Review Papers Claim by Claim

Visual summary — co‑morbidity loadings across the three-class model

Quick facts (from the paper)

• Sample analysed: 706 OCD-affected individuals (624 definite, 82 probable) from 238 families across 6 US sites (familial, early-onset probands) — see original data OCGS sample summary.
• Primary analytic approach: multilevel latent class analysis (Dirichlet random effects) on eight co‑morbidity indicators; BIC favored 2 classes but authors present a clinically informative 3‑class solution and report intra‑cluster correlations ~0.44 (3-class) as a discrete-class heritability analog MLCA methods and heritability.

Detailed critique and interpretation (visual first, then evidence)

1) What the data show — evidence grounded in the paper

1. Three-class solution (authors' interpretation):
   • Class 1 — 'Simplex' (Avg prevalence ≈0.34): low co‑morbidity; MDD most frequent additional disorder (loading ~0.27) Class1 characteristics.
   • Class 2 — 'Tic‑related' (Avg ≈0.32): high probability of GAD (0.56), Tics (0.41), GrD (0.48); predominantly male; higher conscientiousness and neuroticism in some models Class2 characteristics.
   • Class 3 — 'Affective‑related' (Avg ≈0.30): very high affective loadings: MDD 0.68, GAD 0.67, PD 0.48, Som 0.53, GrD 0.59; predominantly female, younger onset, higher OCPD features, taboo symptoms and higher neuroticism / lower conscientiousness Class3 characteristics.
2. Estimated intra‑cluster correlation (ICC; discrete heritability analogue): 0.44 (three‑class), suggesting familial aggregation of class membership but not proving simple monogenic inheritance Class heritability estimates.

2) Strengths — why this is useful

• Large, multisite familial dataset with careful structured diagnostic assessment (SCID, SADS-LA-R adaptation, SIDP for OCPD) and inter-rater reliabilities reported; this supports internal data quality Assessment and reliability.
• Use of MLCA that models family clustering explicitly (Dirichlet random effects) — appropriate for familial samples and gives a measure analogous to heritability for class membership MLCA with Dirichlet effects.
• Clinically coherent classes that map onto existing clinical constructs (tic-related, affective) and correlate with meaningful covariates (sex, age of onset, personality, symptom factors) — aids hypothesis generation for genetics and treatment stratification Clinical correlations of classes.

3) Key limitations & biases (critical)

1. Ascertainment bias — sample is highly familial and enriched for early-onset probands (probands with onset <18 y) and families with ≥2 affected members; this reduces external generalizability to community or non-familial OCD samples. Findings may reflect familial clustering patterns rather than population structure Sample ascertainment limitation.
2. Indicator selection and grouping choices: only eight Axis I indicators used; some domains were combined (e.g., trichotillomania + nail biting into GrD; BDD + hypochondriasis into Som). Those a priori choices can shape latent classes and may omit relevant discriminators (e.g., ADHD, eating disorders, substance use). Authors acknowledge this subjective selection and the possibility of different structures with different indicators Indicator selection caveat.
3. Model selection ambiguity: formal BIC favored the 2‑class model (BIC: 1031 two‑class vs 1062 three‑class), yet authors chose to present the clinically interpretable 3‑class solution and note BIC underestimates class number in modest samples — this is defensible but means the three‑class structure is not a statistically unequivocal choice and may be sample‑specific Model selection/BIC details.
4. Conditional independence assumption: LCA assumes that, conditional on class, indicators are independent. Violations (residual associations between certain disorders) can induce spurious classes; authors mention residual checks but complex residual dependencies are difficult to eliminate and may bias class enumeration and interpretation Conditional independence caution.
5. Missing data & measurement timing: diagnoses are lifetime and based on a single occasion assessment with some missingness (1.4–8.9% by disorder). No explicit multiple-imputation for covariate regressions; analyses assume missing at random (MAR). If missingness relates to unobserved variables (not MAR), estimates can be biased. The EM algorithm used in MLCA handles partial indicator missingness but downstream regression excluded some subjects when covariates missing, reducing power and possibly introducing selection bias Missing data handling.
6. Cross-sectional, retrospective diagnoses limit causal inference: class membership associations with age at onset or personality could stem from recall bias, cohort effects, or reverse causation (e.g., MDD secondary to long-standing OCD rather than etiologic). Authors interpret MDD in simplex cases as likely secondary but this cannot be proven here Cross-sectional/causality limit.

4) Reproducibility & next-step tests (what would falsify or support the classes)

What would strengthen or falsify this subclassification?

• Replicate MLCA on a large, unselected population-based OCD cohort (not enriched for familial early-onset cases) using the same and expanded indicator sets (including ADHD, eating disorders, substance use, and separate BDD/ hypochondriasis) — if the three classes reproduce with similar loadings and covariate associations, external validity is supported; lack of replication or different class structure would falsify the paper's subclassification claim Replication necessity.
• Prospective longitudinal cohorts that map temporal order of onset (OCD → MDD or MDD → OCD?) would test causality and whether class membership predicts later clinical course, treatment response, or familial transmission patterns.
• Genome-wide or candidate-gene analyses stratified by class membership: finding class-specific genetic signals (e.g., linkage in male tic‑class families vs affective-class families) would support etiologic heterogeneity; failure to find class-specific genetic architecture (after adequate power) would argue against strong etiologic separation.

5) Practical implications

• For research: using co‑morbidity-based subclasses could reduce heterogeneity in genetic and neurobiological studies, increasing power to detect class‑specific effects (authors present heritability analog to justify this). However, use only after replication in independent, representative samples.
• For clinical work: classes flag different clinical profiles (e.g., tic-related group — male, conscientious; affective group — female, earlier onset, OCPD, taboo thoughts) that clinicians can use as hypotheses for comorbidity screening and differential management planning, but not yet for treatment selection because no class-specific treatment data were presented.

6) Bottom-line evaluation

Nestadt et al. (2008) provide a careful MLCA of a large familial OCD cohort that yields clinically interpretable co‑morbidity classes. The work is hypothesis‑generating and valuable for psychiatric genetics. However, ascertainment, indicator choice, conditional-independence and model-selection uncertainty, and missing-data caveats mean the three‑class solution should not be taken as definitive until replicated in broader prospective samples and tested for predictive validity (course, treatment response, genetics).

Selected citations (primary source)

Nestadt et al. 2008 OCD subtyping