In a sample of boys with
ADHD, depression was related to subsequent conduct disorder, but the reverse was not
the case (Drabick, Gadow & Spafkin, 2006).
Thus, the evidence cited suggests that a causal role for externalizing behavior on
depression is plausible, and perhaps more robust among girls. Although less consistent,
some evidence suggests a similar causal effect of depression on externalizing behavior.
Nevertheless, it is important to consider non-causal explanations for these data.
Explaining the Relationship between Depression and Externalizing Behavior: Non-causal
Explanations

The foregoing evidence suggests that depression and externalizing behavior are
prospectively related. Even when risk factors common to both disorders are included in
models, the relation persists. However, from these data, we cannot conclude that a causal
relation exists. Significant evidence suggests that the relation between different disorders
21

is a function of unmeasured third variables, and thus the relation is spurious. Fergusson,
Lynsky and Horwood (1996) tested a series of structural equation models to determine if
risk factors common to depressive disorders and disruptive behavior disorders explain
their longitudinal association. Common risk factors included in the models were
affiliation with delinquent peers, poor parental attachment, stressful life events, early
conduct problems, deficient cognitive ability, family conflict, family history of offending,
and gender. They found that the substantial correlation between affective and conduct
symptoms was explained – approximately two-thirds of the covariation – by shared or
correlated risk factors for both symptom domains. There was no evidence of
unidirectional or reciprocal causal impacts. Genetically informative designs also present
some evidence suggesting non-causal relations. Twin studies, for example, have made
important contributions by modeling the covariation between disorders as a function of
additive genetic, shared environmental and non-shared environmental factors (Neale &
Kendler, 1995). Results from such studies suggest genetic influences exert a powerful, if
not decisive effect on symptom covariation. In a key study with an adult twin population,
Kendler, Prescott, Myers, and Neale (2003) assessed genetic and environmental factors
accounting for seven internalizing and externalizing disorders. The weak pattern of
cross-loadings for the shared and unique environmental internalizing and externalizing
factors was interpreted as evidence that genetic factors best account for covariation
between internalizing and externalizing dimensions. In a sample of adolescents,
O‟Connor, McGuire, Reiss, Hetherington and Plomin (1998) investigated genetic
mediation of the covariation between depression and antisocial symptoms. They found
that approximately 45% of the covariation between symptoms could be explained by a
22

common genetic liability. Gjone and Stevenson (1997) found some evidence for genetic
explanation of covariation, although environmental factors figured more prominently
than in the results of O‟Conner et al. (1998). The sample in this second study, featured
younger participants and genetic influences on comorbidity may increase with age as
individuals exert increasing effects on the selection of their environments (O‟Conner,
Neiderhiser, Reiss, Hetherington & Plomin 1998). One final line of evidence militating
against causal explanations is data reviewed earlier, from Krueger and colleagues
(Krueger, 1999; Markon & Krueger, 2005; Krueger & Markon, 2006). Through
confirmatory factor studies, they advanced evidence for a shared-liability model, whereby
comorbidity is a natural consequence of the latent structure of psychopathology.
According to this explanation, comorbidity is largely –if not entirely – explained by the
higher order psychopathological process factors that underlie diverse symptom
presentations. A correlation of .50 between internalizing and externalizing factors was
derived in a meta-analysis of relevant data (Krueger & Markon, 2006). This factor
structure could be reparameterized to include a third-order factor, representing a global
psychopathological trait. Along these lines, Khan, Jacobson, Gardner, Prescott and
Kendler (2005) have marshaled evidence that the personality dimension of neuroticism
figures prominently in explaining general vulnerability to both internalizing and
externalizing distress.

Deficiencies in Extant Literature: Moderating Variables of Gender and Maltreatment
Affecting Previous Findings & Differing Statistical Methodologies

23

Efforts have generally been directed towards finding a single explanation of
comorbidity that fits the population universally. Although not commonly investigated, it
is possible that a third variable affects the processes by which depression and
externalizing behavior are related. This would represent a case of moderated mediation,
whereby the nature of causal mechanisms varies across group variables. Rutter, Caspi
and Moffitt (2003) provide a comprehensive framework for investigating and interpreting
gender differences. In addition to the genetic implications of being male or female,
though not entirely separable from genetic conditioning, the authors suggest that gender
may alter the matrix of risk and protective factors. Boys and girls may differ in their
level of exposure to particular supportive or destructive environments as a consequence
of risk-taking, novelty-seeking or harm-avoidance tendencies. The genders may also
vary in their capacity to effectively cope with environmental insults. Boys and girls may
feature particular personality characteristics that singularly account for psychopathology
or interact with environmental stressors to exacerbate or attenuate psychopathology.
Some evidence suggests that gender may moderate the relationship between
depression and externalizing behavior. It is plausible that the pathogenic comorbidity
hypothesis is more tenable for girls than for boys. Many of the studies reviewed have
featured single-gender samples, and direct comparisons between genders are not
common. Several studies have documented conduct problems as an important prognostic
indicator among girls. In their review of disruptive behavior disorders, Loeber, Burke,
Lahey, Winters and Zera (2000) identify a „gender paradox,‟ whereby individuals with
gender-atypical diagnoses are at special risk for the development of comorbid disorders.
Thus, girls, who feature lower levels of prevalence for conduct problems, are at
24

especially high risk for other psychiatric complications when conduct problems are
present (Loeber & Keenan, 1994). Several studies provide evidence for the prospective
relationship between externalizing behavior and depression in girls. In the „cascade‟
analysis mentioned above, gender difference tests revealed more concurrent and
prospective association between externalizing, internalizing behavior and academic
competence among girls (Masten et al., 2003). In a large longitudinal study of the
development of psychiatric disorder in North Carolina, Costello, Mujstillo, Erkanli,
Keeler and Angold (2003) found strong evidence of greater heterotypic comorbidity
among girls, where one disorder predicts onset of a second disorder, even after
controlling for other comorbidities. Wiesner (2003) investigated relationships between
delinquency and depression using a 4-wave cross-lagged regression analysis. Although
only controlling for parental income and child age, she found much stronger reciprocal
effects among girls. In a cross-sectional study, Rowe, Maughan and Eley (2007) found
the similar mediational relationships between externalizing behavior and depression for
boys and girls, but the authors lament the dearth of extant data and encourage further
exploration of the moderating role of gender. In sum, the evidence suggests that girls
may suffer more profound consequences subsequent to the development of externalizing
behavior. Given the paucity of data regarding the developmental course of depression and
externalizing behavior in girls, the question of the moderating effect of gender remains
important.
Importance of Integrating Maltreatment into Comorbidity Studies
In comorbidity studies, the inclusion of variables that serve as common risk
factors for the target conditions is important. Studies that fail to include such variables
25

may reach biased conclusions. For example, causal relationships between two disorders
might be spuriously inferred when common risk factors are excluded from models
(Fergusson et al., 1996). Heim, Plotsky and Nemeroff (2004) reviewed evidence linking
early life stress – typically child maltreatment – to neurobiological effects that sensitize
individuals to the development of mood disorders. Their review suggests biologically
distinct subgroups of depression, some of which are distinguished by the presence or
absence of early life stress. They offer a serious caution: “Depending on the distribution
of early life stress across patient and control groups, previous findings on the
neurobiology of depression might be significantly confounded” (Heim et al., 2004, p.
644). Similar conclusions regarding behavioral data are justified.
The experience of maltreatment is relevant for the study of both depression and
externalizing behavior. High rates of co-occurring symptoms have been documented in
maltreated samples (Cicchetti & Toth, 2005) and existing evidence suggests that
maltreatment represents a risk factor for both depression (Weiss, Longhurst, & Mazure,
1999) and externalizing behavior (Smith & Thornberry, 1995; McLeer, Callaghan, Henry
& Wallen, 1994). Several studies demonstrate the salience of maltreatment for studies of
the comorbidity of depression and externalizing behavior. Keiley, Lofthouse, Bates,
Dodge and Pettit (2003), using a methodology borrowed from multitrait-multimethod
approaches, created orthogonal internalizing, externalizing, and covarying factors in a
longitudinal study of non-clinical adolescents. Noting that their finding likely represents
the first of its kind, harsh punishment was related not only to externalizing behavior, but
to covariation between externalizing and internalizing behavior. Another method
employed to study comorbidity is the use of „pure‟ versus comorbid cases. In this
26

approach, correlates of single-disorder subjects are compared with the correlates of those
with comorbid symptomatology. Simic and Fombonne (2001) find that children with
depression and conduct disorder are distinguished from children with conduct disorder by
their exposure to recent physical abuse, and a history of sexual abuse. High parental
hostility poses increased risk for co-occurring depressive and behavioral symptoms as
compared with either symptom individually (Ge, Best, Conger, & Simons, 1996). As
compared with depressed adolescents, a history of physical or sexual abuse differentiated
the group with co-occurring depression and comorbid conduct problems (Meller &
Borchardt, 1996; Withbeck, Hoyt & Bao, 2000). Widom, DuMont and Czaja (2007)
followed a cohort of children into adulthood, assessing depression and comorbid
conditions. As might be expected, maltreatment victims showed earlier onset and higher
prevalence of adult depression, as compared with a demographically matched non-
maltreated group. More relevant however, is the finding that among those with
depression, individuals with maltreatment histories were more likely to present with a
current comorbid diagnosis and to have had more lifetime episodes of other disorders.
Thus, evidence suggests that maltreatment serves as a common risk factor for
depression and externalizing behavior. Maltreatment experience may be a confounding
variable in studies of psychopathology. These findings strongly encourage incorporating
measurement of maltreatment into comorbidity research as its exclusion may lead to
spurious conclusions about causal connections between depression and externalizing
behavior.
Differing Statistical Methodologies Employed in Comorbidity Studies
27

Aside from the substantive concerns that warrant inclusion in comorbidity studies,
methodological strategies must be considered in interpreting findings. Different
statistical methods for the analysis of longitudinal comorbidity data have been employed.
Examples include models suited for highly skewed data such as generalized estimating
equations and hierarchical generalized linear modeling employing Poisson models. A
variety of structural equation models have also been employed, including cross-lagged
regression models, latent growth curve models, and certain growth mixture models. One
precedent for latent change models applied to comorbidity data has been identified (King,
King, McArdle, Doron-LaMarca, 2009). It is important to note that different statistical
methods depict different underlying models of change (Ferrer & McArdle, 2003).
Parameters from differing models need to be interpreted carefully when being compared
with the findings from the extant literature. Some of the conflicting findings from the
body of research may be related to different methodologies utilized. This dissertation
examined different models of growth and change and examined the convergence of
findings.

Present Study
The present study is an investigation of the comorbidity of depression and
externalizing behavior in adolescence. As discussed above, the implications of
comorbidity are important for a more empirically sound classification of
psychopathology and for refining intervention strategies. The present investigation fills
in gaps in literature described above by mapping the developmental directions of
depression and externalizing in a gender-balanced sample where maltreatment experience
28

is carefully operationalized. The hypothesis of pathogenic comorbidity will be assessed
for boys and girls as evidence suggests that gender may moderate this association. The
inclusion of maltreatment, a risk factor for both depression and externalizing behavior is
critical in examining the possibility that the association between the two symptom
clusters is spurious. Further, findings from different statistical methods were examined
for points of convergence and difference. Such findings may be an important component
in explaining conflicting findings in the extant literature.

29

Chapter 2: Research Design and Methods
Procedure
Following approval from the University of Southern California Institutional
Review Board and the Los Angeles County Juvenile Court, a collaborative effort with
Los Angeles County Department of Child and Family Services (DCFS) was undertaken
to acquire a sample of maltreated adolescents. With LADCFS and University Park
Institutional Review Board approval, potential participants were contacted via postcard
and asked to indicate their willingness to participate. Phone calls were made to those
participants indicating willingness to participate. A trained research assistant interviewed
the adolescent and caretaker in separate rooms. Various kinds of data were collected
including biological, observational, cognitive and self-report questionnaires. Caretakers
completed standardized measures assessing their own functioning, behavior, and attitudes
and also completed measures reporting on the child in their care. All modules are
completed in a one-day session which lasted between three and five hours. After the
interview, the participants were debriefed and the adolescent and caretaker were paid
following guidelines specified by the National Institutes of Health Normal Volunteer
Program.

Participants
Beginning in 2002, and completed in 2004, 303 young adolescents were recruited
based on traditional power analysis considerations. The sample was recruited according
to the following inclusion criteria: (1) a DCFS case being opened in the preceding month
due to the credible report of maltreatment or the extreme threat of maltreatment, as
30

assessed by DCFS, (2) age between 9 and 13 years, (3) child identified as Latino, African
American, Caucasian or biracial, (4) child currently residing within 10 specified zip
codes within Los Angeles County. The zip codes were chosen to be accessible to the
research site, to contain significant numbers of children of the 3 ethnicities, and to have
substantial numbers of maltreated children. As displayed in Table 1, 41% percent of the
maltreated children were African-American, 35% Latino, 11% Caucasian, and 13%
biracial. Fifty percent were female.
A comparison sample of 151 adolescents was collected from the same zip codes,
using a list of family names obtained from a marketing firm. Thirty two percent of the
comparison children were African-American, 47% Latino, 10% Caucasian, and 11%
biracial. Forty percent were female Twelve months after the initial assessment, families
returned for the second wave of data collection. Eighty-seven percent of the sample –
393 participants – was retained for Time 2. Eighteen months after the Time 2 assessment,
324 (71%) returned for the Time 3 assessment. The vast majority of attriters became
unreachable by phone, mail, or through family and friends listed as contacts at Time 1.
Fewer than 15 formally requested to leave the study. At Time 3, the mean age was 13.7
years. Some of the maltreating families who had been previously separated as a result of
DCFS involvement, were reunited. At Time 1, 68% of the sample was residing with a
biological or step-parent, 17% in non-relative foster or adoptive care, and 15% in kin care.
By Time 3, the percentage of youth with a biological or step-parent was 73%, 8% were in
foster or adoptive homes and 15% in kin care. These data are more fully explored and
tabled in the results section.

31

Maltreatment Experience Description
The Juvenile Court gave permission to access the children‟s maltreatment records
and as part of the consent, parents and children gave consent and assent to review these
records. A systematic review of these records was done and evidence of maltreatment
was coded according to chronicity, severity, type of maltreatment, and the strength of
substantiating evidence. This method of characterizing maltreatment yields dramatically
more information and different estimates compared to the summary statistics provided by
LACDCFS (See Trickett, Mennen, Kim, & Sang, 2009 for more details of this process).
According to the review, the sample has the following characteristics: the mean number
of referrals for suspicion of maltreatment to a child protective agency was 4.9 (standard
deviation = 3.3) with a range from one to 17. Seventy-two percent experienced general
or severe neglect, 49% physical abuse, 20% sexual abuse, 48% emotional abuse, 52%
caretaker incapacity, and 53% „substantial risk‟. Substantial risk was indicated when the
child‟s sibling had been maltreated but definitive evidence that the subject had been
maltreated was lacking. Of those experiencing “substantial risk,” 91% experienced an
additional and specific form of maltreatment. Fifty-six percent of children were
classified as having multiple classifications.

Measures
Child Self Report Measures
Depression. The Children‟s Depression Inventory (CDI) is a 27 item self report measure
representing a downward extension of the Beck Depression Inventory (Kovacs, 1981).
The items are rated on a 3-point scale assessing the frequency of an emotional state or
32

behavior. Myers and Winters (2002) review the CDI and report generally strong
psychometric properties. The measure is supported by its prognostic power (Ialongo et
al., 2001) and through its associations with self-esteem, cognitive distortions,
attributional style, locus-of-control and underachievement (Myers & Winters, 2002). The
CDI includes an interpersonal problems subscale tapping aspects of irritability and
aggression. Although legitimate arguments can be made that irritability in youth
represents one manifestation of depression, Myers and Winters (2002) suggest that the
CDI‟s discriminant validity suffers as a result of the broad definition of the construct.

Aggression and Delinquency. The Youth Self-Report (YSR; Achenbach, 1991) is an
extensively researched tool for adolescent assessment that has been translated into over
60 languages (Achenbach & Rescola, 2001). It represents the self-report version of a
multi-informant strategy that includes assessment of teacher and parental views of the
child. Items are rated on a 3-point scale assessing the frequency of occurrence: 0 – not
true; 1 somewhat or sometimes true; 2 – very or often true. In the current study, only the
aggression and delinquency subscales are utilized. One item was changed for the first
two measurement periods. In the questions used for Time 1 and Time 2, the original item
“I am mean to others” is absent but “I am mean to animals” is included. While the
attention subscale of the YSR is typically considered as representative of the
externalizing construct, previous work suggests attention problems represents a „mixed
syndrome‟ (Lambert, Schmitt, Samms-Vaughan, An, Fairclough & Nutter, 2003). In
contrast, the aggression and delinquency subscales have previously been found to load
33

unequivocally on an externalizing factor (O‟Keefe, Mennen & Lane, 2006; Lambert et al.,
2003).

Peer Delinquency. Affiliation with delinquent peer groups was assessed with a self-report
computerized questionnaire: the Adolescent Delinquency Questionnaire (ADQ; adapted
from Huizinga & Morse, 1986). The ADQ is a 110-item measure that assesses the
frequency of delinquent acts and substance use for the past 12 months and lifetime for the
respondent and affiliation with delinquent peers. Fifty items ask respondents to indicate
how many of their „friends or people you know‟ engage in the activities described.
Answer options include “I don‟t know anyone my age who has done this,” “one or two
people my age that I know do this,” or “lots of people my age that I know do this.” These
items will be utilized as the indicator of delinquent peer affiliation.

Parent Report Measures
Aggression, Delinquency and Depression. The Child Behavior Checklist (CBCL) was
utilized to assess depression and externalizing behavior (Achenbach, 1991). The CBCL
has been used as a caretaker-report measure yielding scores for subscales including
depressed/withdrawn behavior, aggression and rule-breaking behavior. It features three
response options for 112 questions asking how frequently the behavior has been observed
in the past six months. Items are rated on a 3-point scale assessing the frequency of
occurrence: 0 – not true; 1 somewhat or sometimes true; 2 – very or often true. For the
aggression subscale, items include, for example, “attack people” and “argues”. The
delinquency subscale includes items such as “sets fires” and “truant.” The measure is
34

arguably the most extensively normed measure available for assessing child problem
behavior (Lambert, Schmitt, Samms-Vaughan, An, Fairclough & Nutter, 2003). The
measure features high test-retest reliability and high internal consistency. Validity is
supported by studies which have found that the CBCL is related to important functional
outcomes (Achenbach, 1991).

Parental Depression. The caretaker‟s depressive symptoms were assessed using the Brief
Symptom Inventory (BSI; Derogatis & Melisaratos, 1983). The BSI is a well-validated
measure of adult psychological symptoms and has been used profitably in studies of child
maltreatment (Cox, Kotch, & Everson, 2003). The measure includes six items tapping
the depression construct that are rated on a 5-point scale from 0 (not at all) to 4
(extremely). Respondents rate how frequently they have experienced the item description
– “feeling lonely,” for example – in the past seven days.

Harsh Parenting. Harsh parenting was assessed with the Adult-Adolescent Parenting
Inventory-2 (AAPI-2; Bavolek & Keene, 1999). The AAPI-2 is a 40 item measure that
assesses child-rearing habits and beliefs of individuals parenting adolescents. The AAPI-
2 provides five subscale scores measuring the following parenting attitudes: (1)
inappropriate expectations of children (2) parental lack of empathy towards children‟s
needs (3) belief in the use of corporal punishment as a means of discipline (4) reversing
parent–child role responsibilities and (5) oppressing children‟s power and independence.
To operationalize harsh parenting, the lack of empathy and corporal punishment
subscales were utilized. Items are scored on a 5-point scale, ranging from strongly agree
35

to strongly disagree. The AAPI-2 has been useful in discriminating abusive and non-
abusive parents (Bavolek & Keene, 1999) and psychometric studies have provided
additional support for construct validity (Conners, Whiteside-Mansell, Deere, Ledet, &
Edwards, 2006).

