Estimating the Relative Treatment Effects of Natural Clusters of Adolescent Substance Abuse Treatment Services: Combining Latent Class Analysis and Propensity Score Methods
Embargo until
Date
2013-12-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
Objectives: The motivating substantive aim of this dissertation was to identify common clusters of drug treatment services that adolescents receive in practice that are effective in terms of improving substance use outcomes. We first identified clusters of drug treatment services that adolescents in outpatient treatment report receiving, as well as examined factors associated with each class of treatment services (Chapter 2). Our statistical approach for estimating the effect of treatment service classes on outcomes was latent class regression with a distal outcome; we review various statistical methods for implementing latent class regression with a distal outcome in Chapter 3. Addressing potential confounding arising from baseline differences among youth receiving different classes of treatment services was a key concern; Chapter 4 describes emerging methods to address confounding in the context of latent class regression with a distal outcome, highlighting the challenges that arise when the treatment of interest is a latent variable.
Methods: Chapters 2 and 4 used data on 5,527 adolescents receiving drug treatment services through treatment providers funded through the Substance Abuse and Mental Health Services Administration’s Center for Substance Abuse Treatment. Latent class analysis was used to identify classes of substance use treatment services reported by youth. A simulation study to compare 5 statistical methods for latent class regression with a distal outcome was performed in Chapter 3. An additional simulation study to compare 3 methods for addressing confounding in this context was performed in Chapter 4; these methods were also applied to our adolescent data.
Results: Distinct classes of outpatient treatment services received by adolescents were empirically identified using latent class analysis; youth receiving different classes of treatment services were found to be significantly different on numerous baseline characteristics. Statistical performance varied notably across methods for latent class regression with a distal outcome. Finally, failing to account for potential confounding in this setting can lead to significantly biased estimates of the association between the latent class and the distal outcome; the 1-step method we examined performed particularly well in terms of reducing bias.
Conclusions: Emerging methods for modeling the treatment of interest as a latent variable are quite relevant for social and behavior researchers. However, like studies with fully observed variables, care must be taken to address potential confounding; future work should continue to develop methods to address confounding in this context.
Description
Keywords
latent class analysis, causal inference