INTERPARENTAL CONFLICT, ADOLESCENT BEHAVIORAL PROBLEMS, AND ADOLESCENT COMPETENCE: CONVERGENT AND DISCRIMINANT VALIDITY

EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT 10.1177/0013164403258462 EPSTEIN ET AL. ARTICLE INTERPARENTAL CONFLICT, ADOLESCENT BEHAVIORAL PROBLEMS, AND ADOLESCENT COMPETENCE: CONVERGENT AND DISCRIMINANT VALIDITY MONICA K. EPSTEIN University of South Florida KIMBERLY RENK University of Central Florida AMY M. DUHIG Yale University GEORGETTA L. BOSCO Counseling Center of Tampa Bay, Inc. VICKY PHARES University of South Florida To address the lack of studies examining the convergent and discriminant validity of cross-informant ratings, several statistical approaches were used in this study to evaluate the convergent and discriminant validity for ratings of interparental conflict, adolescent behavioral problems, and adolescent competence. A total of 272 adolescent-motherfather triads completed measures assessing interparental conflict, adolescent behavioral problems, and adolescent competence. The validity of these measures appeared questionnable when relationships were examined using Campbell and Fiske s criteria and confirmatory factor analysis, respectively. In contrast, analyses demonstrated adequate fit of cross-informant ratings in a correlated uniqueness model, supporting convergent and discriminant validity. Overall, results suggested that informant effects should be analyzed when examining interparental conflict and adolescent adjustment. These findings emphasize the importance of considering the measurement issues of cross-informant ratings when examining interparental conflict, adolescent behavioral problems, and adolescent competence. Keywords: cross-informant ratings; conflict; behavior problems; competence Educational and Psychological Measurement, Vol. 64 No. 3, June 2004 475-495 DOI: 10.1177/0013164403258462 2004 Sage Publications 475

476 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT The relationships among interparental conflict, adolescent behavioral problems, and adolescent competence have been well documented in the research literature (as reviewed by Cummings & Davies, 1994). In general, findings have suggested that interparental conflict is distressing for children and adolescents. Furthermore, interparental conflict has been linked closely to increased emotional and behavioral problems as well as to decreased competence in children and adolescents. Although a number of studies have examined the interrelationships among these variables, few researchers have examined the validity of these constructs directly when using multiple informants. Issues of the convergent and discriminant validity of constructs measured with the ratings of multiple informants as well as issues pertaining to potential informant bias should be addressed if research in this area is to be advanced. The purpose of this study was to evaluate these validity issues using a variety of statistical approaches and to examine the impact of these issues on the assessment of interparental conflict, adolescent behavioral problems, and adolescent competence. Research has suggested that multiple informants should be used when assessing conflict and violence in families (e.g., Sternberg, Lamb, & Dawud- Noursi, 1997). Thus far, most studies use parental reports only to assess the level of conflict in families rather than employing information from multiple informants. Parents may underestimate or overestimate the amount of conflict their children have witnessed (Grych, Seid, & Fincham, 1992) relative to the reports of the children themselves, however. Parents may provide underestimates when they are unaware of the amount of conflict their children have witnessed, whereas they may provide overestimates when they define interactions that are not salient to children in the family as conflictual (Grych et al., 1992). As a result, Grych and colleagues (1992) suggested that more accurate estimates of children s exposure to conflict would be provided by including measures that assess directly such exposure from children s perspectives. Similarly, many researchers have advocated for the use of multiple informants in the rating of emotional and behavioral functioning in children and adolescents (e.g., Achenbach, McConaughy, & Howell, 1987; Duhig, Renk, Epstein, & Phares, 2000). Mothers and fathers are called on often to provide ratings of such functioning. Such ratings, however, have been criticized as displaying bias and a lack of construct validity when multiple informants from the same family do not demonstrate a substantial degree of congruence in their ratings (e.g., Noller & Guthrie, 1991). These findings have suggested that there is no gold standard in the ratings of emotional and behavioral This research was supported by the National Institute of Mental Health, Grant No. R29 49601-03. Correspondence concerning this article should be addressed to Kimberly Renk, Ph.D., University of Central Florida, Department of Psychology, P.O. Box 161390, Orlando, FL 32816; e-mail: krenk@pegasus.cc.ucf.edu.

EPSTEIN ET AL. 477 problems in children and adolescents and that it is very important that the validity of cross-informant ratings of such behavior be examined. Because there is no gold standard when examining cross-informant ratings of interparental conflict, adolescents emotional and behavioral problems, and adolescents competence, the examination of the validity of these constructs should be continued. Research concerning construct validity typically has emphasized the degree to which any measure or method determines the establishment of convergence and discrimination. Agreement among multiple informants regarding measured theoretical constructs or traits demonstrates convergent validity (e.g., Bryant, 2000). The ability of a construct or trait to be distinguished from other underlying constructs or traits, especially when measured by the same informants, is evidence of discriminant validity (e.g., Bryant, 2000). Historically, Campbell and Fiske (1959) advocated using convergent and discriminant validity to establish construct validity through the use of a multitrait-multimethod (MTMM) matrix. They further stated the need for multiple informants (or methods) and multiple traits in determining convergent and discriminant validity. Although researchers have documented good convergent validity in parental ratings of internalizing and externalizing behavioral problems in children and adolescents, there also has been evidence of weaker discriminant validity and large informant effects (Stanger & Lewis, 1993). A major difficulty with the establishment of construct validity has been the degree to which ratings are influenced by informant (or method) variance. Informant (or method) variance may distort the observed correlations between measures (Campbell & Fiske, 1959). Because each informant contributes a considerable amount of variance not accounted for by other informants (i.e., unique variance), the optimal approach to assessing the relationships among interparental conflict, adolescent behavioral problems, and adolescent competence involves the use of multiple sources of information for each construct (e.g., Grych & Fincham, 1990). For example, moderate degrees of consistency between maternal and paternal reports of emotional and behavioral problems in children and adolescents have been found (Achenbach et al., 1987; Duhig et al., 2000). The agreement of children and adolescents with their parents, however, has been found to be considerably lower than parent-parent agreement, especially with regard to internalizing behavioral problems in children and adolescents (Achenbach et al., 1987). In an effort to address such validity issues when using multiple informants, three different methods for examining the convergent and discriminant validity of interparental conflict, adolescent behavioral problems, and adolescent competence were examined in this study. Campbell and Fiske s (1959) original work with MTMM matrices assessed construct validation through the visual inspection of the correlations between each of the traits and methods or informants. Campbell and Fiske (1959) proposed that con-