Analytic Plan
Treatment of Missing Data and Outliers. Item-level missingness is generally low with
most variables exhibiting less than 1% missing. For these instances, multiple imputation
using Schafer‟s (1999) NORM software program were used to complete the data set.
Multiple imputation is a model-based approach that simulates a specified number of
potential values for the missing data based on parameters in the model and then
arithmetically averages the range of values and includes variation due to expected
sampling variation. Multiple imputation assumes that data are ignorable, or, missing at
random (MAR; Schafer & Olsen, 1998). MAR does not imply that the pattern of
missingness is entirely unpredictable, or „random‟. Rather, this assumption entails that
the patterns of missingness can be accounted for by data included in the model rather
than missingness being correlated with unobserved variables.
Analyses can determine the nature and predictability of the attrition. Logistic
regression is commonly used to identify predictors of drop-out. Separate logit analyses
will be conducted for Time 2 dropout and Time 3 dropout utilizing a range of
demographic and psychosocial predictors. Where possible, variables strongly related to
attrition will be included in models so that parameter estimates optimally adjusted
(Collins, Schafer & Kam, 2001). The substantive analyses will be conducted with the
36

latent variable software AMOS 5.0 (Arbukle, 2003) and implement full information
maximum likelihood procedures to incorporate participants lacking data for an entire
measure. This procedure also assumes MAR. As maximum likelihood procedures
estimate parameters based on the probability density of the observed values, MAR data
does not bias parameter estimates or compromise efficiency as the patterns of
missingness are essentially reflected within the available data (Collins, Schafer & Kam,
2001). The attrition rate of 15% at Time 2 and 29% at Time 3 is acceptable, given the
highly distressed nature of the sample.
The data were also be inspected for influential observations that could artificially
distort the model for the vast majority of the data. McClelland (2000) argues that there
are principled, statistically informed procedures for determining outlying values which
emerge from a different subpopulation and tell a different „story‟ than the rest of the
observations. In any analysis where outlying values exert a substantial effect on the
conclusions of the analysis, decisions to exclude or trim particular data points will be
made explicit.
Incorporating Child Self-Report and Parent-Report. A persistent question for
psychopathology research regards the optimal way to operationalize a construct.
Researchers have alternatively sought to identify ideal informants or aggregate
information across informants. Substantial evidence suggests that different informants
represent valid, if only moderately correlated perspectives (Achenbach, McConaughy &
Howell, 1987). Preliminary analyses suggested that child and parent responses were
weakly correlated and therefore could not be integrated through factor analysis or a
method such as the one suggested by Kraemer, Measelle, Ablow, Essex, Boyce and
37

Kupfer (2003). For the growth models, child and parent report data will be modeled
separately, and the results will be contrasted.
Factor Models of Depression and Externalizing Behavior. Measurement models
representing each of the two constructs will be constructed and analyzed in a multi-group
structural equation modeling approach (MGSEM). As the analyses are interested in
differences between boys and girls, it is necessary to first determine if the constructs of
interest are conceptually equivalent across groups. Although sometimes overlooked in
analytic practice, interpretations based on measures that do not function equivalently
across groups cannot be made unambiguously (Horn & McArdle, 1992; Vandenberg &
Lance, 2000). The MGSEM approach allows the researcher to test the equivalency of
measurement across groups. This was achieved by conducting a series of nested models
whereby cross-group equality constraints are progressively imposed. The likelihood ratio
test, distributed as a chi-square statistic, assesses the relative decrement of model fit with
the equality constraints, given the associated gain in degrees of freedom. Several levels
of increasingly restrictive invariance are tested. Guided by recommendations of Byrne
and Stewart (2006), the current study sought at least weak metric invariance between
groups. Weak metric invariance is obtained when the pattern and strength of factor
loadings are equivalent across groups. When invariance does not obtain, respecification
of the measurement model can be guided by relating an indicator to a different factor,
deleting an indicator, or utilizing correlated measurement errors (Anderson & Gerbing,
1988). These decisions should not be made entirely on the basis of statistical
considerations. Rather, model modifications must be guided by theoretical
considerations (Byrne, 2001).
38

The factor structure for child-report depression was indicated by subscales of the
CDI. For parent report depression and externalizing behavior, and child-report
externalizing behavior, items from the CBCL or YSR were combined randomly to form
item parcels. This form of item parceling has been defended as item level modeling can
compromise reliability, features a larger ratio of unique-to-common factor variance and is
more likely to evidence distributional violations (Little, Cunningham, Shahar &
Widaman, 2002). Little and colleagues (2002) suggest that when the researcher is
interested in relations between latent constructs rather than explorations of the
psychometric properties of the measure, as in the current investigation, the obscuration of
„nuisance‟ factors at the item level do not compromise the interpretability of the parceled
results.

Specific Aim #1: Cross-Lagged Regression Models of Depression and Externalizing
Behavior Contrasted with Latent Change Score Models

1A. Using traditional cross-lagged analytic strategies, determine the
longitudinal relationship between depression and externalizing behavior for
the full sample of girls and boys

1B. Using multi-group analyses, determine if gender moderates the
relationship between depression and externalizing behavior

Latent Variable Cross-Lagged Regression Model. Cross-lagged regression
analyses share some features with latent change score models (LCS) described later,
although the parameters test less dynamic hypotheses. Whereas LCS models represent
the parameters of key interest – the investment of one variable in the longitudinal change
of a second variable – cross-lagged models are less ambitious in their depiction of causal
39

processes. The purpose of this aim is to compare and contrast the conclusions reached
from LCS and cross-lagged models.
Cross-lagged models were derived from time-series analysis and are most clearly
interpretable when the structure of the data is consistent with a complex set of model
constraints (McArdle, 2009). As a first step in the model fitting, the full sample was fit
with constraints that, if well-fitting, suggest that causal processes are stable over time.
That is, adequate fit would suggest that the change processes have reached equilibrium.
These model constraints include equivalence of factor loadings, disturbance variances,
the autoregressive effects for each factor, and the crossed paths for each factor.
In cross-lagged regression models, distinct constructs are each presumed to
influence themselves over time, and possibly, exert an influence on another construct,
longitudinally. The autoregressive parameters indicate the degree of longitudinal
stability and coefficients from the crossed path indicate that one common factor may
exert a causal impact on the other. A significant crossed path suggests that one factor can
explain variance above and beyond the variance explained by the autoregressive effect.
Depression and externalizing behavior were modeled following guidelines of Martens
and Haase (2006), with correlated error terms for manifest indicators and disturbance
terms for the endogenous variables. Longitudinal and group invariance constraints were
also applied, as determined by psychometric model testing.
Common risk factors, including delinquent peer affiliation, child maltreatment,
parental depression and harsh parenting were included as covariates. Additionally, age
was included as a covariate. These variables were modeled with non-zero covariances.
Depression and externalizing behavior at each of the three time points were regressed on
40

the covariates. First, a model with the full sample was fit and, following confirmation of
adequate fit, a series of multi-group analyses, across gender, were tested. As described
below, weak metric invariance was obtained across gender, and thus cross-group equality
constraints were imposed on factor loadings. First, simple autoregressive models were
tested without equality constraints across groups. Decrements in model fit following
constraints on autoregressive parameters were assessed with the chi-square difference test.
Next, each crossed path was assessed by first freely estimating the parameters in one
group while constraining the path to zero in the second group and then comparing
improvement over simple autoregressive models. If the crossed path was significant,
equality constraints were imposed on the crossed effect and the model fit was compared
with the unconstrained model. This method determined if non-zero paths were necessary
in both groups and if the strength of the effect was consistent across groups. Comparison
between the strength of effects from x to y and y to x can only be made when there is the
same scale of measurement so this direct comparison was not made.
The traditional approach is to fit these models scaled by time of measurement,
rather than age. In such models, if the sample features variability in terms of age, that
variable is included as a covariate. This approach will be taken in the current research.
However, a second series of models will be fit with an age-based scaling. The data will
be reorganized so that measurement of all 10 year olds will be categorized together, and
so on for the additional ages. Of course there are more ages than there are measurement
points, so there will be substantial missing data. Models will be fit where there is
adequate coverage to estimate the models. Thus, data at the tails of the distribution – 9,
14, 15, and 16 years olds were eliminated from the model. As described below, the LCS
41

models were age-scaled so this strategy provides an opportunity to directly compare the
models.

Specific Aim #2: Predicting the Trajectory of Depression and Externalizing Behavior

2A. Characterize the trajectories of depression and externalizing behavior
separately for boys and girls with unconditional latent growth curve analyses

2B. Determine if baseline depression predicts level or growth of externalizing
behavior, before and after accounting for common risk factors (peer
delinquency, child maltreatment, parental depression, harsh parenting)

2C. Determine if the baseline externalizing behavior predicts level or growth
of externalizing behavior, before and after accounting for common risk
factors (peer delinquency, child maltreatment, parental depression, harsh
parenting)

2D. Assess the moderating role of gender in the longitudinal association of
depression and externalizing behavior

Latent Growth Curve Modeling. Growth curve analyses, utilizing a structural
modeling framework, will test the longitudinal association between depression and
externalizing behavior for boys and girls. Growth curve models assume that a common
factor explains the repeated measures of a construct over time, and that this factor is
representative of the latent trajectories for the sample across time (McArdle & Epstein,
1987). The latent growth model is particularly well suited to test the hypothesis of
pathogenic comorbidity as it represents an important innovation in longitudinal data
analysis. By simultaneously incorporating data from multiple data points, growth curve
analyses offer more reliable assessment of change as compared with traditional methods.
Further, the method avoids the inflated Type I error rate associated with repeated
measures analysis of variance (Beauchaine, Webster-Stratton & Reid, 2005). This test of
42

pathogenic comorbidity examines if a causal relationship between the constructs is
plausible, even if it does not singularly explain the association. Analyses will seek to
first determine if there is, in fact, a longitudinal association and then examine if that
association remains after accounting for risk factors that are common to depression and
externalizing behavior, as described below. Tests will be conducted separately for girls
and boys. Initially, unconditional growth curve analyses will be fit for the entire sample.
Following conventions, the first model to be estimated will be a „no growth‟ model which
features a fixed basis curve for the intercept parameter, freely estimated variance, and no
growth factor (McArdle & Grimm, 2008). This model is not expected to fit the data well,
but can serve as a basis against which more informative models can be compared. Next,
a linear growth model will be fit. The linear growth model incorporates a slope factor
mean and variance and includes an estimate of the covariation between initial level and
growth parameter. The linear scaling is achieved by fixing the loadings for the slope
factor to a linear pattern. A model incorporating quadratic growth will also be tested.
This will determine if change is better characterized as an accelerating pattern or if linear
models best describe the growth. Choices regarding the scaling of time have important
implications for interpretation of the models (Preacher, Wichman, MacCallum & Briggs,
2008). In the current case, scaling based on age is most developmentally appropriate,
since subjects entered the study with ages ranging from nine to thirteen.

43

Figure 1. Path Diagram of Latent Growth Model Organized by Age
Note. Latent growth model, as given in Ferrer, Hamagami and McArdle,
(2004). μ0=Mean of intercept (initial score at age 9); μ
s
= mean of slope
(rate of change); λ[t] = fixed to be 0 through 7, representing linear change;
σ
e
= variance of residual. Manifest variables are represented as a circle within
a square, suggesting the missingness due to the organization of data by age,
rather than assessment point.

If there is substantial inter-individual variability in growth trajectories – as
indicated by significant variance in the slope factor, time-invariant covariates will be
added to the model. Although multilevel modeling is better known for this language,
conditional growth models are often called class-level interactions as age (level 1) is
essentially interacting with the covariate (level 2) in order to predict the outcome variable
(Preacher, Wichman, MacCallum & Briggs, 2008). For the model of depression,
externalizing behavior at Time 1 will first be added to determine if there is longitudinal
44

association with the growth of depressive symptoms. Next, seven risk factors common to
both depression and externalizing behavior will be added to the model: four types of child
maltreatment, delinquent peer affiliation, parental depression, and harsh parenting. Child
maltreatment will not be dichotomous, but instead indicate the presence or absence of
documented physical abuse, sexual abuse, emotional abuse and neglect. The addition of
these risk factors is necessary in assessing the causal relationships between depression
and externalizing behavior. If a relationship between depression and externalizing
behavior persists, more support is marshaled for causality.
Parameter estimates will be assessed for boys and girls in a multiple-group
framework. Constraints on intercept and slope terms will determine the similarity or
difference of the trajectories for each gender. As noted above, more evidence supports
the hypothesis of pathogenic comorbidity among girls. Furthermore, there is more
evidence that externalizing behavior has a cascading effect on depression than the reverse.
It is hypothesized that, after accounting for common risk factors, the effect of baseline
externalizing behavior will predict changes in depressive symptoms for girls only. In
neither group is depression expected to be predictive of changes in the growth of
externalizing behavior.

Specific Aim #3: Dynamic Modeling of Latent Change Scores

3A. Characterize the course of depression and externalizing behavior
separately with univariate latent change modeling

3B. Determine the dynamic relationships between symptom change and
previous psychological status with bivariate latent change score models

3C. Contrast these model findings with those of the cross-lagged analysis

45

Latent Change Score Models. LCS models enable empirical examination of
theoretically important hypotheses about growth and change. These models incorporate
aspects of cross-lagged regression models and latent growth curve models (McArdle,
2009; Ferrer & McArdle, 2004). LCS models begin with the classical test theory
principle that observed measurements represent a composite of common and unique, or
error variance. Common factors represent the „true score‟ once the unique variance has
been separated out. Status at Time t is represented by effectively residualizing latent
state at t on latent state at t-1. The failure of true score at t-1 to perfectly predict true
score at t can be considered latent change. Longitudinal change in true scores is
expressed as a function of an initial state plus the accumulated latent changes. The
trajectory equation expresses that the score Y at at time t for person i reflects an initial
status on true score plus the accumulation of latent changes, and an error term:
t
Y
it
= y
i0
+ (
yki
) + e
yit

k=1
One source of change from t-1 to t, is expressed by a self-feedback coefficient,
y
,
representing the effect of the construct on the subsequent change. This parameter
expresses the expectations of change, depending on previous levels of the construct. This
is not however, the only source of variability in change scores. The accumulation of
latent changes is also explained by a slope or constant change factor, with a fixed loading.
In latent growth curve analyses, equivalence of factor loadings for the slope factor
implies a level-only model. However, in latent change models, this factor is not directly
depicting the change in true scores but represents the accumulation of proportional latent
changes. Thus, changes in y are expressed as a function of which is the factor loading
46

associated with the slope of the latent change factor, y
is
. In a bivariate model, a coupling
parameter, carries the effect of x
[t-1]
on Δy
[t]
. Including an error term, , the model for
latent change scores in a construct, y, in a bivariate model with construct x, as given in
Ferrer, McArdle, Shaywitz, Holahan, Marchione, and Shaywitz (2007) is:
Δy
it y
y
is
+
y
y
it-1
+
y
x
it-1
+
Δyit

It is important to note that certain LCS models have included a latent change score that is
deterministic, meaning that it does not include the error term in the previous equation
(e.g. Ferrer & McArdle, 2004) which is how the models were fit in the current analysis.
Of note is the coupling parameter, which is of key interest. In the current analysis,
models fit with and without coupling parameters assessed the influence of depression on
changes in externalizing behavior and the influence of externalizing behavior on changes
in depression. Significant improvements in model fit with the inclusion of one or both
sets of coupling parameters, suggests a unidirectional or reciprocal causal relation
between depression and externalizing behavior. The imposition of stable change
processes was achieved through several model constraints. The slope factor loadings
were equivalent over time, as were the self-feedback coefficients. The auto-regressive
effects in traditional cross-lagged models function differently in LCS models, and thus
the coefficient is fixed to unity. Error variances for each of the indicators for the
constructs are also constrained to equality and assumed to be unrelated to the
accumulation of latent changes (McArdle, 2009). In the bivariate models, within time
error terms were correlated. Before bivariate models were fit, univariate models
examined depression and externalizing behavior separately. Figure 2 and Figure 3 depict
the path diagrams for univariate and bivariate LCS models.
47

is

i
2
i

s

s
2

d d d d

d
2
d
2
d
2
d
2
d
2

Figure 2. Path Diagram of Univariate Latent Change Model of Depression
Note.
d
represents self-feedback parameter, held constant across time.
i
represents mean
of the initial depression factor, and
s
is the latent mean of the slope of the differences factor,
each of which have associated variance,
i
2
and
s
2
.These factors are permitted to covary,
is.
The loadings for the slope factor , are also invariant over time. Latent factors for depression
at each age are modeled with a single indicator, the total of the Children‟s Depression
Inventory (CDI). The variance of each residual variable,
d
2
,
is set to equality over
time. Triangle = constant (1). All parameters without a specified parameter are set to 1.
DEP
10
DEP
12

DEP
11

DEP
11

DEP
13

DEP
14

DEP
12

Slope of
DEP

Initial
DEP
DEP
13

DEP
14

CDI
10

CDI
11

CDI
13

CDI
14

CDI
12

e10 e11

e12

e13

e14

1

Figure 3. Path Diagram of Bivariate Latent Change Model
e e e

ds,e0 es,e0
μ
e0 e e e

e e e
d0,e0

es
e
ext
,e
dep

μ
ds
d d d

d0,es

ds,d0
μ
d0 d d d
d d d

DEP
10
DEP
12

DEP
11

DEP
11

DEP
13

DEP
13

DEP
14

CDI
12

CDI
11

CDI
13

CDI
10

EXT
11

EXT
12
EXT
13
EXT
10

EXT
11

EXT
12

EXT
13

EXT
10

EXT
12

EXT
11

EXT
13

DEP
s

EXT
s

EXT
0

DEP
0

1
48

Note that in Figure 3, the covariance between constant change slope factors and
initial levels is estimated. This affords important insight into the dynamics between
depression and externalizing behavior. LCS models enable empirical examination of
hypotheses that cannot be tested with traditional cross-lagged regression analyses.
Specifically, significant crossed paths, labeled
d
and
e
in Figure 3, represent a test of
pathogenic comorbidity. Traditional cross-lagged model do not capture the
developmental growth in the construct and thus cannot be evaluated as a test of dynamic
hypotheses (Ferrer & McArdle, 2007)
49
50

Chapter 3: Results
Preliminary Analyses
The boys and girls, and maltreated and comparison groups were compared in two
series of comparisons. Both sets of analyses showed that the groups evidenced relatively
similar ages, ethnic and gender representation. As LADCFS must often remove children
from dangerous environments, the proportion of maltreated youth living with a biological
parent (52%) was notably lower than the living arrangements of comparison youth (93%
with biological parent). There was no difference between boys and girls. From Time 1
to Time 2, the percentage of maltreated families in biological care increases as a result of
selective attrition and DCFS reunification. Basic demographic variables from each of the
three measurement periods are displayed in Table 1. A similar trend is seen among boys,
who dropped out at higher rates than the girls. Comparisons between the genders are
displayed in Table 2.

51

Table 1.
Sample Characteristics for Maltreated & Comparison Subjects at Time 1, Time 2 and Time 3
Demographic Variable Group
Maltreated

Comparison
Time 1 Time 2 Time 3

Time 1 Time 2 Time 3
N 303 250 195

151 142 142
Age (std deviation)
10.84
(1.15)
12.02
(1.21)
13.58
(1.25)

11.11
(1.15)
12.28
(1.26)
13.81
(1.48)
Gender (%)
Male 50 48 46 60 60 57
Female 50 52 54 40 40 43
Ethnicity (%)
African American 40 40 47 32 32 33
Latino 35 36 30 47 45 43
White 12 11 8 10 11 11
Mixed Biracial 13 13 15 11 12 12
Living Arrangement (%)
With Parent 52 63 63 93 94 93
Foster Care/Extended Family 48 37 37 7 6 7

51
52

Table 2.
Sample Characteristics for Girls and Boys at Time 1, Time 2 and Time 3
Demographic Variable Group

Girls

Boys

Time 1 Time 2 Time 3

Time 1 Time 2 Time 3
N 212 185 160

242 206 164
Age (std deviation)
10.85
(1.16)
12.02
(1.20)
13.72
(1.34)

11.00
(1.16)
12.17
(1.22)
13.64
(1.50)
Maltreatment %
Comparison 29 31 34 37 41 45
Maltreated 71 69 66 63 59 55
Ethnicity (%)
African American 37 37 41 38 38 42
Latino 40 41 37 38 38 34
White 9 10 8 12 12 11
Mixed Biracial 14 13 14 12 12 14
Living Arrangement (%)
With Parent 65 71 71 66 68 68
Foster Care/Extended
Family
35 29 29 32 32 32

52
53

Longitudinal designs almost invariably feature missing data. Data collection was
carefully conducted and therefore item-level missingness was minimal. Most items
featured rates of missing less than one percent and no variables featured missingness
greater than five percent. For missingness at such low rates, a single imputation is
sufficient to yield accurate estimates of parameter values. The percentage of missing data
was very low – typically less than one percent. Substantial attrition was witnessed across
the measurement periods, with 87% of the sample being retained at T2 and 71% of the
sample at T3. Attrition data, sorted on the basis of demographic variables, are displayed
in Table 3. Logistic regression analyses were conducted in order to help determine the
patterns of missingness and examine the degree to which observed variables could
explain attrition. Analyses first examined the loss of data from T1 to T2 and next,
examined attrition from T2 to T3. The model for T2 attrition was unable to explain the
data loss. A range of psychosocial variables, including parent and child report data,
along with demographic variables were entered in the model. Even applying a more
liberal standard for determining statistical significance (p <.10), only maltreatment
experience was associated with attrition. The maltreated group featured an odds ratio of
5.0 (90% CI: 2.1-12.1), suggesting that this group difference was strongly related to
attrition. The Cox & Snell R-Square, however, was a mere .05. In the analysis of data
loss from T2 to T3, psychosocial variables from T2 were added to the model. Several
variables evidenced significant predictive power. Male gender significantly predicted
attrition, with an OR of 1.9 (90% CI 1.1-3.4) as did Latino ethnicity (OR 3.6; 90% CI:
1.3 -9.9). Again, maltreated experience predicted attrition (OR 5.8; 90% CI: 2.6.-13.1).
54

Cox and Snell R-Square statistic was .13. These findings suggest that models including
the variables that were associated with attrition would contribute to the functioning of
FIML in providing unbiased parameter estimates.
53

Table 3.

Attrition Information

Completed Time 1
Only
Completed

Completed
Full Sample Time 1 and 2 Time 1 and 2 and 3 (or T1 & T3)
N N N % Retained N % Retained
Group
Maltreated 303 53 250 82.5% 195 64.4%
Comparison 151 9 142 94.0% 129 85.4%

Gender
Male 242 36 206 85.1% 164 67.8%
Female 212 26 186 87.7% 160 75.5%

Ethnicity
African
American 171 25 146 85.4% 134 78.4%
White 50 7 43 86.0% 30 60.0%
Latino 177 24 153 86.4% 114 64.4%
Biracial 56 6 50 89.3% 46 82.1%
Note. Group difference at Time 2 for maltreatment versus comparison (p<.01). Group differences at T3 for all three variables (p<.01).

55

56

The child self-report scores for the constructs of interest can be found in Table 4.
Table 4 displays the means, standard deviation and skewness for the aggression and
delinquency subscales of the YSR and the total score of the CDI. Aggression and
delinquency will not be analyzed separately but instead be used to model the broader
construct of externalizing behavior. However, as the items selected were based on
Achenbach‟s original classification, the particular subscales are listed below. The scores
are displayed for each gender separately, and across the three measurement points.
Visual inspection suggests relative stability in mean values and significant positive
skewness, particularly for the delinquency construct.
Table 4.
Descriptive Statistics for Child-Report Aggression, Delinquency and Depression
Time 1 Time 2 Time 3
Males Females Males Females Males Females
YSR Aggression
N 242(54%) 210(46%) 205(53%) 185(47%) 163(51%) 156(49%)
Mean 10.69 10.04 10.24 10.03 8.49 8.67
Std Deviation 6.25 5.56 5.81 6.09 5.53 6.55
Skewness 1.19 1.22 1.14 1.26 0.74 1.19

YSR Delinquency
N 242 210 205 185 163 156
Mean 2.29 1.57 2.11 1.79 3.99 3.56
Std Deviation 3.41 2.18 2.68 2.71 3.19 3.27
Skewness 2.84 1.95 1.91 3.54 1.06 1.60

CDI Depression
N 242 211 205 185 161 158
Mean 8.97 9.64 7.70 8.29 7.57 8.68
Std Deviation 7.18 7.22 5.80 6.76 5.86 6.73
Skewness 1.41 1.38 0.99 1.04 1.04 0.99

Note. YSR = Youth Self Report; CDI = Children's Depression Inventory.

