|Healthcare Training Institute - Quality Education since 1979CE for Psychologist, Social Worker, Counselor, & MFT!!
Factor Structure of the Eyberg Child Behavior Inventory: A Parent Rating Scale of Oppositional Defiant Behavior Toward Adults, Inattentive Behavior, and Conduct Problem Behavior. By: Burns, G. Leonard; Patterson, David R.. Journal of Clinical Child Psychology, Dec2000, Vol. 29 Issue 4, p569-577, 9p, 4 charts; (AN 4701614)
The Eyberg Child Behavior Inventory (ECBI) is used to measure disruptive behavior problems in children and adolescents. A controversy exists, however, on the dimensional structure of the ECBI. To evaluate this issue, an exploratory factor analysis was first performed on a sample of 1,263 children and adolescents. This analysis identified 3 meaningful factors (i. e., Oppositional Defiant Behavior Toward Adults, Inattentive Behavior, and Conduct Problem Behavior) and a fourth, poorly defined factor. A confirmatory factor analysis (CFA) evaluated the fit of the 3 meaningful factors in a second sample of 1,264 children and adolescents. The 3-factor model with 2 correlated errors provided a excellent fit. This 3-factor model also provided a significantly better fit than 2- and 1-factor models. Multiple group CFA indicated that the factor pattern, item-factor loadings, factor correlations, and correlated errors were equivalent across the samples. The CFA on sex yielded similar results. Initial normative information is presented for boys (n = 1,322) and girls (n = 1,205) within 4 age ranges (i.e., 25, 6-9, 10-13, 14-17) for the 3 factors. The use of these 3 factors, especially Oppositional Defiant Behavior and Conduct Problem Behavior, should make the ECBI more useful as a screening and outcome measure.
The Eyberg Child Behavior Inventory (ECBI) is a parent rating scale widely used to measure disruptive behavior problems in children and adolescents (McMahon & Estes, 1997). Although the ECBI has positive psychometric properties, a controversy exists on the dimensional structure of the measure. Whereas Eyberg (1992) considers the ECBI a unidimensional measure of conduct problem behavior, others (McMahon & Estes, 1997) view the ECBI as a multidimensional measure of disruptive behavior. Although the ECBI contains items similar to the symptoms of oppositional defiant disorder (ODD), conduct disorder (CD), and attention deficit hyperactivity disorder (ADHD), the controversy continues about the dimensional structure of the measure (Eyberg & Colvin, 1994).
Our goal was to reexamine the structure of the ECBI in a more sophisticated manner than the previous studies (Burns & Patterson, 1990, 1991; Burns, Patterson, Nussbaum, & Parker, 1991; Eyberg & Colvin, 1994; Eyberg & Robinson, 1983; Robinson, Eyberg, & Ross, 1980). In an earlier study (Burns & Patterson, 1991), for example, we examined the ECBI's structure in a sample of 1,526 children from five pediatric clinics from four states and in a random sample of 1,003 children from the Seattle School District. We performed a principal components analysis with varimax rotation on each sample. We also limited the number of factors a priori to three because the ECBI contained ADHD, ODD, and CD type items. Although the three dimensions roughly approximated ODD, CD, and ADHD, we did not examine the degree of model fit or the equivalence of the results across the samples or sex. In addition, because we restricted the exploratory factor analysis to three factors, the analysis was not really exploratory. Finally, given the high levels of comorbidity among ADHD, ODD, and CD (Quay & Hogan, 1999), it is questionable whether orthogonal (varimax) rotation was an appropriate decision. These early studies have thus involved a series of questionable statistical decisions and, given the complexities of factor analyses, it is not surprising the studies yielded different conclusions on the factor structure of the ECBI (e.g., Burns & Patterson, 1990, 1991; Eyberg & Colvin, 1994).