478 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT struct validation could be determined by following several rules. First, Campbell and Fiske (1959) proposed that the monotrait-heteromethod correlations (i.e., the same traits rated by different informants) would differ significantly from zero and would be sufficiently large if convergent validity was demonstrated. In other words, measures of the same construct should be correlated highly if they assess a common construct (Bagozzi, 1993), even when rated by multiple informants. Next, Campbell and Fiske (1959) proposed that discriminant validity also should be assessed. For discriminant validity to be demonstrated, the monotrait-heteromethod coefficients (i.e., same traits rated by different informants) should be higher than their corresponding heterotraitheteromethod coefficients (i.e., different traits rated by different informants; Campbell & Fiske, 1959). Such a finding would demonstrate that efforts to measure the same concept by using different informants should yield higher correlations than would efforts to measure different concepts by different informants (Bagozzi, 1993). In addition, the monotrait-heteromethod coefficients should be higher than their corresponding heterotrait-monomethod coefficients (i.e., different traits rated by the same informant; Campbell & Fiske, 1959). Such a finding would demonstrate that efforts to measure the same concept with ratings by different informants produce higher correlations than do efforts to measure different concepts rated by the same informant (Bagozzi, 1993). Finally, the pattern of correlations should be the same among the heterotrait-monomethod and heterotrait-heteromethod correlations (Campbell & Fiske, 1959). Because Campbell and Fiske s (1959; Fiske & Campbell, 1992) procedure does not provide precise standards for determining how well these criteria should be met and does not consider the importance of the magnitudes of differences between pairs of coefficients (Bagozzi, 1993), a confirmatory factor analysis model has been proposed as an alternative for assessing convergent and discriminant validity (e.g., Wothke, 1996). This model hypothesizes that the total variation in the measurement of specific constructs stems from the combination of trait, method or informant, and error effects (e.g., Jöreskog, 1974). Such a model has several advantages such as allowing for the estimate of the reliability of each measure, yielding measures of goodness-of-fit for the overall model, and allowing for the comparison of alternate explanatory models (Bryant, 2000). This method, however, requires that at least three traits and three methods, four traits and two methods, or two traits and four methods be incorporated into the model (Bagozzi, 1993). Furthermore, the use of confirmatory factor analysis often results in technical problems such as improper solutions, with parameter estimates falling outside of the permissible range (e.g., Marsh, 1989). Due to these problems, a variant of the confirmatory factor analysis approach known as the correlated

EPSTEIN ET AL. 479 uniqueness model analyzed with a latent variable approach has been proposed. Kenny and Berman (1980) suggested that such a latent variable approach could be used to examine the ratings of multiple informants for a variety of constructs (also discussed as multiperspective-multidyad confirmatory factor analysis by Bray, Maxwell, & Cole, 1995). In such a model, the ratings of multiple informants are specified as imperfect indicators of the specific constructs of interest. The relationship of the ratings of specific constructs provided by multiple informants is measured using factor pattern coefficients, with high loadings of ratings on their respective construct indicating convergent validity. Because it is expected that each rating provided by informants also will include a certain amount of variability beyond the true relationship between the construct and the rating provided by the informants, residuals are accounted for in the model (Cook & Goldstein, 1993). These residuals allow for the examination of informant effects. Finally, the constructs are allowed to correlate in this model, providing a measure of the relationships among these constructs after removing the influence of measurement error and systematic informant effects (Cook & Goldstein, 1993). These correlations should reflect the discriminant validity of the constructs of interest (Cook & Goldstein, 1993). Thus, this model is important because it allows researchers to examine the unique and common perspectives of multiple informants while also examining additional variability, such as informant (or method) effects. This latent variable approach can be examined by using a correlated uniqueness model in covariance structural modeling. Applying each of the statistical models discussed here, this study evaluated the relationships among the observed measures of the hypothesized latent constructs (interparental conflict, adolescent internalizing and externalizing behavioral problems, and adolescent competence) provided by different informants (adolescents, mothers, and fathers). The specified models proposed that ratings of different informants that tap the same constructs would demonstrate convergent validity. Thus, it was expected that correlations among ratings of the same constructs by different informants would be high in the Campbell and Fiske (1959) model. For the confirmatory factor model and the correlated uniqueness model, it was expected that the ratings of interparental conflict, adolescent internalizing and externalizing behavioral problems, and adolescent competence, whether rated by adolescents, mothers, or fathers, would load highly on their respective constructs. It also was expected that residuals across the ratings provided by the same informant would be correlated in the correlated uniqueness model. This correlation should occur due to systematic informant effects across the ratings of different constructs. Finally, it was expected that the constructs examined here would demonstrate discriminant validity. In other words, the correla-