57

Table 5 depicts the same statistics for the parental report on the child. The scores are not
directly comparable to the child-report scores as the items on the CBCL and YSR are not
identical. Again, mean values are generally stable with significant skewness observed
across the three subscales.

Table 5.
Descriptive Statistics for Parent-Report Aggression, Delinquency and Depression
Time 1 Time 2 Time 3
Males Females Males Females Males Females
CBCL Aggression
N 242 210 206 186 166 157
Mean 8.83 8.30 7.42 7.67 7.08 6.98
Std Deviation 7.81 7.66 6.68 6.99 6.71 6.59
Skewness 1.06 1.25 1.48 1.31 1.29 1.37

CBCL Delinquency
N 242 210 206 186 166 157
Mean 3.07 2.59 2.53 2.65 2.79 2.62
Std Deviation 3.39 3.09 3.09 3.38 3.50 3.29
Skewness 2.00 1.57 2.74 2.04 2.19 1.86

CBCL
Depression
N 242 210 206 186 166 157
Mean 1.43 1.59 1.20 1.68 1.16 1.23
Std Deviation 2.07 2.16 1.85 2.17 1.83 1.71
Skewness 1.87 1.82 2.06 1.71 2.06 1.57

Note. CBCL = Child Behavior Checklist. Depression is selection of 6 items from Anxious/Depressed
subscale.

Table 6 displays the correlations between all constructs of interest across the three
time measurement periods. To create an externalizing variable, the delinquency and
aggression subscales were combined from the YSR and CBCL. As expected, the table
58

evidences moderate inter-rater correlations, typically in the range of .20 to .30. This is
consistent with previous reports examining different informants reporting on child
behavior problems (Achenbach, 1991). Longitudinal stability within informant is
generally found to be between .35 and .55. The reliabilities of the scales, assessed with
the Cronbach alpha, are also displayed in Table 6. As expected, high alpha statistics were
obtained for the externalizing variables and the child depression variable as they feature a
large number of items. However, the alphas associated with the 6-item parent depression
scale were adequate – ranging from .73 to .76.

Determining Factor Structure for Child-Rated Depression and Externalizing Behavior

The psychometric characteristics of the CDI were examined before conducting the
substantive analyses. Specifically, analyses investigated if the CDI functioned as the
original authors specified, and if the measure functioned similarly for boys and girls. As
growth curve analyses assume that the construct being measured is consistent across time,
longitudinal measurement invariance was assessed.
The five subscales originally specified for the CDI were assessed as manifest
indicators of a latent variable, depression, at Time 1. Unit loading for the first indicator
was specified to identify the model. The model fit the data adequately ( χ
2
=14.9, df=5)
with a CFI of .988 and RMSEA of .066 (90% CI=.029-.106). All loadings were
significant and ranged from .66 to .77. Next, a series of multiple-group models were
conducted in order to assess measurement invariance across gender. At Time 1, a model
including separate groups of boys and girls was fit. Constraints were progressively
imposed until the decrement in model-fit was significant. Loadings were equivalent for
59

boys and girls for all indicators except negative self-esteem. Freeing that parameter
yielded a model with excellent fit ( χ
2
=20.4, df=13, CFI=.991, RMSEA=.035, 90%
57

Table 6.
Correlations between Time 1, Time 2 and Time 3 Outcome
Variables

T1
Parent
External
T1
Parent
Depress
T1
Child
External
T1
Child
Depress
T2
Parent
External
T2
Parent
Depress
T2
Child
External
T2
Child
Depress
T1 Parent External .93
T1 Parent Depress .57* .76
T1 Child External .25* .19* .86
T1 Child Depress .21* .20* .46* .85
T2 Parent External .58* .33* .19* .15* .93
T2 Parent Depress .31* .38* -.02 .11* .52* .76
T2 Child External .16* .07 .33* .20* .16* .05 .85
T2 Child Depress .20* .23* .43* .51* .13* .19* .40* .83
T3 Parent External .53* .34* .58* .15* .77* .46* .18* .17*
T3 Parent Depress .35* .38* .10* .24* .47* .57* .09* .22*
T3 Child External .21* .09 .14* .14* .21* .16* .53* .38*
T3 Child Depress .18* .14* .27* .44* .18* .20* .27* .57*
Note. *p<.05. Reliabilities (Cronbach alpha) along diagonal are bolded. Parent External = CBCL aggression and
delinquency sum score; Parent Depress = CBCL Depression items from Anxiety/Depression aggression and
delinquency sum score. Child External = YSR aggression and delinquency sum score; Child Depress = CDI sum
total

60
58

Table 6 Continued.
Correlations between Time 1, Time 2 and Time 3 Outcome
Variables

T3
Parent
External
T3
Parent
Depress
T3
Child
External
T3
Child
Depress
T1 Parent External
T1 Parent Depress
T1 Child External
T1 Child Depress
T2 Parent External
T2 Parent Depress
T2 Child External
T2 Child Depress
T3 Parent External .93
T3 Parent Depress 60* .73
T3 Child External .35* .26* .89
T3 Child Depress .23* .32* .49* .84
Note. *p<.05. Reliabilities (Cronbach alpha) along diagonal are
bolded. Parent External = CBCL aggression and delinquency sum score;
Parent Depress = CBCL Depression items from Anxiety/Depression
aggression and delinquency sum score. Child External = YSR aggression
and delinquency sum score; Child Depress = CDI sum total

61
62

CI=.000-.064). Standardized loadings ranged from .68 to .78 for boys and .66-.78 for
girls. In assessing for a stronger form of invariance, decrements in fit were observed
suggesting that intercepts were not equivalent for boys and girls.
Testing for Time 2 proceeded in a similar manner. Again, all factor loadings were
equivalent except one – in this case, interpersonal problems. Constraining the other four
loadings to equality yielded a model fit as follows ( χ
2
=32.5, df=13, CFI=.966,
RMSEA=.058, 90% CI=.033-.083). Loadings for boys ranged from .61 to .76 and .59
to .76. Again, intercepts varied across group. At T3, all factor loadings were equivalent
for girls and boys ( χ
2
=21.4, df=14, CFI=.983, RMSEA=.034, 90% CI=.000-.062).
Loadings for boys ranged from .41 to .77 and .62 to .75 for girls. Longitudinal invariance
was assessed across the three time periods in a similar manner. Similar findings were
obtained with all factor loadings, except anhedonia, which did not exhibit measurement
invariance. Model fit was good ( χ
2
=62.3, df=31, CFI=.983, RMSEA=.027, 90% CI=.017-
.037). Intercepts, as across gender, were not invariant. These findings suggest that
minimal criteria for assessing gender differences across time have been met.
After determining the adequacy of this measurement strategy, the sample was
divided on the basis of age. The scaling of time for growth models has been suggested by
McArdle and Hamagami (1992) and Metha and West (2000). This work has warned
against assessment based on measurement points, as subjects who begin the assessment at
different ages may lead to biases in the estimation of covariance between the slope and
intercept factor and bias in the variance of the initial level. Data points represented as
Depression 9 represent participants who were measured between 8.50 years and 9.49.
63

Table 7 displays the means and standard deviations of the full sample across all ages.
Tables 8 and 9 displays these statistics for boys and girls, respectively. Table 10
displays the correlations between depression, measured across age. Most of the
correlations range from about .4 to .6 and where sample size is adequate, most
correlations are significant (p<.01).
The psychometric characteristics of the aggression and delinquency scales of the
YSR were examined similarly. As previously noted, only the aggression and delinquency
subscales have been found to load unequivocally on an externalizing factor (O‟Keefe,
Mennen & Lane, 2006; Lambert, Schmitt, Samms-Vaughan, An, Fairclough & Nutter,
2003). Analyses investigated if the aggression and delinquency subscales converge to
represent the construct of externalizing behavior. For each of the two subscales, two
parcels composed of randomly selected items were created. This provided four manifest
indicators of the hypothesized latent construct of externalizing behavior. Model-fitting
results at each of the three time points indicated this provided very poor fit to the data.
Chi-square statistics were all significant and RMSEA statistics ranged from .12 to
over .20.
To identify an adequate measurement strategy, a principal axis factor analysis
with an oblique rotation (Promax) was conducted on the entire sample for Time 1 data.
Although researchers have cautioned against the exclusive reliance on eigenvalues
(Fabrigar, Wegener, MacCallum & Strahan, 1999), preliminary consideration of factors
was based on eigenvalues greater than 1.0. On the basis of these results which suggested
two factors, a second factor analysis was conducted that specifically extracted two factors.
64

The first factor explained 21% of the total variance and featured an eigenvalue of 7.0.
The second factor captured 4.5% of the variance with an eigenvalue of 1.5. Nine items
65

Table 7

Means and Standard Deviations for Full Sample on Children's Depression Inventory by Age

Depression Depression Depression Depression Depression Depression Depression Depression Depression
9 10 11 12 13 14 15 16 17
N 52 165 226 269 220 135 56 19 16
Mean 10.56 8.26 8.2 8.63 8.37 8.7 7.82 9.26 11.69
Std. Deviation 7.3 6.19 7.02 6.77 6.75 6.46 6.03 6.85 7.42

Table 8

Means and Standard Deviations for Girls on Children's Depression Inventory by Age

Depression Depression Depression Depression Depression Depression Depression Depression Depression
9 10 11 12 13 14 15 16 17
N 27 79 113 126 96 66 30 11 4
Mean 10.52 8.48 8.76 8.82 8.69 9.58 8.33 10.55 9.25
Std. Deviation 7.80 6.36 7.50 6.95 6.73 6.98 6.36 7.93 3.20

Table 9

Means and Standard Deviations for Boys on Children's Depression Inventory by Age

Depression Depression Depression Depression Depression Depression Depression Depression Depression
9 10 11 12 13 14 15 16 17
N 25 86 113 143 124 69 26 8 12
Mean 10.60 8.06 7.65 8.46 8.13 7.86 7.23 7.50 12.50
Std. Deviation 6.88 6.05 6.48 6.62 6.77 5.84 5.68 4.99 8.33

65

66

Table 10

Correlations between Mean Scores on Children's Depression Inventory by Age

Depression
9
Depression
10
Depression
11
Depression
12
Depression
13
Depression
14
Depression
15
Depression
16
Depression 9

1

N 52

Depression 10 Correlation 0.69* 1

N 37 165

Depression 11 Correlation 0.1 0.39* 1

N 17 100 226

Depression 12 Correlation 0.51* 0.46* 0.56* 1

N 18 75 130 269

Depression 13 Correlation 0.95* 0.65* 0.58* 0.48* 1

N 4 48 90 113 220

Depression 14 Correlation - 0.35 0.52* 0.58* 0.67* 1

N - 10 47 83 70 135

Depression 15 Correlation - 0.42 0.67 0.11 0.46* 0.52* 1

N - 5 6 30 42 21 56

Note.. *Correlation is significant at the 0.05 level. Ns vary substantially because there were 3 measurement periods, not annual assessments.

66
67

from the first factor with substantial loadings (greater than .45) no cross-loadings greater
than .20, were selected as indicators of the externalizing construct. The items included
the following: I disobey my parents, I physically attack people, I am mean to animals, I
set fires, I cut classes or skip school, I steal at home, I steal from places other than home,
I run away from home, I use alcohol. Communalities ranged from .22 to .43.
Next, the items were subjected to a confirmatory or restricted factor analysis
(CFA). Of course, this is not, confirmatory, in the ordinary sense. Preceding a CFA with
an exploratory factor analysis means that the CFA must be interpreted differently. It was
thus anticipated that the model would fit well, at least for the first time period, which was
the data on which the exploratory analysis was based. However, it was not clear that
longitudinal invariance or invariance across gender would hold.
Testing proceeded in a similar manner as the work with the depression construct.
The nine items derived from the exploratory factor analysis were randomly assigned to
three parcels. The parcels were modeled as manifest indicators of a single latent
construct of externalizing behavior. The full sample was fit at Time 1 with a unit loading
indicator on the first parcel and the loadings of the second and third indicator set to
equality to identify the model. Model fit was adequate ( χ
2
=3.3, df=1, CFI=.995,
RMSEA=.072, 90% CI= .000-.163). Next, multiple group models were examined at each
time point. The models tested increasingly stringent equality constraints, including factor
loadings, intercepts and residual terms. For Time 1, boys and girls evidenced equal
factor loadings and intercept parameters ( χ
2
=4.4, df=5, CFI=1.0, RMSEA=.000, 90%
CI= .000-.061). Residual variances, however, were not invariant. Loadings ranged
68

from .59 to .91 for girls and .68 to .89 for boys. Time 2 data was assessed similarly and
reached the same conclusions ( χ
2
=7.0, df=5, CFI=.993, RMSEA=.000, 90% CI= .000-
.061) although the Δχ
2
= 8.7(3) for a model constraining residual variances was nearly
non-significant (p=.03). This suggests similar factor structures between the boys and
girls. Loadings ranged from .63 to .78 for girls and .62 to .75 for boys. Time 3 showed
similar equality ( χ
2
=6.8, df=5, CFI=.993, RMSEA=.029, 90% CI= .000-.076). Loadings
ranged from .65 to .78 for boys and .69 to .80 for girls.
Longitudinal invariance tests proceeded in the same manner as with depression.
Tests of configural invariance were met ( χ
2
=3.4, df=4, CFI=1.0, RMSEA=.000, 90%
CI= .000-.040). However, a significant decrement in fit was evidenced when intercept
terms were constrained across time. Loadings ranged from .59 to .89. These findings
suggest that minimal criteria for assessing gender differences across time have been met.
Following the establishment of the measurement models, the data was organized
according to age. Data points represented as Externalize 9 represent participants who
were measured between 8.50 years and 9.49. Table 11 displays the means and standard
deviations of the full sample across all ages. Tables 12 and 13 displays these statistics for
boys and girls, respectively. Table 14 displays the correlations between externalizing
behavior, measured across age. Most of the correlations range between about .1 to .4 and
where sample size is adequate, most correlations are significant (p<.01). These
correlations are notably smaller than the correlations seen for depression.

69

Table 11
Means and Standard Deviations for Full Sample on Child-Report Externalizing Behavior by Age
Externalize 9 Externalize 10 Externalize 11 Externalize 12 Externalize 13 Externalize 14 Externalize 15 Externalize 16
N 51 164 225 266 226 133 58 19
Mean 1.63 0.86 0.70 0.95 1.23 1.68 1.84 2.21
Std. Deviation 3.03 1.69 1.45 1.56 2.10 2.71 2.79 2.23

Table 12
Means and Standard Deviations for Boys on Child-Report Externalizing Behavior by Age
Externalize 9 Externalize 10 Externalize 11 Externalize 12 Externalize 13 Externalize 14 Externalize 15 Externalize 16
N 25 86 112 142 127 69 27 8
Mean 1.80 1.00 0.69 1.11 1.20 1.48 2.04 1.88
Std. Deviation 3.61 1.91 1.53 1.79 2.34 2.32 3.61 2.36

Table 13
Means and Standard Deviations for Girls on Child-Report Externalizing Behavior by Age
Externalize 9 Externalize 10 Externalize 11 Externalize 12 Externalize 13 Externalize 14 Externalize 15 Externalize 16
N 26 78 113 124 99 64 31 11
Mean 1.46 0.71 0.71 0.77 1.27 1.91 1.68 2.45
Std. Deviation 2.42 1.40 1.38 1.23 1.77 3.08 1.85 2.21

69
70

Table 14
Correlations between Mean Scores on Child Report Externalizing Behavior by Age

Externalize
9
Externalize
10
Externalize
11
Externalize
12
Externalize
13
Externalize
14
Externalize
15
Externalize
16
Externalize
9
Pearson
Correlation 1
N 51
Externalize
10
Pearson
Correlation 0.45* 1
N 37 164
Externalize
11
Pearson
Correlation -0.12 0.15 1
N 16 98 225
Externalize
12
Pearson
Correlation 0.65* 0.17 0.38* 1
N 17 73 128 266
Externalize
13
Pearson
Correlation 0.78 0.11 0.42* 0.45* 1
N 4 49 91 116 226
Externalize
14
Pearson
Correlation - -0.18 0.24 0.26* 0.47* 1
N - 10 46 82 71 133
Externalize
15
Pearson
Correlation - -0.38 -0.26 0.34* 0.16 0.19 1
N - 5 7 30 45 21 -

70
71

Determining Factor Structure for Parent-Rated Depression and Externalizing
Behavior

Parent-Report Depression. Before substantive analysis, the psychometric characteristics
of the depression subscale of the CBCL was examined. Depression is tapped by two
subscales, but neither subscale purely represents the construct of depression. The
„withdrawn-depressed‟ subscale includes items that do not clearly tap the depression
construct such as „refuses to talk‟ and „Secretive, keeps things to self.‟ The „anxious-
depressed‟ subscale, naturally, features items that deal more explicitly with anxiety:
“fears certain animals, situations, or places, other than school” and “fears he/she might
think or do something bad.” Although there is evidence that depression and anxiety are
strongly correlated, they can be distinguished by this age (Cole, Truglio & Peeke, 1997).
Depression is not synonymous with internalizing behavior, so it was deemed
inappropriate to include items tapping the broad-band internalizing construct. Therefore,
items representative of the traditional construct of depression were selected. This
included the following six items: Complains of loneliness, Cries a lot, Feels or complains
that no one loves him/her, Feels worthless or inferior, Feels too guilty, Unhappy sad or
depressed. These items were examined before conducting the substantive analyses.
Specifically, analyses investigated if the items functioned as a coherent way of measuring
child depression, and if the measure functioned similarly for boys and girls.
The six items were assessed as manifest indicators of a latent variable, depression,
at Time 1. Unit loading for the first indicator was specified to identify the model. The
model fit the data poorly with a significant χ
2
and an RMSEA in excess of .15. Two
items were responsible for the misfit: „cries a lot‟ and „complains of loneliness.‟ These
72

two items were deleted from the measurement model and the resulting model fit was
good ( χ
2
=4.3, df=2, CFI=.995, RMSEA=.050, 90% CI=.000-.116). All loadings were
significant and ranged from .58 to .81. Next, a series of multiple-group models were
conducted in order to assess measurement invariance across gender. At Time 1, a model
including separate groups of boys and girls was fit. Constraints were progressively
imposed until the decrement in model-fit was significant. Loadings were equivalent for
boys and girls for all indicators and intercepts for the indicators were equivalent.
Constraining these parameters resulted in a non-significant decrement of fit (Δχ²=4.0 on
Δdf =9, p>.05). The error variances, however, were not invariant.
Testing for Time 2 proceeded in a similar manner. Unconstrained model fit was
good ( χ
2
=6.9, df=2, CFI=.993, RMSEA=.040, 90% CI=.000-.089). Constraining both the
loadings and intercepts to equality led to a non-significant decrement in fit (Δχ²=7.2 on
Δdf =9, p>.05). Loadings for boys ranged from .57 to .95 and .51 to .80 for girls. Again,
error variances were not invariant. At T3, all factor loadings were equivalent for girls
and boys ( χ
2
=21.4, df=14, CFI=.983, RMSEA=.034, 90% CI=.000-.062). Loadings for
boys ranged from .41 to .77 and .62 to .75 for girls. For Time 3 data, unconstrained
model fit was good ( χ
2
=6.5, df=2, CFI=.987, RMSEA=.037, 90% CI=.000-.087).
Constraining both the loadings and intercepts to equality led to a non-significant
decrement in fit (Δχ²=10.9 on Δdf =9, p>.05). Loadings for boys ranged from .39 to .73
and .48 to .82 for girls. Again, error variances were not invariant.
Longitudinal invariance was assessed across the three time periods in a similar
manner. Unconstrained model fit was good ( χ
2
=10.9, df=9, CFI=.995, RMSEA=.024, 90%
73

CI=.000-.047). All loadings were in excess of .45 at all three measurement periods.
Loadings were invariant across time period as witnessed by the non-significant
decrement in fit (Δχ²=2.5 on Δdf =3, p>.05). These findings suggest that minimal criteria
for assessing gender differences across time have been met.
After determining the adequacy of this measurement strategy, the sample was
divided on the basis of age in the same manner described above. Data points represented
as Depression 9 represent measurements of participants who were measured between
8.50 years and 9.49 years. Table 15 displays the means and standard deviations of the
full sample across all ages. Tables 16 and 17 displays these statistics for boys and girls,
respectively. Table 18 displays the correlations between depression, measured across
age. Most of the correlations range between about .3 to .6 and where sample size is
adequate, most correlations are significant (p<.01).
Parent-Report Externalizing Behavior. Next, the psychometric characteristics of the
externalizing subscales of the CBCL were examined. The aggression and delinquency
subscales were selected, as these subscales have previously been shown to load
unequivocally on an externalizing factor. Analyses investigated if the items functioned
as a coherent way of measuring child externalizing behavior, and if the measure
functioned similarly for boys and girls and across time.
Item analyses suggested that four items functioned poorly. The deleted items
include “Thinks about sex too much,” “Threatens people,” “Gets in many fights” and
“Vandalism.” The 28 remaining items were grouped randomly in three parcels and were
assessed as manifest indicators of a latent variable, externalizing behavior, at Time 1.
74

Unit loading for the first indicator was specified to identify the model. The model fit was
good ( χ
2
=0.1, df=1, CFI=1.0, RMSEA=.000, 90% CI=.000-.086). All

75

Table 15

Means and Standard Deviations for Full Sample on CBCL Depression by Age

Depression
9
Depression
10
Depression
11 Depression 12 Depression 13 Depression 14 Depression 15 Depression 16
N 52 163 226 270 226 130 57 19
Mean 0.71 0.96 1.05 1.06 0.77 0.82 0.96 1.05
Std. Deviation 1.24 1.49 1.68 1.56 1.32 1.26 1.67 1.65

Table 16
Means and Standard Deviations for Boys on CBCL Depression by Age

Depression
9
Depression
10
Depression
11 Depression 12 Depression 13 Depression 14 Depression 15 Depression 16
N 25 86 113 144 128 66 27 7
Mean 0.52 0.86 0.98 1.06 0.79 0.73 0.93 0.29
Std. Deviation 0.82 1.40 1.64 1.53 1.41 1.02 1.73 0.76

Table 17
Means and Standard Deviations for Girls on CBCL Depression by Age
Depression 9 Depression 10 Depression 11 Depression 12 Depression 13 Depression 14 Depression 15 Depression 16
N 27 77 113 126 98 64 30 12
Mean 0.89 1.08 1.12 1.06 0.74 0.91 1.00 1.50
Std. Deviation 1.53 1.59 1.71 1.61 1.20 1.47 1.64 1.88

75
76

Table 18

Correlations between Mean Scores on CBCL Depression by Age

Depression
9
Depression
10
Depression
11
Depression
12
Depression
13
Depression
14
Depression
15
Depression
16
Depression 9 1
N 52
Depression 10 Pearson Correlation 0.54* 1
N 37 163
Depression 11 Pearson Correlation 0.33 0.47* 1
N 17 98 226
Depression 12 Pearson Correlation 0.03 0.65* 0.45* 1
N 19 74 131 270
Depression 13 Pearson Correlation 0.95* 0.16 0.49* 0.48* 1
N 4 49 91 117 226
Depression 14 Pearson Correlation - 0.28 0.25* 0.35* 0.67* 1
N - 10 46 80 69 130
Depression 15 Pearson Correlation - 0.98* 0.46 0.21 0.46* 0.85* 1
N - 4 6 30 46 19 57

Note. *Correlation is significant at the 0.05 level. Ns vary substantially because there were 3 measurement periods, not annual assessments.

