|Home | About | Journals | Submit | Contact Us | Français|
The appropriateness of the Diagnostic and Statistical Manual of Mental Disorders – Fourth Edition (DSM-IV) nosology for classifying preschool mental health disturbances continues to be debated. To inform this debate, we investigate whether preschool psychopathology shows differentiation along diagnostically specific lines when DSM-IV symptoms are aggregated statistically.
One thousand seventy-three parents of preschoolers aged 2–5 years attending a large pediatric clinic completed the Child Behavior Checklist 1.5–5. A stratified probability sample of 193 parents of high scorers and 114 parents of low scorers were interviewed with the Preschool Age Psychiatric Assessment (PAPA). Confirmatory factor analysis was performed on symptoms from seven DSM disorders.
Comparison of competing models supported the differentiation of emotional syndromes into three factors: social phobia (SOC), separation anxiety (SAD), and depression/generalized anxiety (MDD/GAD), and the differentiation of disruptive syndromes into three factors: oppositional defiant/conduct syndrome (ODD/CD), hyperactivity/impulsivity, and inattention. Latent syndrome correlations were moderately high after accounting for symptom overlap and measurement error.
Psychopathology appears to be differentiated among preschoolers much as it is among older children, and adolescents. We conclude that it is as reasonable to apply the DSM-IV nosology to preschoolers as it is to apply it to older individuals.
The Diagnostic and Statistical Manual of Mental Disorders taxonomy (DSM-IV; American Psychiatric Association, 2000) is increasingly being used to diagnose psychopathology among preschoolers (see Angold & Egger, 2004). Several studies have shown that, when DSM diagnostic algorithms are used to aggregate symptoms from general population-based preschool samples, prevalence and comorbidity rates are within the range reported for older children (Earls, 1982; Egger et al., 2006; Keenan, Shaw, Walsh, Delliquadri, & Giovannelli, 1997; Lavigne et al., 1996; see Egger & Angold, 2006 for a review). Yet, the appropriateness of the DSM-IV for preschool phenomenology continues to be debated (e.g., McClellan & Speltz, 2003; Wilens, Biederman, Spencer, & Monuteaux, 2003). Validation of the DSM-IV for preschoolers involves demonstrating that preschoolers’ mental health disturbances are really differentiated along the lines of DSM-IV syndromes, rather than appearing in alternative or undifferentiated precursor forms, with DSM-IV ‘disorders’ being merely arbitrary impositions.
Few studies have tested the differentiability of DSMIV syndromes in preschoolers, although some studies of mostly older children have included small samples of children down to age 2 (e.g., Burns et al., 1997). DSM syndrome differentiation is conventionally tested by constructing a model that partitions DSM symptoms according to DSM syndromes and testing its fit to the data using confirmatory factor analysis. Such studies of older children and adolescents have indicated that certain areas of psychopathology are well fit by the DSM-IV criteria, but that some others are not. There are three main areas of poor or inconsistent fit (reviewed in Lahey et al., 2004): (1) lack of differentiation between generalized anxiety (GAD) and major depression (MDD); (2) lack of differentiation between oppositional defiant disorder (ODD) and conduct disorder (CD); and (3) the differentiation of attention deficit-hyperactivity disorder (ADHD) into two dimensions.
Whereas the differentiability of a broad range of DSM symptoms according to DSM syndromes has been tested among older children and adolescents, the focus in preschoolers has mostly been on deriving ‘empirical’ syndromes (e.g., Achenbach & Rescorla, 2000). For several reasons, such investigations cannot tell us whether the DSM-IV provides reasonable working descriptions of preschool psychopathology (that has not been their goal). Although some empirically derived syndromes resemble DSMIV syndromes descriptively (see Achenbach, Dumenci, & Rescorla, 2003), the item pools employed have provided limited coverage of the scope of psychopathology, and the symptom information they represent has been insufficiently detailed to represent the approach and content of the DSM-IV. For example, an attempt to classify Child Behavior Checklist (CBCL 1.5–5; Achenbach & Rescorla, 2000) items into ‘DSM-oriented syndromes’ noted that there were too few anxiety symptoms assessed to permit estimation of separate GAD, separation anxiety, and social phobia syndromes (Achenbach et al., 2003, p. 333). When only a few symptoms of a particular DSM syndrome are represented on a checklist, there will often be insufficient power to recover that syndrome in a factor analysis (see MacCallum, Widaman, Zhang, & Hong, 1999). That syndrome’s symptoms can instead end up loading on secondary syndromes (MacCallum et al., 1999). So, for instance, it is not surprising that the few SAD symptoms included in the CBCL 1.5–5 do not cohere as a separate factor, but cross-load on other syndromes (Achenbach et al., 2003).