480 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT tions between the constructs would be low to moderate, demonstrating that these constructs were related but different. Method Participants A total of 272 biological adolescent-mother-father triads were recruited from a southeastern suburban area to participate in the study. Of the families, 142 had a male adolescent, whereas 130 families had a female adolescent. Adolescents ages ranged from 11 to 18 years, with a mean age of 13.52 years (SD = 1.84 years). Mothers ranged in age from 27 to 55 years (M = 40.13 years, SD = 5.06 years), and fathers ranged in age from 25 to 70 years (M = 42.77 years, SD = 6.43 years). Regarding ethnicity, the sample was primarily Caucasian (74.6%), with some families of African American (12.9%), Latino/Latina (10.7%), Asian American (0.4%), Native American (1.1%), and other (0.4%) ethnicities. A majority of the parents reported being married to their adolescent s other parent (83.1%). The remainder were divorced from their adolescent s other parent and remarried (5.9%), were divorced but not remarried (8.1%), had never married their adolescent s other parent (0.7%), were widowed (0.4%), or had some other family configuration (0.4%). Information regarding marital status was unavailable for four (1.5%) of the families. Although parents were recruited from the community, several mothers and fathers were experiencing varying degrees of psychological symptoms. Based on information obtained through structured clinical interviews using the Mini Structured Clinical Interview for the Diagnostic and Statistical Manual (SCID) and complete modules of the SCID (First, Gibbon, Williams, & Spitzer, 1995) conducted by trained researchers, 108 mothers met diagnostic criteria for at least one current or previous diagnosis during their adolescent s lifetime (22.4% met criteria for major depressive disorder, 31.3% met criteria for an anxiety disorder, and 16.9% met criteria for alcohol, substance abuse, or dependence). A total of 106 fathers met diagnostic criteria during their adolescent s lifetime (11.0% met criteria for major depressive disorder, 10.7% met criteria for an anxiety disorder, and 40.5% met criteria for alcohol or substance abuse or dependence). A total of 50 families consisted of mothers with a psychiatric diagnosis and nonclinical fathers, 52 families had fathers with a psychiatric diagnosis and nonclinical mothers, and 49 families had mothers and fathers who both met clinical diagnostic criteria. In 98 families, neither parent met criteria for a psychiatric diagnosis. A total of 70 (25.7%) mothers were receiving individual psychological or psychiatric services, whereas 49 (18.0%) fathers were receiving individual psychological or psychiatric services. In all, 11 (4.0%) mothers and 11 (4.0%)

EPSTEIN ET AL. 481 fathers were receiving marital or family treatment. Thus, the population of families examined here were diverse in their experience of psychological symptoms. Procedure As part of a larger, prospective study, families were recruited via advertisements throughout the community. To meet eligibility for participation, mothers and fathers had to be the biological parents of adolescents who were between the ages of 11 and 18 years. Parents with more than one adolescent were asked to select one of their adolescents for participation. Thus, parents had complete control of which adolescent was brought to the research session. If both parents were not living currently in the same house, adolescents were required to have at least monthly face-to-face contact with noncustodial parents. Families who contacted the researchers were invited to attend a 3- hour session at a southeastern university and subsequently were given $60 for their participation. After obtaining informed consent and permission from both parents and assent from adolescents at the beginning of the session, each participant was administered individually the research protocol by trained graduate-level research assistants in clinical psychology. Each parent completed the structured clinical interview mentioned previously (Mini-SCID and SCID modules; First et al., 1995) and questionnaires designed to assess interparental conflict, adolescent behavioral problems, and adolescent competence along with additional questionnaires not included in this study. Adolescents completed parallel measures of interparental conflict and their own behavioral problems and competence as well as other measures not included in this study. Participants were provided with a brief explanation of the study and were given the opportunity to have any questions answered following the completion of their participation. Measures Interparental conflict. Adolescents perceptions of interparental conflict were measured using the six-item Marital Discord subscale of the Child s Perception Questionnaire (CPQ; Emery & O Leary, 1982). Mothers and fathers completed the O Leary-Porter Scale Revised (OPS-R; Porter & O Leary, 1980), which was developed to assess the frequency of overt marital hostility observed by adolescents. The scores for both these measures have had adequate reliability and validity (e.g., Emery & O Leary, 1982; Porter & O Leary, 1980). In this study, the scores for both measures demonstrated adequate internal consistency (.90 for the CPQ and.83 for the OPS when completed by mothers;.81 for the OPS when completed by fathers).