76
77

loadings were significant and ranged from .86 to .93. Next, a series of multiple-group
models were conducted in order to assess measurement invariance across gender. At
Time 1, a model including separate groups of boys and girls was fit. Constraints were
progressively imposed until the decrement in model-fit was significant. Loadings were
equivalent for boys and girls for all indicators and intercepts for the indicators were
equivalent. Constraining these parameters resulted in a non-significant decrement of fit
(Δχ²=8.0 on Δdf =4, p>.05). The error variances, however, were not invariant.
Testing for Time 2 proceeded in a similar manner. Constraining both the loadings,
intercepts and residual variances led to a well-fitting model ( χ
2
=12.4, df=8, CFI=.995,
RMSEA=.035, 90% CI=.000-.071). Loadings for boys ranged from .86 to .93 and .88
to .94 for girls. For Time 3 data, unconstrained model fit was good. Constraining the
loadings, intercepts and residual variances led to a well-fitting model ( χ
2
=12.8, df=8,
CFI=.991, RMSEA=.036, 90% CI=.000-.072). Loadings for boys ranged from .68 to .94
and .68 to .94 for girls. Longitudinal invariance was assessed across the three time
periods in a similar manner. Similar findings were obtained. Unconstrained model fit
was good ( χ
2
=0.1, df=1, CFI=1.0, RMSEA=.000, 90% CI=.000-.050). All loadings were
in excess of .86 at all three measurement periods. Loadings were invariant across time
period as witnessed by the non-significant decrement in fit (Δχ²=2.8 on Δdf =3, p>.05).
These findings suggest that minimal criteria for assessing gender differences across time
have been met.
After determining the adequacy of this measurement strategy, the sample was
divided on the basis of age in the same manner described above. Table 19 displays the
78

means and standard deviations of the full sample across all ages. Tables 20 and 21
displays these statistics for boys and girls, respectively. As there were some notable
outlying values for this scale, scores more than four standard deviations above the mean
were trimmed to the upper threshold of four standard deviations. This resulted in a total
of three scores being downwardly adjusted. However, this did not affect the substantive
findings for the analysis, so the data results for the untrimmed data are reported. Table 22
displays the correlations between externalizing behavior, measured across age.
79

Table 19

Means and Standard Deviations for Full Sample on CBCL Externalizing Behavior by Age

Externalize
9
Externalize
10
Externalize
11
Externalize
12
Externalize
13
Externalize
14
Externalize
15
Externalize
16
N

51 164 226 270 227 133 56 20
Mean

10.35 10.3 10.44 10.3 10.39 8.8 7.55 12.55
Std. Deviation 8.35 9.24 9.58 9.29 8.51 8.59 8.22 13.1

Table 20
Means and Standard Deviations for Boys on CBCL Externalizing Behavior by Age

Externalize
9
Externalize
10
Externalize
11
Externalize
12
Externalize
13
Externalize
14
Externalize
15
Externalize
16
N

25 86 113 144 128 68 26 8
Mean

10.4 10.53 10.13 10.15 10.92 8.76 7.23 13.25
Std. Deviation 9.03 9.42 9.33 8.83 8.91 8.22 9.34 16.43

Table 21
Means and Standard Deviations for Girls on CBCL Externalizing Behavior by Age

Externalize
9
Externalize
10
Externalize
11
Externalize
12
Externalize
13
Externalize
14
Externalize
15
Externalize
16
N

26 78 113 126 99 65 30 12
Mean

10.31 10.04 10.75 10.46 9.71 8.85 7.83 12.08
Std. Deviation 7.82 9.09 9.87 9.82 7.96 9.01 7.26 11.15

79
80

Table 22
Correlations between Mean Scores on Children's CBCL Externalizing Behavior by Age

Externalize
9
Externalize
10
Externalize
11
Externalize
12
Externalize
13
Externalize
14
Externalize
15
Externalize
16
Externalize
9 1
N 51
Externalize
10 Pearson Correlation 0.81* 1
N 37 164
Externalize
11 Pearson Correlation 0.70* 0.70* 1
N 16 99 226
Externalize
12 Pearson Correlation 0.83* 0.81* 0.53* 1
N 18 75 131 270
Externalize
13 Pearson Correlation 0.71 0.54* 0.44* 0.66* 1
N 4 49 91 119 227
Externalize
14 Pearson Correlation - 0.85* 0.54* 0.65* 0.68* 1
N - 10 46 81 72 133
Externalize
15 Pearson Correlation - 0.61 0.50 0.36* 0.62* 0.54* 1
N - 4 6 30 45 20 56

Note. *Correlation is significant at the 0.05 level. Ns vary substantially because there were 3 measurement periods, not annual assessments.

80

81

Specific Aim #1
Cross-Lagged Panel Analysis of Depression and Externalizing Behavior
The first series of models fit were cross-lagged regression models. Although
meaningful in themselves, these models primarily serve as a comparison point with
findings from the LCS models in Aim #3. As many investigations of comorbidity have
utilized cross-lagged regressions, comparing the results of these models with models that
test more dynamic hypotheses of change is worthwhile. It is conceivable that
methodological differences explain some of the conflicting findings.
Utilizing the measurement models composed as previously described, a series of
cross-lagged regressions of common factors were fit. These models, derived from time-
series analysis, are most clearly interpretable when the structure of the data is consistent
with a complex set of model constraints (McArdle, 2009). These model constraints
include equivalence of factor loadings, disturbance variances, the autoregressive effects
for each factor, and the crossed paths for each factor. A model with an adequate number
of measurement points that fits well given these constraints, suggests that causal process
are stationary. The present data was modeled according to these specifications, with one
deviation. In preliminary longitudinal invariance analyses the anhedonia indicator was
not invariant and was permitted to freely vary across the three measurement periods. The
first set of models was scaled, as is typically done, by measurement point rather than age.
Model fit was as follows: χ
2
=573.2, df=236, CFI=.906, RMSEA=.056, 90% CI= .050-
.062. This compares unfavorably with a model that relaxed the assumption of stationarity
82

(Δχ²=80 on Δdf =8, p<.05). This suggests that the causal processes have not fully
stabilized and that therefore, interpretations must be made cautiously.
Before multiple group analyses, the full sample was fit to a model imposing
longitudinal invariance constraints without predictor variables, other than autoregressive
factors. Error terms for the same indicators were permitted to covary across time. The
lagged and crossed paths were all freely estimated. The model fit was reasonable
( χ
2
=492, df=228, CFI=.927, RMSEA=.051, 90% CI= .044-.057). Constraining the paths
from depression to externalizing behavior resulted in a significant decrement in fit
(Δχ²=15 on Δdf =8, p<.05). Upon further inspection, this effect was restricted to the
crossed path from Time 2 depression to Time 3 externalizing. Externalizing behavior
however, appeared not to exert an effect on depression at Time 2 or Time 3. These
findings were next examined in the context of covariates including age, parent depression,
harsh parenting, delinquent peer affiliation and four forms of maltreatment including
neglect, sexual, physical and emotional abuse. Results were consistent with the previous
model. Model fit was acceptable although the model featured a large number of
parameters given the sample size ( χ
2
=661, df=357, CFI=.918, RMSEA=.043, 90%
CI= .038-.048) Autoregressive effects ranged from .31 to .72. There was no evidence of
a downstream effect of externalizing behavior on depression, but Time 2 depression
predicted increases in Time 3 externalizing behavior (b= .158, S.E.=.045, =.26, p<.05).
Parameter estimates for the best fitting model are depicted in Figure 4.
83

Figure 4. Parameter Estimates for Best-fitting Cross-Lagged Regression Model with Full Sample

Note. Covariates, manifest indicators and their associated error terms are omitted for simplicity of
presentation. All parameters are standardized and significant at p<.05.

T1
Depression
T2
Depression
T3
Depression
T1
Externalizing
T2
Externalizing
T3
Externalizing
β=.31
β=.72
β=.42
β=.26
d
d
d
d
d
d
.26 .43
.31
β=.57
83

84

Next, a multi-group modeling strategy was untaken to examine if the longitudinal course
was similar among boys and girls. Due to the large number of parameters being
estimated, covariates were eliminated from the model if they evidenced no significant
relationships with either depression or externalizing behavior at any of the three time
points. Age, physical abuse, neglect, parental depression, and peer delinquency were
retained. First, a model including only autoregressive parameters with no equality
constraints was fit. Model 2, nested within model 1, assessed if the stability of
depression and externalizing behavior was consistent across gender. The chi-square
difference test suggested that gender showed similar stability across this period of
adolescence (Δχ²=7 on Δdf =4, p>.05). Model 3 freed the crossed paths from
externalizing behavior to subsequent depression. The inclusion of these two parameters
resulted in a non-significant improvement in model fit (Δχ²=7 on Δdf =4, p>.05).
Crossed paths from depression to externalizing behavior were examined in the same way
in model 4. The inclusion of these parameters significantly improved model fit (Δχ²=11
on Δdf =4, p<.05), suggesting that for one or both genders, there was a downstream effect
of depression. In order to determine if this was consistent for boys and girls, model 5
constrained the crossed paths from depression to zero among boys and was freely
estimated for girls. Testing this model against model 4, a non-significant decrement was
noted (Δχ²=7.3 on Δdf =2, p>.05). This evidence supports the claim that early
depression does not exert an effect on subsequent depression for boys but such an effect
exists for girls. Finally, model 6 assessed if this parameter was needed for both time
points for the girls. Model fitting results suggest that only the crossed path from Time 2
85

depression to Time 3 externalizing behavior was needed for the girls. Thus, only a single
parameter of interest was different between boys and girls. Examining these findings in
light of the single group model described above, it can be seen that the significant crossed
path was carried by the girls in the multi-group model (b= .17, S.E.=.05, =.28, p<.05),
and was non-significant for the boys. Figure 5 depicts the findings from the best-fitting
multi-group model. As most parameters were judged invariant, the model was re-run as a
single-group model with the truncated list of covariates, and coefficients have been
standardized within the entire sample for ease of interpretation. Model fit is listed in
Table 23. Thus, no evidence for pathogenic comorbidity was found for boys, and weak
evidence was identified for girls. This pattern of gender differences is consistent with
previous evidence. Against predictions from Capaldi‟s (1991) failure model, depression
appeared to be causally related to externalizing behavior, rather than the reverse.
86

Figure 5. Parameter Estimates for Best-fitting Gender Multi-Group Cross-Lagged Regression Model

Note. Covariates (age, physical abuse, neglect, parental depression, and peer delinquency), manifest indicators and their associated error
terms are omitted for simplicity of presentation. All parameters are standardized within the full sample for ease of interpretation and
significant at p<.05. All parameters are invariant across gender except where noted.

T1
Depression
T2
Depression
T3
Depression
T1
Externalizing
T2
Externalizing
T3
Externalizing
β=.31
β=.72
β=.42
βgirls=.28
βboys=ns
d
d
d
d
d
d
.26 .43 .31
β=.57
86

87

Table 23. Model Fit Statistics for Multi-Group Cross-Lagged Regression Models
Model χ2 df RMSEA Model Comparison Δχ2 (Δdf) p Δχ2
Model 1: Autoregressive paths only without equality constraints 1151 652 0.041 - - -
Model 2: Autoregressive paths with equality constraints 1158 656 0.041 2 vs 1 7(4) ns
Model 3: Autoregressive paths with equality constraints, 1151 652 0.041 3 vs 2 7(4) ns
Externalizing to Depression freely estimated for both groups
Model 4: Autoregressive paths with equality constraints, 1147 652 0.041 4 vs 2 11(4) 0.02
Depression to externalizing freely estimated for both groups
Model 5: Autoregressive paths with equality constraints; Depression 1150 654 0.041 5 vs 4 2(2) ns
to externalizing constrained to 0 for boys
Model 6: Autoregressive paths with equality constraints, T1 Depression 1150 655 0.041 6 vs 5 .1(1) ns
to T2 externalizing constrained to 0 for both genders
Note. DF = degrees of freedom; RMSEA = root mean square error of approximation; CFI=Confirmatory Fit Index

87
88

Next, the data was reorganized so that a cross-lagged model could be fit to
subjects each measured at the same age. As there were only three measurement points
and ages nine through 17 represented in the sample, coverage was inadequate to feature
each age in the model. As the LCS models included ages 10 through 13, these ages were
included. The first model converged but there was a problem fitting the saturated model,
even though this is the model where all possible covariances are freely estimated.
Consequently, measures of model fit which compare the analyst‟s model with the
saturated or independent models could not be generated. Nevertheless, parameter
estimates could be examined in a preliminary manner, before models based on more
restricted ages were fit. In this first model with the full sample, two crossed paths were
significant and positive. Externalizing behavior at age 10 was associated with increases
in depression at age 11, after controlling for previous level of depression. Additionally,
externalizing behavior at age 11 was associated with increases in depression at age 12.
These two crossed effects are notably different from the time-scaled models, where age
was merely included as a covariate. Of note, the age-scaled models could not incorporate
the other covariates because of data coverage issues. However, the inclusion of the
covariates did not have a substantial impact on the time-scaled models, so it is unlikely
that their exclusion substantially distorts the findings from the age-scaled models. In the
time-scaled models, it was depression that was linked with subsequent externalizing
behavior while the reverse direction was evidenced in the age-scaled models.
To further examine this finding for only ages where data coverage was adequate,
a model featuring only ages 11, 12 and 13 was fit. For the full sample, model fit was
89

adequate other than a marginal CFI ( χ
2
=446, df=222, CFI=.895, RMSEA=.047, 90%
CI= .041-.053). Autoregressive effects ranged from .53 to .71. There was no evidence of
a downstream effect of depression on externalizing behavior , but externalizing behavior
at age 11 predicted increases in age-12 depression (b= .659, S.E.=.295, =.25, p<.05).
Models were then fit for girls and boys but, only the boys model, which has a larger n,
converged. All crossed effects were approximately zero, suggesting that girls were
carrying the downstream effect of externalizing behavior on depression.

Specific Aim #2
Latent Growth Model of Depression – Child Report
Before fitting a linear growth model, a no-growth model was fit for the full
sample. This model proposes that an intercept term that differs from zero, variability
around the initial level, and residual variance, adequately accounts for the data. The
model tested depression from ages 9 to 17. However, as there was not adequate coverage
at the tails of the measurement period, the model did not converge. Data from the 9 year
olds, composed of 52 data points and ages 16 and 17, composed of only 19 and 16 data
points, were dropped. The model successfully converged and the no-growth model
adequately captured the variation in depression from age 10 through 15. Model fit was
good ( χ
2
=23.6, df=19, CFI=.973, RMSEA=.023, 90% CI= .000-.050). The intercept
mean was significant as was the variability around the initial level estimate (μ
0
=8.4,
S.E.=.275, p<0.01; σ
2
0
=24.2, S.E.=2.3, p<0.01). Even though the fit of the model was
adequate, a linear growth model was assessed in order to determine if model fit would be
90

improved by the inclusion of linear growth in the structural equations. The growth was
specified as linear by setting the basis coefficients from 0 to 5 for the six data points
measured. The inclusion of the slope factor, which was permitted to covary with the
intercept factor, and have variability around it‟s mean, did not significantly improve
model fit (Δχ²=6.3 on Δdf =3, p>.05). Nevertheless, significant variability in the slope
was observed, suggesting that intercept-only models were not adequate for all subjects.
Accordingly, variability in initial depression level and changes in depression was
investigated by including Time 1 externalizing behavior as a predictor of both variables.
Time 1 data, rather than age 10 data, was utilized as only a minority of the sample was
assessed at age 10. Time 1 Model fit was good ( χ
2
=24.6, df=20, CFI=.985,
RMSEA=.023, 90% CI= .000-.049). Baseline externalizing behavior significantly
predicted both the intercept and slope variables. Higher externalizing behavior, was, as
expected, associated with elevated depression ( =1.5, S.E.=.17, p<.05) but was
negatively related to the slope factor ( =-.18, S.E.=.06, p<.05). This suggests, against the
initial hypothesis, that individuals beginning the study with elevated externalizing
behavior can expect to see downward trajectories in their depression levels. The finding
that higher externalizing subjects initially feature elevated depression suggests a complex
relationship between externalizing behavior and depression during this period of
adolescence.
Interpreting these findings in the absence of relevant covariates is potentially
misleading. Accordingly, variables with established relationships to psychopathology
were added to the models in order to determine if the effect of initial externalizing
91

behavior persists in the presence of additional common risk factors. Harsh parenting,
parental depression, delinquent peer affiliation, and maltreatment experience and gender
were included in the models. Model fit was good ( χ
2
=48.0, df=40, CFI=.979,
RMSEA=.021, 90% CI= .000-.040). Even when adjusting for these additional variables,
initial externalizing behavior was associated with higher initial levels of depression ( =
1.43, S.E.=.17, p<.05). Against predictions, elevated externalizing behavior remained a
significant predictor of declines in depressive symptoms ( =-.18, S.E.=.06, p<.05).
Next, growth models were tested for girls and boys in a multiple-group
framework. Models including all ages did not converge as there was not adequate
coverage for each of the genders. Consequently, a model was fit for ages 11 through 15.
A no-growth model fit the data well ( χ
2
=32.1, df=26, CFI=.967, RMSEA=.023, 90%
CI= .000-.046). Including linear growth factor, however, significantly improved model
fit (Δχ²=17 on Δdf =6, p<.05). The latent mean of the slope factor was non-significant
for girls, suggesting that there was no appreciable increase or decline in depression over
this period of adolescence. Findings for the boys suggested a modest decline in
depression (μ
s
=.-.540, S.E.=.196, p<.05). Estimated growth trajectories are displayed in
Figure 6. Table 24 displays model fit statistics. While the parameter estimates from
freely estimated multi-group models suggest different developmental trajectories for boys
and girls, it is unclear if those trajectories can be statistically distinguished. Constraining
the parameters for initial level and slope to equality did not improve fit over a constrained
model (Δχ²=1.3 on Δdf =2, p>.05). There was no evidence of heterogeneity in growth
trajectories among the boys, but girls featured heterogeneous patterns of growth
92

(σ
2
s
=3.00, S.E.=1.49, p<.05). The next series of model testing assessed if initial
externalizing behavior levels predicted initial depression and change in depression over
time for girls. The slope and intercept factors were regressed on Time 1 externalizing
behavior. As boys featured no variability in growth trajectories, estimating the effect of
exogenous variables is moot. Baseline externalizing behavior and important risk factors
included harsh parenting, parental depression, delinquent peer affiliation, and child
maltreatment were included in the conditional model. The maltreatment variable was not
dichotomous but instead featured variables reflecting the diversity of traumatic
experience. Thus, sexual abuse, physical abuse, emotional abuse and neglect were
considered separately. This strategy allows for more precise adjustment of parameter
estimates by acknowledging the heterogeneity of maltreatment experiences.
1

1
Additional forms of maltreatment were initially modeled, including caretaker incapacity and at-risk
sibling, but models did not converge. However, about 90% of subjects in these categories were positive for
another form of maltreatment included in the models.
93

Figure 6. Estimated Growth Trajectories for Child-Report Depression

With the full range of predictor variables, there were problems with convergence.
Accordingly, the model was tested with a narrower range of ages included – ages 11
through 14. Model fit for the multi-group model with covariates was good ( χ
2
=65.0,
df=52, CFI=.962, RMSEA=.024, 90% CI= .000-.040). Importantly, baseline depression
was a significant predictor of initial level of externalizing behavior in both girls and boys,
even when considered in the presence of several other important predictor variables. The
strength of this effect was the same in both genders.
0
2
4
6
8
10
12
14
11 12 13 14 15
Child-Report Depression
Age
Boys
Depression
Girls
Depression
94

Table 24

Linear Latent Growth Curve Analyses of Child-Report Depression
χ
2
df Δχ²/Δdf Δp RMSEA μ
s
μ
0
σ
2
s
σ
2
0

0,s
Single Group LGM 17.3 16 - - 0.013 -.16(.13) 8.83 (.38)* 1.11 (.56)* 27.34 (4.63)* -1.86(1.44)

Unconditional Multi-Group
LGM 15.2 20 - - 0.000
Freely Estimated μ
s
and μ
0

Girls 0.04(.25) 8.76(.57)* 3.00(1.49)* 30.81(7.81)* -4.67(3.09)
Boys -0.54(.20)* 9.03(.53)* 1.62(.98) 42.30(6.49)* 5.03(2.19)*

μ0 constrained to equality 15.4 21 .2/1 ns 0.000
Girls 0.00(.22) 8.76(.57)* 3.04(1.50)* 30.91(7.81)* -4.72(3.09)
Boys -0.51(.18)* 8.76(.57)* 1.62(.98) 42.17(6.48)* 5.00(2.19)*

μs constrained to equality 19.8 22 4.4/1 .04 0.000 a
Note. a = not reported as there was significant decrement in fit. χ
2
= chi-square value; df= degrees of freedom; Δχ²/Δdf=change in chi-square/change
In degrees of freedom; Δp= effect on model fit probability;; μ
s
= slope mean; μ
0
= intercept mean; σ
2
s
= variance of slope; σ
2
0
= variance of intercept.
RMSEA = root mean square error of approximation.

