scmBayesPost v0.3.0

scmBayesPost v0.3.0 — Major performance release

The problem

On datasets with large numbers of treated units (J0 ~ 1879 in the
motivating application), the Gibbs samplers for the Bayesian probit
first stage (selection_probit_bayes) were taking approximately
21 seconds per iteration — around 6 hours for a 2,000-iteration
run. Profiling identified three dominant costs:

1. Cholesky factorisation of a [K*J0 x K*J0] matrix (e.g. 1879 x 1879) at every iteration, despite this matrix being block-diagonal
2. Sparse matrix-vector product X_block %*% beta_bd over a
   [N_stacked x K*J0] matrix at every iteration
3. Per-iteration data.table merge to align nu_hat with X_block
   row ordering

What changed

The samplers now exploit the block-diagonal structure of X_block
throughout. Key changes:

• The [K*J0 x K*J0] Cholesky is replaced by J0 independent
  [K x K] factorisations — for K = 1 (the common single-treatment
  case) this collapses to scalar arithmetic with no matrix operations
• X_block %*% beta_bd is replaced by a unit-by-unit loop over
  precomputed row ranges and cached column slices
• The data.table merge is replaced by a one-time integer index map
• Truncated normal draws use a fast inverse-CDF sampler with
  precomputed treated/untreated splits
• rho and sigma2 updates use direct vector operations rather than
  sparse matrix products

Expected speedup

The theoretical reduction in the Cholesky step alone is from
O((K*J0)^3) to O(J0 * K^3) — for K=1, J0=1879 this is a
factor of 1879^2 ≈ 3.5 million. In practice the overall per-iteration
speedup depends on which operations remain dominant after the fix
(primarily X_fs %*% delta, a 292k × 162 dense matrix-vector product).
Multi-threaded BLAS is recommended for further acceleration:

RhpcBLASctl::blas_set_num_threads(parallel::detectCores(logical = FALSE))

API unchanged

No changes to function signatures, arguments, or output. Existing code
runs without modification.

Installation

remotes::install_github("andrebonfrer/scmBayesPost@v0.3.0")