482 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT Adolescent behavioral problems and competence. To assess the adolescents behavioral problems and competence, adolescents completed the Youth Self-Report (YSR; Achenbach, 1991b). The YSR is a measure designed to provide standardized descriptions of emotional and behavioral problems and competence in children and adolescents. A total of 118 items measure self-reported internalizing, externalizing, and total behavioral problems. Self-reported involvement in sports, activities, organizations, and family and social interactions as well as the quality of these activities and interactions were used to assess adolescents competence. In addition, mothers and fathers completed the Child Behavior Checklist (CBCL; Achenbach, 1991a) independently. The CBCL is a parentcompleted measure similar to the YSR. This measure consists of 120 problem behavior and 20 competence items. The CBCL has been shown to discriminate between clinic-referred and nonreferred children (Achenbach, 1991a). T-scores from both the YSR and the CBCL for adolescent internalizing behavioral problems, adolescent externalizing behavioral problems, and adolescent competence were used in this study. Although T-scores may result in a truncated distribution for some measures, Achenbach (1991a) reported that the T-scores for the Internalizing Behavior Problems, Externalizing Behavior Problems, and Competence scales were not truncated and should yield results similar to those obtained with raw scores. The Total Behavior Problems scores from the YSR and CBCL were not used in this study due to the item overlap between this scale and those for the Internalizing and Externalizing Behavior Problems scales. Adequate reliability and validity for scores on both the YSR and the CBCL have been documented (Achenbach, 1991a, 1991b). Results Differences Among Ratings Provided by Adolescents, Mothers, and Fathers Descriptive statistics for the measures used in this study are provided in Table 1. To examine the potential differences in the ratings provided by adolescents, mothers, and fathers, a series of paired-samples t tests were conducted. With regard to interparental conflict, adolescents differed significantly from their mothers, t (df = 269) = 29.84, p <.000, d = 3.64, and their fathers, t (df = 260) = 30.21, p <.000, d = 3.75, in their ratings. In contrast, maternal and paternal ratings of interparental conflict did not differ significantly, t (df = 261) =.03, not significant, d =.00. With regard to adolescent internalizing behavioral problems, adolescents differed significantly from their mothers, t (df = 267) = 2.03, p <.04, d =.25, but not from their fathers, t

EPSTEIN ET AL. 483 Table 1 Descriptive Statistics M SD Skew Kurtosis Interparental conflict Adolescent report 1.34 0.20 1.33 0.98 Maternal report 12.24 6.16 0.65 0.73 Paternal report 12.12 5.89 0.61 0.11 Internalizing problems Adolescent report 48.75 9.90 0.49 0.37 Maternal report 50.31 10.40 0.24 0.25 Paternal report 48.82 9.92 0.23 0.03 Externalizing problems Adolescent report 50.59 10.31 0.02 0.39 Maternal report 50.28 10.05 0.39 0.16 Paternal report 50.47 9.80 0.18 0.28 Competence Adolescent report 48.86 10.67 0.30 1.00 Maternal report 47.62 9.41 0.16 0.14 Paternal report 46.75 9.19 0.11 0.04 (df = 264) =.20, not significant, d =.02, in their ratings. Mothers and fathers also differed significantly in their ratings of their adolescents internalizing behavioral problems, t (df = 261) = 2.12, p <.04, d =.26. With regard to externalizing behavioral problems, adolescents did not differ significantly from their mothers, t (df = 267) =.43, not significant, d =.05, or from their fathers, t (df = 264) =.07, not significant, d =.00, in their ratings. Similarly, mothers and fathers did not differ significantly in their ratings, t (df = 261) =.37, not significant, d =.05. Finally, with regard to adolescent competence, adolescents differed significantly from their mothers, t (df = 246) = 2.30, p <.02, d =.29, and from their fathers, t (df = 244) = 2.89, p <.004, d=.37, in their ratings. Mothers and fathers, however, did not differ in their ratings, t (df = 240) = 1.71, not significant, d =.22. Application of the Campbell and Fiske Criteria Campbell and Fiske s (1959) criteria were applied to the correlation matrix in Table 2. To assess convergent validity, the validity diagonals were examined. Although each of the correlations in the validity diagonals were significantly different from zero, the diagonal values were not all large. The correlations ranged from.15 to.58, with those correlations measuring the internalizing behavioral problems of adolescents being the smallest in magnitude (.25,.15, and.38). These correlation magnitudes were consistent with previous findings, which state that cross-informants tend to show a smaller amount of agreement on measures of internalizing behavioral problems in children and adolescents (e.g., Achenbach et al., 1987; Duhig et al., 2000).

Table 2 Multitrait-Multimethod Correlation Coefficient Matrix for Study Measures 1 2 3 4 5 6 7 8 9 10 11 12 Adolescent report 1. Interparental conflict 2. Internalizing problems.39 a ** 3. Externalizing problems.36 a **.54 a ** 4. Competence.12 a.15 a *.14 a * Maternal report 5. Interparental conflict.43 b **.16**.19**.06 6. Internalizing problems.17**.25 b **.15*.17**.23 a ** 7. Externalizing problems.20**.20**.43 b **.13*.30 a **.56 a ** 8. Competence.23**.08.19**.43 b **.20 a **.31 a **.40 a ** Paternal report 9. Interparental conflict.36 b **.15*.15*.10.42 b **.13*.19**.15** 10. Internalizing problems.10.15 b *.08.29**.03.38 b **.22**.36**.23 a ** 11. Externalizing problems.08.07.33 b **.14**.07.23**.53 b **.35**.27 a **.61 a ** 12. Competence.12.20**.16**.39 b **.11.24**.24**.58 b **.12 a.29 a **.28 a ** Note: Numbers represent heterotrait-heteromethod discriminant validity coefficients unless otherwise indicated. a. Heterotrait-monomethod discriminant validity coefficients. b. Convergent validity coefficients. *p <.05. **p <.01. 484

