Changelog • epidist

epidist 0.5.0

Models

Added an exponentially growing primary event distribution to the latent model. as_epidist_latent_model(primary = "expgrowth") makes the growth rate a distributional parameter, so it takes a brms formula and prior and can vary by covariate. See #489 and #618.

Features

Added delay_summary_draws(), which wraps the three usual post-processing steps into one call. It builds one row per unique combination of the predictors with epidist_strata(), draws the delay distribution parameters for each with delay_parameter_draws(), and adds the natural scale mean and standard deviation, and any quantiles asked for, with add_summaries(). Each step is still available on its own. See ?delay_summary_draws and #667.
Added epidist_newdata(), which builds the newdata needed to predict from a fitted model. It expands the variables you give it into a grid and adds the response and observation process variables the model uses, so you no longer have to know the column names each model expects. The defaults give the delay distribution with no censoring and no truncation, and arguments set the censoring windows, the relative observation time and the minimum delay. The result works with brms::posterior_epred(), delay_summary_draws() and the tidybayes draw functions. See ?epidist_newdata and #280.
Added delay_parameter_draws() and add_delay_parameter_draws(), which return posterior draws of the delay distribution parameters in the long format used by tidybayes.
Added delay_parameter_draws() and add_delay_parameter_draws(), which return posterior draws of the delay distribution parameters in the long format used by tidybayes. The draws come back with .row, .chain, .iteration and .draw columns alongside the columns of newdata. predict_delay_parameters() and predict_dpar() are removed in their favour. See #471.
Added add_summaries(), which adds the mean, the standard deviation and quantiles of the delay distribution implied by each draw of its parameters. It uses the analytic solution for the lognormal, gamma and Weibull families, and simulates from any other family, so it works for every family brms can predict from. add_mean_sd() is removed in its favour. See #471.
Added epidist_strata(), which returns one row of the model data per unique combination of the variables that predict the delay distribution parameters. Passing it to add_delay_parameter_draws() draws each set of parameters once rather than once per observation. See #471.
epidist data objects now check themselves when they are modified. Every object also carries a shared epidist_data class with methods for subsetting, replacement, rbind() and the dplyr verbs. These re-check the object and drop any epidist class whose requirements it no longer meets, warning about what was dropped and why. An object that still carries an epidist class is therefore a valid object of that class. dplyr::group_by() and results with no columns are exceptions, both documented in ?epidist_data. See ?epidist_data and #399.
Dropped the checks in epidist_stancode() and in the conversions between linelist and aggregate data, which ran on objects that had already been checked. The conversions from linelist data to a model still check their input, because new_epidist_linelist_data() does not. See #399.
epidist_transform_data_model() now checks the object it builds for the marginal and naive models. That object was never checked before, which only showed once the check in epidist_stancode() was removed. See #399. ## Package
Reworded the message as_epidist_marginal_model() gives when it sets relative observation times to Inf. The message now names relative_obs_time and orig_relative_obs_time. It explains that the impact on accuracy is small because these observation times cause very limited right truncation. It also points at obs_time_threshold for users who do not want the behaviour. See #536.
epidist_prior() no longer warns about user priors on parameters that are in the model but not in the epidist default set. The warning now checks user priors against the brms default priors for the model as well as the epidist ones, so a prior on a regression coefficient no longer looks unmatched. See #483.
Simplified the internals of prior handling. .replace_prior() now only merges priors, with the warning about unmatched priors moved to .warn_unmatched_prior() and the merge argument handled in epidist_prior(). The latent model checks of the event window priors moved to R/latent_model.R and dispatch on the data class, so a model can now state its own prior requirements. The returned priors, and the Stan code they produce, are unchanged. See #483.
The warning about unmatched priors now lists each prior with the parameter it applies to, rather than printing the internal join it came from. See #483.
Documented the return value of every exported function.
Declared scales in Suggests, which the FAQ vignette loads but nothing declared.
Fixed four typos that inst/WORDLIST was masking, corrected two moved URLs, title cased the Title field, and set Language: en-GB.
Fixed inst/CITATION rendering the year as NULL.
epidist() now restores the PKG_CPPFLAGS and PKG_LIBS environment variables it found before fitting. The rstan backend sets both while compiling and never restores them. The leaked PKG_CPPFLAGS made the next pkgbuild::has_build_tools() check fail, which printed a spurious fatal error: cmath: No such file or directory before the model compiled and fitted successfully. See #532.
Added simulate_dates(), which turns simulated event times into the censored dates an analyst would receive.
Removed the calls to unexported brms functions that R CMD check --as-cran flags. R/brms-compat.R now holds small internal helpers reproducing the narrow behaviour epidist relied on from brms:::validate_family(), brms:::validate_formula(), brms:::validate_data(), brms:::dpar_bounds() and brms:::log_lik_weight(). The helpers are written against the public brms interface rather than copied from brms. tests/testthat/test-brms-compat.R checks each one against the brms internal it replaces. Those checks are skipped on CRAN, since they reach into brms internals. Credit for the original behaviour goes to the brms authors. See #420 and paul-buerkner/brms#1676.
Removed the Remotes field from DESCRIPTION so dependencies resolve from CRAN. cmdstanr is now found through Additional_repositories and the development version of brms is no longer used. See #592.
Turned off evaluation of the approximate inference vignette. It uses pathfinder, which needs an unreleased brms fix. This release resolves brms from CRAN. See #579.
Added a brms (>= 2.23.0) floor, the version the compatibility helpers were checked against.
Pointed the CI workflows at the Stan r-universe with extra-repositories. Dropping Remotes means pak can no longer resolve cmdstanr. pak does not read Additional_repositories.
Raised the minimum R version to 4.1.0. The package uses the native pipe and the lambda shorthand. Both need R 4.1.0.
Added the copyright holder role to Sam Abbott in DESCRIPTION.
Guarded the shared test fits and the tests that use them so the suite runs without cmdstanr.
Wrapped the epidist() and epidist_diagnostics() examples in \donttest{}. Both fit a model. They ran for 118 and 110 seconds against CRAN’s 5 second guidance.
Anchored the brms links in the documentation so R CMD check no longer reports Rd cross-references with missing package anchors.
Dropped a stale fix entry from the declared global variables.
Updated the brms documentation URL, which had moved.
Added cran-comments.md.
Rewrote the generic epidist_gen_log_lik() method so it evaluates the brms log likelihood once per delay rather than once per delay per posterior draw. A single brms call already returns the cdf for every draw, so the results are cached and reused. The method also calls primarycensored::pcens_cdf() directly instead of primarycensored::dpcens(), which revalidates the distribution function at random points on every call and so would defeat the cache. Cost is now linear rather than quadratic in the number of draws. For 500 draws this is around 80 times faster, and the log likelihoods are unchanged. The guard that dpcens() applied when the delay upper bound exceeds the relative observation time is reproduced explicitly, since this no longer goes through dpcens(). Left truncation is carried through the rewritten path: the density is normalised over the interval from delay_min to the relative observation time. See #476.

