Skip to contents

Changelog

Source: NEWS.md

mostr (development version)

Breaking changes

  • sim_markov() replaces sim_trajectories_markov(), sim_actt1_markov(), and sim_actt2_markov(). Choose a model with preset: "actt1", "actt2", "violet", or NULL to supply your own baseline_data, intercepts, and lp_function. preset has no default. Any model argument you supply, such as intercepts or baseline_data, replaces that part of the preset and keeps the rest. Each preset’s model is described in its own help topic: ?actt1_preset, ?actt2_preset, and ?violet_preset. sim_markov("actt2", ...) gives the same trajectories as sim_actt2_markov(...) did with the same arguments and seed.

  • sim_markov() simulates exactly the patients in a supplied baseline_data and stops if you also supply n_patients or treatment_prob, which only sample patients from a preset. Previously the trial wrappers recycled the columns of a supplied baseline_data over n_patients generated patients, so a 2-patient baseline_data with the default n_patients = 1000 gave 1,000 patients.

  • preset = "actt1" uses the ACTT-2 transition model with the starting states reported for ACTT-1 (13.1%, 41.4%, 18.4%, and 27.1% in states 4 to 7) and a threshold for death raised from -22.0693 to -21.4071, so that 18% of control patients have died by day 28. The ACTT-1 model based on Rohde et al. (2024) that sim_actt1_markov() used has been removed, and vignette("partial-po-vglm-sops") now simulates VIOLET data.

  • An lp_function now receives treatment_effect, the treatment effect on the current day after treatment_effect_decay, instead of parameter, so treatment_effect_decay also works for your own models. With preset = NULL, treatment_effect defaults to 0, extra_params to NULL, absorbing_states to the worst state (length(intercepts) + 1), covariate_names to every column of baseline_data except id and yprev, and follow_up_time to 28 days (previously 60). lp_violet() has been removed.

New features

  • New preset = "violet" simulates 4-state daily trajectories (home, in hospital, on a ventilator or with ARDS, dead) from the Markov model that Frank Harrell fitted to the VIOLET trial of vitamin D3 to create his simulated trial simlongord. Patients are sampled from the 1,351 VIOLET patients’ starting states, ages, and SOFA scores. The default treatment odds ratio is 0.75, as in simlongord, and the daily state proportions agree with simlongord within simulation error.

  • violet_baseline now holds the 1,351 VIOLET patients as 858 distinct combinations of starting state, age, and SOFA score with their counts n (previously 250,000 resampled patients in a modified 6-state coding). Its states follow sim_markov("violet").

Minor improvements

  • sops(), avg_sops(), avg_time(), and avg_comparisons() now warn when times skips fitted visits. SOPs are calculated one transition per value of times, so with daily visits times = c(1, 7, 14) gives the probabilities after 1, 2, and 3 transitions, labelled 1, 7, and 14, not the probabilities on days 7 and 14. To get SOPs at selected visits, request every visit, such as times = 1:14, and keep the rows you need. The functions also warn when times is not in increasing order, and, for the starting profiles stored on a model fitted by orm_markov(), blrm_markov(), or vglm_markov(), when times does not begin at the first follow-up visit. There is no warning when gap_var is supplied, because each transition then uses its own elapsed time, or for times after the last fitted visit. The help for times now explains that it lists consecutive visits.

  • sim_markov("actt1") and sim_markov("actt2") are much faster for large samples, because they now calculate each day’s linear predictor for all patients in one step instead of one patient at a time, as sim_markov("violet") already did. The trajectories for a given seed are unchanged. A model that you supply through lp_function is still called once per patient.

Bug fixes

  • Wald intervals from simulation or bootstrap draws (conf_type = "wald" in inferences()) are now centered on the point estimate for SOP and time-in-state comparisons from avg_comparisons() and for SOPs from interpolate_sops(). These intervals were previously centered on the mean of the draws, which differs from the point estimate, so the reported estimate was not at the middle of its interval. Standard errors and percentile intervals are unchanged.

mostr 0.1.1

  • sim_actt2_markov() uses a model refitted with states 1 and 2 merged, because state 2 was mostly recorded only at study visits. The old fit rehospitalized about 29% of control patients (2% observed); the new one about 2.7%. The default treatment_effect of sim_actt2_markov() and sim_actt1_markov() is now -0.0730.

mostr 0.1.0

  • sim_actt1_markov() now adds the average age and sex contribution of an ACTT-1 patient (0.2253 on the log-odds scale) to every threshold. The reported thresholds describe a woman aged 0, so simulated patients previously moved to better states too quickly; placebo mortality by day 28 rises from 9.5% to 13.4%, closer to the reported 15.2% by day 29.

