Multilevel latent class models with dirichlet mixing distribution.
ABSTRACT: Latent class analysis (LCA) and latent class regression (LCR) are widely used for modeling multivariate categorical outcomes in social science and biomedical studies. Standard analyses assume data of different respondents to be mutually independent, excluding application of the methods to familial and other designs in which participants are clustered. In this article, we consider multilevel latent class models, in which subpopulation mixing probabilities are treated as random effects that vary among clusters according to a common Dirichlet distribution. We apply the expectation-maximization (EM) algorithm for model fitting by maximum likelihood (ML). This approach works well, but is computationally intensive when either the number of classes or the cluster size is large. We propose a maximum pairwise likelihood (MPL) approach via a modified EM algorithm for this case. We also show that a simple latent class analysis, combined with robust standard errors, provides another consistent, robust, but less-efficient inferential procedure. Simulation studies suggest that the three methods work well in finite samples, and that the MPL estimates often enjoy comparable precision as the ML estimates. We apply our methods to the analysis of comorbid symptoms in the obsessive compulsive disorder study. Our models' random effects structure has more straightforward interpretation than those of competing methods, thus should usefully augment tools available for LCA of multilevel data.
Project description:BACKGROUND:Multiple drug and sexual risk behaviors among people who inject drugs (PWID) in intimate relationships increase the risk of HIV and HCV transmission. Using data on PWID in intimate partnerships in Almaty, Kazakhstan, this study performed latent class analysis (LCA) on drug and sexual risk behaviors and estimated associations between dyadic relationship factors and membership in latent classes. METHODS:LCA was performed on a sample of 510 PWID (181-females/FWID, 321-males/MWID) to identify levels of drug and sexual risk behaviors. Generalized structural equation modeling with multinomial regressions estimated associations between relationship factors (length risk reduction communication, risk reduction self-efficacy) and class membership after adjusting for substance use severity, overdose, depression, binge drinking, intimate partner violence, structural factors, and sociodemographic characteristics. Models were sex-stratified to include FWID and PWID. RESULTS:A 3-class model best fit the data and consisted of low, medium, and high-risk classes. GSEM found that greater injection self-efficacy was associated with a lower likelihood of membership in the high-risk class for PWID and FWID. For MWID, greater length of the relationship was associated with a lower likelihood of membership in the medium-risk class. Greater relationship communication was associated with increased risk of membership in the high-risk latent class for MWID. CONCLUSIONS:Future research must investigate if increasing risk reduction and safe sex self-efficacy could reduce drug and sexual risk behaviors and HIV transmission among PWID and their intimate partners. Interventions are needed that reduce power inequities within relationships as a method of increasing self-efficacy, particularly among women.
Project description:BACKGROUND:Disrupting women's employment is a strategy that abusive partners could use to prevent women from maintaining economic independence and stability. Yet, few studies have investigated disruptions in employment among victims of intimate partner violence (IPV) in low-income and middle-income countries. Moreover, even fewer have sought to identify which female victims of IPV are most vulnerable to such disruptions. METHODS:Using baseline data from 947 women in Mexico City enrolled in a randomised controlled trial, multilevel latent class analysis (LCA) was used to classify women based on their reported IPV experiences. Furthermore, multilevel logistic regression analyses were performed on a subsample of women reporting current work (n=572) to investigate associations between LCA membership and IPV-related employment disruptions. RESULTS:Overall, 40.6% of women who were working at the time of the survey reported some form of work-related disruption due to IPV. LCA identified four distinct classes of IPV experiences: Low Physical and Sexual Violence (39.1%); High Sexual and Low Physical Violence class (9.6%); High Physical and Low Sexual Violence and Injuries (36.5%); High Physical and Sexual Violence and Injuries (14.8%). Compared with women in the Low Physical and Sexual Violence class, women in the High Physical and Sexual Violence and Injuries class and women in the High Physical and Low Sexual Violence and Injuries class were at greater risk of work disruption (adjusted relative risk (ARR) 2.44, 95% CI 1.80 to 3.29; ARR 2.05, 95% CI 1.56 to 2.70, respectively). No other statistically significant associations emerged. CONCLUSION:IPV, and specific patterns of IPV experiences, must be considered both in work settings and, more broadly, by economic development programmes. TRIAL REGISTRATION NUMBER:NCT01661504.
Project description:We apply latent class analysis (LCA) to quantify multidimensional patterns of weight-loss strategies in a sample of 197 women, and explore the degree to which dietary restraint, disinhibition, and other individual characteristics predict membership in latent classes of weight-loss strategies. Latent class models were fit to a set of 14 healthy and unhealthy weight-loss strategies. BMI, weight concern, body satisfaction, depression, dietary disinhibition and restraint, and the interaction of disinhibition and restraint were included as predictors of latent class membership. All analyses were conducted with PROC LCA, a recently developed SAS procedure available for download. Results revealed four subgroups of women based on their history of weight-loss strategies: No Weight Loss Strategy (10.0%), Dietary Guidelines (26.5%), Guidelines+Macronutrients (39.4%), and Guidelines+Macronutrients+Restrictive (24.2%). BMI, weight concerns, the desire to be thinner, disinhibition, and dietary restraint were all significantly related to weight-control strategy latent class. Among women with low dietary restraint, disinhibition increases the odds of engaging in any set of weight-loss strategies vs. none, whereas among medium- and high-restraint women disinhibition increases the odds of use of unhealthy vs. healthy strategies. LCA was an effective tool for organizing multiple weight-loss strategies in order to identify subgroups of individuals who have engaged in particular sets of strategies over time. This person-centered approach provides a measure weight-control status, where the different statuses are characterized by particular combinations of healthy and unhealthy weight-loss strategies.
