NEWS.md
This is a maintenance release. It fixes a number of defects in the framework functions, most of which produced wrong results silently rather than failing, and adds input validation and unit tests so that the same mistakes are caught in future.
The calibrated parameters are unchanged. Re-running the IMIS calibration reproduces the published posterior bit-for-bit: m_calib_post and v_calib_post_map are identical to version 0.1.0, and tables/03_summary_posterior.csv agrees to 14 significant digits. The framework functions also return the same values as 0.1.0 at the parameters shipped with the package (log-posterior at the MAP = 19.00374). The cost-effectiveness results do change; see Regenerated results below.
R/03_calibration_functions.R)log_lik() assigned -Inf to the whole v_llik_overall vector instead of to element j when the decision model could not be evaluated at a parameter set. A single failing parameter set collapsed the vector to length 1 and then re-grew it, so the log-likelihoods already computed for valid parameter sets were replaced by -Inf and NA. Any Bayesian calibration that encountered one infeasible draw returned a corrupted likelihood surface with no warning. A failing parameter set now gets -Inf on its own and the others are untouched.
log_lik() read the third calibration target as SickSicker_targets$PropSick and l_model_res$PropSick, which only resolved because $ partial-matches PropSicker on lists. Any additional target whose name began with PropSick would have silently changed which target was used. All target names are now exact.
calibration_out() took model outputs from the hard-coded cohort-trace rows c(11, 21, 31). With a time horizon shorter than 30 cycles it returned NA outputs and the calibration proceeded on them. Target times are now the v_target_times argument (default c(10, 20, 30)), matched against the named rows of the trace and validated against n_t.
prior(), likelihood(), log_post() and posterior() accepted only v_params, so prior bounds, target weights and l_params_all passed to them were silently dropped and the defaults were used instead. They now take and forward the same arguments as log_prior() and log_lik().
analysis/03_calibration.R computed the MAP estimate as which.max(exp(log_post(...))). exp() underflows to zero for every draw once the log-posterior is strongly negative, in which case which.max() silently returns the first draw. The MAP is now taken on the log scale.
R/05b_deterministic_analysis_functions.R)calculate_ce_out() discounted costs at d_e and QALYs at d_c — the two rates were crossed over. The repository ships d_c = d_e = 0.03, so the base case and the published results are unaffected, but any user who set different rates for costs and effects got each stream discounted at the other’s rate.
owsa_det() and twsa_det() validated FUN with a test (is.na(sum(is.na(jj)))) that is FALSE both when FUN works and when it fails, so the guard could never fire. A broken FUN failed later with an unrelated message. The check now works and reports the underlying error.
owsa_det() and twsa_det() consumed ranges positionally while checking its names with sum(parms == names(ranges)), which errors on a length mismatch and gives a misleading message when the list is unnamed. ranges is now matched to the parameter names, so a list written in a different order than parms no longer varies a parameter over another parameter’s range.
twsa_det()’s progress report compared floating point numbers with ==.
R/05a_probabilistic_analysis_functions.R)generate_psa_params() overwrote its own n_sim argument with nrow(m_calib_post) on the first line of the body, so it always returned 1000 parameter sets. analysis/05a_probabilistic_analysis.R sized its output matrices from its own n_sim, so raising n_sim above 1000 there left the PSA loop reading past the end of the input data frame and feeding rows of NA parameters to the model.
generate_psa_params() assigned its seed argument to an unused local (set_seed <- seed) instead of calling set.seed(), so the PSA dataset was not reproducible and two identical calls returned different data.
analysis/05a_probabilistic_analysis.R called data("l_psa") immediately after building the PSA object, silently replacing the object just simulated with the archived one, so every figure and table below came from the packaged data rather than from the run. It is now commented out.
R/02_decision_model_functions.R)check_sum_of_transition_array() did not check what it documented. It summed all the row sums of a cycle and compared that total with the number of states, so row sums that were wrong in opposite directions cancelled out: an array whose rows summed to 1.25 and 0.75 passed. It also used exact == equality on floating point numbers, so a valid array could be rejected for rounding alone. Every row of every cycle is now checked against one within a tolerance, and the failing rows, cycles and sums are reported.
check_transition_probability() selected invalid entries with which(a_P < 0 | a_P > 1). Comparing NA or NaN with < or > gives NA, which which() drops, so an array with undefined transition probabilities passed and the model went on to produce a cohort trace full of NA.
R/01_model_inputs_functions.R)update_param_list() appended misspelled parameter names to l_params_all and left the parameter the user meant to change at its old value. The model then ran to completion and returned base-case results with no warning. Unknown names now raise an error listing them; allow_new = TRUE restores the previous behaviour.
load_all_params() built its parameter list with the self-assignment df_params_init <- df_params_init and then appended df_params_init at the end, which was always the package data. A user’s file.init was therefore partly ignored.
load_all_params() ended on an assignment, so it returned invisibly and printed nothing at the console.
v_age_names was one label shorter than the n_t + 1 rows of the cohort trace it labels.
dplyr::select(df, .data$Total) is deprecated in tidyselect >= 1.2.0 and warned on every call to load_all_params().