  • Bootstrap refits (inferences(method = "bootstrap") and method = "fwb") now always resample the data stored by orm_markov() or vglm_markov(): the rows used for fitting, after any subset, with their fitting weights. The refit_data argument of sops(), avg_sops(), avg_comparisons(), avg_time(), plot_transitions(), plot_correlation(), and plot_variogram() has been removed; drop it from existing calls. newdata still sets the patient profiles used for prediction and can contain any profiles. Refit bootstraps of models fitted without these wrappers now error.

  • bootstrap_model_coefs() no longer has a data argument. It resamples the patients stored on a model fitted with orm_markov() or vglm_markov(), and id_var now defaults to the patient ID column given when fitting.

  • fast_group_bootstrap(), materialize_bootstrap_sample(), apply_to_bootstrap(), bootstrap_analysis_wrapper(), and relevel_factors_consecutive() are no longer exported. Use inferences() or bootstrap_model_coefs() for bootstrap calculations.

  • Bootstrap refits use starting values (from use_coefstart = TRUE or a coefstart passed to vglm_markov()) only when the resampled data contain every outcome state. When a state is missing, the refitted model has fewer intercepts, so it is fitted without starting values. A refit with starting values that fails, for example because a covariate level is absent from the resample, is retried once without them.

  • avg_sops() now accepts variables = NULL to average over patients using their observed starting states and covariates, with the existing analytical, posterior, simulation, and bootstrap inference methods.

  • avg_time() estimates average total time in separate or pooled states, using visit sums or mapped elapsed-time integration. It supports conditional and unconditional analytical variance, posterior intervals, and all existing simulation and bootstrap methods. Use vcov() for analytical covariance and get_draws() for retained time-total draws.

  • With time_map and no target_times, avg_comparisons(estimand = "time_in_state") (and "time_benefit") and time_in_state() for SOP data frames now integrate from baseline_time (default 0) to the last mapped visit, including the interval before the first visit. Previously, the omitted grid started at the first mapped visit. Set baseline_time = NULL or supply target_times to choose the period. avg_time() uses the same default.

  • When variables sets the starting state (for example variables = list(yprev = c("1", "2"))), real-time summaries from avg_time(), avg_comparisons(estimand = "time_in_state"), and interpolate_sops() now start each scenario in the set state at baseline_time. Previously they started from the patients’ observed starting states. These results support only conditional variance: inferences(method = "delta") requires vcov = "conditional" (or a coefficient covariance matrix), and method = "mvn" remains available; unconditional delta, score-bootstrap, bootstrap, and FWB inference now error.

  • plot_time() plots avg_time() results. type = "bar" draws one horizontal stacked bar per scenario, split into the average time in each state (“Grotta bars”); scale = "proportion" shows shares of follow-up. type = "pointrange" draws one point per state and scenario with its vertical confidence interval. Scenarios are drawn in black and successively lighter shades of gray.

  • inferences() now tests avg_comparisons() results against “no effect” by default: 0 for differences and 1 for ratios. The result gains statistic, p.value, and s.value columns. The new default is null = "auto"; pass a number to test against another value, or null = NULL to skip the test. SOP and average-time results are still not tested unless you supply null.

  • plot_correlation() and plot_variogram() now default to Spearman rank correlation for observed data and fitted models, accounting for tied states. Use method = "pearson" to retain the previous calculation.

  • plot_correlation() now uses a 0-1 color scale when all available correlations are nonnegative, retaining -1 to 1 when negative correlations are present. Set fill_limits explicitly to keep a common scale across plots.

  • Speed up row assembly, grouped summaries, patient bootstrap materialization, and baseline selection using data.table internally. Results remain data frames.

  • Initial release of Markov Ordinal State Transition modeling tools, migrated from markov.misc with its original authorship and GPL license preserved.

  • Retain Markov data generation, first- and second-order transition models, state occupancy probabilities, treatment comparisons, time in state, diagnostics, and analytical, posterior, MVN, and bootstrap inference.

  • Remove non-Markov generators, simulation-study and power workflows, Arrow sampling, endpoint conversions, competing-risk analyses, and coefficient tidiers. The simulation-truth helper calc_time_in_state_diff() is also excluded; model-based summaries remain available through time_in_state().

  • Use Markov-generated data throughout examples, tests, and all nine vignettes.

  • sim_actt2_markov() supports custom follow-up, including 60 days; the separate Brownian-calibrated 60-day generator is excluded.