Only two studies of preschoolers have tested the fit of DSM syndromes using confirmatory factor analysis of DSM symptoms, and these dealt only with anxiety. In large (N > 700), community samples, Eley et al. (2003) and Spence, Rapee, McDonald, and Ingram (2001) found that dimensions representing SAD, social anxiety, obsessive compulsive symptoms, fears, and generalized anxiety were differentiable.
Given these gaps in the current literature, our first aim is to examine the degree to which the DSM-IV adequately represents preschool psychiatric symptomatology using: (a) a reasonably representative preschool sample, (b) precise DSM-IV symptom measurement, and (c) a wider range of common syndromes – attention deficit hyperactivity disorder (ADHD), major depression (MDD), social phobia (SOC), oppositional defiant disorder (ODD), separation anxiety (SAD), conduct disorder (CD), and generalized anxiety disorder (GAD). Our second aim is to determine whether the three areas that are poorly or inconsistently fit by the DSM-IV criteria among older children and adolescents are similarly poorly fitting among preschoolers.
Details of the study’s methods can be found in an earlier publication (Egger et al., 2006). We present a summary pertinent to the analyses conducted here.
Subjects were recruited from the Duke Children’s Primary Care Clinic to take part in a one-week test–retest reliability study of the Preschool Age Psychiatric Assessment (PAPA; Egger & Angold, 2004), a structured psychiatric diagnostic interview for use with parents of preschoolers aged 2–5 years. An overview of study design is presented in Figure 1. Data from the first of the two repeated interviews were employed in the present analyses. The study used a psychopathology screen-stratified design, with oversampling of those with high screen scores, and stratification by gender, age (2-, 3-, 4- and 5-year-olds), and race (African American and non-African American). The use of sampling weights in these analyses permits unbiased pediatric clinic estimates to be computed from such a stratified sample.
Over the 18 months of data collection, 1,191 parents were approached by a screener, who explained the study and sought informed consent for completion of the Child Behavior Checklist (CBCL) 1.5–5 (Achenbach & Rescorla, 2000). Of these, 1,073 were screened: 20 refused, and 98 were excluded, either because the accompanying adult was not English speaking (n = 48) or could not provide legal consent (n = 21), or the child had autism, mental retardation, or another pervasive developmental disorder (n = 14), or had a sibling already enrolled (n = 15). Three hundred seven children ‘screened high’ (i.e., obtained a CBCL total symptom T-score ≥ 55). Seven hundred seventy-six ‘screened low’ (i.e., obtained a CBCL total symptom T-score < 55). Stratifying by age, gender, and race, 80% of screen highs were randomly selected for PAPA interviews, as well as well as 20% of ‘screen lows.’ We continued to request participation in the interview phase from members of each age × gender × ethnicity group until their particular cell was full. PAPA interviews took place with interviewers blind to the parent’s screen status, usually in the participant’s home.
Demographic characteristics of the screened sample, the interviewed sample, and surrounding Durham County where the study was conducted are presented in Table 1. Eighty-six percent of the interviews were conducted with the child’s biological mother. No significant differences in gender, age, or Medicaid status emerged between screen-refusers and participants or study-refusers and participants (see Egger et al., 2006), so the sample weights were not adjusted for study-refusal.
