Beath (2017) randomLCA: An R Package for Latent Class with Random Effects Analysis

Summary

A Journal of Statistical Software paper introducing the R package randomLCA, which fits latent class models with random effects. The core problem it addresses is conditional dependence — heterogeneity of outcome probabilities within a true class violates the standard latent-class conditional-independence assumption and forces an uninterpretably large number of classes. The fix is to add a normally-distributed subject-level random effect, so outcomes are conditionally independent given both the class and the random effect; a further period-level random effect handles subjects observed repeatedly over time. Demonstrated on myocardial-infarction symptom classification, dentist diagnostic-agreement for dental caries, and longitudinal infant respiratory/allergy symptoms.

Key Claims

• Conditional dependence is the practical obstacle. Standard latent class assumes local (conditional) independence of indicators given class; when that fails, one either adds spurious classes (harming interpretation) or gets biased estimates — exactly the conditional-dependence problem of Vacek (1985) and the reason for the dependence-structure models in diagnostic accuracy.
• Random-effects latent class (the GRE model). Add a subject random effect $b_i\sim\mathcal N(0,1)$: $\Pr(Y_{ij}=1\mid d_i,b_i)=\Phi(\beta_{j,d_i}+\sigma_{d_i}b_i)$ — outcomes are conditionally independent given class $d_i$ and $b_i$. This is the Gaussian-random-effects (Qu-Tan-Kutner 1996) dependence structure, now available as software.
• Two-level (longitudinal) extension. For subjects measured over multiple periods, a second period-level random effect captures within-subject, across-time association on top of the subject effect — a two-level random-effects latent class model.
• Model-based classification and selection. Because it is a finite mixture, class number and structure are chosen by model-selection criteria (information criteria) rather than ad hoc symptom weighting; the package computes posterior class membership, marginal and conditional outcome probabilities, and (for diagnostic applications) sensitivity/specificity.
• Estimation. Marginal maximum likelihood integrating out the random effects by adaptive Gauss-Hermite quadrature, with the usual EM / direct-optimization machinery.

Concepts Introduced or Extended

• Latent Class Model
• Diagnostic Accuracy

Entities Mentioned

• Ken J. Beath

Quotes

"A major difficulty with the use of latent class models is the presence of heterogeneity of the outcome probabilities within the true classes, which violates the assumption of conditional independence... A solution is to include a normally distributed subject level random effect in the model so that the outcomes are now conditionally independent given both the class and random effect."

My Take

This is the applied, made-usable counterpart to the theory papers just ingested: Vacek (1985) shows conditional dependence biases no-gold-standard accuracy estimates optimistically, and Albert-Dodd (2004) shows the dependence-structure models are hard to tell apart with few tests — randomLCA is where you actually fit the Gaussian-random-effects version of that model, and adds the two-level (subject + period) structure that longitudinal diagnostic data need. The random-effect trick is the same one that turns a latent class model into a mixed model: absorb residual within-class association into a continuous latent $b_i$ rather than proliferating discrete classes, which keeps the class solution interpretable. For the wiki it grounds the GRE dependence structure with a concrete implementation and links the latent-class / diagnostic-accuracy thread to the random-effects machinery; as a software paper its value is reproducibility rather than new theory.