EPSTEIN ET AL. 485 To examine discriminant validity, the validity diagonals were first compared with the heterotrait-heteromethod coefficients. This assessment involved the comparison of each of the 12 correlations in the validity diagonal with each of the 36 heterotrait-heteromethod coefficients. This comparison revealed several heterotrait-heteromethod coefficients that were larger than those in the validity diagonal. In particular, the correlations between adolescent-mother reports of internalizing behavioral problems, adolescentfather reports of interparental conflict, adolescent-father reports of internalizing behavioral problems, and adolescent-father reports of externalizing behavioral problems were larger than were those found in the validity diagonal. As the proportion of failures of this first criterion was p = 24 / 432 =.055, which is at the same level as would be expected by chance (p =.05), it appeared that this criterion was met overall. Next, the validity diagonals were compared with the heterotraitmonomethod coefficients. This assessment involved the comparison of each of the 12 correlations in the validity diagonal with each of the 18 heterotraitmonomethod coefficients. Again, several violations were noted. In this case, the proportion of failures was p = 61 / 216 =.28, which is greater than would be expected by chance. As a result, the second criterion for discriminant validity, in which coefficients in the validity diagonal should be larger than heterotrait-monomethod coefficients, was not supported. Finally, the pattern of correlations across the heterotrait-heteromethod and the heterotrait-monomethod correlations was examined using the rank order of coefficients in each triangle. The pattern of correlations in the heterotrait-heteromethod and the heterotrait-monomethod triangles varied greatly. As a result, the third criterion for discriminant validity, in which the pattern of correlation across the heterotrait-heteromethod and the heterotraitmonomethod triangles should be the same, was not supported. Application of the Confirmatory Factor Analysis Model Due to the subjectivity of Campbell and Fiske s (1959) criteria for MTMM matrices, a confirmatory factor analysis model was used to determine the relationships among the measured variables (assessed by the CPQ, OPS-R, YSR, and CBCL) in reference to the latent constructs (interparental conflict, internalizing behavioral problems, externalizing behavioral problems, and competence as well as adolescent ratings, maternal ratings, and paternal ratings). In the confirmatory factor analysis model, both trait factors and method factors are specified and are allowed to correlate. Thus, this model consisted of four intercorrelated trait factors (Interparental Conflict, Internalizing Behavioral Problems, Externalizing Behavioral Problems, and Social Competence) and three intercorrelated method factors (Adolescent Ratings, Maternal Ratings, and Paternal Ratings), each predicting three mea-

486 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT sured variables. More specifically, the measured variables completed by adolescents, mothers, and fathers were allowed to load on their respective latent constructs (e.g., the adolescents report on the CPQ and maternal and paternal reports on the OPS-R were allowed to load on the interparental conflict construct). If the measures completed by the different informants were to display convergent validity, the factor pattern coefficients of these measures should be higher on each of their respective constructs than for other constructs. In addition, it would be expected that ratings completed by each informant regardless of the construct measured also should load highly on each respective method factor. Finally, each of the latent variables should demonstrate low to moderate correlations rather than large correlations if they exhibit discriminant validity. Model results. Based on the results for the proposed model, it was determined that the solution was not proper. Several of the parameter estimates were out of range, consistent with problems noted in previous research (e.g., Marsh, 1989). As a result, the fit indices of this model were of questionnable value and are not reported here. Application of the Correlated Uniqueness Model Due to the subjectivity in identifying relationships in MTMM matrices using the criteria provided by Campbell and Fiske (1959) and the statistical problems demonstrated with confirmatory factor analyses, a correlated uniqueness model also was used to determine the relationships of the measured variables (assessed by the CPQ, OPS-R, YSR, and CBCL) to the latent constructs (interparental conflict, internalizing behavioral problems, externalizing behavioral problems, and competence). As mentioned previously, a correlated uniqueness model is consistent with the measuring of the latent variable approach proposed by Kenny and Berman (1980). In correlated uniqueness models, only trait factors rather than trait and method factors are specified in the model (Kenny & Kashy, 1992). The trait factors or constructs of interest are allowed to correlate. To accommodate possible informant (or method) variance, the disturbance terms for each measure completed by the same informant (adolescent, mother, or father) are allowed to covary (Kenny & Kashy, 1992). The advantage of using this type of approach is that it allows for multidimensional effects of method (Kenny & Kashy, 1992) and has the ability to produce an identified solution. In other approaches, separate method and trait factors are specified for MTMM models, sometimes resulting in empirical underidentification of the model. Thus, the unique and common contribution of multiple informants to the ratings of a construct of interest as well as possible variance accounted for by using the same informants across constructs can be measured.

EPSTEIN ET AL. 487 The model tested consisted of four intercorrelated latent variables (interparental conflict, adolescent internalizing behavioral problems, adolescent externalizing behavioral problems, and adolescent competence), each predicting three measured variables (adolescent report, maternal report, and paternal report). More specifically, the measured variables completed by adolescents, mothers, and fathers were allowed to load on their respective latent constructs (e.g., the adolescents report on the CPQ and maternal and paternal reports on the OPS-R were allowed to load on the interparental conflict construct; e.g., Bray et al., 1995). Furthermore, all measured variables completed by the same informant were allowed to covary in an effort to account for informant (or method) variance (e.g., the scores from the CPQ and the YSR scales completed by adolescents were allowed to covary). Finally, interparental conflict, adolescent internalizing behavioral problems, adolescent externalizing behavioral problems, and adolescent competence (the latent constructs or variables) were allowed to intercorrelate. These latent variables should demonstrate low to moderate correlations rather than large correlations if they exhibit discriminant validity. A correlated uniqueness model was then tested using LISREL 8. Model results. Based on the fit measures for the proposed model, the fit of the data to the model was good. Although the χ 2 of exact fit was significant (χ 2 = 52.56, df = 30, p <.007), indicating that the proposed model was not a plausible model, this χ 2 index tends to be sensitive to sample size and to departures from multivariate normality of variables (Schumacker & Lomax, 1996). As a result, other indices of fit were examined to derive a better estimate of the fit of the model. The χ 2 of not a close fit (χ 2 =.05, df = 30) indicated that the proposed model was a good estimate. In addition, the root mean square error of approximation was small (.05), the normed fit index was large (.95), and the nonnormed fit index was large (.95). These measures indicated that the model demonstrated good fit to the data. Thus, the model appeared to account adequately for the relationships among the measured and latent variables. The residuals examining informant effects, factor pattern coefficients, factor structure coefficients, and intercorrelations of the latent constructs are provided in Tables 3 and 4. If the ratings provided by each of the informants in the family are reliable, these ratings should be related to their respective latent variables. In this model, such a relationship was demonstrated through the significant factor pattern coefficients of the measured ratings on each of the respective latent variables. Each of the factor pattern coefficients in this model was greater than.33. Although the factor pattern coefficients indicate the presence of convergent validity, none of these factor pattern coefficients were perfect indicators of their latent variables. This finding suggested that each informant also was providing a unique component in his or her ratings. In addi-