The PAPA (Egger & Angold, 2004) is a parent-report, interviewer-based, structured psychiatric assessment involving a range of mandatory questions and probes, supplemented by further detailed exploratory probing to ensure that the ratings appropriately represent the child’s problems. When symptoms (e.g., depressed mood, irritability) were reported, their frequency, duration and onset dates were also collected for a three-month primary period, in order to determine whether they met the criteria for the symptoms of various DSMIV diagnoses. Symptom algorithms implementing the DSM-IV criteria were programmed using SAS software. These SAS algorithms were applied to the raw frequency, duration, and onset PAPA data to generate DSM-IV psychiatric symptom data. As far as possible, the PAPA symptom algorithms followed the Child and Adolescent Psychiatric Assessment (CAPA; Angold & Costello, 2000) algorithms. However, (1) DSM-IV symptoms that are not applicable to young children were excluded (e.g., for CD, 5 out of the 15 possible symptoms were excluded), (2) Research Diagnostic Criteria – Preschool Age (Task Force on Research Diagnostic Criteria: Infancy and Preschool, 2003) developmentally modified DSM-IV criteria were used for depression (e.g., suicidal themes in play could serve as endorsement of the symptom of suicidality; Luby et al., 2003), and (3) the high prevalence of certain behaviors in preschoolers indicated a need to modify the CAPA’s cut-points for several symptoms. For instance, because the frequency of ODD symptoms such as ‘often loses temper’ is higher in preschoolers than in older children, the ODD algorithm was modified so that each ODD symptom reflected the top 10% of frequency for preschoolers based on PAPA data. Thus, we maintained the CAPA’s 90th percentile frequency cutoff conceptualization of ODD symptomatology (Angold & Costello, 1996) by modifying the criteria frequency levels. A similar approach was taken for the CD symptoms of assaults and lying. (See Egger et al., 2006 for further details about the algorithms.) The one-week test–retest symptom scale reliabilities ranged from .61 (GAD) to .81 (ADHD), comparable to those of interviews for older children and adults (Egger et al., 2006).
We employed confirmatory factor analysis (CFA) rather than exploratory factor analysis (EFA) because we were testing a pre-formed diagnostic system (DSM-IV), rather than deriving a descriptive system empirically. CFA treats observed symptoms as probabilistic indicators of latent DSM-IV syndromes and assumes that these underlying DSM-IV syndromes are continuously distributed; (empirical and theoretical justifications of this assumption are provided in van den Oord, Pickles, & Waldman, 2003).
A series of four emotional-syndrome models and four disruptive-syndrome models were estimated separately, rather than as a series of composite models, in order to better test the dimensionality of the disorders in each domain. Such issues are hard to resolve in nested tests of large, composite models, because poor performance in one domain of the composite model can affect the validity of dimensionality testing in another domain (Hartman et al., 2001).
Each series of four models included: a ‘DSM-IV model’; two models representing differentiation patterns most commonly found in older children/adolescents (‘Child–Adolescent models’) and an ‘Undifferentiated (one-factor) model’. The four emotional-syndrome models were: (i) the DSM-IV model (GAD, SAD, SOC, and MDD); Child- Adolescent models (ii) SAD and SOC as separate factors but MDD/GAD combined and (iii) MDD as one factor and all anxiety disorders as the other; and (iv) an Undifferentiated (1-factor) model. The four disruptive-syndrome models were: (i) the DSM-IV model (ADHD, ODD, and CD); Child–Adolescent models (ii) hyperactivity/impulsivity [HI], inattention [I], ODD, and CD and (iii) HI and I as separate factors but ODD/CD combined; and (iv) an Undifferentiated (one-factor) model.
A composite model was formed by combining the preferred model from the set of emotional-syndrome models with the preferred model from the set of disruptive- syndrome models and allowing all factors to correlate. In the final model, we focused on descriptive correlations among all of the latent syndrome factors, which depict the extent of association among all syndromes accounting for measurement error and shared symptom effects.