Project description:The increasing availability of software with which to estimate multivariate multilevel models (also called multilevel structural equation models) makes it easier than ever before to leverage these powerful techniques to answer research questions at multiple levels of analysis simultaneously. However, interpretation can be tricky given that different choices for centering model predictors can lead to different versions of what appear to be the same parameters; this is especially the case when the predictors are latent variables created through model-estimated variance components. A further complication is a recent change to Mplus (Version 8.1), a popular software program for estimating multivariate multilevel models, in which the selection of Bayesian estimation instead of maximum likelihood results in different lower-level predictors when random slopes are requested. This article provides a detailed explication of how the parameters of multilevel models differ as a function of the analyst's decisions regarding centering and the form of lower-level predictors (i.e., observed or latent), the method of estimation, and the variant of program syntax used. After explaining how different methods of centering lower-level observed predictor variables result in different higher-level effects within univariate multilevel models, this article uses simulated data to demonstrate how these same concepts apply in specifying multivariate multilevel models with latent lower-level predictor variables. Complete data, input, and output files for all of the example models have been made available online to further aid readers in accurately translating these central tenets of multivariate multilevel modeling into practice.
Project description:Injury is a leading cause of morbidity and mortality in the paediatric population and exhibits complex injury patterns. This study aimed to identify homogeneous groups of paediatric major trauma patients based on their profile of injury for use in mortality and functional outcomes risk-adjusted models. Data were extracted from the population-based Victorian State Trauma Registry for patients aged 0-15 years, injured 2006-2016. Four Latent Class Analysis (LCA) models with/without covariates of age/sex tested up to six possible latent classes. Five risk-adjusted models of in-hospital mortality and 6-month functional outcomes incorporated a combination of Injury Severity Score (ISS), New ISS (NISS), and LCA classes. LCA models replicated the best log-likelihood and entropy > 0.8 for all models (N = 1281). Four latent injury classes were identified: isolated head; isolated abdominal organ; multi-trauma injuries, and other injuries. The best models, in terms of goodness of fit statistics and model diagnostics, included the LCA classes and NISS. The identification of isolated head, isolated abdominal, multi-trauma and other injuries as key latent paediatric injury classes highlights areas for emphasis in planning prevention initiatives and paediatric trauma system development. Future risk-adjusted paediatric injury models that include these injury classes with the NISS when evaluating mortality and functional outcomes is recommended.
Project description:Prevention scientists use latent class analysis (LCA) with increasing frequency to characterize complex behavior patterns and profiles of risk. Often, the most important research questions in these studies involve establishing characteristics that predict membership in the latent classes, thus describing the composition of the subgroups and suggesting possible points of intervention. More recently, prevention scientists have begun to adopt modern methods for drawing causal inference from observational data because of the bias that can be introduced by confounders. This same issue of confounding exists in any analysis of observational data, including prediction of latent class membership. This study demonstrates a straightforward approach to causal inference in LCA that builds on propensity score methods. We demonstrate this approach by examining the causal effect of early sex on subsequent delinquency latent classes using data from 1,890 adolescents in 11th and 12th grade from wave I of the National Longitudinal Study of Adolescent Health. Prior to the statistical adjustment for potential confounders, early sex was significantly associated with delinquency latent class membership for both genders (p?=?0.02). However, the propensity score adjusted analysis indicated no evidence for a causal effect of early sex on delinquency class membership (p?=?0.76) for either gender. Sample R and SAS code is included in an Appendix in the ESM so that prevention scientists may adopt this approach to causal inference in LCA in their own work.
Project description:BACKGROUND: The norovirus group (NVG) of caliciviruses are the etiological agents of most institutional outbreaks of gastroenteritis in North America and Europe. Identification of NVG is complicated by the non-culturable nature of this virus, and the absence of a diagnostic gold standard makes traditional evaluation of test characteristics problematic. METHODS: We evaluated 189 specimens derived from 440 acute gastroenteritis outbreaks investigated in Ontario in 2006-07. Parallel testing for NVG was performed with real-time reverse-transcriptase polymerase chain reaction (RT2-PCR), enzyme immunoassay (EIA) and electron microscopy (EM). Test characteristics (sensitivity and specificity) were estimated using latent class models and composite reference standard methods. The practical implications of test characteristics were evaluated using binomial probability models. RESULTS: Latent class modelling estimated sensitivities of RT2-PCR, EIA, and EM as 100%, 86%, and 17% respectively; specificities were 84%, 92%, and 100%; estimates obtained using a composite reference standard were similar. If all specimens contained norovirus, RT2-PCR or EIA would be associated with > 99.9% likelihood of at least one test being positive after three specimens tested. Testing of more than 5 true negative specimens with RT2-PCR would be associated with a greater than 50% likelihood of a false positive test. CONCLUSION: Our findings support the characterization of EM as lacking sensitivity for NVG outbreaks. The high sensitivity of RT2-PCR and EIA permit identification of NVG outbreaks with testing of limited numbers of clinical specimens. Given risks of false positive test results, it is reasonable to limit the number of specimens tested when RT2-PCR or EIA are available.