Documentation

Added an extending-epidist vignette covering why you might build your own model type, the six generics a model type implements, a worked example, and a table of the packages that already extend epidist.
Precomputed the ebola, faq and approx-inference vignettes. All three fit models and need cmdstanr, so they were excluded from the build by .Rbuildignore and never reached anyone who installed the package. They are now knitted from a .Rmd.orig source into a committed .Rmd holding static output, so they ship without needing cmdstanr or a model fit at build time.
Gave each precomputed vignette its own figure prefix. ebola and approx-inference both wrote to figures/epidist-, which would collide once more than one is precomputed.

Package

Made epidist_family_param() internal. It is reached through epidist_family(), and a custom model supplies its family through epidist_family_model() instead. See #79.
Exported epidist_gen_log_lik(), which was the only one of the three post-processing generators not exported. See #79.
Made epidist_transform_data() internal. It is a wrapper that dispatches to epidist_transform_data_model(), which is the generic an extension implements and which remains exported. See #79.

Models

Added left truncation support via a delay_min parameter in as_epidist_marginal_model(). This passes the L (left truncation) argument through to the primarycensored likelihood. The default of 0 reproduces the previous behaviour. See #588 and #596.

CI

Added a render-vignettes workflow that rebuilds the precomputed vignettes and opens a pull request with the result.