Several rare symptoms, each belonging to the same syndrome, i.e., the same dimension, were ‘parceled’ (see Sass & Smith, 2006) prior to model estimation to prevent low variance from reducing their loadings, and thereby reducing common factor variance (see Hartman et al., 2001). Specifically, four uncommon SAD symptoms were summed to form two indicators: (1) worry about untoward events and loss or harm of attachment figures, and (2) sleep reluctance or refusal and separation nightmares. Besides improving the distributional characteristics of rare items, parceling also improves model parsimony when applied to highly correlated items. Hence, three highly-correlated hyperactivity symptoms were summed to form one indicator: leaves seat; runs and climbs; on the go. Additionally, social anxiety symptoms were combined into two indicators, (1) social or performance fear and (2) anxiety or avoidance of social contact or public performance. Two physical cruelty CD symptoms (to animals and people) were combined, as were the two property destruction CD symptoms (deliberate property damage and firesetting).
To reduce the possibility of artifactual covariation, prior to model estimation, ‘specificity’ terms for overlapping symptoms contributing to different disorders were correlated to account for shared method factors intrinsic to that symptom (irritability for MDD and ODD; blaming/lying for ODD and CD; fatigue, irritability, sleep, and concentration for GAD and MDD). In factor analytic terms, ‘specificity’ is that part of a symptom’s variance not explained by the syndrome factor, nor by measurement error, and therefore is unique to that symptom. Allowing correlated specificities is preferable to deleting one symptom out of each overlapping pair and allowing the remaining symptom to crossload, because, unless the two symptoms are perfectly interchangeable, the latter approach sacrifices syndrome-specific nuances of their measurement. In our study, overlapping symptoms’ bivariate correlations never exceeded r = .85 because each disorder’s criteria required slightly different behaviors, duration, impairment, and/or onset criteria.
Because the observed indicators were dichotomous, a tetrachoric correlation matrix was analyzed, with a robust estimator appropriate for non-normal data (Weighted Least Squares with Mean and Variance adjustment, WLSMV, as implemented in Mplus 4.1 (Muthén & Muthén, 1998–2006).
Each model was evaluated according to four criteria: (1) absolute fit (mean- and variance-adjusted chi-square statistic that corrects for non-normality); (2) relative fit indices (Comparative Fit Index, CFI; Tucker-Lewis Index, TLI); (3) the Root Mean Square Error of Approximation (RMSEA), which estimates discrepancy between model-implied and population covariance matrices; and (4) residual-based fit indices (Weighted Root Mean Squared Residual, WRMR). With binary indicators and N ≥ 250, suggested guidelines demarcating good fit are: CFI ≥ .96, TLI ≥ .95 RMSEA ≤ .05; WRMR ≤ 1.0 (Yu & Muthén, 2002). Component fit was also evaluated (i.e., size and significance of standardized loadings).
Next, a sequence of inferential nested model tests was performed on the set of emotional-syndrome models, and on the set of disruptive-syndrome models, to identify the best-fitting model in each domain. We performed the minimum possible number of inferential tests that still enabled us to fulfill our study’s second aim. In all, three tests were performed on the set of emotional-syndrome models, and three tests were performed on the set of disruptive-syndrome models. Specifically, for each set, first the DSM-IV model was compared to the more complex of the two Child–Adolescent Models (with more complex referring to having more factors). Second, the better-fitting model of these two was then compared to the second, less complex, Child–Adolescent Model. Third, the better-fitting of these two was compared to the Undifferentiated (one-factor) model. Model selection employed mean- and variance- adjusted χ2 difference tests to compare the more restricted (i.e., less differentiated) model to the less restricted model within which it was nested. A significant difference in the χ2’s indicates that the more differentiated model fits the data better than the less differentiated.
The final composite model would ideally then be cross-validated in an independent sample, but no such sample was available. We, therefore, adopted the next best alternative – evaluating the composite model descriptively only. In theory, application of inferential tests to the composite model, such as the RMSEA or the chi-square test of absolute fit, could inflate Type I error due to incremental improvements in fit arising from comparing and selecting the best-fitting sub-models before fitting the composite model (capitalization on chance). Although inflated Type I error is of greatest concern when initial models are poorly fitting and model comparisons result in large improvements in fit (neither of which are the case here), some inflation of Type I error would still be likely.