Project description:Although oppositional defiant disorder (ODD) is usually considered the mildest of the disruptive behavior disorders, it is a key factor in predicting young adult anxiety and depression and is distinguishable from normal childhood behavior. In an effort to understand possible subsets of oppositional defiant behavior (ODB) that may differentially predict outcome, we used latent class analysis of mother report on the Conners' Parent Rating Scales Revised Short Forms (CPRS-R:S).Data were obtained from mother report for Dutch twins (7 years old, n = 7,597; 10 years old, n = 6,548; and 12 years old, n = 5,717) from the Netherlands Twin Registry. Samples partially overlapped at ages 7 and 10 years (19% overlapping) and at ages 10 and 12 years (30% overlapping), but not at ages 7 and 12 years. Oppositional defiant behavior was measured using the six-item Oppositional subscale of the CPRS-R:S. Multilevel LCA with robust standard error estimates was performed using the Latent Gold program to control for twin-twin dependence in the data. Class assignment across ages was determined and an estimate of heritability for each class was calculated. Comparisons with maternal report Child Behavior Checklist (CBCL) scores were examined using linear mixed models at each age, corrected for multiple comparisons.The LCA identified an optimal solution of four classes across age groups. Class 1 was associated with no or low symptom endorsement (69-75% of the children); class 2 was characterized by defiance (11-12%); class 3 was characterized by irritability (9-11%); and class 4 was associated with elevated scores on all symptoms (5-8%). Odds ratios for twins being in the same class at each successive age point were higher within classes across ages than between classes. Heritability within the two "intermediate" classes was nearly as high as for the class with all symptoms, except for boys at age 12. Children in the Irritable class were more likely to have mood symptoms on the CBCL scales than children in the Defiant class but demonstrated similar scores on aggression and externalizing scales. Children in the All Symptoms class were higher in both internalizing and externalizing scales and subscales.The LCA indicates four distinct latent classes of oppositional defiant behavior, in which the distinguishing feature between the two intermediate classes (classes 2 and 3) is the level of irritability and defiance. Implications for the longitudinal course of these symptoms, association with other disorders, and genetics are discussed.
Project description:This paper illustrates two psychometric methods, latent class analysis (LCA) and taxometric analysis (TA) using empirical data from research probing children's mental representation in science learning. LCA is used to obtain a typology based on observed variables and to further investigate how the encountered classes might be related to external variables, where the effectiveness of classification process and the unbiased estimations of parameters become the main concern. In the step-wise LCA, the class membership is assigned and subsequently its relationship with covariates is established. This leading-edge modeling approach suffers from severe downward-biased estimations. The illustration of LCA is focused on alternative bias correction approaches and demonstrates the effect of modal and proportional class-membership assignment along with BCH and ML correction procedures. The illustration of LCA is presented with three covariates, which are psychometric variables operationalizing formal reasoning, divergent thinking and field dependence-independence, respectively. Moreover, taxometric analysis, a method designed to detect the type of the latent structural model, categorical or dimensional, is introduced, along with the relevant basic concepts and tools. TA was applied complementarily in the same data sets to answer the fundamental hypothesis about children's naïve knowledge on the matters under study and it comprises an additional asset in building theory which is fundamental for educational practices. Taxometric analysis provided results that were ambiguous as far as the type of the latent structure. This finding initiates further discussion and sets a problematization within this framework rethinking fundamental assumptions and epistemological issues.
Project description:Latent class models have become a popular means of summarizing survey questionnaires and other large sets of categorical variables. Often these classes are of primary interest to better understand complex patterns in data. Increasingly, these latent classes are reified into predictors of other outcomes of interests, treating the most likely class as the true class to which an individual belongs even though there is uncertainty in class membership. This uncertainty can be viewed as a form of measurement error in predictors, leading to bias in the estimates of the regression parameters associated with the latent classes. Despite this fact, there is very limited literature treating latent class predictors as measurement error models. Most applications ignore this issue and fit a two-stage model that treats the modal class prediction as truth. Here, we develop two approaches-one likelihood-based, the other Bayesian-to implement a joint model for latent class analysis and outcome prediction. We apply these methods to an analysis of how acculturation behaviors predict depression in South Asian immigrants to the United States. A simulation study gives guidance for when a two-stage model can be safely implemented and when the joint model may be required.