Models

Added the meta model, for fitting to summarised and potentially biased published estimates, jointly with individual level data. Published estimates are forward modelled from the study’s own estimation procedure, so summaries that did not adjust for right truncation or that treated interval censored data as continuous can still contribute unbiased information, given correct metadata describing what each study did. The meta model is experimental and its interface may still change. See as_epidist_meta_model(). See #620.
Added as_epidist_estimates_data() for preparing published summary estimates, with documentation of the study metadata the meta model needs. See #620.
The meta model supports studies that stopped collecting at a calendar date through the trunc_design field of as_epidist_estimates_data(), which weights the estimand by the follow up available to each delay rather than conditioning on a single cohort cutoff. See #620.
The meta model supports midpoint imputation, where a study assigned each delay to the centre of the interval it was observed in, as cens_adjusted code 3. See #620.
A standard error supplied for a quantile row of as_epidist_estimates_data() is now interpreted on the delay scale, as studies report it, and the row is fitted on that scale against the implied quantile. See #620.
as_epidist_estimates_data() now rejects a reported quantile at or beyond the largest delay its study could have seen, which would otherwise contribute a constant to the likelihood rather than information. See #620.
The meta model now fits the summaries a study computed from the same delays jointly rather than as independent terms. A mean with a standard deviation uses the asymptotic bivariate normal of the pair, and a set of quantiles uses the multinomial mass of the delays falling between them. This removes the over-weighting of a study reporting a median with an interquartile range. One observation is now a group of summaries, so log_lik() and loo() work at that level. See #620.
A quantile reported by a study that took integer date differences from a cohort now costs three distribution function evaluations rather than one per grid cell, because the grid is normalised by the distribution function at its top. On the quantile reporting studies of the meta vignette this cuts the evaluations they need per gradient from 480 to 36. The shortcut does not apply to a study that stopped collecting at a calendar date, which reweights each cell before renormalising and so keeps the full grid. See #620.
The implied quantile of a continuous estimand is inverted exactly where the family quantile function exists and refined by Newton steps otherwise, so covariance rows with quantile members and quantile rows with a delay scale standard error are no longer limited by the quadrature spacing. See #620.
The meta model gained several further speed ups. Softmax normalisation of the cohort grid mass is replaced by division where the normaliser is already known. A zero growth rate accrual design skips its exponential terms. R post-processing batches and caches implied summaries across grouped rows sharing a design. See #620.
The quadrature resolution used for truncated continuous moments is now set by options(epidist.meta_n_quad = ), defaulting to 100 intervals. This is the floor of the resolution chosen per study, see below. See #620.
Added as_epidist_multivariate(), which summarises draws of a set of parameters by their mean vector and covariance matrix, over an optional trajectory index. Passing the result to as_epidist_estimates_data() gives a vector of reported summaries with the covariance between them, fitted as a multivariate normal. This is the format we recommend when a study cannot share its delays, because it keeps the correlation between the quantities it reports. Draws of the natural parameters of a fitted distribution are pushed through to the summaries the distribution implies, so no linearisation is used. See #620.
Added epidist_estimates_summaries() and epidist_estimates_parameters(), which take one study’s contribution in the shape it reported it. epidist_estimates_parameters() converts the parameters of a distribution a study fitted into the summaries that distribution implies, carrying any reported parameter standard errors onto that scale by the delta method as a covariance over the summaries, which is fitted jointly and carries the study’s information about its parameters exactly. The family a study fitted need not match the family being fitted to it. See #620.
as_epidist_estimates_data() combines contributions passed in a list, so studies reporting in different shapes assemble into one object. See #620.
The meta model gained a fifth censoring adjustment code, cens_adjusted = 4, for a study that placed the primary event at the midpoint of its window and integrated the secondary interval. Its estimand is that of cens_adjusted = 2 moved down the delay axis by half a primary window, so its reported mean loses the half window bias while its spread keeps the primary window’s variance. See #620.