R CMD build failed: knitr::opts_knit$set(root.dir = '..') inside the figure chunks of two vignettes pointed the following chunk at the package’s parent directory, where ../figs/*.png does not resolve. No source tarball with vignettes could be produced, so R CMD check could not run to completion. R CMD check now reports Status: OK.
open_guide() was broken for anyone using darthpack as an installed package. The guide lives in the top-level report directory, which is not installed, so system.file() returned "" and the function opened the root of the filesystem. It now falls back to the online guide.
SickSicker_targets.Rd and df_posterior_summ.Rd were malformed and lost their \format and \description sections: under markdown = TRUE, the \% in the roxygen block became \\% in the Rd file, where the % starts a comment that swallowed the rest of the line.
sample_prior() follows the DARTH snake_case naming convention. sample.prior() is retained as an exported alias, because IMIS::IMIS() resolves the sampling function by that exact name.
log_lik(), log_post() and posterior() gain a v_weights argument for the per-target weights that the comments previously told users to set by editing the function body.
log_post() no longer runs the decision model for parameter sets outside the support of the prior, where the log-posterior is -Inf whatever the likelihood is. This avoids evaluating the model at infeasible parameter values and speeds up calibration.
calibration_out() gains v_target_times; check_sum_of_transition_array() gains tolerance and defaults n_states and n_t to the dimensions of the array; twsa_det() gains progress; update_param_list() gains allow_new; generate_psa_params() accepts seed = NULL; open_guide() gains online.
The two check_ functions return their report invisibly, so it can be inspected without parsing warning text.
The test suite grows from 27 to 294 assertions and now covers components 01, 02, 03, 05a and 05b. Each of the defects above has a test that fails against version 0.1.0. One pre-existing assertion read all(a_P >= 0) | all(a_P <= 1), which passed whenever either bound held on its own, and is now &. tests/testthat/test_05a_deterministic_analysis_function.R was renamed to test_05b_deterministic_analysis_functions.R: it tested a component 05b function under an 05a name.
plotrix moved from Imports to Suggests; no package function uses it.bookdown, matrixStats, psych, scatterplot3d, shiny, rstudioapi and usethis. Users who followed the README and installed everything in DESCRIPTION were missing them.BiocManager and rlang, which nothing uses.IMIS has been archived on CRAN; analysis/03_calibration.R now documents installing it from the archive with devtools::install_version().All the artefacts in data/, output/, figs/ and tables/ were regenerated by running analysis/_master.R end to end against the fixed code, so the repository is now internally consistent: every committed figure and table is what the code in the repository actually produces.
The calibration is unchanged. The IMIS run reproduces the published posterior exactly, so m_calib_post, v_calib_post_map and the posterior summary carry the same values as before (the summary table differs only in the last one or two of 15 significant digits).
The deterministic CEA table changed substantially, because it had been stale. tables/05b_deterministic_cea_results.csv previously held No Treatment 159232.1 / 22.08072 and Treatment 293747.6 / 22.80879, giving an ICER of 184755.6. Those numbers were produced by an earlier version of calculate_ce_out() and were never regenerated, so they did not match what the code in the repository produced even before this release. The table now holds 115244.8 / 20.0233 and 214165.8 / 20.7226, with an ICER of 141457.4. The discrepancy was not caused by the discounting fix, which cannot change these numbers while the repository sets d_c = d_e.
A useful cross-check: the deterministic and probabilistic ICERs now agree (141457 and 141799). Previously they differed by 29% (184756 against 143201) for the same model, which is what flagged the stale table.
The PSA results moved slightly, because the sampling is now reproducible. generate_psa_params() now sets the seed it is given, so the draws differ from the unseeded ones behind the old l_psa. The probabilistic ICER moves from 143201.5 to 141798.7, within Monte Carlo error for 1000 samples. Re-running analysis/05a_probabilistic_analysis.R now reproduces l_psa exactly.
tables/05a_probabilistic_cea_results.csv no longer has the unnamed row-number column that write.csv() added without row.names = FALSE.
All 21 figures in figs/ were regenerated.
To keep the package data objects in step with the analysis output, _master.R now also sources data-raw/03_imis_output.R and data-raw/05a_psa_dataset.R, which refresh m_calib_post, v_calib_post_map and l_psa. Running only the analysis/ scripts, as the previous _master.R did, left those .rda objects holding the old calibration while output/ held the new one — and because generate_psa_params() reads m_calib_post from the package data, the probabilistic analysis would then have run on the stale posterior. The data-raw/ scripts now pass overwrite = TRUE so they can be re-run at all.
Note that _master.R re-runs the IMIS calibration, which takes a few minutes. The previous version said the calibration and validation components were commented out, but the source() calls were live.
Archived the release that accompanies the published article in zenodo: https://zenodo.org/record/3445451.
Version included in following manuscript:
Added a NEWS.md file to track changes to the package.