Introducing Psychometric Approaches to Analyze Neuropsychological Measures: Mixture Modeling Association for Psychological Science Annual Meeting, New York City, Friday, May 26, 2006 Gregory Anderson, Ph.D., XTRIA Futoshi Yumoto, MA, ABT Daisy Wise, MA, University of Maryland
Introduction: Neuropsychological Tests • Neuropsychological tests are generally brief measures that are used to identify dysfunction/function in some brain region. • Research has identified that problems with frontal lobe functioning, especially anterior cingulate and orbital frontal functioning, is associated with ADHD. • Therefore, a number of neuropsychological measures were assembled to assess individuals with ADHD and to discriminate them from typically developing students and students with the common comorbidity of Learning Disabilities (LD). • In the past, neuropsychological measures have been found to be fairly poor at discriminating students with ADHD from typically developing students
Introduction: Mixture Modeling • It was hypothesized that identifying latent traits and examining how they predicted group membership might improve diagnostics. • There are special challenges in analyzing neuropsychological measures due to their unique types of scores, including: positive and negative scores, error counts, time, and correct responses. • Combining Confirmatory Factor analysis with a Latent Class model provided us a way to identify different patterns of responses on the neuropsychological tests given by different subject types. • This Mixture Modeling approach provides us a way to assess different types of clinical subjects on the basis of their neuropsychological profiles.
Methods: Measures • The neuropsychological measures were assessments utilized in the field trials of the Attention Test Linking Assessment and Services (ATLAS), a comprehensive test for ADHD by Anderson & Post. • The neuropsychological measures used in this study were: Trails A, Trails B, Divided Attention between Trails A & Cancellation, Verbal Superspan Memory (1st & 3rd time), Digits Forward, Digits Reverse, Serial Subtraction and Error Counts on several of these measures • Criterion measures were parent reported diagnoses of ADHD and LD.
Methods: Subjects • Subjects were 220 individuals assessed as a part of the field trials of the ATLAS. • The sample contained three times as many males as females. This was due the greater frequency of males with LD & ADHD. • The sample included subjects from 8 to 18 years of age from across the nation and gathered by field researchers.
Methods: Sequence of Analysis • A theoretical structure used in the confirmatory factor analysis was hypothesized by Anderson and Yumoto. • Confirmatory Factor Analysis • Add covariate (MIMIC) • Mixture Analysis • Analyze latent class: Using 2 to 5 class models • Add criterion measure • Examine relationships between latent classes and manifest groupings (possible diagnoses) • MPlus (Muthen & Muthen, 2004) was used to conduct the analysis.
Results: Factor Structure • A confirmatory factor analysis was conducted first which established our three factor model. It had a very adequate fit with the data. The factors are listed below: • Factor 1: Visual Sequencing and Tracking - Trails A, Trails B, Divided Attention Trails A & Divided Attention Cancellation • Factor 2: Memory – Word List Memory Time 1, Word List Memory Time 3, Digit Memory Forward, Digit Memory Backward, Serial Subtraction • Factor 3: Impulsive Errors - Cancellation Commission Error, Cancel Omission Errors, Trails A Errors, Trails B Errors
Results: Identify Latent Classes • Using the Factor Structure above, we incorporated the factors into latent class models using two, three, four and five latent classes. • Models were selected on the basis of fit statistic indices: likelihood, Akaike Information Criteria (AIC), Bayesian Information Criteria (BIC) & Likelihood Ratio Test.
Results: Fit Statistics (Model Selection) Models Fit Stat. 2 3 4 5 LH -4649 -4613 -4597 -4586 AIC 9390 9334 9318 9312 BIC 9546 9516 9528 9549 P 46 54 62 70 Based on the fit statistics and interpretability we selected the 4 class model.
Result: Selected Model Statistics Group Proportion 1 2 3 4 1 .55 .878 .001 .022 .099 2 .08 .004 .850 .105 .041 3 .11 .052 .058 .827 .062 4 .26 .163 .012 .015 .810 Probability of a subject being assigned to a group, given their true group membership.
Results: Model Parameters Factor Means (Intercepts) Factor Group1 Group2 Group3 Group4 1 -3.828 3.606 0.093 0.000 2 0.490 0.418 -0.876 0.000 3 1.427 -1.796 -0.310 0.000 * For factor 1 negative numbers are better. For factors 2 and 3 positive numbers are better.
Results: Model Parameter-Effect of Covariate Covariate (Grade)/Covariate Factor Typical LD ADHD LD/ADHD 1 -.036 .000 -.088 -.139 2 .174 .034 .356 .283 3 -.314 -.255 -.601 -.591
Results: Classification Table In the LDADHD column 0=Typical sample, 1=LD, 2=ADHD, 3=LD & ADHD. Several of the students in the typical group had previously been assessed for ADHD but had not met the required criteria. The most severe ADHD students are in the ADHD class and not LD & ADHD. Latent Class Group 2 was the most severe/pathological group. This agreed with the univariate studies where some individuals with just ADHD were found to be the most severe on the majority of these measures.
Discussion • LD & ADHD are notoriously hard to differentiate because: Both disorders are found on a continuum, both have several variants, LD is primarily a school categorical (notoriously variable) and ADHD is a psychiatric diagnosis, high comorbidity (up to 50%, Brown, 2005) almost to the level of being symptoms. • In many studies (i.e. Solanto, 2004), neuropsychological measures did not provide clear diagnostic identification and we used a limited set of measures, none of which were likely to be used to identify LD. • In this analysis, both diagnoses were reported by parents, though other criterion measures were gathered. • In spite of these difficulties, this approach provided a reasonably good prediction of group membership.
Implications • The current study provides evidence for the value of utilizing Latent Class Mixture Modeling approaches for the identification of disorders from neuropsychological assessment data. • Indeed, with this sample, the identification of ADHD could probably be made more accurately than from parent report of prior diagnosis. • With the addition of measures for LD we would expect a significant improvement in group differentiation and with other measures of ADHD (i.e. additional neuropsychological measures, observations, parent report, measures of ODD, etc.) we would expect that it would surpass traditional clinical interpretation.