The meta model supports left truncation through delay_min, on both individual level rows and summary rows. A study that only counted delays above a minimum has every implied summary conditioned on the delay exceeding it. See #596 and #620.
A single quantile reported by a study that summarised integer day delays is now fitted as the cell in which the empirical distribution function crossed its probability, the exact event a rounded quantile stands for, rather than with a multinomial on the continuity corrected grid whose claimed precision keeps growing with the sample size. Two quantiles reported at the same value are accepted and merged into one cell, and as_epidist_estimates_data() warns when a large study reports several such quantiles, whose joint likelihood is still overconfident. See #620.
A multinomial cell that underflows is floored rather than sent to zero, so the R and Stan log likelihoods are both finite for a badly misfitting draw and loo() keeps working. See #620.
as_epidist_estimates_data() refuses more summaries from a fitted family than it has parameters, and a covariance over reported summaries that is singular to within a relative eigenvalue of 1e-4, because such a row charges any error in the implied summaries against a vanishing eigenvalue. See #620.
as_epidist_estimates_data() warns when the relative standard error of a reported standard deviation, at the kurtosis its mean and standard deviation imply under a lognormal delay, exceeds a quarter, which is where the normal sampling likelihood of a standard deviation stops being calibrated. See #620.
The meta model takes a primary argument for its individual level rows, as the marginal model does. With primary = "expgrowth" the growth rate of primary events is estimated as the pgrowth distributional parameter. Summary rows are unchanged and keep the growth_rate metadata of their study as a known tilt. See #620.
Added an epidist_model_prior() method for the meta model, which puts a normal(1, 1) prior on the intercept of mu where it is on the log scale, the scale of the lognormal family prior. The centre is fixed rather than taken from the reported values, because a prior chosen from the data would put the posterior of a small review where the data already sit. The same method puts a half normal prior with a standard deviation of 0.25 on the between study standard deviation of any group level term, where the brms default is a half Student t with scale 2.5. Closes #683. Without it a Gamma or Weibull fit to summaries alone took the brms default, which is centred on the response column and so on a delay of zero, because that column is a placeholder on summary rows. See #620.
Added an epidist_newdata() method for the meta model, which builds an individual level row with the same arguments as the marginal model method, so predicting from a meta model fit no longer means copying a summary row out of the model data. See #620.
as_epidist_estimates_data() now rejects a reported mean at or beyond the observation time of a study that did not adjust for right truncation, a standard error of zero, a standard deviation of zero, and a cens_adjusted code that is not a whole number from 0 to 4. It warns, rather than messages, when no trunc_adjusted column is supplied and a study is therefore assumed to have adjusted for right truncation. The growth_rate documentation now separates its within window tilt from the accrual weight it applies under trunc_design = "accrual", and points to the primary = "expgrowth" option of the marginal model for individual level data. See #620.
as_epidist_estimates_data() accepts NA censoring windows for a study that fully adjusted for censoring (cens_adjusted = 1), since none of its estimands read them, and rejects them with a message naming the study for every other code. See #620.
The advisory messages of as_epidist_estimates_data() are now two short sentences naming the studies, and the input row where a single summary is meant, and point at a new Checks section of its documentation that carries the reasoning. They run once, when the estimates are built, rather than again when the object is passed to as_epidist_meta_model(). See #620.
The advisory messages of as_epidist_estimates_data() now say what each problem is in plain language, run once on the studies combined in a list rather than once per study, and are followed by a single pointer to the Checks section, with simulate_study() leaving them to the combined object and advise = FALSE skipping them. See #620.
The midpoint imputation codes of the meta model (cens_adjusted 3 and 4) now move delay_min with the midpoint shift, so a study that dropped reported delays below a minimum is left truncated at the right point. Before this the implied mean was 5 to 15% low for code 4 and up to 24% high for code 3 with a wide secondary window. See #620.
