A Novel Approach for Sensitivity Analyses and Validation for Mixture Models

• Activity Number: 281 - New Methods with Applications in Mental Health Statistics
• Type: Contributed
• Date/Time: Tuesday, August 9, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Douglas Gunzler* and Alessandro De Nadai and Adam Perzynski and Jarrod Dalton and Kristen Berg and Farren Briggs
• Companies: Case Western Reserve University and Texas State University and Case Western Reserve University and The Cleveland Clinic Foundation and Case Western Reserve University and Case Western Reserve University
• Keywords: mixture modeling; latent variable; latent class growth analysis; patient-reported outcomes; sensitivity analyses; model validation

Abstract

There is currently no widely used approach to assess the appropriateness of a mixture model specification and to appreciate the strength of the conclusions being drawn from such a model. Mixture modeling is a multivariate clustering technique using latent variables. Our team has developed a method for application in the analyses of a mixture model for two main objectives: (1) conducting sensitivity analyses to evaluate the importance of specific observed indicators or time periods (2) validating/reproducing the model using different observed indicators or time periods. Most likely cluster membership can be summarized using a nominal categorical variable. The method calculates concordance after permuting a contingency table to maximize the sum of the diagonal for most likely cluster membership cross-tabulated across two competing models. The method can be broadly applied even for rectangular matrices. We illustrate two different applications of the method using latent class growth analysis for validation of patient-reported outcomes for mobility impairment using large observational databases of people living with multiple sclerosis (N = 8,687) and Parkinson’s disease (N = 16,863).