Table 3 Correlations Among the Residuals for Examining Informant Effects in the Correlated Uniqueness Model CPQ OPSM OPSD YSRINT CBINTM CBINTD YSREXT CBEXTM CBEXTD YSRCOMP CBCOMPM CBCOMPD CPQ. OPSM.. OPSD.... YSRINT.27**... CBINTM..12*... CBINTD...13*... YSREXT.23**...46**... CBEXTM..13**...35**.. CBEXTD...15**...39**. YSRCOMP.01...04...05 CBCOMPM..04...05...08 CBCOMPD...01...04..03 CPQ = adolescent report of conflict; OPSM = maternal report of conflict; OPSD = paternal report of conflict; YSRINT = adolescent report of internalizing; CBINTM = maternal report of internalizing; CBINTD = paternal report of internalizing; YSREXT = adolescent report of externalizing; CBEXTM = maternal report of externalizing; CBEXTD = paternal report of externalizing; YSRCOMP = adolescent report of competence; CBCOMPM = maternal report of competence; CBCOMPD = paternal report of competence. *p <.05. **p <.01. 488

EPSTEIN ET AL. 489 Table 4 Factor Pattern Coefficients and Intercorrelation of Constructs in the Correlated Uniqueness Model Conflict Internalizing Externalizing Competence Factor pattern coefficients CPQ.57 OPSM.66 OPSD.64 YSRINT.33 CBINTM.55 CBINTD.66 YSREXT.52 CBEXTM.73 CBEXTD.71 YSRCOMP.55 CBCOMPM.81 CBCOMPD.72 Factor structure coefficients CPQ.57.18.19.18 OPSM.66.20.21.20 OPSD.64.20.21.20 YSRINT.10.33.16.21 CBINTM.17.55.26.35 CBINTD.20.66.31.42 YSREXT.17.25.52.26 CBEXTM.24.35.73.37 CBEXTD.23.34.71.36 YSRCOMP.17.35.27.55 CBCOMPM.25.52.41.81 CBCOMPD.22.46.36.72 Intercorrelations of constructs Conflict Internalizing.31 Externalizing.32.47 Competence.31.64.50 Note. All factor pattern coefficients and interfactor correlations are statistically significant at.01 level. CPQ = adolescent report of conflict; OPSM = maternal report of conflict; OPSD = paternal report of conflict; YSRINT = adolescent report of internalizing; CBINTM = maternal report of internalizing; CBINTD = paternal report of internalizing; YSREXT = adolescent report of externalizing; CBEXTM = maternal report of externalizing; CBEXTD = paternal report of externalizing; YSRCOMP = adolescent report of competence; CBCOMPM = maternal report of competence; CBCOMPD = paternal report of competence. tion, with regard to ratings of interparental conflict, adolescent internalizing behavioral problems, and adolescent externalizing behavioral problems, it appears that the residuals across measures completed by the same informant are related moderately to each other and are significant in a statistical sense, indicating a systematic bias of informant ratings.

490 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT This model also demonstrated the discriminant validity of the constructs of interest through the moderate but not large correlations among the latent variables. Furthermore, the correlations among adolescent competence and the other latent variables were in the expected direction. Overall, these relationships demonstrated that these constructs were different but related. Discussion This study examined the validity of the cross-informant ratings of interparental conflict, adolescent behavioral problems, and adolescent competence using several statistical procedures. The findings of this study supported the convergent validity of interparental conflict, adolescent internalizing behavioral problems, adolescent externalizing behavioral problems, and adolescent competence as assessed by the ratings of adolescents, mothers, and fathers in the same family. Although the convergent validity of these constructs was supported, the discriminant validity of these constructs was not supported strongly. The presence of common as well as unique components in the cross-informant ratings of family members supported the use of multiple informants in determining the levels of interparental conflict, adolescent behavioral problems, and adolescent competence. With regard to the use of Campbell and Fiske s (1959) criteria in the determination of construct validity, convergent validity was supported, whereas divergent validity was more questionable. Adolescents and their mothers and fathers tended to demonstrate significant agreement in their ratings of interparental conflict, adolescent behavioral problems, and adolescent competence. As demonstrated in other studies (e.g., Achenbach et al., 1987; Duhig et al., 2000), these cross-informants tended to show a smaller amount of agreement on measures of internalizing behavioral problems in children and adolescents. Because internalizing behavioral problems are less observable and more internal in experience, it is not uncommon for cross-informants to demonstrate lower concordance in such instances (Achenbach et al., 1987; Kenny, 1991; Kenrick & Stringfield, 1980). When using Campbell and Fiske s (1959) criteria, several violations were noted with regard to the discriminant validity of the MTMM matrix examined in this study. These findings suggested that each of the informants in this study were contributing unique variance across their ratings of the constructs examined. For example, in some cases, ratings of different constructs provided by different informants and ratings of different constructs provided by the same informant were larger than were ratings of the same construct by different informants. As demonstrated by other researchers (e.g., Stanger & Lewis, 1993), there has been evidence of weak discriminant validity and informant effects in the ratings of such variables. Overall, these results suggested that it is vitally important to continue an examination of these problems in the cross-informant ratings of family and adolescent variables.