A study that adjusted for right truncation and counted only delays above delay_min now has its moments left truncated analytically, so they describe the same estimand as its quantile rows and no longer depend on max_delay. For a heavy tailed delay the implied standard deviation was 6 to 17% low before this. See #620.
The accrual weight on the discrete grid now cuts each reporting cell at the primary windows it spans, so it is exact for unequal censoring windows whenever the collection window is a multiple of pwindow. Weighting at the cell’s lower edge put the implied mean 12 to 36% low for a daily primary and weekly secondary window. See #620.
The meta model now chooses its quadrature resolution per study from the spread the study reported, so that the node spacing is at most a quarter of that spread, with options(epidist.meta_n_quad) as the floor and 2000 intervals as a cap the option lifts when set above it. The number travels with each summary row as its n_quad slot rather than being compiled into the Stan code. Before this a fixed 100 intervals over the default max_delay put the implied standard deviation of a delay with a coefficient of variation of 0.05 out by a factor of two and pinned its kurtosis at the floor, which made a reported standard deviation almost infinitely precise. as_epidist_estimates_data() now warns only when the cap leaves a study unresolved. See #620.
A draw whose implied moments overflow is rejected on every meta model moment row rather than returning NaN for an ungrouped standard deviation. The analytic moments themselves now reject such a draw in Stan and return the failure vector in R, so a covariance row rejects a very wide draw instead of carrying an infinite gradient, and the posterior predictive of such a draw is NA rather than a NaN standard error. See #620.
The meta model no longer evaluates the primary censored distribution function at grid or quadrature nodes deep in the lower tail of a narrow delay, where the Stan function of primarycensored has a finite value with a NaN gradient. A node whose log distribution function is certainly below -100, decided from a closed form bound on the parameters, is treated as holding no mass in both implementations. CmdStan’s gradient diagnostic now passes on every design of the meta vignette at a log standard deviation of 0.03, where four of them could not start before. See #620.
Fixed two gaps in the Stan mirror of a left truncated midpoint code: the delay scale quantile path sized its nodes from the unshifted delay_min, and the Newton refinement of an implied quantile normalised from delay_min rather than the moved left truncation point of a code 4 study. The Stan crossing cell of a single integer day quantile also takes its binomial upper tail through the accurate side, which had put it 5e-3 away from R at a poorly fitting draw. See #620.
The crossing cell likelihood of a single integer day quantile is now taken as the difference of two binomial tails on the side where both are small, with each tail summed term by term once it is far out, and as a sum over the count below the cell where the difference still underflows, in R and in Stan. Before this the difference cancelled or underflowed to -Inf when the implied distribution put the reported quantile far into its tail, which stopped chains initialising on the Ebola fit of the meta vignette, and Stan’s binomial distribution function had non finite partial derivatives in the same region, which trapped a chain of the shared test fixture in one run in four. CmdStan’s gradient diagnostic now matches finite differences at every probed point. See #620.
as_epidist_estimates_data() rejects a cens_adjusted = 4 study whose delay_min plus half its pwindow reaches the grid cutoff. See #620.
The meta model intercept prior for the lognormal family is now on the log scale, as it is for a log link. The lognormal mu has an identity link but is meanlog, so it was treated as the delay itself. A calibration check of a study reporting a mean with its standard error from 25 delays found it, where the 90% intervals covered the truth about half of the time. See #620.
Added opt in simulation checks of the meta model: EPIDIST_META_RECOVERY=true fits one meta model per censoring adjustment code and truncation design, and EPIDIST_META_CALIBRATION=true fits forty replicates of two study designs and checks interval coverage and the rank of the truth. See #620.
Added simulate_study(), which applies the observation and estimation procedure of one published study to a simulated line list and returns the summaries that study would have reported as an epidist_estimates_data object. It covers every censoring adjustment code, both truncation designs, a minimum delay and a subsample, and reports a mean and standard deviation, quantiles, a mean with a standard error, or a multivariate mean and standard deviation with their bootstrap covariance. Closes #672.

