durmodtools

Convenience tools and broom-style tidiers for the durmod package.

durmod (Simen Gaure) estimates a piecewise-constant mixed proportional hazard competing risk model with the NPMLE. The numerics are excellent but the surrounding ergonomics — building the formula, choosing control parameters, validating the data, extracting tidy results — are sparse, so most users end up writing the same boilerplate in every project.

durmodtools packages that boilerplate as a thin layer on top. It does not modify or fork durmod; it depends on it.

Installation

durmod is no longer maintained on CRAN, so install it directly from Simen Gaure's GitHub:

# install.packages("remotes")
remotes::install_github("sgaure/durmod")

durmod has a C++ backend (Rcpp + OpenMP), so you'll need a working C++ toolchain:

• Windows — install Rtools matching your R version.
• macOS — install Xcode command-line tools (xcode-select --install) and an OpenMP-enabled clang (the easiest path is via Homebrew: brew install libomp).
• Linux — most distributions already have everything.

Then install durmodtools:

remotes::install_github("forlop/durmodtools")

Quick tour

Before estimation

library(durmodtools)            # auto-loads durmod
data(durdata, package = "durmod")

# Build the formula declaratively (no need to remember ID() / D() / S() / C())
f <- build_formula(
  outcome     = "d",
  id          = "id",
  duration    = "duration",
  state       = "alpha + 1",
  covariates  = c("x1", "x2"),
  conditional = list(job = "alpha")
)

# Validate the design before launching a long mphcrm run
risksets <- list(unemp = c("job", "program"), onprogram = "job")
validate_design(f, durdata, risksets = risksets)

# Bundled control-parameter presets
ctrl <- control_preset("default", threads = 4)
# or: control_preset("fast")     # 3 iters, no Fisher — for testing the spec
# or: control_preset("careful")  # 50 iters + numerical Hessian — for inference

# Configurable logging callback (replaces hand-written ones)
ctrl$callback <- logging_callback(show = c("default", "distribution"))

fit <- mphcrm(f, data = durdata, risksets = risksets, control = ctrl)

After estimation

# Information criteria across iterations
model_info(fit)

# Pick the best iteration by AIC, BIC, or logLik
best <- select_iter(fit, by = "AIC")

# Tidy coefficient table (broom / modelsummary / gt compatible)
tidy(best, conf.int = TRUE)

# One-row fit summary
glance(best)

# Wald confidence intervals
confint(best, level = 0.95)

# Per-level factor estimates with SE/CI (no more grep + geninv + sqrt(diag(...)))
factor_effects(best, transition = "job", factor_name = "x1")

Stability check across seeds

res <- mphcrm_robust(f, data = durdata, risksets = risksets,
                     n_runs = 5,
                     control = control_preset("default", threads = 4))
res$summary    # AIC / BIC / npoints across the 5 fits

What's in the box

Pre-estimation

function	what it does
build_formula(...)	declarative constructor for mphcrm's special-syntax formula
validate_design(...)	pre-flight check of formula, data, and risksets
control_preset(...)	mphcrm.control() presets: "fast" / "default" / "careful"
logging_callback(...)	configurable progress + intermediate-save callback
mphcrm_robust(...)	run mphcrm N times with different seeds; return all + summary

Post-estimation

function	what it does
model_info(fit)	tibble of logLik / AIC / BIC / df / nobs across all iterations
select_iter(fit, by = ...)	pick one iteration by AIC, BIC, logLik, or npoints
tidy(opt, conf.int = TRUE)	broom-style coefficient table
glance(opt)	broom-style one-row fit summary
confint(opt, level = 0.95)	Wald confidence intervals for all coefficients
factor_effects(opt, tr, f)	per-level estimates with SE/CI for one (transition × factor)
predict(opt, newdata)	placeholder — interface defined, implementation pending

Vignettes

• vignette("pre-estimation", package = "durmodtools") — formula building, validation, control presets.
• vignette("extracting-results", package = "durmodtools") — accessors, tidiers, factor effects.
• vignette("mphcrm-output-reference", package = "durmodtools") — what a fitted mphcrm object carries: every metric, the source-of-truth R formula behind it, and how to interpret it.

Relationship to durmod

durmodtools does not modify durmod. It depends on it (pinned in DESCRIPTION to a tested minimum version) and registers S3 methods on durmod's mphcrm.opt and mphcrm.list classes. If durmod is updated upstream, durmodtools should continue to work provided the output structure is unchanged; the test suite is designed to catch silent breakage.

License

Artistic-2.0, matching durmod. Original durmod © Simen Gaure. durmodtools is a companion package, written with the goal that its contents could be folded into durmod upstream at any time.