Prior analyses of this dataset found no prominent differences in comorbidity, diagnostic prevalence, or scale score reliabilities across gender, race, or age (Egger et al., 2006), and so invariance testing across gender, race, and age was not conducted. Such multiple- group models, moreover, would have been infeasible because our moderate sample size and low frequency of many clinical symptoms gives rise to empty cells.
Overall absolute, relative, and residual-based fit, as well as component fit, for each of the four nested emotional-syndrome models and four nested disruptive- syndrome models is presented in Table 2. All DSM symptoms in all models loaded positively and significantly on their designated DSM syndromes.
The DSM-IV model (SOC,- SAD, GAD, MDD) and the 3-factor Child–Adolescent model (SOC, SAD, GAD/MDD) both had good absolute fit (χ2, p > .05). Good absolute fit implies (a) factor loadings fixed to zero – including all possible cross loadings – are evidently near-zero in the sample data, (b) high factor loadings are not an artifactual byproduct of high inter-factor correlations; and (c) the null hypothesis of perfect model fit cannot be rejected. Both of these models also had good relative fit (CFI ≥ .96, TLI ≥ .95) residual fit (WRMR ≤ 1.0), and low error of approximation (RMSEA < .05). The overall fit of the 2-factor Child-Adolescent model (MDD, all-anxiety-disorders) and the Undifferentiated model were mixed: both showed borderline relative fit (CFI of .96), and the Undifferentiated model showed poor absolute fit. Residual-based fit, component fit, and RMSEA were still acceptable, however. While these indices generally preferred the DSM-IV or 3-factor models over the 2-factor or Undifferentiated models, we employed inferential nested chi-square difference testing to select the best fitting model.
When the DSM-IV model (SAD, SOC, MDD, GAD) was compared to the 2-factor (MDD, all-anxiety-disorders) model, the DSM-IV model showed significantly better fit (χ2 (4, N = 307) = 19.042, p = .0008). This indicates that there was no support for collapsing the three anxiety disorders into a unidimensional construct. When the DSM-IV model was compared to the 3-factor (SOC, SAD, GAD/MDD) model, the DSM-IV model did not show significant improvement in fit. This indicates that generalized anxiety could not be statistically differentiated from depression (χ2 (2, N = 307) = .243, p = .88) after accounting for overlapping symptoms. This result corroborates a correlation, within rounding of 1.0, between GAD and MDD latent factors in the 4-factor model. In the preferred 3-factor model, latent syndrome correlations were: GAD/MDD with SAD, r = .77; GAD/MDD with SOC, r = .75; and SAD with SOC, r = .65. All these correlations were significant (p < .01). Since the Undifferentiated model fit significantly worse than the preferred 3-factor model (χ2 (3, N = 307) = 18.18, p = .0004), we concluded that the 3-factor model for emotional syndromes was preferable statistically.
Good relative fit, residual-based fit, and error of approximation was found for the DSM-IV model (ODD, CD, ADHD), the 4-factor model (ODD, CD, HI, I), and the 3-factor model (ODD/CD, HI, I), as shown in Table 2. Yet, we could reject the null hypothesis of perfect absolute fit for all these models (χ2 p = .03 or .02). Although χ2 is adjusted for binary outcomes, strong non-normality (as found for the CD symptoms physical cruelty to animals or people) could still lower type I error and lead to an over-rejection of properly specified models (Yu & Muthén, 2002). This may explain why these models were found to be well-fitting according to all other overall fit criteria. In contrast, the Undifferentiated model showed poor RMSEA, poor residual-based fit, and borderline relative fit (CFI of .96).