EPSTEIN ET AL. 491 Unfortunately, the use of confirmatory factor analysis did not assist in the elucidation of the relationships among the constructs measured using crossinformant ratings in this study. Consistent with the findings of other researchers (Bagozzi, 1993; Marsh, 1989; Marsh & Bailey, 1991), who have noted the number of problems that seem to plague confirmatory factor analyses when used to interpret MTMM data, the solution for the confirmatory factor analysis model examined in this study exhibited out-of-range values for several of the parameter estimates. As a result, the solution for this model could not be used adequately to draw conclusions about the validity of the crossinformant ratings examined in this study. With regard to the correlated uniqueness model used in this study, the construct validity of the ratings of interparental conflict, adolescent behavioral problems, and adolescent competence was supported. With regard to convergent validity, factor pattern coefficients loaded significantly on their respective latent variables. Although these coefficients were not equal across all informants, they were consistently moderate to high for each latent variable. Consistent with previous research (e.g., Achenbach et al., 1987; Duhig et al., 2000), loadings for mothers and fathers were relatively higher and more consistent than were those provided by the adolescents themselves. Overall, this model also supported the divergent validity of the ratings of interparental conflict, adolescent behavioral problems, and adolescent competence. The latent variables examined in this study were related but different. In most cases, interparental conflict, internalizing behavioral problems, externalizing behavioral problems, and competence were correlated moderately, as suggested by several researchers (e.g., Cummings & Davies, 1994; Grych & Fincham, 1990). Although the ratings of cross-informants using the correlated uniqueness model in this study provided support for the construct validity of interparental conflict, adolescent behavioral problems, and adolescent competence, these ratings were not perfect. The residuals in the correlated uniqueness model were related moderately and significantly in the cases of interparental conflict, adolescent internalizing behavioral problems, and externalizing behavioral problems. Because adolescents and their mothers and fathers were administered parallel forms of measures assessing these constructs, it may be concluded that the ratings provided by these cross-informants also were affected by measurement effects and by systematic informant bias. In contrast, significant relationships were not noted in the ratings of adolescent competence, possibly resulting from the multifaceted way in which adolescent competence was measured in this study. The competence scores supplied by the YSR and the CBCL are composed of a variety of components, including activities in which children participate as well as family and peer interactions (Achenbach, 1991a, 1991b). These findings suggested that cross-informant ratings of adolescent competence, in particular, deserve further study.

492 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT Ratings of interparental conflict, adolescent behavioral problems, and adolescent competence may be related to informant characteristics for several reasons. First, all family members provide their own unique individual perspectives of events that occur in their families (e.g., Achenbach et al., 1987). Second, all of the family informants may provide slightly different information based on different experiences in their own families and on differential knowledge about family members. For example, adolescents may not have witnessed all of the instances of interparental conflict in their families, which may result in differing reports across informants in the family (Grych et al., 1992). Finally, family members can witness the behaviors of themselves and other individuals in the family but are only privy to their own internal states and perceptions. The added information from such internal states may result in differences in cross-informant ratings of an individual s own behavior versus that of other family members, especially in the case of adolescent internalizing behavioral problems. As a result, using different statistical procedures to demonstrate construct validity, the findings of this study supported the continued use of multiple informants in the assessment of family functioning. Only by considering the perceptions of multiple family members can clinicians and researchers reach a gold standard in assessment of families. The results of this study also supported the continued use of new and innovative statistical methods to decipher the components of construct validity. It should be noted that each of the statistical methods used in this study has its own weaknesses. For example, the analysis of MTMM matrices using Campbell and Fiske s (1959) proposed criteria is somewhat subjective in nature (Bagozzi, 1993). Furthermore, as demonstrated in this study, the use of confirmatory factor analysis for such data also can be problematic (Bagozzi, 1993; Marsh, 1989; Marsh & Bailey, 1991). The correlated uniqueness model was used in this study to explore a latent variable approach because it represents the data more accurately when the effects of a particular method are not unidimensional and because it results often in well-defined and proper solutions. The correlated uniqueness model has its own limitations, however. For example, Byrne and Goffin (1993) noted that solutions of correlated uniqueness models typically yield trait loadings and correlations that are higher than are those for a general factor analysis model. This characteristic of the correlated uniqueness model may limit our findings by providing stronger evidence of convergent validity and weaker evidence of discriminant validity. Another limitation to the correlated uniqueness model is that it requires the theoretical factor pattern coefficients to be equal in order to eliminate a biasing effect. Although the theoretical factor pattern coefficients may be equal in magnitude for each of the latent constructs, the factor pattern coefficients