Documentation

Added a “The meta model” section to the model guide vignette, with the forward model and sampling likelihoods used for published summary estimates. See #620.
Added a vignette showcasing the meta model on simulated data. Its case study builds nine studies, each applying a different estimation procedure to the same line list, so the recovery result tests every bias the model adjusts for. See #620.
The meta vignette is now precomputed from a .Rmd.orig source, so it ships with the package. See #619 and #620.
The meta vignette now works through the published Ebola onset to death estimates collated by epireview. It adjusts for the phase of the outbreak each estimate was made in, taking the retrospective studies as the reference so the population level estimate is the one least affected by right truncation, and reports the phase bias as a marginal effect. It reports the population level posterior of the Gamma shape and scale alongside the natural mean and standard deviation, and compares the result with a modern re-analysis of one of the same line lists. See #620.

Bug fixes

.delay_family() now strips the meta_ prefix alongside latent_ and marginal_. Without it add_summaries() could not find the delay distribution of a meta model fit, because the family is named meta_gamma rather than gamma. See #620.
Meta model objects now carry the shared epidist_data class, as the latent and marginal model objects do, so they are re-checked when modified and the helpers that dispatch on that class no longer skip them. Closes #684.
The coarse quantile warning of as_epidist_estimates_data() now keys on the smallest quantile a study reported rather than the largest, because that is the one nearest the edge of the discrete grid and it biased the fitted spread of a midpoint study that escaped the warning. Closes #682.
The short grid cutoff warning of as_epidist_estimates_data() now judges max_delay by a lognormal matched to the reported mean and standard deviation, or to the median and largest quantile where only quantiles are reported, and fires when more than 2% of its second moment lies beyond the cutoff. The previous mean + 10 sd yardstick missed a heavy tailed study whose implied standard deviation was 6.5% low at the default cutoff. Closes #681.
The default max_delay of as_epidist_estimates_data() is now the delay beyond which one percent of the second moment of a lognormal matched to the study’s summaries lies, rounded up to a whole number of secondary windows with a floor of ten, in place of twenty times the largest reported value. This is the yardstick of the short cutoff check, so the default never trips it, and it gives a delay of mean 7 and standard deviation 3.6 a cutoff of 31 days rather than 140. A study reporting a single quantile or a mean with a standard error, where nothing can be matched, gets five times its largest reported value. See #620.
Added a missing Jacobian adjustment to the latent model for observations whose primary and secondary censoring windows overlap. Without it the latent model did not target the same likelihood as the marginal model. Under daily censoring the affected observations are the zero-delay cases. See #606.
Declared reformulas in Suggests and skipped the marginaleffects integration test when it is absent. insight needs reformulas to read the formula of a brmsfit, but only suggests it, so the test failed on a clean library. See #601.

Documentation

Added a left-truncation vignette showing how to use delay_min. See #596.

CI

Passed the coverage report to codecov/codecov-action through files rather than file. file is not an input the action accepts, so with disable_search set it found no report and the test-coverage job failed on main.
Pinned the precommit hooks to a revision whose lockfile uses digest 0.6.39. The tagged v0.4.3 lockfile pins digest 0.6.36, which calls Calloc and Free. Those were removed from the R API in R 4.5, so the hook environment failed to build and the pre-commit job failed on every pull request. See #578.

Documentation

Documented installing from CRAN in the README, with r-universe as the route to the latest version.

epidist 0.4.1

Bug fixes

Fixed Stan compilation failure with primarycensored >= 1.4.0 by adding the new L (left truncation) parameter to the primarycensored_lpmf call in the marginal model. See #583.
Added primarycensored (>= 1.4.0) version bound to DESCRIPTION.
Updated test expectations for changed primarycensored error handling.
Re-enabled approximate inference vignette evaluation using dev brms with pathfinder path fix. See #579.

Package

Load only required primarycensored Stan functions (primarycensored_lpmf and ODE/distribution helpers) with pcd_load_stan_functions(dependencies = TRUE) instead of loading all functions. See #582.

CI

Extended check-cmdstan workflow to also check marginal model Stan syntax.

epidist 0.4.0

Package

Enforce line length and use cli for latent prior checks. See #580.
Removed CodeDepends from DESCRIPTION dependencies.

Documentation

Restructured pkgdown reference with higher-level categories. See #574.
Updated FAQ to recommend pp_check with expanded data. See #575.
Clarified weight parameter documentation in as_epidist_marginal_model(). See #565.
Fixed pathfinder parameter usage in approximate inference vignette. See #573.

epidist 0.3.1

Hotfix release to patch a change in how the grepl function works in new versions of R.

epidist 0.3.0

This release adds support for a wider range of distributions in the marginal model, improves documentation with new vignettes and FAQ sections, enhances the getting started guide with clearer examples of model comparison, and fixes several bugs related to parameter bounds and likelihood calculations.

Models

Added Stan-side support for fitting all distributions supported by primarycensored in the marginal model. See #540.
Added R-side analytical likelihood support for Lognormal, Gamma, and Weibull distributions. See #540.

Package

Remove caching of vignettes. See #533.

Documentation

Added a new vignette “Guide to the statistical models implemented in epidist”. See #514.
Added a new FAQ section showcasing how to use the posterior package with epidist models, particularly for working with random variables (rvars) to propagate uncertainty in calculations. See #547.
Added a new FAQ section on how to use the marginaleffects package with epidist models. See #547.
Reduced the focus on simulating data in the getting started vignette to make it more accessible. See #549.
Made the entry to the package friendlier with clearer examples and improved documentation. See #549.
Added a schematic to explain right truncation more clearly to the getting started vignette. See #549.
Added a comparison of fitting naive and marginal models in the getting started vignette to highlight the importance of accounting for biases. See #549.
Added examples showing how to extract estimated parameters and plot them against true values to evaluate model performance. See #549.