94

95

Table 25 displays the results of the conditional models. As can be seen, higher baseline
externalizing behavior predicted higher baseline depression. However, against
predictions, it predicted declines among girls.
Although not the focal interest in the analyses, the effects of the covariates are
notable. First, it can be seen that girls appears more sensitive to the effects of
psychosocial risk variables. Whereas depression for boys was not explained by any
variable other than externalizing behavior, additional effects were witnessed for girls.
Harsh parenting and emotional abuse were associated with elevations in baseline
depression, and delinquent peer affiliation at age 10 was predictive of increased
depression over the course of the next few years. In contrast, emotional abuse was
predictive of declines in depression symptoms.
Table 25. Effects of Predictors in Multi-Group Model of Child-Report Depression
Predictors for Girls μ
s
(S.E.) μ
0
(S.E.)
Baseline Externalizing Behavior -0.59(.18)* 2.11(.33)*
Delinquent Peer Affiliation ns 0.09(.03)*
Parental Depression ns ns
Harsh Parenting ns ns
Sexual Abuse ns ns
Physical Abuse ns ns
Emotional Abuse 1.89(.63)* -2.75(1.19)*
Neglect ns ns

Predictors for Boys μ
s
(S.E.)
a
μ
0
(S.E.)
Baseline Externalizing Behavior - 1.35(.22)*
Delinquent Peer Affiliation
-
ns
Parental Depression
-
ns
Harsh Parenting
-
ns
Sexual Abuse
-
ns
Physical Abuse
-
ns
Emotional Abuse
-
ns
Neglect
-
ns
Note. * = p<.05, ** μ
s
.= Mean of slope factor. μ
0
= Mean of
intercept factor. S.E. = standard error of estimate.
a
No variability
in slope among boys so predictive effects were not estimated

96

Latent Growth Model of Externalizing Behavior – Child Report
A similar strategy was followed for modeling externalizing behavior. Before
fitting a linear growth model, a no-growth model was fit for the full sample. This model
proposes that an intercept term that differs from zero, variability around the initial level,
and residual variance, adequately accounts for the data. The model tested externalizing
behavior from ages 9 to 17. However, again, as there was not adequate coverage at the
tails of the measurement period, and the model did not converge. Data from the 9 year
olds, composed of 52 data points and ages 16 and 17, composed of only 19 and 16 data
points, were dropped. The model successfully converged and the no-growth model failed
to capture the variation in externalizing behavior from age 10 through 15. Model fit was
poor ( χ
2
=61.2, df=19, CFI=.414, RMSEA=.07, 90% CI= .051-.090). Next a linear
growth model was assessed in order to determine if model fit would be improved by the
inclusion of linear growth factor. The inclusion of the slope factor, which was permitted
to covary with the intercept factor, and have variability around its mean, significantly
improved model fit (Δχ²=34.6 on Δdf =3, p<.05). Model fit was reasonable, except for a
low CFI ( χ
2
=25.6, df=16, CFI=.866, RMSEA=.036, 90% CI= .000-.062). Thus, the
inclusion of a quadratic factor was explored. This model, however, did better fit the data.
Next, the variability in initial externalizing behavior level and changes in externalizing
behavior by including Time 1 depression as a predictor of both variables was investigated.
The matrices were nonpositive definite. This may have been a function of the
skewness in the indicated variables. Thus, the indicators for externalizing behavior were
97

transformed via log transformation and the model was fit again. The model successfully
converged and fit was good ( χ
2
=21.6, df=20, CFI=.986, RMSEA=.014, 90% CI= .000-
.044). Significant variability in both intercept and slope factors was evidenced, as
displayed in Table 26. Baseline depression significantly predicted both the intercept and
slope variables. Higher depression was, as expected, associated with elevated depression
( =.031, S.E.=.004, p<.05) but was negatively related to the slope factor ( =-.004,
S.E.=.002, p<.05). This suggests that individuals beginning the study with elevated
depression can expect to see downward trajectories in their externalizing behavior. Next,
variables with established relationships to psychopathology were added to the models in
order to determine if the effect of initial externalizing behavior persists in the presence of
additional common risk factors. Harsh parenting, parental depression, delinquent peer
affiliation, and maltreatment experience and gender were included in the models. Even
after adjusting for these additional variables, initial depression remained associated with
higher initial levels of externalizing behavior and predicted declines in externalizing
symptoms over time.
Next, growth models were tested for girls and boys in a multiple-group
framework. Models including all ages did not converge. Consequently, a model was fit
for ages 11 through 14. A no-growth model fit the data poorly. Including linear growth
factor, however, significantly improved model fit (Δχ²=49 on Δdf =6, p<.05). Model fit
was marginal ( χ
2
=27.8, df=10, CFI=.772, RMSEA=.063, 90% CI= .036-.091).
Constraining slope and initial level to equality did not compromise model fit (Δχ²=0.2 on
Δdf =2, p>.05) so results are depicted for the more parsimonious model. Estimated
98

growth trajectories are displayed in Figure 7. The latent mean of the slope factor was
significant, suggesting that boys and girls can expect a modest linear increase in
externalizing behavior during mid-adolescence (μ
s
=.12, S.E.=.02, p<.05). There was no
evidence of heterogeneity in growth trajectories among the girls, although a marginally
precise estimate for boys suggested heterogeneous patterns of growth (σ
2
s
=.04, S.E.=.02,
p=.07). Table 26 displays parameter estimates and model fit statistics. The next series of
models tested if baseline depression predicted initial externalizing behavior for boys or
girls or change in externalizing behavior over time for boys.

Figure 7. Estimated Growth Trajectories for Child-Report Externalizing Behavior.
Note. Patterns of growth are identical for boys and girls.

0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
11 12 13 14
Child-Report Externalizing Behavior
Age
Boys
Externalizing
Behavior
Girls
Externalizing
Behavior
99

Note. a = not reported as there was significant decrement in fit. χ2 = chi-square value; df= degrees of freedom; Δχ²/Δdf=change in chi-square/
change in degrees of freedom; Δp= effect on model fit probability; μ
s
= slope mea μ
0
= intercept mean; σ
2
s= variance of slope; σ
2
0= variance of
intercept. RMSEA = root mean square error of approximation.

Table 26
Linear Latent Growth Curve Analyses of Child-Report Externalizing Behavior
χ
2
df Δχ²/Δdf Δp RMSEA μ
s
μ
0
σ
2
s
σ
2
0

0,s
Single Group LGM 25.6 16 - - 0.036 .23(.05)* .57(.10)* .18 (.08)* .76 (.33)* -.11(.148)

Unconditional Multi-Group LGM 27.8 10 - - 0.063
Freely Estimated μ
s
and μ
0

Girls 0.13(.03)* .31(.05)* .02(.02) .15(.06)* .03(.03)
Boys 0.12(.03)* .33(.05)* .04(.02) .14(.07)* -.01(.03)

μs and μ0 constrained to equality 28.0 12 .2/2 ns 0.054
Girls 0.12(.02)* .32(.03)* .02(.02) .15(.06)* -.03(.03)
Boys 0.12(.02)* .32(.03)* .04(.02) .14(.07)* -.02(.03)

99
100

The slope and intercept factors were first regressed solely on Time 1 depression. Next,
the model added the important covariates including harsh parenting, parental depression,
delinquent peer affiliation, sexual abuse, physical abuse, emotional abuse and neglect.
Model fit for the full model was reasonable except the low CFI, which was perhaps
sensitive to the skewed data ( χ
2
=75.5, df=51, CFI=.866, RMSEA=.033, 90% CI= .015-
.047). Baseline depression, as expected, predicted initial status on externalizing behavior
in girls ( .024, S.E.=.006, p<.05) and boys ( .023, S.E.=.006, p<.05). Higher levels of
depression were not associated with changes in externalizing symptoms in the full model
for boys ( -.004, S.E.=.004, p>.05). The effects of the covariates are included in Table
27.

Table 27.
Effects of Predictors in Multi-Group Model of Child-Report Externalizing Behavior
Predictors for Girls μ
s
(S.E.)
a
μ
0
(S.E.)
Baseline Depression - .024(.006)*
Delinquent Peer Affiliation - ns
Parental Depression - ns
Harsh Parenting - ns
Sexual Abuse - ns
Physical Abuse - ns
Emotional Abuse - ns
Neglect - ns

Predictors for Boys μ
s
(S.E.) μ
0
(S.E.)
Baseline Depression ns .023(.006)*
Delinquent Peer Affiliation ns 0.007(.002)*
Parental Depression ns ns
Harsh Parenting ns ns
Sexual Abuse ns ns
Physical Abuse ns ns
Emotional Abuse ns ns
Neglect ns ns
Note. * = p<.05, μ
s
.= Mean of slope factor. μ
0
= Mean of intercept
factor. S.E. = standard error of estimate. .
a
No variability in slope
among boys so predictive effects were not estimated

101

Latent Growth Model of Depression – Parent Report
Before fitting a linear growth model, a no-growth model was fit for the full
sample. Indicator scores were transformed with a square root transformation plus a
constant. As there was not adequate coverage at the tails of the measurement period, the
model was fit for ages 10 through 15. The model successfully converged although the no-
growth model poorly captured the variation in depression from age 10 through 15
( χ
2
=57.5, df=19, CFI=.802, RMSEA=.067, 90% CI= .047-.087). A linear growth factor
was added to the model in order to determine if model fit would be improved by the
inclusion of a linear growth factor. The inclusion of the slope factor, which was
permitted to covary with the intercept factor, and have variability around it‟s mean, did
not significantly improve model fit (Δχ²=6.8 on Δdf =3, p>.05) and model fit was still
poor. Addition of a quadratic factor led to an inadmissible solution. The error variance
term for the age-15 indicator variable was negative. Gerbing and Anderson (1987)
suggested that negative variances could be constrained to zero without substantially
distorting the model results. This was done, and the model successfully converged, and
appeared to improve model fit substantially (Δχ²=10.3 on Δdf =2, p<.05). Nevertheless,
model fit was still not adequate ( χ
2
=40.4, df=14, CFI=.864, RMSEA=.064, 90%
CI= .042-.088). A cubic term was added and based on the nested Δχ² difference test, the
improvement was substantial (Δχ²=10.2 on Δdf =2, p<.05). However, other model fit
indices suggested that model fit was adversely affected. CFI was .860 and the RMSEA
was .079 (90% CI= .044-.115). Furthermore, the cubic term featured a mean not
different from zero and evidenced no heterogeneity. Thus, it was concluded that the
102

improved model fit was an artifact of the data and not interpretable. As model fit was not
sufficient, no further efforts to include cross-level effects were made. It is possible that
misspecification was specific to one gender. Therefore, analyses were conducted to
determine if an adequate unconditional growth model could be specified for either gender.
Growth models were tested for girls and boys in a multiple-group framework.
Models including all ages did not converge as there was not adequate coverage for each
of the genders. Consequently, a model was fit for a narrower range ages (ages 11-15),
featuring greater coverage. However, results suggested misspecified models. Models
returned inadmissible solutions or featured terrible model fit (i.e. CFI < .80 and RMSEA
> .10). Various attempts were made to impose model constraints, employ indicator
variables that were transformed in various ways, specify different iteration routines, or
restricting the age ranges modeled. However, adequate models were not identified. As
there were no additional parent-report measures of depression, hypotheses could not be
fully explored.

Latent Growth Model of Externalizing Behavior – Parent Report
The observed data for parent-reported externalizing behavior did not exhibit a
clear pattern of growth. A no-growth model was first fit for the full sample. Data from
the 9, 15, 16 and 17 year olds were dropped as there was inadequate data to fit the models.
The model successfully converged but the no-growth model failed to optimally capture
the variation in externalizing behavior from age 10 through 15 ( χ
2
=41.3, df=13, CFI=.912,
RMSEA=.069, 90% CI= .046-.089). The inclusion of a linear slope factor, significantly
103

improved model fit (Δχ²=13.2 on Δdf =3, p<.05). Model fit was reasonable, but not
excellent ( χ
2
=28.1, df=13, CFI=.945, RMSEA=.063, 90% CI= .036-.092) and thus we
explored the inclusion of a quadratic factor. The inclusion of a quadratic factor, however,
did provide a more adequate description of the data. Statistics from the linear growth
model suggest a downward trajectory of growth indicated by a negative latent slope mean
(μ
s
=-.45, S.E.=.18, p<.05). Heterogeneity was observed in both the initial level
(σ
2
0
=76.00, S.E.=9.11, p<.05) and slope factor (σ
2
s
=3.22, S.E.=1.56, p<.05). Next, the
variability in initial externalizing behavior level and changes in externalizing behavior by
including Time 1 depression as a predictor of both variables was investigated. Baseline
depression was a significant predictor of both initial status and trajectory when it was the
sole predictor in the conditional growth model. Depression, as expected, predicted initial
status on externalizing behavior ( 3.56, S.E.=.29, p=.000). Against expectations, but
consistent with findings from child report data, higher levels of depression were
associated with declines in externalizing symptoms ( -.53, S.E.=.10, p=.000). This
effect was observed when depression was modeled as the only predictor variable, but
persisted even in the context of the important covariates including harsh parenting,
parental depression, delinquent peer affiliation, sexual abuse, physical abuse, emotional
abuse and neglect. Inclusion of the covariates did not even diminish the direct effect of
baseline depression on the trajectory of externalizing behavior. The full model fit was
reasonable ( χ
2
=76.4, df=37, CFI=.929, RMSEA=.048, 90% CI= .033-.064). This result
suggests that against predictions, individuals beginning the study with elevated
depression can expect to see downward trajectories in their externalizing behavior.
104

Multi-group models including all ages did not converge so a model was fit for
ages 10 through 13. A no-growth model fit the data adequately (χ
2
=28.9, df=16,
CFI=.949, RMSEA=.042, 90% CI= .015-.067). Including linear growth factor, however,
significantly improved model fit (Δχ²=12.6 on Δdf =6, p<.05) leading to good model fit
( χ
2
=16.3, df=10, CFI=.975, RMSEA=.037, 90% CI= .000-.069). Constraining slope and
initial level to equality did not compromise model fit (Δχ²=0.9 on Δdf =2, p>.05) so
results are depicted for the more parsimonious model. Estimated trajectories are depicted
in Figure 8.

Figure 8. Estimated Growth Trajectories for Parent-Report Externalizing Behavior.
Note. Patterns of growth are identical for boys and girls.

The latent mean of the slope factor was non-significant, suggesting that parent-rated
externalizing behavior is stable over this period of mid-adolescence (μ
s
= -.276, S.E.=.209,
p>.05). There was minimal evidence of heterogeneity in growth trajectories among boys
0
5
10
15
20
10 11 12 13
Parent-Report Externalizing
Behavior
Age
Boys
Externalizing
Behavior
Girls
Externalizing
Behavior
105

and girls, with both variance estimates being marginally precise (p<.10). Model fitting
results and parameter estimates are displayed in Table 29.
Although variance estimates for the slope term for both genders was marginally
precise, conditional growth models were examined. Baseline depression was a significant
predictor of both initial status and trajectory when it was the sole predictor in the
conditional growth model. This effect persisted when depression was modeled in the
context of the important covariates. Baseline depression, as expected, predicted initial
status on externalizing behavior in girls ( 2.533, S.E.=.424, p<.05) and boys ( 4.339,
S.E.=.456, p<.05). Against expectations, but consistent with the child self-report models,
higher levels of depression were associated with declines in externalizing symptoms for
girls ( -.419, S.E.=.183, p<.05) and boys ( -.910, S.E.=.177, p<.05). No other
predictive effects were observed for boys, but for girls, more delinquent peers, parental
depression and harsh parenting was associated with higher initial externalizing status.
The effects of the predictor variables are included in Table 28. Once again, girls
evidenced more sensitivity to the effects of risk factors than do boys.

106

Table 28.
Effects of Predictors in Multi-Group Model of Parent-Rated Externalizing Behavior
Predictors for Girls μ
s
(S.E.) μ
0
(S.E.)
Baseline Depression -0.419(.183)* 2.533(.424)*
Delinquent Peer Affiliation ns 0.077(.037)*
Parental Depression ns .564(.178)*
Harsh Parenting ns .147(.074)*
Sexual Abuse ns ns
Physical Abuse ns ns
Emotional Abuse ns ns
Neglect ns ns

Predictors for Boys μ
s
(S.E.) μ
0
(S.E.)
Baseline Depression -0.910(.177)* 4.339(.456)*
Delinquent Peer Affiliation ns ns
Parental Depression ns ns
Harsh Parenting ns ns
Sexual Abuse ns ns
Physical Abuse ns ns
Emotional Abuse ns ns
Neglect ns ns
Note. * = p<.05. μ
s
.= Mean of slope factor. μ
0
= Mean of intercept
factor. S.E. = standard error of estimate.

107

Table 29.
Linear Latent Growth Analyses of Parent-Report Externalizing Behavior
χ
2
df Δχ²/Δdf RMSEA μ
s
μ
0
σ
2
s
σ
2
0

0,s
Single Group LGM 28.1 10 - 0.063 -.45(.18)* 11.24(.54)* 76.00 (9.17)** 3.22 (1.16)* -8.29(2.83)

Unconditional Multi-Group
LGM
a
16.3 10 - 0.037
Freely Estimated μ
s
and μ
0

Girls -0.27(.31) 10.63(.82)* 5.03(2.83) 82.32(14.04)*
-
13.90(5.26)*
Boys -0.29(.28) 11.42(.78)* 3.87(2.31) 79.19(14.07)* -7.24(4.83)

μ0 and μs constrained to
equality 17.2 12
.9/2
(ns) 0.031
Girls -0.28(.21)

11.02(.56)* 5.06(2.83)

82.64(14.09)*
-
13.98(5.27)*
Boys -0.28(.21) 11.02(.56)* 3.79(2.31) 78.61(14.05)* -7.03(4.83)
Note. Note. a = not reported as there was significant decrement in fit. χ2 = chi-square value; df= degrees of freedom; Δχ²/Δdf=change in chi-square
/change in degrees of freedom; Δp= effect on model fit probability; μ
s
= slope mean; μ
0
= intercept mean; σ
2
s= variance of slope; σ
2
0= variance
of intercept. RMSEA = root mean square error of approximation. * = p<.05.

107

108

Specific Aim #3
Univariate Latent Change Score Models
All of the previous measurement analyses served as the basis for the latent change
models, which were also organized by age. As with the latent growth curve models, a
univariate model of depression did not converge when including ages 9 through 16. This
was predictable, as the tails of the age distribution feature excessively high missingness.
Accordingly, the model was scaled back and successfully converged when ages 10
through 14 were included. As displayed in Table 30, model fit was good with a non-
significant χ
2
value. As expected, initial status on depression was significantly different
from zero, and very similar to the estimate in the latent growth model of depression
(LCS=8.60; LGM=8.83), and featured significant variability. The slope mean was
positive, suggesting that increases in the true score accumulated over time. The
proportional coefficient assessed the relation of prior depression level to the amount of
subsequent change. A negative coefficient, as in the case of depression (
d
= -1.38,
p<.01), suggested that those with elevated depression can be expected to decline at the
following time period. The proportional coefficient is not the only parameter that
conveys information about the relation between prior status and future growth. In a
univariate LCS model, the constant change factor and initial level factors are permitted to
covary. For depression, the slope and initial level factor were positively correlated. This
suggests that an unmeasured variable affects both the initial depression level and the
change expected over the next four years. While the proportional coefficient represents a
form of auto-regressive effects (the effect of t-1 on t), the slope-initial level covariance
109

echoes latent growth models and depicts divergent trajectories that are a function of
initial depression. Although atypical, these conflicting findings are plausible and suggest
a complex relation between initial depression and longitudinal change.
The univariate LCS model for externalizing behavior included the same ages as
the depression model, as a model including ages 9 through 16 did not converge. Model
fit, as displayed in Table 30 was good.

Table 30
Estimates from Univariate Latent Change Score Models for Depression & Externalizing Behavior
Parameter Depression Externalizing Behavior

Fixed Parameters

Initial mean μ
0
8.60* 0.71*
Slope mean μ
s
11.83* -0.42
Proportional -1.38* 0.72*

Random Parameters

Initial variance σ
2
0
18.49* .86*
Slope variance σ
2
s
48.70* 0.61*
Error variance σ
2
e
20.35* 1.73*

Covariances

Slope-Level
s,0
28.57* -0.69

Model Fit

χ
2
(df) 14.79 (13) 22.60 (13)
CFI 0.99 0.85
RMSEA 0.02 0.04
90% RMSEA CI .00-.05 .01-.07
Note. * = p<.05. RMSEA = root mean square error of approximation
CFI = Comparative Fit Index.

Again, as expected, initial externalizing level was different from zero, and featured
significant variability. The slope, or constant change term, was not different from zero,
suggesting no systematic pattern of growth in externalizing behavior. The positive
110

proportional coefficient suggests that elevations in symptoms predict subsequent
increases ( .72, p<.01). This parameter is consistent with a pattern whereby more
distressed individuals experience an exacerbation of symptoms. This finding is
consistent with a class of individuals who are markedly different from their peers with
respect to externalizing behavior (Moffitt, 1993).