First, the fit of the DSM-IV model (ODD, CD, ADHD) was compared to the 4-factor Child–Adolescent model (ODD, CD, HI, I). The 4-factor model showed significantly better fit than the DSM-IV model (χ2 (3, N = 307) = 14.88, p = .002), indicating that ADHD is not a unitary syndrome in preschoolers. However, the 4-factor model did not show significantly better fit than the 3-factor Child-Adolescent model (ODD/CD, HI, I), indicating that ODD and CD could not be statistically differentiated from each other, even after accounting for their overlapping symptom of lying/blaming (χ2 (2, N = 307) = 4.31, p = .11). This result is statistically unsurprising given that the ODD and CD latent syndromes from the 4-syndrome model had a correlation, within rounding, of 1.0. The preferred 3- factor model evidenced latent syndrome correlations of: I and HI, r = .94; ODD/CD and I, r = .71; ODD/CD and HI, r = .79 (All p < .01). Finally, the 3-factor model fit significantly better than the Undifferentiated model (χ2 (2, N = 307) = 35.11, p < .0001). Hence, a 3-factor model for disruptive syndromes was preferable statistically.
The preferred 3-factor model of emotional syndromes was combined with the preferred 3-factor model of disruptive syndromes. Correlations between each of the resultant six latent syndromes are shown in Table 3. We briefly present model fit statistics for descriptive purposes: relative fit (CFI = .95; TLI = .95; RMSEA = .05; free parameters = 117; χ2 (50, N = 307) = 82.59, p = .003; all loadings significant at p < .05). Since the highest correlations between latent syndromes in the final model were between the sub-dimensions of ADHD, and in earlier analyses we showed that combining these sub-dimensions into a single factor significantly decreased fit, there was no need for exploratory comparisons to see whether other factors in the final model could be combined without decrementing fit.
MDD/GAD evidenced near-equivalent correlations with emotional and disruptive syndromes. Since several researchers (Luby et al., 2003; Egger et al., 2006; Lahey et al., 2004) have suggested that the observed strong relations between mood and disruptive disorders in young children are driven to a large degree by the inclusion of irritability in the criteria for both types of disorder, we reanalyzed our data without taking into account this symptom overlap. When symptom overlap was ignored, ODD and MDD appeared unidimensional (undifferentiated) in a bivariate analysis (χ2 (1, N = 307) = .49, p = .48); but when symptom overlap was accounted for, they were differentiable (χ2 (1, N = 307) = 4.62, p = .03). Furthermore, significant loadings for irritability remained on each factor after controlling for the symptom’s shared method variance.
Our first aim here was to shed light on debates about the appropriateness of classifying preschool phenomenology according to the DSM-IV nosology (McClellan & Speltz, 2003; Wilens et al., 2003). We tested whether symptom dimensions underlying seven DSM disorders were differentiated along diagnostically specific lines among preschoolers. We found no support for the idea that preschool psychopathology is essentially undifferentiated, or limited to precursor ‘internalizing’ and/or ‘externalizing’ forms. Instead, our results not only indicated that psychopathology in preschoolers is largely differentiated according to DSM syndromes, but also showed that the ways in which preschooler syndrome differentiation departs from the DSM-IV nosology are strikingly similar to those found in older children and adolescents: a lack of separation between GAD and MDD and between ODD and CD, and the splitting of ADHD into two separable (but highly correlated) sub-components – hyperactivity/impulsivity and inattention. Overall the DSM-IV provided as good a description of preschool psychopathology as it does for the mental health problems of older children.
Paralleling factor analytic studies of older children (Lahey et al., 2004) and adults (Kendler et al., 2003; Vollebergh et al., 2001), three emotional syndromes were distinguished among preschoolers (SOC, SAD, GAD/MDD). The unidimensionality of GAD/MDD was not an artifact of overlapping symptomology, since it persisted after controlling for shared method variance in the overlapping symptoms. Interestingly, however, if we had lacked the symptom coverage required to estimate three separate anxiety disorders (as did Achenbach et al., 2003; Cole et al., 1997; Ollendick & Yule, 1990), we might have erroneously concluded that anxiety and depression are undifferentiated as a result of the r = .95 correlation between the depression and the nonspecific-anxiety factor resulting from our 2-factor model. Our more precise measurement of each anxiety syndrome shows that this high intercorrelation is the by-product of a still higher correlation between GAD and depression.
