First release.
pcdreg() fits the semiparametric proportional rate model for panel count
data with time-varying covariates, using the EM algorithm of Sun et al. (2024).
pcdreg(model = "mean") fits the proportional means model by the estimating equation of
Hu, Sun and Wei (2003). It is the comparator used in the application of Sun et
al. (2024), and is included so that the two model families can be compared on
the same data. Note that the two are not reparametrisations of each other when
covariates vary over time, and that the fitted mean function is not
constrained to increase.
Covariance estimation for the rate model follows the paper directly. vcov()
supports type = "robust" (the sandwich estimator, consistent without the
Poisson assumption), type = "information" (the efficient information
estimator), and type = "profile" (the Murphy--van der Vaart profile
likelihood estimator, computed only when requested via
pcdreg(profile = TRUE)). The means model has no likelihood, so only the
sandwich estimator is available for it.
Model specification uses pcd() on the left hand side of the formula,
following the Surv() convention: pcd(id, time, count) for
examination-time rows and pcd(id, tstart, tstop, count) for
counting-process rows carrying time-varying covariates.
Both fitters return an object inheriting from "pcdfit", so coef(),
vcov(), confint(), summary(), baseline(), predict() and plot()
work the same way for either.
tidy(), glance() and augment() methods are provided for the broom
generics. tidy() takes the same type argument as vcov(), so the tidy
path reports the same standard errors as summary(). augment() adds the
fitted mean, the observed cumulative count and their difference to each row.
sim_pcd() simulates from the design of the paper's simulation study.
autoplot() methods return ggplot objects for the data, drawn as one tile
per examination interval shaded by its count, and for the fitted baseline.
plot() draws them and returns the object invisibly.
baseline(), predict(type = "mean") and sim_pcd() return tibbles.
Errors, warnings and messages use cli and carry condition classes, so a
particular failure can be caught with tryCatch() rather than by matching on
message text.