NEWS

MRCCC 0.99.4

V is now a required argument of mr_ccc().
sign_reversal() uses a scale-free cutoff for near-zero interaction draws: a draw is dropped when its standardised interaction coefficient is at most 1e-8 in absolute value.
Additional input-validation tests.

mr_ccc() now centres every column of the instrument matrices G, H and the covariate matrix V, as it already did for X, Z and Y. Neither first stage of the model has an intercept, so uncentred designs (for example genotype dosages coded 0/1/2) forced each first-stage fit through the origin and shifted the projections X* and Z* and the fixed prior scale. The documentation of mr_ccc_gibbs(), which does not centre its inputs, now states that all six inputs must be centred.
mr_ccc() stops with an explicit message when cbind(G, V) or cbind(H, V) is rank deficient after centring.
mr_ccc() warns when any Gelman-Rubin R-hat exceeds 1.01, and issues a message when only one chain is run or when the Monte Carlo standard error of the PIP cannot be estimated.
mr_ccc() and simulate_mrccc() restore the caller's random number state after running under a supplied seed. The fitted object records the g-prior scale used in settings$g.
mr_ccc_gibbs() validates its arguments, reports numerically singular matrices with an informative error, and can be interrupted from R.

The priors on the second-stage coefficients (beta_X, beta_XZ) and beta_Z are now scaled by a fixed diagonal computed once from the least-squares first stage, instead of by the current X*, Z* at every sweep. Because X* and Z* are functions of the first-stage parameters, the prior used in 0.99.0 entered their full conditionals, which the sampler did not account for. With the scale fixed the first-stage updates are exact as written and every update remains conjugate. The scale is returned in settings$prior_scale.

Initial Bioconductor submission.
mr_ccc() fits the MR-CCC model to one ligand-receptor-pathway triplet with one or more Gibbs chains, returning the posterior inclusion probability with its Monte Carlo standard error, effect estimates on the raw and standardised scales, pooled posterior draws, the joint log-likelihood of every retained draw, and the Gelman-Rubin statistic.
mr_ccc() also reports first-stage instrument strength: the partial F of each instrument block after the covariates, with a message when either block is weak by the conventional two-stage least squares threshold.
credible_intervals(), sign_probabilities(), sign_reversal() and plot_effect_curve() summarise the posterior draws.
bayes_fdr() and bfdr_threshold_set() provide Bayesian false discovery rate control across many triplets.
simulate_mrccc() generates data from the working model with an unmeasured confounder.
mr_ccc_gibbs() exposes the compiled RcppArmadillo sampler.