Post-Hoc Mixture Model. Following the suggestion from the externalizing behavior LCS
model, a growth mixture model of externalizing behavior was fit as a supplemental
analysis. While latent growth models assume the population distribution of growth
trajectories is best described by a continuous distribution function, mixture models seek
to identify latent groupings of individuals with discrete developmental pathways (Muthen
& Muthen, 2000; Nagin, 1999). The goal of a mixture model is to identify parameters
that define the shape of the discrete trajectories and the probability of membership in a
particular group. It is not a variant of cluster analysis but instead utilizes maximum
likelihood estimation and thus carries the desirable statistical properties of this estimation
procedure (Nagin, 2005). The models implemented here permit inter-individual
differences in the slope and intercept within each trajectory group, which can be
contrasted with Nagin‟s (1999; 2005) approach where within-class homogeneity is
imposed. Decisions regarding the selection of the number of components in growth
mixture models is notoriously difficult (Nagin, 2005) and prone to overextraction (Bauer
& Curran, 2003). The Bayesian information criterion, representing a balance between
explanatory power and parsimony, is commonly employed. In addition to model fit
111

indices, posterior probability estimates can inform the assessment of a model‟s adequacy.
It is important to recall that in these „indirect‟ mixture models, we are not attempting to
recover true groups – girls and boys, for example. Instead, the models represent an
approximation of group membership and, as always, are best understood as awaiting
falsification. McArdle (2009) has suggested that these models best function to highlight
hypotheses for further examination.
The models were fit with the full sample with data from ages 10 through 13.
Analyses suggested two discrete trajectory groups, consistent with the implications from
the LCS model. The first group can be considered a low-symptom group, representing
the vast majority of the sample (91%). A second group is moderately symptomatic and
features an escalating trajectory. This is a period in adolescence where more troubling
developmental pathways may become evident. Of course, more data is needed to
differentiate adolescent-limited delinquency from delinquency that precedes a pattern of
persistent antisocial behavior. Girls and boys were equally represented in the
symptomatic group, but other notable differences emerged. The escalating group was
more likely to have experienced neglect and featured higher levels of affiliation with
delinquent peers. Additionally, they evidenced more depression at each of the four time
points examined in the mixture model. While these findings are intriguing, they are best
considered as suggestive in nature, and can provide the basis for future hypotheses that
can be explicitly tested. Specifically, it would be interesting to investigate if the
mechanisms of comorbidity are similar for individuals in different trajectory groups. In
112

the current example, only one group was symptomatic, but it may be possible that
different clinical populations feature different sources of comorbidity.

Figure 9. Estimated Trajectories of Externalizing Behavior from Growth Mixture Models

Bivariate Latent Change Score Models
The bivariate LCS model was initially fit with five waves of data – from ages 10
through 14. This model did not converge as the number of data points for age 14 was
limited and coverage was not adequate. A model featuring ages 10 through 13 converged,
however, some inadmissible values were identified. The problems appeared to be
restricted to specific variables, rather than global problems with the model. Accordingly,
three parameters were fixed based on estimates from the univariate models. Although in
a different modeling context – dual trajectory growth mixture modeling – importing
0
2
4
6
8
10
10 11 12 13
Child-Report Externalizing
Behavior
Age
Low
Symptom
Group (91%)
Escalating
Symptom
Group (9%)
113

parameters from univariate models to bivariate models has been justified by Wiesner and
Kim (2006). The residual variance term for the first depression indicator at age 10 was
fixed based on the estimate from the univariate model. Additionally, the slope mean and
variance for the externalizing behavior slope variable was fixed. These minor
accommodations led to successful convergence and the absence of implausible parameter
estimates.
As displayed in Table 31, model fit was good. Parameter estimates from the
bivariate model were very similar to the estimates from the two univariate models. An
exception is the finding of a significant negative relationship between baseline
externalizing behavior and its associated constant change term. This finding, however,
may be a function of the necessity of fixing the mean and variance from the univariate
estimates for the externalizing change factor.

114

Table 31
Estimates from Bivariate Latent Change Score Model of Depression & Externalizing Behavior
Fit Indices

χ
2
(df) 26.22(23)
CFI 0.99
RMSEA (90% RMSEA CI) 0.018 (.00-.04)

Parameter Estimates Depression Externalizing Behavior
Initial status means μ
d0,
μ
e0
8.69* .81*
Initial status variances σ
2
d0,
σ
2
e0
36.48* .65*
Slope mean μ
ds
μ
es
9.90* 0
a

Slope variance σ
2
ds,
σ
2
es
36.17* 0.61*
a

Proportional Change
d, e
-1.21* 1.20*
Initial Mean covaried with Slope Mean
ds,d0 es,e0
25.17* -.52*
Depression-Externalizing Behavior associations

Initial Depression covaried with Initial Externalizing Behavior
d0,e0
3.01*
Initial Depression covaried with Externalizing Behavior Slope
d0,es
-0.91
Depression Slope covaried with Initial Externalizing Behavior
ds,e0
2.57*
Crossed Effect of Depression on Subsequent Externalizing Change
d
-.11*
Crossed Effect of Externalizing on Subsequent Depression Change
e
.50
Note. * = p<.05.
a
These parameters were fixed to facilitate convergence. They represent the parameter
estimates from the univariate latent change score model of externalizing behavior. RMSEA = root mean
square error of approximation. CFI = Comparative Fit Index. 90% RMSEA CI = Confidence Interval
for root mean square error of approximation.

The advance of the latent change model lies in its capacity to test complex hypotheses
about the longitudinal relation between two constructs. The relation between depression
and externalizing behavior is manifest in several different parameters. Information is
conveyed by the covariance between initial depression and initial externalizing behavior
(
d0,p0
). As expected, this was positive and significantly different from zero. Baseline
externalizing was positive and associated with the constant change of depression (
ds,p0
=2.57,p<.01), but the reverse was not the case. Of special relevance for the investigation
of the dynamics of comorbidity is the coupling parameter. The coupling parameters,
d

and
e
, differ from the traditional crossed paths in cross-lagged regression models. For
depression, the coupling parameters represent the regression of the change in true score
115

of externalizing behavior between t-1 and t, on the true score of depression at t-1.
Coupling parameters are estimated for each of the gaps between measured periods and
can be tested as constrained to equality or freely estimated. Thus, in the model for
adolescents between 10 and 13, three coupling parameters were estimated for depression
and three parameters were estimated for externalizing behavior.
There was no evidence that externalizing behavior catalyzed subsequent changes
in depression (
e
= 0). In contrast, changes in externalizing behavior were predicted by
previous level of depression. The coupling coefficient was significant and negative, (
d
=
-.11, p<.05), suggesting that elevations in depression at t-1 are associated with decreases
in externalizing behavior at t. Testifying to the robustness of this estimate, the decrement
in model when the coupling paths were constrained to zero, was significant when
compared to a model where the coupling paths were constrained to be equal ( χ
2
=5.5,
df=1, p<.05) and where the paths were freely estimated ( χ
2
=10.6, df=3, p<.05). A
path diagram with significant unstandardized estimates are displayed in Figure 10. Model
fit statistics for the nested bivariate models are displayed in Table 32.
116

Figure 10. Results from Bivariate Latent Change Score Model.
DEP
10
DEP
12

DEP
11

DEP
11

DEP
13

DEP
13

DEP
14

EXT
11

EXT
12
EXT
13
EXT
10

EXT
11

EXT
12

EXT
13

DEP
s

EXT
s

EXT
0

DEP
0

116

117

Table 32

Fit for Bivariate Latent Change Score Models with and Without Coupling Parameters
Δχ² df p Δχ²/Δdf Δp
RMSEA (p
close-fit)
No Coupling Effects 31.8 25 0.16 - - 0.024 (.969)
Depression Invested in Externalizing Behavior 26.3 24 0.34 5.5/1 0.02 0.015 (.989)
Externalizing Behavior Invested in Depression 31.6 24 0.14 .2/1 ns 0.026 (.956)
Full Model (vs. Depression Investment Only) 26.2 23 0.29 .1/1 ns 0.018 (.984)
Note. RMSEA = root mean square error of approximation.

As this finding of a negative relation between depression and externalizing
change over the next year was unexpected given the preponderance of evidence, it was
further investigated. Specifically, it is possible that relations between the constructs exist
only for minimal-symptom participants. Those with very low levels of psychopathology
may feature interesting interrelationships but if all of these relations exist below the
threshold of even sub-syndromal depression, the implications of the finding remain in
doubt. The negative coupling effect was further investigated by segmenting the sample
at the median depression score, and examining model parameters for the more
symptomatic participants. The same model was fit with the higher symptom group and in
this model, the inverse coupling effect became non-significant. Although this could be a
function of power, it appears that the initial finding of the negative coupling effect was a
product of individuals with very low levels of psychopathology.

118

Chapter 4: Discussion
This research sought to elucidate the developmental relations between depression
and externalizing behavior among adolescent boys and girls. Specifically, analyses tested
the plausibility of pathogenic comorbidity – the hypothesis that the frequent co-
occurrence of depression and externalizing behavior is a function of a causal relationship
between the two symptom clusters. The complex relation between the two constructs
was assessed with three statistical methodologies, each featuring a different model of
change. Furthermore, certain analyses utilized both self-report and parent-report data.
The strengths of the study lie in its attention to sampling, multi-informant longitudinal
data, careful measurement of risk factors common to depression and externalizing
behavior, and utilization of statistical models that can test complex developmental
hypotheses. Contrasting results from the different statistical treatments provide some
insight into the conflicting findings within the extant literature.

Assessing Pathogenic Comorbidity
Latent growth curve models assessed the growth trajectories for boys and girls
and test the hypothesis that baseline symptoms in one domain would predict escalating
trajectories in the other domain. Specifically, it was hypothesized that only girls would
evidence a positive relation between baseline externalizing behavior and subsequent
escalation of depressive symptoms. These models also assessed the cross-sectional
relationship between depression and externalizing behavior at baseline. Longitudinal and
concurrent relations reveal important developmental aspects of the relationship. Here,
119

growth trajectories are presumed to be evenly distributed throughout the population and
are expected to vary as a function of both initial level and other exogenous variables.
The optimal shape of the estimated growth trajectory is determined empirically, although
a particular shape may be hypothesized a priori.
As a group, boys reported declines in depression while girls were stable in self-
reported depressive symptoms during early adolescence. However, these effects could
not be statistically distinguished. During puberty, rates of depression between the
genders begin to diverge (Costello, Mustillo, Erkanli, Keeler, & Angold, 2003). While
results provided suggestive evidence for this divergence, an increase was expected for
girls. It is possible, however, that the developmental period in which we would expect
clear gender differences falls after the time window analyzed in the growth models. In
epidemiological surveys from three countries, Wade, Cairney and Pevalin (2002) found
that gender differences only emerged after age 14 while data analyzed in the growth
models ranged from 11 to 15. As has been noted in the literature on depression, age-
related gender divergence may actually serve as a proxy variable for pubertal
development (Angold, Costello, Erkanli, & Worthman, 1999). The girls in the current
sample were relatively early in pubertal development at age 11, when the mean Tanner
stage for breast development was less than two. Angold, Costello and Worthman (1998)
found that only after Tanner stage III do escalations of female depression begin to
emerge. It is likely that escalations in female depression in the present sample will
become evident when the data from late adolescence is collected and analyzed.
120

Among girls, elevated baseline externalizing behavior was associated with
declines in self-rated depression over the next several years. These results persisted even
when adjusting for several non-specific predictors of child psychopathology, including
delinquent peer affiliation, harsh parenting and parental depression and four types of
child maltreatment. The growth factor variance estimate for boys was non-significant
and thus obviated the need to test the effects of exogenous variables while the trajectory
for the girls was flat. These features of the data urge some caution in interpretation. The
result should not over-interpreted to suggest that externalizing behavior serves as a
buffering factor against the development of depression. It is more reasonable to conclude
that no evidence for pathogenic comorbidity emerged in these analyses. This finding
contrasts with some previous findings, where positive causal linkages were identified,
especially among girls (Masten, Roisman, Long, Burt, Obradovic, Riley, Boelcke-
Stennes & Tellegen, 2003; Wiesner, 2003; Little & Garber, 2005; Measelle, Stice &
Hogansen, 2006).
The failure to find a positive prospective relationship between externalizing
behavior and changes in depression does not imply that depression is not prospectively
related to externalizing behavior. Some previous findings have suggested unidirectional
rather than reciprocal relations whereby depression exacerbates externalizing behavior
over time, rather than the reverse (Beyers & Loeber, 2003). Child self-report and parent-
report data provided important insights about the course of externalizing symptoms as a
function of depression. Child report data suggested modest increases in externalizing
behavior, and this pattern was similar for boys and girls. Initial levels of externalizing
121

behavior were generally low, and similar for both genders. For both boys and girls,
baseline depression, as expected, was related to the estimated initial level of externalizing
behavior. This finding replicates the findings from the growth models of depression,
albeit with a different measurement strategy for the variables. As girls did not evidence
heterogeneity of growth trajectories, the exogenous variables were not included in the
model. Among boys, baseline depression was not associated with changes in
externalizing behavior over time. An effect was not witnessed even in the absence of the
seven common risk factors.
Parent-report data on externalizing behavior provided additional evidence against
the pathogenic comorbidity hypothesis. Baseline externalizing behavior was strongly
associated with depression at age ten for boys and girls but associated with declining
trajectories over the next three years. This finding is in conflict with research that has
documented prospective relations between depression and escalations of externalizing
behavior (e.g. Wiesner, 2003).
While no positive prospective relation was identified in either series of models,
cross-sectional associations between depression and externalizing behavior were
consistently strong, and positive for both genders. This concurrent association of
depression and externalizing behavior suggests a complex relationship. These results are
consistent with the work of Lahey et al. (2002) which has challenged Capaldi‟s (1991;
1992) failure model. Data in that clinical sample of boys, found more evidence for
concurrent fluctuations rather than a causal relationship whereby conduct problems
cascade and lead to subsequent conduct problems. In the current data, despite the
122

absence of positive prospective relationship, substantial within-time correlations were
observed. The weighted mean average correlation between depression and externalizing
behavior was .34 – an effect witnessed for both boys and girls. The size of this effect is
almost identical to the cross-sectional associations in the work by both Lahey et al. (2002)
and Fergusson, Lynskey and Horwood (1996).
Thus, latent growth models yielded no evidence for pathogenic comorbidity in
boys or girls. While the existing evidence for boys was weak, it was hypothesized that
girls would show a positive prospective relationship between baseline externalizing
behavior and the development of depression. However, the structure of the data may
have compromised the ability to detect these relationships. Specifically, the growth
trajectories for both child and parent report data were flat or relatively flat making it more
difficult to detect relationships. During this early phase of adolescence – most models
included only participants who were 14.5 years or younger – we would not expect the
same escalations of psychopathology that is more typical during later adolescence (e.g.
Measelle, Stice & Hogansen, 2006). It is possible that with the escalation of symptoms,
new relations would have been uncovered.
Further exploration of the developmental relation between depression and
externalizing behavior was implemented relying on a different depiction of the
mechanisms of longitudinal change. Growth models represent development as a function
of deviations from an initial level, summarized in a growth factor. While borrowing
concepts from growth models to assess non-stationarity, LCS models integrate another
process of underlying change. Specifically, interactions between constructs at adjacent
123

measurement points, as in cross-lagged regression models, are included. However, the
dynamic exchanges between constructs are depicted in more complex and meaningful
ways than in traditional cross-lagged models. The findings from the LCS models served
to augment and elucidate findings from the growth models. The univariate model of
depression provided some clues regarding the surprising finding of an inverse
relationship between baseline externalizing behavior and change in depression. The
proportional coefficient, signifying the relationship between a construct at time t and its
change over the next year, was negative. This suggests that those with elevated
depression can be expected to decline in symptoms one year later – an estimate that could
be interpreted as a form of regression to the mean. Given the concurrent associations
between the two symptom domains, those with elevated depression – and thus elevated
externalizing behavior – would be expected to decline over time. Therefore, the negative
effect witnessed for externalizing behavior in predicting depressive change might be
expected. The apparent inverse relationship between baseline externalizing behavior and
depressive trajectories does not imply a causal linkage.
The bivariate LCS model is well-situated to further investigate the effects of each
symptom cluster on the other, while adjusting for both autoregressive effects and non-
stationarity. The coupling coefficient carries information about the relation between the
latent change and the antecedent latent score for the second variable. This effect has been
partialled from unmeasured factors of change. The bivariate models suggested no
downstream effects of externalizing behavior on depression. Consistent with the findings
from the growth models, a coupling effect was evidenced where depression negatively
124

predicted change over the next year in externalizing behavior. The robustness of this
finding was supported by both the significance of the effect and the improvement in
model fit when the coupling effects were included. This effect was witnessed when the
individual coupling effects were freely estimated and when they were constrained to
equality. Again, it was unclear how to interpret the negative finding, given the
preponderance of evidence for either a positive or null relation. Sample size seemed to
preclude fitting models for each gender separately. However, further investigation
segmented the sample at the median depression score, and examined model parameters
for the more symptomatic participants. These models successfully converged. In this
model, the inverse coupling effect vanished. While it is possible this is merely a function
of power, it appeared that the negative coupling effect was a product of individuals with
very low levels of psychopathology. These results militate against the hypothesis that
comorbidity is a function of causal relationships between the two symptom clusters.
While it would be unwise to entirely dismiss the patterns of very low-symptom
individuals, it has less bearing on understanding developmental psychopathology. Future
investigations would benefit from systematically examining individuals at stratified
symptom levels. It is possible that the mechanisms of diagnostic covariation vary as a
function of the level of psychopathology. Along these lines, both dimensional and
categorical depictions of psychopathology could be utilized in the same study. Although
it is clear that dichotomization sacrifices information, when a coherent point of
dichotomization exists, important relations may be uncovered by utilizing discrete
categories.
125

The failure to identify patterns of pathogenic comorbidity for either gender,
coupled with the robust within time association, suggest at least two possible
explanations. First, the observed covariation between depression and externalizing
behavior may be a function of the nature of risk. Risk factors could predispose to the
individual to both disorders or, it is possible that certain risk factors are specific to one
disorder but are positively associated with risk factors for the other disorder. Fergusson,
Lynsky and Horwood (1996) tested this possibility in a series of structural models with
two measurement points in a large birth cohort. The covariation between depression and
conduct problems during adolescence was primarily explained by common risk factors
including affiliation with delinquent peers, parental attachment, stressful life events, early
conduct problems, cognitive ability, family conflict, family history of offending, and
gender. They found that the correlation between affective and conduct symptoms was
explained – approximately two-thirds – by shared or correlated risk factors for both
symptom domains. There was no evidence of unidirectional or reciprocal causal impacts.
Other lines of research have investigated the genetic influences on symptom covariation.
Twin studies, for example, have expanded the range of common risk factors that can be
studied. Behavioral genetic studies decompose influences into genetic and environmental
components. In one notable study of adults, Kendler, Prescott, Myers, and Neale (2003)
modeled seven internalizing and externalizing disorders and found that genetic risk best
explained the covariation between disorders. In samples of adolescents, comorbidity has
arisen as a function of genetic risks predisposing individuals to both internalizing and
126

externalizing behavior (O‟Connor, McGuire, Reiss, Hetherington & Plomin, 1998; Gjone
& Stevenson, 1997).
In the current sample, the bivariate LCS model highlighted a relevant clue
regarding explanations based upon overlapping or correlated risk factors. In a bivariate
model, one of the parameters depicts the relation between initial externalizing behavior
and change in depression that is unexplained by the either autoregressive effects or
coupling (crossed) effects. The information conveyed by this parameter is absent in
traditional cross-lagged models but carries important meaning. A strong positive
correlation was evidenced between baseline externalizing behavior and unexplained
changes in depression. This correlation suggests that a third and unmeasured variable,
related to both symptom clusters, may account for the cross-sectional associations
between depression and externalizing behavior. Candidate variables might include
genetic factors or the behavioral tendency towards negative emotionality or neuroticism
(Khan, Jacobson, Gardner, Prescott & Kendler, 2005).
A second plausible explanation for the results relates to the latent structure of
psychopathology. This explanation posits that diagnostic formulations fail to fully
acknowledge the commonalities of discrete disorders. Krueger and Markon (2006) have
marshaled substantial evidence demonstrating that the pervasive nature of comorbidity is
a natural consequence of the latent structure of psychopathology. Accordingly, what is
shared by many disorders may be more fundamental than their phenomenological
differences. This evidence arises primarily from factor analytic studies that have
established the relevance of higher-order factors in the explanation of psychopathology
127

(Krueger, 1999; Vollebergh Iedema, Bijl, Graaf, Smit & Ormel, 2001; Slade & Watson,
2006). In their meta-analysis, a correlation of .50 between internalizing and externalizing
factors was derived – large enough to suggest the existence of a global
psychopathological trait (Krueger & Markon, 2006). This explanation is consistent with
the study‟s findings of cross-sectional relationship coupled with the absence of direct
casual relations. Future investigations should seek to explicitly test competing
explanations of comorbidity, although sample size requirements may be prohibitively
large (Rhee, Hewitt, Corley, & Willcutt, 2005).

Methodological Effects in the Assessment of Pathogenic Comorbidity: Conflicting
Findings & Future Directions

Methodological choices can exert significant effects on the evaluation of
pathogenic comorbidity. Cross-lagged regression models have been commonly
employed in this line of research. Despite their utility, the suitability of these models for
evaluating complex developmental hypotheses has been questioned (McArdle, 2009).
The current study fit traditional cross-lagged models, in addition to the latent change
score models. This permits a direct comparison of the conclusions suggested by each
model. While results from the growth curve models were consonant with findings from
the latent change score models, the conclusions reached in the cross-lagged models were
quite different. Specifically, in the time-scaled cross-lagged models, evidence emerged
for a downstream effect of depression on externalizing behavior. Depression at the
second time point predicted escalations in externalizing behavior at the third
measurement point. Gender analyses revealed that girls carried this effect, while the boys
128

featured no crossed effects. Time 1 featured a range of ages – from nine to 13. To
address this variability, age was employed as a covariate. However, this practice may
identify relations that would not be evident in alternative modeling strategies. It is
possible that the crossed path in the cross-lagged model represents a cohort effect rather
than a developmental effect. Scaling a cross-lagged model using age might lead to
different results. For more direct comparison with the LCS models, age-scaled cross-
lagged models were also fit. These models evidenced significant and positive crossed
effects, but unlike the age-scaled models, externalizing behavior was identified as the
cause of subsequent depression, rather than the reverse. This effect was evident at two of
the three crossed paths that were included, and appeared to be carried exclusively by girls.
This is in striking contrast with the findings from all of the other models, which
uncovered null or negative relations between the two symptom clusters. Methodological
factors may be responsible for the discrepancy. Crossed paths are most clearly
interpretable in the context of stationarity. As King, King, McArdle, Shalev, and Doron-
LaMarca (2009) have noted, cross-lagged models may not be optimal when certain
instability in growth patterns are evidenced. Psychopathology during adolescence is an
example where stationarity cannot be assumed. Failing to control for this effect may lead
to the interpretation of a crossed path as a „coupling effect‟ – responsible for change –
when the parameter actually represents as an association with a general pattern of growth.
The present study highlights the conflicting conclusions that may arise on the basis of
modeling decisions. Future investigations are cautioned to specify the process of change
explicitly, and find a modeling strategy that is consistent with that process.
129

Measurement of Maltreatment
The inclusion of maltreatment in the present analyses is important as
maltreatment may serve as a confounding variable in analyses of psychopathology (Heim,
Plotsky & Nemeroff, 2004). This is especially true in studies of comorbidity where
maltreatment often functions as a common risk factor for a range of diverse symptoms.
Some studies of psychopathology in youth omit this variable, despite the over-
representation of maltreated youth in clinical populations, while others employ a
dichotomous variable that attempts to summarize the diverse meanings and impacts of
maltreatment experience. One of the strengths of the current study lies in its extensive
efforts to characterize the heterogeneous experiences of maltreatment. The utility of
carefully differentiating physical abuse, sexual abuse, neglect and emotional abuse was
highlighted by model findings. In the growth curve models and the cross-lagged
regression models, these variables were employed as covariates. Distinct patterns of
relationships emerged between these variables and the measures of psychopathology.
While this study does not represent a systematic effort to catalogue the effects of diverse
maltreatment experiences, it is important to note that some maltreatment variables
evidenced few or no relationships in the models while others featured numerous
associations. It is unclear how a dichotomized variable would have functioned in the
current analyses. It is possible if not probable that some of the noise from the non-
significant variables would have substantially diluted the variable and obscured the
relationships observed in the current study. These findings serve as a caution to future
investigations where maltreatment may play an important developmental role and suggest
130

that greater attention must be paid to the delineation of forms of maltreatment (Trickett,
Mennen, Kim & Sang, 2009).