Bugs

Fixed a vector length issue for censoring that was causing problems in some likelihood calls. See #540.
Fixed a bug in the preprocessing of the Weibull family. See #540.
Fixed a bug where bounds were not set for mu parameters in custom families. See #549.
Fixed a bug in predict_delay_parameters() where it couldn’t detect brms families when used directly. See #549.

epidist 0.2.0

This release adds a new marginal model based on primarycensored which provides a more efficient approach for fitting delay distributions compared to the existing latent model. We’ve also improved data handling by adding support for aggregated data across all models, added comprehensive examples using real world data, and enhanced documentation based on user feedback. The package has also undergone significant internal improvements including generalised Stan reparameterisation and improved data transformation methods.

As part of this release we have moved from @athowes maintaining the package (who led the initial package development, implementation of the S3 infrastructure, implementation of the core models, and wrote the first versions of the getting started vignette, Ebola case study, FAQ section, and the approximate inference vignette) to @seabbs maintaining the package.

Models

Added a marginalised likelihood model based on primarycensored. This can be specified using as_epidist_marginal_model(). This is currently limited to Weibull, log-normal, and gamma distributions with uniform primary censoring but this will be generalised in future releases. See #426.
Added user settable primary event priors to the latent model. See #474.
Added a marginalised likelihood to the latent model. See #474.
Added a weight argument to as_epidist_marginal_model() to allow for weighted data (for example count data) to be used in the marginal model. See #509.
Added a epidist_aggregate_data method to as_epidist_marginal_model() to allow straightforward use of the marginal model with aggregated data. See #510.
Added a epidist_aggregate_data method to as_epidist_latent_model() to allow straightforward use of the latent model with aggregated data. See #510.
Added a epidist_aggregate_data method to as_epidist_naive_model() to allow straightforward use of the naive model with aggregated data. See #510.
Updated the naive model to internally transform the data to be optimally aggregated as for the marginal model. See #510.

Package

Remove the default method for epidist(). See #473.
Added enforce_presence argument to epidist_prior() to allow for priors to be specified if they do not match existing parameters. See #474.
Added a merge argument to epidist_prior() to allow for not merging user and package priors. See #474.
Generalised the Stan reparametrisation feature to work across all distributions without manual specification by generating Stan code with brms and then extracting the reparameterisation. See #474.
Added a transform_data S3 method to allow for data to be transformed for specific models. This is specifically useful for the marginal model at the moment as it allows reducing the data to its unique strata. See #474.
Added new epidist_aggregate_data class to handle pre-aggregated line list data. See #510.
Added a as_epidist_aggregate_data() method for epidist_linelist_data objects to allow for easy conversion to aggregate data. See #510.
Added a as_epidist_linelist_data() method for epidist_aggregate_data objects to allow for easy conversion to linelist data. See #510.
Added an example dataset sierra_leone_ebola_data to the package. See #510.
Added examples to most functions to show usage of the package. See #510.
Added improved documentation explaining how the epidist_transform_data() methods work for the marginal and naive models. See #510.

Documentation

Brings the README into line with epinowcast standards. See #467.
Switched over to using the marginal model as default in the documentation. See #426.
Added a helper functions for new variables to avoid code duplication in vignettes. See #426.
Improved the Ebola case study vignette to use truncated data and to reduce the focus on exploratory data analysis. See #510.

Bugs

Switched to using a patched of primarycensored that doesn’t make use of size(). This fixes some Mac compilation edge cases. See #524.

epidist 0.1.0

This is the first minor release of epidist intended for early test users of the package. As some features may change, the package is marked as experimental. We expect to release a stable 1.0.0 version shortly.

The epidist package implements models for epidemiological delay distributions. It uses brms to perform Bayesian inference.

One data format is currently available:

The linelist data format

Two statistical models are currently available:

The naive model: which models the delay directly using brms
The latent model: which implements a latent variable model to correct for biases in the data

The package is readily extensible to additional models via an S3 class based system. In particular, model fitting with [epidist()] is possible using S3 classes for custom:

We provide functionality for post-processing. Alternatively, users may directly use tidybayes for specific families.

Three vignettes are available. There is also a frequently asked questions section.