Review Papers Claim by Claim

Visualization-first summary (text + key numeric highlights)

• Data: UK Biobank and Generation Scotland relatives only (1st–3rd degree): 81,326 pairs (UKB) and 38,006 pairs (GS:SFHS); 130 continuous phenotypes analyzed.BIGFAM methods & cohorts
• Key outputs (paper-reported): Among 81 phenotypes with robust estimates: mean V_G = 0.353 (SD 0.145), mean V_S = 0.043 (SD 0.036); 29 phenotypes show significant V_X (mean V_X = 0.011, SD 0.005). BIGFAM-V_G correlations with genotype/summary methods r^2 ≈ 0.73–0.79; V_X correlates with GCTA V_X (r ≈ 0.66).BIGFAM main results
• Core conceptual innovation: Use of decay-rate differences across degrees of relatedness (genetic decay fixed at factor 2, shared-env decay w_S estimated) plus relationship-specific X correlations (t_l residuals) to separate V_G, V_S and estimate V_X using phenotype-only FR-reg coefficients.BIGFAM methodological principle

Strengths (what the paper does well)

1. Clear, reproducible pipeline: FR-reg generation, slope test, grid-search + cross-validation, and L2-penalized V_X — code and source data linked and archived (GitHub, Zenodo DOI linked in paper).BIGFAM code & data availability
2. Robust validation: extensive simulations under plausible shared-environment scenarios (nuclear-family-specific, maternal-effect, second-degree-dominant) demonstrating where the method succeeds and where it struggles (e.g., w_S ≈ 2 makes decomposition impossible) — authors transparently report failure modes and CI behavior.
3. Practical value: enables variance decomposition and chromosome-specific (X) heritability without genotypes — broadening use of large phenotyped cohorts lacking genotyping or where genotype sharing is restricted.
4. Concordance with genotype-based methods: authors quantify correlation with multiple orthogonal approaches (GCTA, LDSC, LDpred2, RDR, HE), increasing confidence that V_G is not purely an artifact of modelling choices.BIGFAM validation vs other methods

Limitations, blind spots, and potential biases

• Dependence on correct relationship inference: UKB familial-relationship calls are inferred using kinship, age and X-kinship heuristics; misclassification (e.g., mis-assigned PO vs sib, misreported ages, non-paternity) would bias FR-reg patterns and thus V_G/V_S decomposition. Authors note this explicitly.Relationship inference caveat
• No ancestry principal-component correction: BIGFAM uses phenotype residuals adjusted for age/sex but not genetic PCs; population stratification (ancestry differences correlated with familial structures) can inflate V_G or masquerade as V_S. Paper acknowledges lack of direct ancestry correction as a limitation.
• Model assumptions that can fail: key assumption: shared-environment decays with degree via a single parameter w_S (constant across degrees). Simulations show that heterogeneity of S across degrees or relationship-specific S (maternal effects) can bias estimates — especially when decay approximates genetic decay (w_S ≈ 2) where decomposition is statistically impossible.
• X-chromosome confounding: When relationship-specific shared environment correlates with X-patterns (e.g., maternal caregiving correlated with mother-daughter X sharing), V_X can be overestimated — authors simulated this and observed modest overestimation in maternal-effect scenarios.
• Binary traits and gene–environment interaction: BIGFAM currently targets continuous traits; liability-threshold/binary models and gene–environment interaction (including genetic nurture) are not handled and would require additional modelling (authors mention this limitation).
• Assortative mating: Not modelled explicitly; can inflate V_G from pedigree contrasts. Authors argue decay factor 2 remains for FR-reg but acknowledge effect on absolute V_G magnitude.

Technical critique and reproducibility checklist

1. FR-reg input generation: authors include bootstrap (randomly selecting one relationship per individual per resample) to reduce non-independence — good practice; code publicly available to reproduce this exact resampling.
2. Model identifiability: authors explicitly show (analytically and by simulation) that identifiability fails when w_S = 2; they therefore implement a slope test to detect identifiability — sound and transparent. Practical takeaway: only report decomposed V_G/V_S when slope test rejects w_S=2.
3. Parameter selection and cross-validation: grid-search ranges for w_S informed by slope test, with tenfold resampling repeated 10× — this cross-validation reduces overfitting risk. Authors provide resampling-derived CIs.
4. L2 penalty for V_X: authors tune α by simulation and choose α=2 to reduce CI widths while avoiding downward bias — appropriate, but users should sensitivity-check α in new cohorts.
5. Validation: multi-method validation (summary-based, genotype-based, pedigree-based) across two cohorts is a strong design choice; bootstrapped correlation CIs are reported.

Practical recommendations for users and next studies

• Run the slope test first — only proceed to VG/VS decomposition when slope CI excludes 1 (i.e., w_S ≠ 2).
• Sensitivity analyses: (a) vary α (L2 weight) for V_X estimation, (b) simulate relationship misclassification to bound biases, (c) repeat with ancestry-stratified subsamples or include PCs where possible (e.g., in datasets where PCs can be shared in privacy-compliant ways).
• When possible, validate with genotype-based subsets (if available) to confirm concordance; if unavailable, compare with published SNP heritabilities or LDSC estimates for plausibility checks.
• For maternal-effect suspicion, explicitly model relationship-specific S or remove maternal-only pairs to check robustness of V_X and V_S estimates.
• Do not apply to binary traits without extending to liability models and checking for bias from close-relative ascertainment — authors note this limitation and recommend future work.

Conclusion — balanced, evidence-weighted

BIGFAM is a strong methodological contribution: it supplies a transparent, reproducible, and well-validated way to estimate V_G, V_S and V_X from relatives-only phenotypes, with clear failure modes and validation against multiple orthogonal methods. It is especially useful when genotype data are missing or inaccessible. However, users must beware of relationship inference errors, uncorrected ancestry structure, and scenarios where shared environment decays like genetic relationships (identifiability problem). Many of these caveats are explicitly discussed in the paper and testable with the provided code and Source Data.BIGFAM limitations & data

Author reviews (one-click)