NEWS

hhdynamics 1.3.3 (2026-04-08)

Fixed .effective_sample_size() error on constant MCMC chains. When a parameter has zero variance (chain never moved), acf() returns NaN, causing if (pair_sum < 0) to fail with "missing value where TRUE/FALSE needed." Now returns ESS = 1 for constant chains and handles any remaining NaN autocorrelation values defensively. Fixes CRAN check ERROR on r-devel-linux-x86_64-fedora-gcc.

Limit RcppParallel threads to 2 during tests for CRAN compliance. Previously unconstrained parallelism caused a CPU/elapsed time ratio of 22x on Debian, triggering a NOTE.
Share MCMC fit fixtures across plot and table tests instead of fitting independently in each test. Reduces total MCMC fits from ~32 to ~12.
Reduce MCMC iterations in structural tests from 3000 to 500. These tests check output structure (column names, dimensions, class), not convergence.
Mark expensive recovery/comparison tests with skip_on_cran(): onset imputation parameter comparison (5000 iterations x 2 fits), NA covariate attack rate table, and SI summary format check.

Use configure script for platform-specific LAPACK/BLAS linking. On macOS, these libraries are provided by the R framework and do not need explicit linking; on Linux, $(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS) are required to resolve symbols like dpotrf_.

Replaced C-style variable length arrays (VLAs) with std::vector<double> in C++ MCMC code to comply with ISO C++ standard. Fixes CRAN check WARNING on Windows (GCC 14) and installation failure on Debian Linux.
Removed obsolete CXX_STD = CXX11 from Makevars/Makevars.win and SystemRequirements: C++11 from DESCRIPTION. C++11 support has been removed in R-devel.
Added LAPACK/BLAS linking flags to src/Makevars (previously only in Makevars.win), fixing undefined symbol: dpotrf_ installation error on Debian Linux.

Forest plot for covariate effects. plot_covariates(fit) produces a publication-ready forest plot grouped by Susceptibility and Infectiousness sections, with reference level labels, alternating row shading, and a dashed null line at RR = 1. Optional file argument saves a PDF with auto-sized dimensions based on the number of covariate rows. Custom variable and level labels can be supplied via the labels argument.
Summary tables. Three new exported functions return clean data frames: table_parameters(fit) (posterior mean/median/CrI with ESS and acceptance), table_covariates(fit) (covariate effects on the log and exponentiated scale), and table_attack_rates(fit, by) (secondary attack rates with Wilson CIs, optionally stratified by a covariate).

Fixed table_attack_rates() NA handling: rows with NA in the stratifying variable are now correctly excluded from strata and counts (previously x == NA returned NA rather than FALSE, poisoning subset sizes).
Fixed .effective_sample_size() lag indexing: Geyer (1992) pairing now correctly starts at lag 1 (index 2 of R's acf() output) rather than lag 0, which was inflating the integrated autocorrelation time by ~2 and causing ESS to be systematically underestimated by roughly half.

Optional serial interval estimation. Set estimate_SI = TRUE in household_dynamics() to jointly estimate the serial interval distribution as a Weibull(shape, scale) alongside other model parameters. Two additional MCMC parameters (si_shape, si_scale) are sampled. Priors: shape ~ Uniform(0.1, 10), scale ~ Uniform(0.1, 20). The existing serial_density() C++ function computes the discretized PMF from the Weibull parameters at each MCMC iteration.
Missing onset time imputation. Infected household contacts with missing (NA) onset times are now automatically imputed during MCMC via Bayesian data augmentation, drawing uniformly from the follow-up window and accepting/rejecting via the full likelihood. An informational message reports the number and percentage of missing onsets. Index case onset times must still be non-missing.
SI defaults to bundled influenza serial interval. The SI argument in household_dynamics(), run_MCMC(), and simulate_data() now defaults to NULL, which loads the bundled flu SI from Tsang et al. (2014). Users no longer need to call data(SI) and pass it explicitly.

simulate_data() now forces single-threaded execution to avoid thread-unsafe R::runif calls in parallelFor. This fixes correlated random draws that attenuated covariate effects in simulated data.

Added testthat test suite (47 tests across 4 files).
run_MCMC() now warns when with_rm = 1 (experimental random-effects branch with known per-household recording limitation).

Missing covariate imputation via Bayesian data augmentation. Factor covariates with missing values (NA) are now automatically imputed during MCMC, jointly with onset time imputation. The MCMC draws from the full conditional distribution at each iteration using a uniform categorical proposal over factor levels. No user action is required — simply pass data with NAs in factor covariates. An informational message reports the number and percentage of missing values per covariate.

Only categorical (factor) covariates are imputed. Continuous covariates with missing values will still produce an error.
Interaction terms (~sex*age) are not supported with missing covariate data.
The same variable cannot appear in both inf_factor and sus_factor if it has missing values.

household_dynamics() now returns an S3 object of class hhdynamics_fit instead of a plain data frame. Use summary(fit) to get the old-style parameter estimates table.
Formula arguments (inf_factor, sus_factor) now use R formula objects (~sex, ~age) instead of character strings ('~sex', '~age'). Use NULL instead of '~' when no covariates are needed.

S3 methods for hhdynamics_fit objects:
- print(): brief model summary (data size, MCMC settings, runtime)
- summary(): parameter estimates with credible intervals
- coef(): named vector of posterior means
Full MCMC output preserved in the fit object. Access via:
- fit$samples: posterior parameter samples
- fit$log_likelihood: log-likelihood trace (for convergence diagnostics)
- fit$acceptance: per-parameter acceptance rates
- fit$update_accept: per-iteration infection-time update acceptance
- fit$imputed_data: final imputed data matrix
Input validation in household_dynamics() catches common errors (missing columns, wrong formula types, bad MCMC settings) with clear messages before the C++ MCMC runs.

Fixed buffer overflow in C++ MCMC when household size exceeds 20. The updateacceptrate matrix is now sized to max_member instead of a hardcoded 20.
para_summary() no longer overwrites the median with the mean, and no longer has plotting side effects (layout(), par()).
run_MCMC() now uses message() instead of print() for runtime output.