Phenomenology of Clinic-Referred Children and Adolescents with Oppositional Defiant Disorder and Comorbid Anxiety

All MANCOVAs were performed in a structural equation modeling (SEM) statistical software program (Mplus Version 6.12) and are often referred to as generalized estimating equation (GEE) analyses. GEE analyses were performed instead of the more traditional MANCOVA for four reasons. First, GEE was used to handle missing data with maximum likelihood estimation via the inclusion of additional variables to form a ‘saturated correlates’ data analysis model that increases the likelihood that the implicit missing data assumption of missing at random (MAR) is met without biasing the parameter estimates of theoretical interest. Specifically, in our analyses, missing data for each MANOVA was handled by including all other response variables not being analyzed (e.g., for MANOVA #1, we included the response variables for MANOVAs 2-5) as missing data correlates in a ‘saturated correlates’ missing data handling model. Second, GEE was used to minimize Type-1 errors that could result from non-normally distributed data by using a maximum likelihood parameter estimation algorithm robust to non-normally distributed data (e.g., MLR). Third, GEE ensures that the traditional MANCOVA assumption of ‘homogeneity of covariate regression slopes’ was met through parameter estimate constraint procedures available in SEM software packages. Lastly, GEE allows for proceeding directly to the between-group tests of mean differences among correlated response variables without first conducting the MANCOVA omnibus F. For these reasons, the analyses performed in this paper are more accurately described as GEE analyses. The analysis logic employed is consistent with traditional MANCOVAs, but many of the expected statistics associated with MANCOVAs (i.e., an omnibus F statistic, such as Wilk’s Lambda) are neither needed nor computed.