Paralleling ADHD and CD/ODD research on older children and adolescents (Glutting et al., 2005; Pillow et al., 1998; Willoughby et al., 2000), three disruptive syndromes were distinguished among preschoolers (hyperactivity/impulsivity, inattention, and ODD/CD). The multidimensionality of ADHD seems not to be a mere artifact of a subcluster of ADHD items that were both highly related and common in our data, because all symptoms loaded significantly and substantially on their appropriate disorders. Our undifferentiated preschool ODD/CD factor is consistent with prior research suggesting that, among older children and adolescents, these are subsets of a unitary construct (as in ICD-10), and that ODD is a developmental precursor of adolescent CD (e.g., Loeber et al., 1993). Our undifferentiated preschool ODD/CD factor appears, at first sight, to contrast with the findings of Burns et al. (1997). However, in their work, CD and ODD syndromes were found to be differentiable only after CD’s symptom set was reduced to the few symptoms that did not significantly correlate with ODD or ADHD (i.e., 4 overt aggression symptoms); if they had retained the 10 DSM indicators of CD used here (encompassing overt aggression, destruction of property or objects, and deceitfulness), our results might have been quite similar. Future item analyses of conduct disorder symptoms among preschoolers are warranted to build consensus on developmentally appropriate CD items and to test whether these differentially predict outcomes above and beyond those predicted by ODD.
When the emotional and disruptive syndromes were combined in our final model, six distinct syndromes were supported, that, if combined further, would have resulted in decrements to model fit. Their moderate-to- high intercorrelations affirm Lahey et al.’s (2004) assertion that ‘substantial intercorrelation’ is an ‘inherent characteristic’ of psychopathology across development (p. 377). These correlations between latent syndromes were higher than previous observed-syndrome correlations (e.g., Lavigne et al., 1996); the latter are attenuated by unreliability.
Since we had only a moderate-sized sample, analyses were limited to common Axis I disorders that had often been used in prior CFA analyses on older children. It would have been desirable to include more anxiety disorders; however, prevalence rates of obsessive compulsive disorder and post traumatic stress disorder were extremely low in our sample (<1%) and their inclusion in analyses would have led to estimation problems. A larger sample would also have permitted us to break our sample into sub-groups and perform measurement invariance testing by age, gender, and ethnicity. Additionally, assessment of overall model fit for the composite emotional and disruptive disorders model awaits cross-validation in an independent sample. Although we have drawn qualitative parallels between our results and previously published findings on older children, statistically quantifying the degree of developmental differences in syndrome differentiation between a best-fitting CFA for preschoolers and a best-fitting CFA for older children is another matter, requiring invariance testing on a cross-age multiple group structural equation model.
Our symptom data were obtained solely from parental report, and the children solely from a pediatric clinic. However, recruited children were seen at the clinic both for well-and sick-child visits, and our sample demographics were similar to those of the area where the clinic is located. Furthermore, the distribution of these children’s CBCL scores closely matched reported general population norms (Achenbach & Rescorla, 2000).
Whereas our results suggest that it is as valid to apply the DSM-IV nosology to preschoolers as it is to apply it to older children or adolescents, construct validity is of course a multifaceted process. We have focused on only one component, symptom differentiation, and one set of behaviors nested within one psychiatric assessment approach. Nonetheless, this study provides initial evidence supporting the use of a common nosological system across a wide age range. This is an important point because use of a common system facilitates exploration of homotypic and heterotypic continuity more readily than would the employment of a preschool-specific classification.
This project was supported by grants from the NIMH (R01 MH-63670 (Dr Angold) and K23 MH-02016 (Dr. Egger)) and a Pfizer Faculty Scholars Award in Clinical Epidemiology (Dr. Egger). The authors thank Kristopher Preacher for helpful comments on an earlier draft.
Conflict of interest statement: Dr Egger was the recipient of a Pfizer Faculty Scholars Award in Clinical Epidemiology. The other authors do not have financial ties with for-profit enterprises. The authors do not receive financial benefit from the instrument described in this paper.