To evaluate the structure of the ECBI in a better manner, we first combined the pediatric and random samples. Our next step was to create two random samples from the total data set, 1,263 children and adolescents in the first sample and 1,264 in the second. We then performed an exploratory factor analysis (EFA) with oblique rotation on the first sample. The goal of this exploratory analysis was to determine the number of clinically meaningful dimensions in the ECBI. The model selected from this EFA was then evaluated with confirmatory factor analysis (CFA) in the second sample to determine if the model provided a good fit as well as a significantly better fit than simpler models. In addition, we performed a multiple group CFA to determine if the factor pattern, item-factor loadings, and factor correlations of the model were equivalent across sex and across the samples. Finally, if meaningful dimensions are identified and replicated, we will present initial normative data on these dimensions to make the ECBI more specific as a screening and outcome measure.
ECBI. The ECBI contains 36 disruptive behavior problems. The parent indicates on a 7-point scale how often each behavior occurs; 1 (never), 2 and 3 (seldom), 4 (sometimes), 5 and 6 (often), and 7 (always). The parent also indicates if the occurrence of the specific behavior is currently a problem by circling "yes" or "no" for each behavior. This results in two summary scores--an intensity score (IS) and a problem score (PS). The IS score represents the total frequency of occurrence of the 36 behaviors (possible range from 36 to 252). The PS represents the total number of the 36 behaviors that are indicated to be problems (possible range from 0 to 36). Table 1 shows the 36 items on the ECBI.
Participants and Procedures
For the pediatric sample, a total of 1,526 ECBIs were completed by parents or guardians in five outpatient pediatric clinics in four northwestern states (Pullman, WA; Seattle, WA; Lewiston, ID; Missoula, MT; and Portland, OR). For the random sample, 300 children were randomly selected on the basis of sex and ethnicity (Asian, African American, and Caucasian) within each grade level for Grades 1 to 12 from the Seattle School District (a total of 3,600 parents were mailed ECBIs). A total of 1,003 completed ECBIs were returned by the parents. This return rate of 28% was similar to a return rate of 29% obtained in a second study in the Seattle School District (Bums et al., 1997). Two of the adolescents were 18 years old and these two ratings were eliminated because they were outside the age range of the ECBI (2-17). This left a total of 1,001 children and adolescents.
Characteristics of the 2,527 Children and Adolescents
The combination of the pediatric and random samples resulted in data on 2,527 children and adolescents. The sample was 52% boys and 48% girls, with an average age of 8.95 years (SD = 4.36, range 2-17). A total of 1,639 (65%) of the children were living with their biological mother and father; 464 (18%) with their mother only; 30 (1%) with their father only; 245 (10%) with their mother and stepfather; 42 (2%) with their father and stepmother; 18 (< 1%) with foster parents; and 89 (4%) with other relatives. In terms of ethnicity, 85% of the children were Caucasian, 5% African American, 4% Asian, 3% American Indian, less than 1% Hispanic, and 4% mixed ethnicity (e.g., 1/2 Caucasian and 1/2 African American). A total of 2,180 (86%) ECBIs were completed by the child's mother, 255 (10%) by the child's father, and 92 (4%) by other relatives or foster parents. The average education of the person who completed the ECBI on the child was 13.91 grades (SD = 2.65). A total of 179 (7%) of the raters had not completed high school; 860 (34%) had obtained a high school degree; 626 (25%) had attended some college; 538 (21%) had obtained a college degree; and 324 (13%) had completed some graduate study. In terms of family income, 335 families (14%) reported a yearly income of less than $10,000; 391 (16%) between $10,000 and $19,999; 557 (22%) between $20,000 and $29,999; and 1,194 (48%) over $30,000. Fifty of the raters did not provide information on family income. In terms of treatment status, 2,335 (92%) of the children were not currently in treatment for learning disabilities or behavioral problems; 86 (3%) were in treatment for learning disabilities; 72 (3%) for behavioral problems; and 34 (1%) for learning and behavioral problems.