Limitations and Future Directions
The results of this study must be considered in light of several limitations.
Sampling decisions must always balance different considerations. The use of a non-
clinical sample may limit the generalizability of the current findings. Maltreated
individuals were oversampled, and are likely to present with elevated psychopathology,
but the sample was not gathered on basis of psychological functioning. There were
suggestions in the current sample that symptom relations differed on the basis of clinical
impairment. Specifically, more depressed participants evidenced distinct patterns of
relations between depression and externalizing behavior. Future investigations could
include a larger number of individuals meeting clinical thresholds.
While this study is strengthened by its use of child self-report and parent report
data, depression and externalizing behavior were not assessed with a traditional clinician-
administered diagnostic interview. The absence of interview data complicates
comparisons of the present work with some of the extant literature.
Structural models effectively specify, a priori, a complex set of hypotheses and
assesses their tenability. However, it is important to recognize the tentative nature of
such conclusions. It is commonly noted that even well fitting models do not permit
definitive conclusions. Tomarken and Waller (2003) demonstrate that even in relatively
simple models, there often exist many models with identical implied covariance matrices
131

and thus, identical fit statistics. Additionally, alternative nonequivalent models exist that
might fit the data as well or better than the specified model. Well-fitting lower-order
components may also mask misfit in higher-order factors (Tomarken and Waller, 2003).
As always, we are cautioned by the inductive nature of scientific inquiry.
The use of maximum likelihood estimation with the ordered categorical variables
of the YSR and CBCL is not optimal. Instead, Muthén‟s mean and variance-adjusted
weighted least squares estimator (WLSMV) would be more appropriate given the
distributional violations (Muthén & Muthén, 2006). The multi-group WLSMV approach
analyzes threshold and factor loadings to accommodate for the categorical nature of the
indicators. However, simulation work of Lubke and Muthén (2004) demonstrate that the
risk of applying a continuous model to ordered categorical data in a multi-group
framework is inflation of Type 1 error rates. Apparent rejection of measurement
invariance creates an important interpretive ambiguity as the misfit may reflect actual
measurement non-equivalence, or misapplication of a continuous model. However, in the
current study, the measurement models were generally invariant across groups, thereby
minimizing concerns about implementing a model for continuous indicator variables.
Furthermore, the substantive questions aggregated the data on the basis of the item
analyses, rather than utilizing single items.
A limitation of the current effort is the lack of measurement of important
variables that might predispose to risk for depression and externalizing behavior. While
the absence of important common risk factors poses more risk when evidence emerges
for pathogenic comorbidity, it nevertheless remains important to include a broad array of
132

common risk factors. Childhood temperamental differences, behavioral dysregulation
among parents, and stressful life events would substantially add to the effective
adjustment for common risk factors. Since some of the participants were in the care of a
foster parent, we could not collect data on the functioning of their biological parent. As
developmental psychopathology seeks multiple levels of explanation, it is important to
acknowledge the benefits of genetically informative designs, which have made important
contributions by modeling the covariation between disorders as a function of additive
genetic, shared environment and non-shared environmental factors (Neale & Kendler,
1995). Results from such studies suggest powerful genetic influences even though
environmental factors remain a significant determinant of symptom covariation after
when accounting for the influence of genetic factors (Burt, Krueger, McGue & Iacono,
2003).

Conclusion
To study psychopathology is to study comorbidity. As the pervasiveness of
comorbidity has been well-established, efforts have been directed to explaining the
phenomenon. Many of those efforts have suggested that causal relations singularly
explain the co-occurrence of depression and externalizing behavior. The present study
found that when adequate methodological safeguards are taken, no support was identified
for this claim in either male or female adolescents. Importantly, the failure to adequately
account for risk factors common to both symptom clusters may lead mistakenly to the
conclusion of a causal connection. Additionally, statistical methodologies must be
133

carefully selected as evidence for pathogenic comorbidity may be an artifact of a
particular modeling strategy. Even if strong causal evidence emerged whereby
depression catalyzed increases in externalizing behavior (or the reverse) – it would still
be unwise to conclude that this is the only mechanism of comorbidity. Unfortunately,
efforts have tended to focus on establishing a singular mechanism while not evaluating
the plausibility of multiple mechanisms. If future studies seek to establish the causal
relations between two clusters of symptoms, it would be useful to design the study so that
competing mechanisms can be directly evaluated. The current evidence – strong
concurrent relations in the absence of longitudinal predictions – is consistent with both
the shared-liabilities explanation of comorbidity and the suggestion that diverse
symptomatology is unified by the higher-order latent structure of psychopathology. The
shared-liabilities model has clear implications for prevention research and interventions.
This model suggests that less effort should be made to find unique paths linking specific
risks and specific outcomes. Instead, it concedes that a diverse range of negative
outcomes may arise as a function of the same risk factor. These key risk factors would be
prioritized in preventative interventions. Although treatment implications of a global
psychopathological trait have not been fully explored in the literature, it seems likely that
the current diversification of treatment manuals would be tempered. Interventions might
ignore some of the apparent symptomatic differences, instead, focusing on more general
processes responsible for the range of diverse symptoms.

134

Bibliography

Achenbach, T.M. & Rescorla, L.A. (2001). Manual for the ASEBA school-age forms and
profiles. Burlington: University of Vermont, Center for Children, Youth &
Family.

Achenbach, T.M. (1991). Integrative guide for the 1991 CBCL/4-18, YSR, and TRF
profiles. Burlington: University of Vermont, Department of Psychiatry.

Achenbach, T.M., McConaughy, S.H. & Howell, C.T. (1987). Child/Adolescent
Behavioral and Emotional Problems: Implications of Cross-Informant
Correlations for Situational Specificity. Psychological Bulletin, 191, 213-232.

Allen, J.P. & Land. D. (1999). Attachment in Adolescence. In J. Cassidy & P.R. Shaver
(Eds.). Handbook of Attachment, (pp. 317-335). New York: Guilford.

American Psychiatric Association. (2000). Practice Guidelines for the Treatment of
Psychiatric Disorders. Washington, D.C.: Author.

American Psychiatric Association. (1994). Diagnostic and Statistical Manual of Mental
Disorders, (4th Ed). Washington, D.C.: Author.

Anderson, J. C, & Gerbing, D. W. (1988). Structural equation modeling in practice: A
review and recommended two-step approach. Psychological Bulletin, 103, 411-
423.

Angold A, Costello EJ, & Erkanli A. (1999). Comorbidity. Journal of Child Psychology
and Psychiatry, 40, 57–87.

Angold, A., Costello, E. J., Erkanli, A., & Worthman, C. M. (1999). Pubertal changes in
hormone levels and depression in girls. Psychological Medicine, 29, 1043-1053.

Angold, A., Costello, E. J. & Worthman, C. M. (1998). Puberty and depression: the roles
of age, pubertal status, and pubertal timing. Psychological Medicine 28, 51-61.

Angold, A., & Costello, E. J. (1993). Depressive comorbidity in children and adolescents:
Empirical, theoretical, and methodological issues. American Journal of
Psychiatry, 150, 1779 –1791.

Arbuckle, J.L. (2003). Amos 5.0 [Computer Software]. Chicago: Smallwaters.

Austrian, S.G. (2005). Mental Disorders, Medications, and Clinical Social Work. New
York: Columbia University Press.

135

Bauer, D.J., & Curran, P.J. (2003). Distributional assumptions of growth mixture models:
Implications for overextraction of latent trajectory classes. Psychological Methods,
8, 338–363.

Bavolek, S. J., & Keene, R. J. (1999). Adult-adolescent Parenting Inventory 2:
Administration and Development Handbook. Park City, UT: Family Development
Resources, Inc.

Beauchaine, T. P., Webster-Stratton, C., & Reid, M. (2005). Mediators, moderators, and
predictors of 1-year outcomes among children treated for early-onset conduct
problems: A latent growth curve analysis. Journal of Consulting and Clinical
Psychology, 73, 371–388.

Belanoff, J.K., Rothschild, A.J., Cassidy, F., DeBattista, C., Baulieu, E., Schold, C.,
Schatzberg, A.F. (2002). An open label trial of C-1073 (mifepristone) for
psychotic major depression. Biological Psychiatry, 52, 386–392.

Beyers, J. M., & Loeber, R. (2003). Untangling developmental relations between
depressed mood and delinquency in male adolescents. Journal of Abnormal Child
Psychology, 31, 247–266.

Biederman J., Faraone S., Mick E., Lelon E. (1995) Psychiatric comorbidity among
referred juveniles with major depression: Fact or artifact? Journal of the
American Academy of Child and Adolescent Psychiatry, 34, 579–590.

Bird, H. R., Canino, G., Rubio-Stipec, M., Gould, M. S., Ribera, J., Sesman, M.,
Woodbury, M., Huertas Goldman, S., Pagan, A. Sanchez-Lacay, A., & Moscoso,
M. (1988). Estimates of the prevalence of childhood maladjustment in a
community survey in Puerto Rico. Archives of Genera1 Psychiatry, 45,1120-
1126.

Bosma, H. & Gerlsma, C. (2003). From Early Attachment Relations to the Adolescent
and Adult Organization of Self. In J. Valsiner & K.J. Connolly (Eds.), Handbook
of Developmental Psychology (pp. 450-490). London: Sage.

Burke, J.D., Loeber, R., Lahey, B.B., & Rathouz, P.J. (2005). Developmental transitions
among affective and behavioural disorders in adolescent boys. Journal of Child
Psychology and Psychiatry, 46, 1200–10.

Burke, J.D., Loeber, R. & Birmaher, B. (2002). Oppositional Defiant Disorder and
Conduct Disorder: A Review of the Past 10 Years, Part II. Journal of the
American Academy of Child and Adolescent Psychiatry, 41, 1275–1293.

136

Burt, A. S., Krueger, R. F., McGue, M., & Iacono, W. (2003). Parent-child conflict and
the comorbidity among childhood externalizing disorders. Archives of General
Psychiatry, 60, 505–513.

Byrne, B.M. & Stewart, S.M. (2006). The MACS Approach to Testing for Multigroup
Invariance of a Second-Order Structure: A Walk Through the Process. Structural
Equation Modeling, 13, 287-321.

Byrne, B.M. (2001). Structural Equation Modeling With AMOS. Mahwah, New Jersey:
Lawrence Erlbaum Associates.

Capaldi, D. M. (1991). The co-occurrence of conduct problems and depressive symptoms
in early adolescent boys: I. Familial factors and general adjustment at grade 6.
Development and Psychopathology, 3, 277-300.

Capaldi. D. M. (1992). Co-occurrence of conduct problems and depressive symptoms in
early adolescent boys: IT. A 2-year follow-up at grade 8. Development and
Psychopathology, 4, 125-144.

Chambers, R.A., Taylor, J.R. & Potenza, M.N. (2003). Developmental neurocircuitry
motivation in adolescence: A critical period of addiction vulnerability. American
Journal of Psychiatry, 160, 1041-1053.

Chen, F.F., Sousa, K.H. & West, S.G. (2005). Testing Measurement Invariance of
Second-Order Factor Models. Structural Equation Modeling, 12, 471-492.

Cicchetti, D. & Toth, S.L. (2005). Child Maltreatment. Annual Review of Clinical
Psychology, 1, 409-438.

Cicchetti, D., & Rogosch, F. A. (2002). A developmental psychopathology perspective
on adolescence. Journal of Consulting and Clinical Psychology, 70, 6–20.

Clark, L. A., Watson, D., & Reynolds, S. (1995). Diagnosis and classification of
psychopathology: Challenges to the current system and future directions. Annual
Review of Psychology, 46, 121-153.

Clarkin, J. E, & Kendall, P. C. (1992). Comorbidity and treatment planning: Summary
and future directions. Journal of Consulting and Clinical Psychology, 60, 904-908.

Cole, D. A., Nolen-Hoeksema, S., Girgus, J., & Paul, G. (2006). Stress exposure and
stress generation in child and adolescent depression: A latent trait-state-error
approach to longitudinal analyses. Journal of Abnormal Psychology, 115, 40 –51.

137

Cole, D. A., Truglio, R., & Peeke, L. (1997). Relation between symptoms of anxiety and
depression in children: A muln'trait-multirnethod-multigroup assessment. Journal
of Consulting and Clinical Psychology, 65, 110-119.

Collins, L. M., Schafer, J. L., & Kam, C. M. (2001). A comparison of inclusive and
restrictive strategies in modern missing data procedures. Psychological Methods,
6, 330–351.

Conners, N. A., Whiteside-Mansell, L., Deere, D., Ledet, T., & Edwards, M.C. (2006).
Measuring the potential for child maltreatment: The reliability and validity of the
Adult Adolescent Parenting Inventory–2. Child Abuse & Neglect, 30, 39–53.

Costello, E.J., Mustillo, S., Erkanli, A., Keeler, G., & Angold, A. (2003). Prevalence and
development of psychiatric disorders in childhood and adolescence. Archives of
General Psychiatry, 60, 837–844.

Cox, C. E., Kotch, J. B., & Everson, M. D. (2003). A longitudinal study of modifying
influence in the relationship between domestic violence and child maltreatment.
Journal of Family Violence, 18, 5-17.

Derogatis LR, Melisaratos N. (1983). The Brief Symptom Inventory: an introductory
report. Psychological Medicine, 12, 595-605.

Dodge, K. A. (1983). Behavioral antecedents of peer social status. Child Development,
54, 1386-1399.

Drabick, D. A., Gadow, K. D., & Sprafkin, J. (2006). Co-occurrence of conduct disorder
and depression in a clinic-based sample of boys with ADHD. Journal of Child
Psychology and Psychiatry, 47 , 766–774.

Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., & Strahan, E. J. (1999). Evaluating
the use of exploratory factor analysis in psychological research. Psychological
Methods, 4, 272–299.

Feehan, M., McGee, R., Raja, S. N., &Williams, S. (1994). DSM-III-R disorders in New
Zealand 18 year olds. Autralian and New Zealand Journal of Psychiatry, 87-89.

Feinstein, A.R. (1970). The pre-therapeutic classification of co-morbidity in chronic
disease. Journal of Chronic Diseases, 23, 455-468.

Fergusson, D. M., Homwood, J., & Lynskey, M. T. (1993). Prevalence and
comorbidity of DSM-III-R diagnoses in a birth cohort of 15 year olds. Journal of
the American Academy of Child and Adolescent Psychiatry, 32, 1127-1134.

138

Ferrer, E., & McArdle, J. J. (2003). Alternative structural models for multivariate
longitudinal data analysis. Structural Equation Modeling, 10, 493–524.

Ferrer, E. & McArdle, J.J. (2004). An experimental analysis of dynamic hypotheses about
cognitive abilities and achievement from childhood to early adulthood.
Developmental Psychology,40, 935–952.

Ferrer, E., Hamagami, F.,&McArdle, J. J. (2004). Modeling latent growth curves with
incomplete data using different types of structural equation modeling and
multilevel software. Structural Equation Modeling, 11, 452–483.

Ferrer, E., McArdle, J.J., Shaywitz, B.A., Holahan, J.M., Marchione, K., & Shaywitz,
S.E. (2007). Longitudinal models of developmental dynamics between reading
and cognition from childhood to adolescence. Developmental Psychology, 43,
1460–1473.

Flett, G.L., Vredenburg, K. & Krames, L. (1997). The Continuity of Depression in
Clinical and Nonclinical Samples. Psychological Bulletin, 121, 395-416.

Ge, X., Best, K.M., Conger, R.D., & Simons, R.L. (1996). Parenting Behaviors and the
Occurrence and Co-Occurrence of Adolescent Depressive Symptoms and
Conduct Problems. Developmental Psychology, 32, 717-731.

Gjone, H., & Stevenson, J. (1997). A longitudinal twin study of temperament and
behavior problems: Common genetic or environmental influences? Journal of the
American Academy of Child and Adolescent Psychiatry, 36, 1448–1456.

Goodyer, I. M., Herbert, J., Tamplin, A. & Altham, P. M. E. (2000). First-episode major
depression in adolescents: affective, cognitive and endocrine characteristics of
risk status and predictors of onset. British Journal of Psychiatry, 176, 142–149.

Goodyer, I.M. (2008) Emanuel Miller lecture: early onset depressions-meaning,
mechanisms and processes. Journal of Child Psychology & Psychiatry, 49, 1239-
1256.

Greene, R. W., Biederman, J., Zerwas, S., Monuteaux, M. C., Goring, J. C., Faraone, S.
V. (2002). Psychiatric comorbidity, family dysfunction, and social impairment in
referred youth with oppositional defiant disorder. American Journal of
Psychiatry,, 159, 1214-1224.

Gross, H.E., Shaw, D.S., Burwell, R.A., & Nagin, D.S. (2009) Transactional processes in
child disruptive behavior and maternal depression: A longitudinal study from
early childhood to adolescence. Development and Psychopathology, 21, 139-156.

139

Helzer, J.E., Kraemer, H.C., Krueger, R.F. (2006). The feasibility and need for
dimensional psychiatric diagnoses. Psychological Medicine, 36, 1671-1680.

Hollon, S. D., DeRubeis, R. J., Shelton, R. C., Amsterdam, J. D., Salomon, R. M., &
O‟Reardon, J. P., et al. (2005). Prevention of relapse following cognitive therapy
vs. medications in moderate to severe depression. Archives of General Psychiatry,
62, 417–422

Heim, C., Newport, D.J., Heit, S., Graham, Y.P., Wilcox, M., Bonsall, R., Miller, A.H.,
& Nemeroff, C.B. (2000). Pituitary–adrenal and autonomic responses to stress in
women after sexual and physical abuse in childhood. JAMA, 284, 592–597.

Heim, C. & Nemeroff, C.B. (2001). The Role of Childhood Trauma in the Neurobiology
of Mood and Anxiety Disorders: Preclinical and Clinical Studies. Biological
Psychiatry, 49, 1023-1039.

Heim, C., Newport, D.J., Bonsall, R., Miller, A.H., & Nemeroff, C.B. (2001). Altered
pituitary–adrenal axis responses to provocative challenge tests in adult survivors
of childhood abuse. American Journal of Psychiatry, 158, 575–581.

Heim, C., Plotsky, P.M., Nemeroff, C.B., 2004. Importance of studying the contributions
of early adverse experience to neurobiological findings in depression.
Neuropsychopharmacology, 29, 641–648.

Heim, C., Newport, D.J., Mletzko, T., Miller, A.H., & Nemeroff, C.B. (2008). The link
between childhood trauma and depression: insights from HPA axis studies in
humans. Psychoneuroendocrinology, 33, 693–710

Hinshaw, S.P. (1992). Externalizing behavior problems and academic underachievement
in childhood and adolescence: causal relationships and underlying mechanisms.
Psychological Bulletin, 111, 127–55.

Horn, J.L. & McArdle, J.J. (1992). A practical and theoretical guide to measurement
invariance in aging research. Experimental Aging Research, 18, 117-144.decade.
Journal of Abnormal Child Psychology, 31, 293-300.

Hofstra, M.B., van der Ende, J. & Verhulst, F.C. (2002). Pathways of self-reported
problem behaviors from adolescence into adulthood. American Journal of
Psychiatry, 159, 401-408.

Ialongo, N.S., Edelsohn, G., & Kellam, S.G. (2001). A Further Look at the Prognostic
Power of Young Children‟s Reports of Depressed Mood and Feelings. Child
Development, 72, 736-747.

140

Judd, C. M., McClelland, G. H., & Culhane, S. E. (1995). Data analysis: Continuing
issues in the everyday analysis of psychological data. Annual Review of
Psychology, 46, 433–465.

Keiley, M.K., Lofthouse, N., Bates, J.E., Dodge, K.A. & Pettit, G.S. (2003). Differential
Risks of Covarying and Pure Components in Mother and Teacher Reports of
Externalizing and Internalizing Behavior Across Ages 5 to 14. Journal of
Abnormal Child Psychology, 31, 267-283.

Kendler K.S. (2005). „A gene for‟: the nature of gene action in psychiatric disorders.
American Journal of Psychiatry, 162, 1243–1252.

Kendler, K.S., Prescott, C.A., Myers, J., & Neale, M.C. (2003). The Structure of Genetic
andEnvironmental Risk Factors for Common Psychiatric and Substance Use
Disorders in Men and Women. Archives of General Psychiatry, 60, 929-937.

Kessler, R.C., Berglund, P., Demler, O., Jin, R., Merikangas, K.R., Walters, E. (2005).
Lifetime Prevalence and Age-of-Onset Distributions of DSM-IV Disorders in the
National Comorbidity Survey Replication. Archives of General Psychiatry, 62,
593-602.

Kessler, R.C., DuPont, R.L., Berglund, P., & Wittchen, H. (1999). Impairment in Pure
and Comorbid Generalized Anxiety Disorder and Major Depression at 12 Months
in Two National Surveys. American Journal of Psychiatry, 156, 1915-1923.