EPSTEIN ET AL. 493 of the measured variables on their respective constructs were consistent but were not equal in the model used in this study. This finding may have been a result of the documented differences in cross-informant ratings (e.g., Achenbach et al., 1987; Duhig et al., 2000). As a result, trait variance and trait-trait covariance may have been overestimated, causing a distortion of estimates of convergent and discriminant validity. Although the correlated uniqueness model has limitations, it has promise for encouraging researchers to refine and evaluate the understanding of these complex constructs (Bray et al., 1995). The findings of this study also should be viewed within the context of its limitations. Although a large sample of adolescent-mother-father triads were used for this study, these families volunteered actively to participate in this study. All families contacted the researchers in response to advertisements displayed in the community and then attended a research session in a university setting. As a result, these families may have been particularly motivated or willing to share information about themselves. Furthermore, the characteristics of this sample of families may reduce the generalizability of the results of this study to other types of families. For example, the majority of these families reported that they were Caucasian and consisted of an intact marital dyad. In addition, many of these mothers and fathers were experiencing varying degrees of psychological symptoms, perhaps making them different from other mothers and fathers in the community. Finally, mothers and fathers were allowed to select which of their adolescents, if there were more than one in the family, would participate. Consequently, the results described in this study may not be representative of those of families with other characteristics. In summary, this study highlighted the importance of using multiple informants to evaluate interparental conflict, adolescent behavioral problems, and adolescent competence. Several statistical methods were used to examine the construct validity of the ratings of these variables. Although the convergent validity of the ratings of these constructs appeared adequate when relationships were examined using Campbell and Fiske s (1959) criteria, discriminant validity was not supported strongly. The use of confirmatory factor analysis was not helpful in the elucidation of the construct validity of the constructs examined as a result of problems that occur commonly with the use of this procedure. Finally, analyses demonstrated adequate fit of the crossinformant ratings in a correlated uniqueness model in this sample of families, supporting convergent and discriminant validity. Overall, results suggested that measurement issues and informant effects should be examined in the ratings of interparental conflict, adolescent behavioral problems, and adolescent competence.

494 EDUCATIONAL AND PSYCHOLOGICAL MEASUREMENT References Achenbach, T. M. (1991a). Manual for the Child Behavior Checklist/4-18 and 1991 profile. Burlington: University of Vermont Department of Psychiatry. Achenbach, T. M. (1991b). Manual for the Youth Self-Report and 1991 profile. Burlington: University of Vermont Department of Psychiatry. Achenbach, T. M., McConaughy, S. T., & Howell, C. T. (1987). Child/adolescent behavioral and emotional problems. Implications of cross-informant correlations for situational specificity. Psychological Bulletin, 101, 213-232. Bagozzi, R. P. (1993). Assessing construct validity in personality research: Applications to measures of self-esteem. Journal of Research in Personality, 27, 49-87. Bray, J. H., Maxwell, S. E., & Cole, D. (1995). Multivariate statistics for family psychology research. Journal of Family Psychology, 9, 144-160. Bryant, F. B. (2000). Assessing the validity of measurement. In L. G. Grimm & P. R. Yarnold (Eds.), Reading and understanding more multivariate statistics (pp. 99-146). Washington, DC: American Psychological Association. Byrne, B. M., & Goffin, R. D. (1993). Modeling MTMM data from additive and multiplicative covariance structures: An audit of construct validity concordance. Multivariate Behavioral Research, 29, 67-96. Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56, 81-105. Cook, W. L., & Goldstein, M. J. (1993). Multiple perspectives on family relationships: A latent variables model. Child Development, 64, 1377-1388. Cummings, E. M., & Davies, P. (1994). Children and marital conflict: The impact of family dispute and resolution. New York: Guilford. Duhig, A. M., Renk, K., Epstein, M. K., & Phares, V. (2000). Interparental agreement on internalizing, externalizing, and total behavior problems: A meta-analysis. Clinical Psychology: Science and Practice, 7, 435-453. Emery, R. E., & O Leary, K. D. (1982). Children s perceptions of marital discord and behavior problems of boys and girls. Journal of Abnormal Child Psychology, 10, 11-24. First, M. B., Gibbon, M., Williams, J. B. W., & Spitzer, R. L. (1995). Instruction manual for the Mini-Structured Clinical Interview for DSM-IV (SCID) Version 2. Washington, DC: American Psychiatric Press. Fiske, D. W., & Campbell, D. T. (1992). Citations do not solve problems. Psychological Bulletin, 112, 393-395. Grych, J. H., & Fincham, F. D. (1990). Marital conflict and children s adjustment: A cognitivecontextual framework. Psychological Bulletin, 108, 267-290. Grych, J. H., Seid, M., & Fincham, F. D. (1992). Assessing marital conflict from the child s perspective: The Children s Perception of Interparental Conflict Scale. Child Development, 63, 558-572. Jöreskog, K. G. (1974). Analyzing psychological data by structural analysis of covariance matrices. In R. C. Atkinson, D.H. Krantz, R. D. Luce, & P. Suppes (Eds.), Contemporary developments in mathematical psychology (Volume 2, pp. 1-56). New York: Freeman. Kenny, D. A. (1991). A general model of consensus and accuracy in interpersonal perception. Psychological Review, 98, 155-163. Kenny, D. A., & Berman, J. A. (1980). Statistical approaches to the correction of correlation bias. Psychological Bulletin, 88, 288-295. Kenny, D. A., & Kashy, D. A. (1992). Analysis of the multitrait-multimethod matrix by confirmatory factor analysis. Psychological Bulletin, 112, 165-172. Kenrick, D. T., & Stringfield, D. O. (1980). Personality traits and the eye of the beholder: Crossing some traditional philosophical boundaries in the search for consistency in all of the people. Psychological Review, 87, 88-104.