Structural Organization of the ECBI
The 2,257 children were randomly separately into two samples, 1,263 in the first and 1,264 in the second. The random assignment was performed so that each sample contained an equal percentage of children from the pediatric clinics and the Seattle School District. The factor analyses were performed on the IS item ratings because the PS item ratings involved a categorical variable (i.e., a "yes" or "no" answer for each item).
EFA on Sample 1. An EFA with maximum likelihood extraction and promax (oblique) rotation was performed on the Sample 1 IS ratings. Seven eigenvalues were greater than one (11.73, 2.47, 2.07, 1.56, 1.34, 1.32, and 1.10). We examined factor solutions of two to seven factors. In the two-factor model, Factor 1 consisted of ODD and CD type items and Factor 2 ADHD type items. In the three-factor model, the first factor consisted of ODD type items, the second factor CD type items, and the third factor ADHD type items. In the four-factor model, the items with weak loadings on the first factor, the ODD factor, separated to form the fourth factor. In this four-factor model, the ADHD and CD factors emerged as Factors 2 and 3. The five-, six-, and seven-factor models primarily resulted in the fourth factor from the four-factor model dividing into smaller factors. The results from the four-factor model were considered to provide the most clinically useful dimensions. The specific reasons for this decision will be discussed after the presentation of the results from the four-factor model.
Table 1 shows the results from the four-factor model. Items with loadings greater than .29 are shown in boldface in the table. The first factor involved oppositional defiant behavior toward adults (i.e., "argues with parents about rules," "acts defiant when told to do something," "refuses to obey until threatened with punishment," "sasses adults," "refuses to do chores when asked," "gets angry when does not get own way," "does not obey house rules on own," "refuses to go to bed on time," "has temper tantrums," and "yells or screams"). Because the items "slow in getting ready for bed" and "has poor table manners" had low loadings on this factor and did not involve an oppositional defiant aspect, these two items were not included in the CFA on the second sample. The elimination of these two items resulted in a clear and strong Oppositional Defiant Behavior factor.
The second factor contained behaviors similar to the symptoms of ADHD. The four items with the highest loadings represented ADHD inattentive symptoms (i.e., "has short attention span," "is easily distracted," "has difficulty concentrating on one thing," and "fails to finish tasks or projects"). These four items had loadings from .71 to .95. The other two items were somewhat similar to ADHD hyperactivity symptoms (i.e., "is overactive or restless" and "has difficulty entertaining himself alone"). These two items also had loadings substantially lower (.43 and .34, respectively) than the first four items. Given the distinction between inattention and hyperactive and impulsivity symptoms in the Diagnostic and Statistical Manual of Mental Disorders (4th ed. [DSM-IV]; American Psychiatric Association, 1994) and the relative low loadings of these two items, they were not included on this factor for the CFA on Sample 2. The elimination of these two items resulted in a clear Inattentive Behavior factor.
The third factor involved overt and covert conduct problem behaviors. The overt aspect involved verbal and physical aggression toward other children (i.e., "teases or provokes other children," "verbally fights with friends his/her own age," "physically fights with friends his/her own age," "verbally fights with sisters and brothers," and "physically fights with sisters and brothers"). The covert aspect involved the behaviors of lying, stealing, and destruction of property. The item with the lowest loading on this factor was "is careless with toys and other objects" (.33). This item also had a relatively high loading (.26) on the ADHD factor. Because the item was also conceptually different from the other items on this factor, it was not included in the CFA on the second sample. The elimination of this item resulted in a clear Conduct Problem Behavior factor.