Kessler R.C., McGonagle K.A., Zhao S., Nelson C.B., Hughes M., Eshleman S.,
Wittchen, Kendler, K.S. (1994): Lifetime and 12-month prevalence of DSM-III-R
psychiatric disorders in the United States: Results from the National Comorbidity
Survey. Archives of General Psychiatry, 51, 8–19.

Khan, A.A., Jacobson, K.C., Gardner, C.O., Prescott, C.A., Kendler, K.S., 2005.
Personality and comorbidity of common psychiatric disorders. British Journal of
Psychiatry, 186, 190–196.

Kim, J. K., Conger, R. D., Elder, G. H., Jr, & Lorenz, F. O. (2003). Reciprocal influences
between stressful life events and adolescent internalizing and externalizing
problems. Child Development, 74, 127–143.

King, D.W., King, L.A., McArdle, J.J., Shalev, A.Y., & Doron-LaMarca, S. (2009).
Sequential Temporal Dependencies in Associations between Symptoms of
Depression and Posttraumatic Stress Disorder: An Application of Bivariate Latent
Difference Score Structural Equation Modeling. Multivariate Behavioral
Research, 44, 437-464.

141

Kirschbaum, C., Pirke, K., & Hellhammer, D.H. (1993). The „„Trier Social Stress Test‟‟:
a tool for investigating psychobiological stress responses in a laboratory setting.
Neuropsychobiology, 28, 76–81.

Klein, D.N. & Riso, L.P. (1993). Psychiatric disorders: problems of boundaries and
comorbidity. In: Costello, C.G. (Ed.), Basic Issues in Psychopathology. Guilford,
New York, pp. 19–66.

Kline, R.B. (2004). Principles and Practice of Structural Equation Modeling (2
nd
Ed.).
New York: Guilford.

Kovacs, M. (1981). Rating scales to assess depression in school-age children. Acta
Paedopsychiatrica, 46, 305-315.

Kraemer, H.C., Measelle, J.R., Ablow, J.C., Essex, M.J., Boyce, W.T., & Kupfer, D.J.
(2003). A New Approach to Integrating Data From Multiple Informants in
Psychiatric Assessment and Research: Mixing and Maxing Contexts and
Perspectives. American Journal of Psychiatry, 160, 1566-1577.

Krueger, R. F. (1999). The structure of common mental disorders. Archives of General
Psychiatry, 56, 921–926.

Krueger, R. F., & Markon, K. E. (2006). Reinterpreting comorbidity: A model-based
approach to understanding and classifying psychopathology. Annual Review of
Clinical Psychology, 2, 111-133.

Krueger, R. F., Caspi, A., Moffitt, T. E., & Silva, P. A. (1998). The structure and stability
of common mental disorders (DSM-III-R): A longitudinal-epidemiological study.
Journal of Abnormal Psychology , 107, 216-227

Krueger, R.F., Hicks, B.M., Patrick, C.J., Carlson, S.R., Iacono, W.R. & McGue, M.
(2002).Etiologic Connections Among Substance Abuse, Antisocial Behavior, and
Personality: Modeling the Externalizing Spectrum. Journal of Abnormal
Psychology, 111, 411-424.

Lahey, B.B., Loeber, R., Burke, J., Rathouz, P.J., & McBurnett, K. (2002). Waxing and
waning in concert: Dynamic comorbidity of conduct disorder with other
disruptive and emotional problems over 7years among clinic-referred boys.
Journal of Abnormal Psychology, 111, 556–567.

142

Lambert, M.C., Schmitt, N., Samms-Vaughan, M.E. An, J.S., Fairclough, M. & Nutter,
C.A. (2003). Is it Prudent to Administer All Items for Each Child Behavior
Checklist Cross-Informant Syndrome? Evaluating the Psychometric
Properties of the Youth Self-Report Dimensions With Confirmatory
Factor Analysis and Item Response Theory. Psychological Assessment,
15, 550-568.

Levy, J.C., Deykin, E.Y. (1989). Suicidality, Depression and Substance Abuse in
Adolescence. American Journal of Psychiatry, 146, 1462-1468.

Lilienfeld, S.O. (2003). Comorbidity between and within childhood externalizing and
internalizing disorders: reflections and directions. Journal of Abnormal Child
Psychology, 31, 285–16.

Lilienfeld, S.O., Waldman, I.D. & Israel, A.C. (1994). A Critical Examination of the Use
of the Term and Concept of Comorbidity in Psychopathology Research. Clinical
Psychology: Science and Practice, 1, 73-81.

Little, S. A., & Garber, J. (2005). The role of social stressors and interpersonal
orientation in explaining the longitudinal relation between externalizing and
depressive symptoms. Journal of Abnormal Psychology, 114, 432– 443.

Little, T.D., Cunningham, W.A., Shahar, G. & Widaman, K.F. (2002). To Parcel or Not
to Parcel: Exploring the Question, Weighing the Merits. Structural Equation
Modeling, 9, 151-173.

Loeber, R., Burke, J.D., Lahey, B.B., Winters, A., & Zera, M. (2000). Oppositional
defiant and conduct disorder: A review of the past 10 years, Part I. Journal of the
American Academy of Child and Adolescent Psychiatry, 39, 1468–1484.

Loeber, R., & Keenan, K. (1994). Interaction between conduct disorder and its comorbid
conditions: Effects of age and gender. Clinical Psychology Review, 14, 497-523.

Lubke, G.H. & Muthén, B.O. (2004). Applying Multigroup Confirmatory Factor Models
for Continuous Outcomes to Likert Scale Data Complicates Meaningful Group
Comparisons. Structural Equation Modeling, 11, 514-534.

Markon, K.E., & Krueger, R.F. (2005). Categorical and continuous models of liability to
externalizing disorders: A direct comparison in NESARC. Archives of General
Psychiatry, 62, 1352–1359.

Martens, M.P., & Haase, R.F. (2006). Advanced applications of structural equation
modeling is counseling psychology research. The Counseling Psychologist, 34,
878–911.
143

Masten, A.S., Roisman, G.I., Long, J.D., Burt, K.B., Obradovic, J., Riley, J.R., Boelcke-
Stennes, K., & Tellegen, A. (2005). Developmental cascades: Linking academic
achievement and externalizing and internalizing symptoms over 20 years.
Developmental Psychology, 41, 733-746.

McArdle, J.J., & Epstein, D. (1987). Latent growth curves within developmental
structural equation models. Child Development, 58, 110–133.

McArdle, J.J. & Hamagami, F. (1992). Modeling incomplete longitudinal and cross-
sectional data using latent growth structural models. Experimental Aging
Research, 18, 145-166.

McArdle, J. J. (1996). Current directions in structural factor analysis. Current Directions
in Psychological Science, 5, 10–17.

McArdle, J.J., & Grimm, K. (2008). Five Steps in Latent Growth Curve Modeling.
Unpublished data. University of Southern California.

McArdle, J.J. (2009). Latent variable modeling of differences and changes with
longitudinal data. Annual Review of Psychology, 60, 577–605.

McLeer, S.V., Callaghan, M. Henry, D., & Wallen, J. (1994). Psychiatric Disorders in
Sexually Abused Children. Journal of the American Academy of Child and
Adolescent Psychiatry, 33, 313-319.

McClelland, G.H. (2000). Nasty Data: Unruly, Ill-Mannered Observations can Ruin Your
Analysis. In H.T. Reis, & C.M. Judd (Eds.) Handbook of Research Methods in
Social and Personality Psychology (pp. 393-411). Cambridge, England:
Cambridge University Press.

Measelle, J.R., Stice, E., & Hogansen J.M. (2006). Developmental trajectories of co-
occurring depressive, eating, antisocial, and substance abuse problems in female
adolescents. Journal of Abnormal Psychology, 115, 524–538.

Metha, P. D., & West, S. G. (2000). Putting the individual back into individual growth
curves. Psychological Methods, 5, 23–43.

Meller, W.H. & Borschardt, C.M. (1996). Comorbidity of major depression and conduct
disorder. Journal of Affective Disorders, 39, 123-126.

Meyers, K. & Winters, N.C. (2002). Ten-Year Review of Rating Scales. II: Scales for
Internalizing Disorders. Journal of the American Academy of Child and
Adolescent Psychiatry, 41, 634-659.
144

Miller, G.E., Chen, E., & Zhou, E.S. (2007). If it goes up, must it come down? Chronic
stress and the hypothalamic-pituitary-adrenocortical axis in humans.
Psychological Bulletin, 133, 25–45.

Mineka. S.. Watson, D.. & Clark. L. A. (1998). Comorbidity of anxiety and unipolar
mood disorders. Annual Review of Psychology, 49, 377-412.

Moffitt, T. (1993). Adolescence-limited and life-course-persistent antisocial behavior: A
developmental taxonomy. Psychological Review, 100, 674–701.

Muthén, L.K. & Muthén, B.O. (2006). Mplus User’s Guide. Fourth Edition, Los Angeles,
CA; Muthén & Muthén.

Muthén, B.O., & Muthén, L.K. (2000). Integrating person-centered and variable-centered
analysis: Growth mixture modeling with latent trajectory classes. Alcoholism:
Clinical and Experimental Research, 24, 882–891.

Myers, M.G., Stewart, D.G. & Brown, S.A. (1998). Progression From Conduct Disorder
to Antisocial Personality Disorder Following Treatment for Adolescent Substance
Abuse. American Journal of Psychiatry, 155, 479-485.

Nagin, D.S. (2005). Group-based modeling of development. Cambridge, MA: Harvard
University Press.

Nagin, D.S. (1999). Analyzing developmental trajectories: A semi-parametric, group-
based approach. Psychological Methods, 4, 139–177.

Neale, M.C., Kendler, K.S., 1995. Models of comorbidity for multi-factorial disorders.
American Journal of Human Genetics, 57, 935–953.

Negriff, S., Fung, M., & Trickett, P. (2008). Self-rated pubertal development, depressive
symptoms and delinquency: measurement issues and moderation by gender and
maltreatment. Journal of Youth and Adolescence, 37, 736–746.

Nelson, J.C., & Davis, J.M. (1997). DST studies in psychotic depression: A meta-
analysis. American Journal of Psychiatry, 154, 1497-1503.

O‟Connor, T.G., McGuire, S., Reiss, D., Hetherington, E.M. & Plomin, R. (1998). Co-
occurrence of Depressive Symptoms and Antisocial Behavior in Adolescence: A
Common Genetic Liability. Journal of Abnormal Psychology, 107, 27-37.

145

O‟Connor, T. G., Neiderhiser, J. M., Reiss, D., Hetherington, E. M., & Plomin, R. (1998).
Genetic contributions to continuity, change, and co-occurrence of antisocial and
depressive symptoms in adolescence. Journal of Child Psychology and Psychiatry
and Allied Disciplines, 39, 323–336.

O‟Keefe, M., Mennen, F.E., & Lane, C.J. (2006). An Examination of the Factor Structure
for the Youth Self Report in a Multiethnic Population. Research on Social Work
Practice, 16, 315-325.

Preacher, K.J., Wichman, A.L., MacCallum, R.C., & Briggs, N.E. (2008). Latent growth
curve modeling. Thousand Oaks, CA: Sage.

Raine, A., Reynolds, C., Venables, P.H., & Mednick, S.A. (1997). Resting heart rate, skin
conductance orienting, and physique. In A. Raine, P.A. Brennan, D.P. Farrington,
& S.A. Mednick (Eds.), Biosocial bases of violence (pp. 107-126). New York:
Plenum.

Raine, A., Reynolds, C., Venables, P.H., & Mednick, S.A., & Farrington, D.P. (1998).
Fearlessness, stimulation-seeking, and large body size at age 3 years as early
predispositions to childhood aggression at age 11 years. Archives of General
Psychiatry, 55,745- 751.

Raine, A. (2002). Biosocial studies of antisocial and violent behavior in children and
adults: A review. Journal of Abnormal Child Psychology, 30, 311–326.

Reich, D.B. & Zanarini, M.C. (2001). Developmental aspects of borderline personality
disorder. Harvard Review of Psychiatry, 9, 294-301.

Renouf, A.G., Kovacs, M., & Mukerji, P. (1997). Relationship of depressive conduct and
comorbid disorders in childhood. Journal of the American Academy of Child and
Adolescent Psychiatry, 36, 998–1004.

Rhee, S.H., Hewitt, J.K., Corley, R.P., & Willcutt, E.G. (2005). Testing hypotheses
regarding the causes of comorbidity: examining the underlying deficits of
comorbid disorders. Journal of Abnormal Pyschology, 114, 346–362.

Rohde, P., Lewinsohn, P. M., & Seeley, J. R. (1991). Comorbidity of unipolar
depression: II. Comorbidity with other mental disorders in adolescents and adults.
Journal of Abnormal Psychology, 100, 214-222.

Rosen, G.M., & Davison, G.C. (2003). Psychology should list empirically supported
principles of change (ESPs) and not credentialed trademarked therapies or other
treatment packages. Behavior Modification, 27, 300–312.

146

Rowe, R., Maughan, B., & Eley, T. C. (2006). Links between antisocial behavior and
depressed mood: The role of life events and attributional style. Journal of
Abnormal Child Psychology, 34, 293–302.

Rutter, M., Caspi, A., & Moos, R. H. (2003). Using sex differences in psychopathology
to study causal mechanisms: Unifying issues and research strategies. Journal of
Child Psychology and Psychiatry and Allied Disciplines, 44, 1092–1115.

Sackett D.L., Richardson W.S., Rosenberg W.S. & Haynes, R.B. (1997). Evidence-based
medicine. New York: Churchill Livingstone.

Sapolsky, R. M. (2000). Glucocorticoids and hippocampalatrophy in neuropsychiatric
disorders. Archives of General Psychiatry, 57, 925–935.

Schafer, J. (1999). NORM: Multiple imputation of incomplete multivariate data under a
normal model, Version 2.03. Department of Statistics, The Pennsylvania State
University.

Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the art.
Psychological Methods, 7, 147–177.

Schafer, J. L., & Olsen, M. K. (1998). Multiple imputation for multivariate missing-data
problems: A data analyst‟s perspective. Multivariate Behavioral Research, 33,
545-571.

Skodol, A. E., Gunderson, J. G., Pfohl, B., Widiger, T. A., Livesley, W. J., & Siever, L. J.
(2002). The borderline diagnosis I: Psychopathology, comorbidity, and
personality structure. Biological Psychiatry, 51, 936 –950.

Slade, T. & Watson, D. (2006). The structure of common DSM-IV and ICD-10 mental
disorders in the Australian general population. Psychological Medicine, 36, 1593-
1600.

Simic, M. & Fombonne, E. (2001). Depressive conduct disorder: symptom patterns and
correlates in referred children and adolescents. Journal of Affective Disorders, 62,
175-185.

Smith, C.A. & Thornberry, T. (1995). The Relationship between Child Maltreatment and
Adolescent involvement in Delinquency. Criminology, 33, 451-481.

Sroufe, L. A. (1979). The coherence of individual development: Early care, attachment,
and subsequent developmental issues. American Psychologist, 34, 834-841.

147

Taylor, S. E., Repetti, R. L., & Seeman, T. L. (1997). Health psychology: What is an
unhealthy environment and how does it get under the skin? Annual Review of
Psychology, 48, 411-447.

Tomarken, A.J. & Waller, N.G. (2003). Potential Problems with “Well Fitting” Models.
Journal of Abnormal Psychology, 112, 578-595.

Trickett, P.K., Mennen, F.E., Kim, K., & Sang, J. (2009). Emotional abuse in a sample
of multiply maltreated, urban young adolescents: Issues of definition and
identification. Child Abuse and Neglect, 33, 27-35.

Vandenberg, R.J. & Lance, C.E. (2000). A Review and Synthesis of the Measurement
Invariance Literature: Suggestions, Practices, and Recommendations for
Organizational Research. Organizational Research Methods, 3, 4-70.

Vollebergh, W.A.M., Iedema J., Bijl R.V., de Graaf R., Smit F., Ormel J. (2001). The
structure and stability of common mental disorders: the NEMESIS study.
Archives of General Psychiatry, 58, 597–603.

Wade T.J., Cairney, J.,& Pevalin, D.J. (2002). Emergence of gender differences in
depression during adolescence: national panel results from three countries.
Journal of the American Academy of Child and Adolescent Psychiatry, 41, 190-
198.

Watson, D. (2005). Rethinking the mood and anxiety disorders: A quantitative
hierarchichal model for DSM-V. Journal of Abnormal Psychology, 114, 522-536.

Watson, D., & Clark, L. A. (1992). Affects separable and inseparable: On the hierarchical
arrangement of the negative affects. Journal of Personality and Social
Psychology, 62, 489-505.

Weiss, E.L., Longhurst, J.G., & Mazure, C.M. (1999). Childhood sexual abuse as a
risk factor for depression in women: psychosocial and neurobiological
correlates. American Journal of Psychiatry, 156, 816-828.

Westen, D., Novotny, C. M., & Thompson-Brenner, H. (2004). The empirical status of
empirically supported psychotherapies: Assumptions, findings, and reporting in
controlled clinical trials. Psychological Bulletin, 130, 631–663.

Whitbeck, L.B., Hoyt, D.R., & Bao, W. (2000). Depressive Symptoms and Co-occurring
Depressive Symptoms, Substance Abuse, and Conduct Problems among Runaway
and Homeless Adolescents. Child Development, 71, 721-732.

148

Widom, C.S., DuMont, K. & Czaja, S.J. (2007). A Prospective Investigation of Major
Depressive Disorder and Comorbidity in Abused and Neglected Children Grown
Up. Archives of General Psychiatry, 64, 49-56.

Wolff, J.C. & Ollendick, T.H. (2006). The Comorbidity of Conduct Problems and
Depression in Childhood and Adolescence. Clinical Child and Family Psychology
Review, 9, 201-220.

Wiesner, M., & Kim, H.K. (2006). Co-occurring delinquency and depressive symptoms
of adolescent boys and girls: A dual trajectory modeling approach. Developmental
Psychology, 42, 1220-1235.

Wiesner, M. (2003). A longitudinal latent variable analysis of reciprocal relations
between depressive symptoms and delinquency during adolescence. Journal of
Abnormal Psychology, 112, 633–645.

Zoccolillo, M., & Rogers, K. (1991). Characteristics and outcome of hospitalized
adolescent girls with conduct disorder. Journal of the American Academy of Child
and Adolescent Psychiatry, 30, 973-981.

Abstract (if available)

Abstract To study psychopathology is to study comorbidity. In both clinical and representative samples, the co-occurrence of two or more disorders is the rule, rather than the exception. While the virtual ubiquity of psychiatric comorbidity has been established, efforts have been redirected to explaining the phenomenon. The current study investigated the sources of covariation between depression and externalizing behavior among adolescent boys and girls. Specifically, it examined the plausibility that one symptom cluster exerted a causal effect on the other cluster of symptoms – a hypothesis known as pathogenic comorbidity. This hypothesis has more support among girls, and for the downstream effects of externalizing behavior on depression. A gender-balanced group of 454 young adolescents was sampled, with 303 of those having met specific criteria for the recent experience of maltreatment. The youth and their caretakers were assessed three times over the period of the next two and a half years. Pathogenic comorbidity was examined utilizing different statistical methods, each with their own unique model of change. Findings employing child and parent-report data were compared.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Examining the longitudinal relationships between community violence exposure and aggressive behavior among a sample of maltreated and non-maltreated adolescents

PDF

The relationship between pubertal timing and delinquent behavior in maltreated male and female adolescents

PDF

Social network influences on depressive symptoms among Chinese adolescents

PDF

The link between maternal depression and adolescent daughters' risk behavior: the mediating and moderating role of family

PDF

The impact of child maltreatment on the mental health of adolescents: a longitudinal study of social anxiety and self-perception

PDF

The role of depression symptoms on social information processing and tobacco use among adolescents

PDF

Trajectories of internalizing problems among maltreated girls and boys: differences by maltreatment type and developmental timing

PDF

Attachment, maltreatment and autonomic nervous system responsivity as predictors of adolescent anxiety and depression

PDF

Maltreated adolescents and their families: a longitudinal examination of family functioning, parenting attitudes, & youth mental health

PDF

Multivariate biometric modeling among multiple traits across different raters in child externalizing behavior studies

PDF

Parenting practices among non-offending mothers of sexually abused girls and its impact on the abused girls' behavioral adjustment: perspectives from a multigenerational, longitudinal study

PDF

Childhood adversity across generations and its impact on externalizing behavior

PDF

Using observed peer discussions to understand adolescent depressive symptoms and interpersonal interactions

PDF

Cultural risk and protective factors for tobacco use behaviors and depressive symptoms among American Indian adolescents in California

PDF

Depression severity, self-care behaviors, and self-reported diabetes symptoms and daily functioning among low-income patients receiving depression care

PDF

Trauma-related treatment gains among women with histories of interpersonal violence and co-occurring mental health and substance abuse disorders

PDF

Using a structural model of psychopathology to distinguish relations between shared and specific features of psychopathology, smoking, and underlying mechanisms

PDF

Neighborhood context and adolescent mental health: development and mechanisms

PDF

Emotion regulation as a mechanism linking parents’ marital aggression to adolescent behavioral problems: a longitudinal analysis

PDF

Factors and correlates of sexual behaviors among young adults from continuation high schools

Asset Metadata

Creator Brensilver, Matthew (author)

Core Title Untangling the developmental relations between depression and externalizing behavior among maltreated adolescents

School School of Social Work

Degree Doctor of Philosophy

Degree Program Social Work

Publication Date 08/07/2010

Defense Date 06/11/2010

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag comorbidity,Depression,externalizing behavior,latent change score models,maltreatment,OAI-PMH Harvest

Place Name California (states), Los Angeles (counties)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Mennen, Ferol (committee chair), McArdle, John J. (committee member), Trickett, Penelope K. (committee member)

Creator Email brensilv@usc.edu,matthewbrensilver@yahoo.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-m3338

Unique identifier UC1452578

Identifier etd-Brensilver-3888 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-378534 (legacy record id),usctheses-m3338 (legacy record id)

Legacy Identifier etd-Brensilver-3888.pdf

Dmrecord 378534

Document Type Dissertation

Rights Brensilver, Matthew

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email cisadmin@lib.usc.edu