zipoisson: Zero-inflated Poisson family for native Laplace models

View source: R/FUN_glmm_families.R

zipoissonR Documentation

Zero-inflated Poisson family for native Laplace models

Description

Constructs a zero-inflated Poisson family with a Poisson log-mean component and a separate fixed-effects logit component for structural zeros. An intercept-only probability is estimated by default by native Laplace fitting.

Usage

zipoisson(pi = 0.1, link = "log", zero = ~1)

Arguments

pi

Starting structural-zero probability, a finite value strictly between zero and one.

link

Link for the Poisson count mean; currently only "log" is supported.

zero

One-sided fixed-effects formula for the logit structural-zero probability. Defaults to ~1, retaining the scalar intercept-only interface. Covariates, factors, transformations and offsets are supported; random effects are not. A nontrivial formula requires data and estimate=TRUE.

Details

For observation i, let \mu_i = \exp(\eta_i) be the Poisson count mean and \pi_i the structural-zero probability. The response is zero with probability \pi_i + (1-\pi_i)\exp(-\mu_i), and a positive count y_i has probability (1-\pi_i)\operatorname{Poisson}(y_i;\mu_i).

For zero=~1, use familyParameters = list(zeroInflation = value, estimate = FALSE) in likelihood to fix its probability. For a nontrivial formula, \operatorname{logit}(\pi_i)=w_i'\gamma+a_i, with fixed design row w_i and optional zero-component offset a_i. The intercept starts at qlogis(pi) and other coefficients at zero; without an intercept, all coefficients start at zero. Formula coefficients are jointly estimated and reported in fit$zeroCoefficients on the logit scale, not as one scalar probability in familyParameters. summary(fit) adds post-fit joint-curvature standard errors. Zero-only missing covariates use naMethodX; in familym() they affect only their ZIP trait's rows. The family is currently supported only by native Laplace ML fitting, including as a named component of familym. PQL fitting is not supported.

The link-scale predictor describes the underlying Poisson count mean. Fitted and response-scale predicted values are the mixture mean (1-\pi_i)\mu_i. Simulation draws the structural-zero indicator and then, when not structurally zero, a Poisson count. Randomized quantile residuals use the zero-inflated response CDF.

Random effects in the zero component, hurdle models and zero-inflated NB2 are not supported. Sparse data and probabilities near zero or one can yield weak identification, unavailable uncertainty or numerical-boundary errors. Convergence alone does not establish a structural-zero mechanism. The fixed-effects formula is a bounded extension: development calibration supported moderate-rate interior settings but did not meet its registered convergence, coefficient-bias and uncertainty-availability criteria in a low-rate setting. It is not a universally validated mixture-inference method.

Value

An object inheriting from "family".

See Also

mmes, familym, nbinom2

Examples

zipoisson(pi = 0.2)
zipoisson(pi = 0.2, zero = ~ treatment)

sommer documentation built on Nov. 14, 2026, 5:07 p.m.