The fourth factor did not represent a meaningful dimension (see Table 1). In addition, there were only three items with substantial loadings on this factor (i.e., "whines" "cries easily," and "dawdles or lingers at mealtime"), with the other six items having loadings of approximately .30. This fourth factor may represent a response bias effect (Nunnally & Bernstein, 1994; C. Parks, personal communication, May 8, 2000). For example, when items with low loadings on the first factor separate from the first factor to form a separate factor consisting of items with low loadings, this suggests response bias and the possibility of a meaningless factor. The inclusion of this fourth factor, however, resulted in stronger and clearer first factor because the items on the fourth factor were removed from the first factor. This was the reason that the four-factor model was selected over the three-factor model. The four-factor model thus resulted in three clinically meaningful factors and one factor that did not represent a meaningful dimension and may also represent response bias (Nunnally & Bernstein, 1994; C. Parks, personal communication, May 8, 2000). Our goal in the CFA phase of the study was to evaluate the fit of the three clinically meaningful factors. The fourth factor was not included in the CFA because it did not represent a meaningful dimension, the items on this factor had low loadings, and the factor may represent response bias.
CFA on Sample 2. The first goal was to determine if the three-factor model resulted in a good fit. The three factors were (a) Oppositional Defiant Behavior Toward Adults, (b) Inattentive Behavior, and (c) Conduct Problem Behavior. The second goal was to determine if the fit of this three-factor model was significantly better than two- and one-factor models. In the two-factor model, the Oppositional Defiant and Conduct Problem Behavior factors were combined into a single factor with the Inattentive Behavior dimension being the second factor. In the one-factor model, the three factors were combined into a single factor.
EQS (version 5.7b, Multivariate Software, Encino, CA; Bentler, 1995) was used to perform the CFA on the second sample. Maximum likelihood estimation was used for these analyses along with robust estimation procedures. The EQS Comparative Fit Index (CFI), EQS Robust Comparative Fit Index (RCFI), LISREL Goodness-of-Fit (GFI), the standardized root mean square residual (SRMR), and the root mean square error of approximation (RMSEA) were used to evaluate model fit. The GFI provides a measure of the relative amount of variance and covariance accounted for by the model, whereas the CFI provides a measure of fit of a particular model relative to another model, usually a null model. Values greater than .90 for the GFI and CFI are usually required to indicate a good fit (Byrne, 1994). The SRMR represents the average of the absolute discrepancies between the observed and hypothesized matrices in correlational units (Bentler, 1995). Values of .05 or lower are suggested as necessary to consider a model a good fit. RMSEA provides a measure of model fit relative to the population covariance matrix when the complexity of the model is taken into account. Values less than .05 for the RMSEA indicate a close fit, with values as high as .08 representing a reasonable fit (Joreskog & Sorbom, 1993).
The three-factor model resulted in a reasonable fit. The CFI, RCFI, GFI, SRMR, and RMSEA values for the three-factor model were .860, .849, .860, .060, and .087 (.90 CI = .083-.090), respectively. The Multivariate Lagrange Multiplier Test, however, indicated that a significant improvement in model fit would occur with two correlated errors, DELTAchi2(2) = 540.46, p < .000001. The two-item pairs were "verbally fights with sisters and brothers" with "physically fights with sisters and brothers" and "steals" with "lies". The similar content of the first two items and the perhaps the high co-occurrence of the second two were possible reasons for the high correlations (i.e., .59 and .44, respectively; see Byrne, 1994). The CFI, RCFI, GFI, SRMR, and RMSEA values for the three-factor model with two correlated errors were .913, .907, .900, .049, and .069 (.90 CI = .065-.072), respectively. These values indicate a good fit. The three-factor model with two correlated errors also provided a significantly better fit than the two- and one-factor models, DELTAchi2(4) = 1030.00, p < .000001, and DELTAchi2(5) = 2390.48, p < .000001, respectively.
Multiple-group CFA across samples. A multiple-group CFA was used to determine if the factor pattern, factor loadings, factor correlations, and two correlated errors were equivalent from Sample 2 to Sample 1 (Byrne, 1994). Although it would have been better to have a third sample for this analysis, the results, nonetheless, allow for a specific test of the equivalence of these parameters from Sample 2 to Sample 1. Because there were 25 constraints imposed in this analysis, the per-comparison alpha was set at .002. None of the constraints were significant (all 25 ps > .02 and 23 ps > .05). The factor pattern, loadings, correlations, and correlated errors were thus equivalent across the two samples. Table 2 shows the factor loadings and Table 3 the factor correlations for Sample 1 and Sample 2. Each of the items had a significant loading on its assigned factor (ps < .0001). The factor correlations were significant as well (ps < .0001).
CFA on sex. For boys, the CFI, RCFI, GFI, SRMR, and RMSEA values for the three-factor model with two correlated errors were .911, .908, .891, .052, and .072 (.90 CI = .069-.075), respectively. For girls, these values were .903, .898, .895, .050, and .069 (.90 CI = .066-.073), respectively. This model also provided a significantly better fit than the two- and one-factor models for boys, DELTAchi2(4) = 1,122.28, p < .000001, and DELTAchi2(5) = 2,781.71, p < .000001, respectively, as well as for girls, DELTAchi2(4) = 803.45, p < .000001, and DELTAchi2(5) = 2,002.79, p < .000001, respectively.
Multiple-group CFA on sex. In the multiple-group CFA across sex, 3 of the 25 constraints were significant (ps < .0001). Three of the items had significantly higher loadings for boys than girls (i.e., "destroys toys and other objects," "verbally fights with friends his/her own age," and "physically fights with friends his/her own age"). With the exception of these three items, there was equivalence of factor pattern, factor loadings, factor correlations, and error correlations across sex. Table 2 show the factor loadings and Table 3 the factor correlations for boys and girls. Each of the items had a significant loadings on its assigned factor with the three factor correlations also being significant for boys and girls (ps < .0001).
Initial Normative Information on the Three ECBI Subscales
Table 4 shows the internal consistency coefficients (Cronbach's alpha), means, standard deviations, and the scores corresponding to the 80th, 90th, and 98th percentiles for the Oppositional Defiant, Inattentive, and Conduct Problem Behavior subscales. These results are presented for boys and girls as well as boys and girls within four age ranges (2-5,6-9,10-13,and 14-17).
Correlations Between ECBI IS and PS Subscales
For boys (n = 1,322), the correlation between the IS and PS Oppositional Defiant Behavior subscales was .77, .70 for the Inattentive Behavior subscales, and .73 for the Conduct Problem Behavior subscales. For girls (n = 1,205), the correlation between the IS and PS Oppositional Defiant Behavior subscales was .77, .70 for the Inattentive Behavior subscales, and .73 for the Conduct Problem Behavior subscales.
The results from the EFA identified three meaningful factors in the ECBI and a fourth, poorly defined factor that probably represents response bias (Nunnally & Bernstein, 1994; C. Parks, personal communication, May 8, 2000).
The three meaningful factors were Oppositional Behavior Toward Adults, Inattentive Behavior, and Conduct Problem Behavior. A CFA on a second sample indicated that these three factors provided an excellent fit. The CFA also indicated that this model provided a significantly better fit than two- and one-factor models. In addition, multiple-group CFA demonstrated the equivalence of the factor pattern, loadings, correlations, and correlated errors across the samples. Similar results also occurred for sex. Three items, however, had significantly higher loadings for boys than girls (i.e., "destroys toys and other objects," "verbally fights with friends his/her own age," and "physically fights with friends his/her own age"). For boys, thus, verbal aggression, physical aggression, and property destruction had a stronger relation to Conduct Problem Behavior construct than for girls.
The attention deficit and disruptive behavior domain (Quay & Hogan, 1999) involves other factors in addition to these three factors in the ECBI. Within the ECBI item set, however, the Oppositional Defiant Behavior Toward Adults, Inattentive Behavior, and Conduct Problem behavior are the three clinically meaningful dimensions. Given the size of our samples as well as the repeatability of our findings across samples, sex, and IS/PS ratings, the same three factors should occur in other similarly large samples.
Usefulness of the Subscales
Although our normative data on the three subscales do not represent national norms and individuals should carefully consider the characteristics of our sample in their use of these norms, this information on 1,322 boys and 1,205 girls from four states as well as urban and rural settings represents the best available normative information on the ECBI. The normative information also addresses a major limitation of the ECBI. When it is important to identify children who are presenting with "pure" CP [conduct problems], one potential solution [to the problem that the ECBI contains ODD, CD, and ADHD items] might be to score only those items related to either ODD (or ODD and CD), which would facilitate the selection of more homogeneous samples of children (but would preclude the use of existing normative data and cut-off scores). (McMahon & Estes, 1997, p. 153)
Our results provide solutions to these issues. Here we have established and replicated Opposition Defiant, Conduct Problem, and Inattentive Behavior subscales, provided initial normative data on the subscales as well as suggestions for cut-off scores for screening for a more complete assessment (e.g., the 90th percentile score on the subscales). The three subscales should make the ECBI more useful as a screening, outcome, and research measure.
In terms of a screening measure for treatment programs, probably the most common research use of the scale (e.g., Webster-Stratton, 1998), the Oppositional Defiant and Conduct Problem subscales should be particularly useful. As noted in the quote, if the goal is to identify children with only high levels of oppositional defiant behavior for parenting therapy interventions, then high scores on the Oppositional Defiant Behavior subscale would provide a better choice than high scores on the complete ECBI. In addition, the Conduct Problem Behavior subscale would be useful to determine the severity of the child's problems. For example, a child with a high score on the Oppositional Defiant and Conduct Problem Behavior subscales would probably represent a child with more severe problems than a child with only high scores on the Oppositional Defiant Behavior subscale (i.e., a child who has advanced further in the disruptive behavior disorder progression). If a child had high scores on the three subscales, then there would probably be an increased likelihood that a more formal assessment would indicate comorbidity for ODD, CD, and ADHD.
The three subscales may also result in the ECBI being more sensitive to interventions. For example, because many parenting therapy programs focus on increasing the child's compliance in parent-child interactions (McMahon & Wells, 1998), such treatment procedures may result in greater changes on the Oppositional Defiant Behavior subscale than the other two subscales. In addition, the behaviors coded in direct observational systems may have a stronger correlation with the Oppositional Defiant Behavior subscale than the other two subscales, especially given the focus on the child's oppositional defiant behavior toward the parent in these coding systems (McMahon & Estes, 1997). Direct observation of the child's behavior in interactions with other children, however, might show a stronger relation to the Conduct Problem Behavior subscale.
The three subscales should make the ECBI more useful for these reasons. Individuals should remember, however, that DSM-IV ODD is measured much better by the Oppositional Defiant Behavior subscale than the DSM-IV ADHD inattentive symptom dimension is measured by the Inattentive Behavior subscale. In a similar fashion, the CP subscale does not assess the full range of the DSM-IV CD symptoms, especially the more serious ones. In spite of these cautions, we believe that researchers and clinicians will find the subscales, especially the Oppositional Defiant and Conduct Problem Behavior subscales, more useful as screening, outcome, and research measures than the total ECBI score.
1 A CFA was not performed on the PS item ratings due to the categorical nature of the ratings (i.e., "yes" or "no"). It was possible, however, to perform a CFA on the PS ratings through the creation of item sets or parcels. For each factor, the items assigned to that factor were combined into two sets. Items were assigned to the sets based on their loadings from the CFA on the IS ratings (i.e., the item with highest loading was assigned to one set, the item with next highest loading to the next set, and so on until each item on the factor was assigned to one of the two sets). Here each factor is represented by two variables (i.e., two sets of items). The CFI, RCFI, GFI values for the three-factor model on Sample 1, Sample 2, boys and girls were .99 or higher. There was also equivalence across the samples and sex. Although these analyses do not allow for the evaluation of individual items, they can be considered to provide additional support for the three-factor model. These results are available from G. Leonard Burns.