Package {mzimp}

mzimp: Zero-Inflated Mixed Poisson Models

Description

Fits and selects zero-inflated mixed Poisson models for multivariate count regression. Supports multiple mixing distributions and variance specifications for flexible marginal modeling, while allowing dependence in the structural-zero components and latent mixing intensities to be modeled separately through copulas. The methodology is described in Nguyen (2026) doi:10.48550/arXiv.2608.12732.

Author(s)

Maintainer: Quang Huy Nguyen huynqtkt@neu.edu.vn [copyright holder]

Fit a Bivariate Mixed Poisson or Zero-Inflated Mixed Poisson Model

Description

Fits a bivariate Mixed Poisson (MP) or Zero-Inflated Mixed Poisson (ZIMP) regression model using inference functions for margins (IFM).

The two marginal models are first fitted separately using fit_zimp(). Dependence between their latent mixing variables is then modeled by a specified bivariate copula. When both margins are zero-inflated, dependence between the Bernoulli activation indicators may also be modeled.

Usage

fit_bizimp(
  formula1 = NULL,
  formula2 = NULL,
  data = NULL,
  Y = NULL,
  X = NULL,
  Z = NULL,
  mixing.distribution.1 = "gamma",
  mixing.distribution.2 = "gamma",
  var.power = NULL,
  mixing.copula = 1L,
  zero.dependence = c("gaussian", "comonotone", "countermonotone"),
  d0 = 10L,
  n_gh = 20,
  gh = NULL,
  ini.par = NULL,
  progress = FALSE,
  trace = 0,
  eval.max = 100,
  rel.tol = 1e-05,
  x.tol = 1e-05,
  iter.max = 100,
  gaussian.sign.tol = 1e-06,
  t.df.grid = c(3, 5, 10, 30)
)

Arguments

Details

fit_bizimp() fits one fully specified bivariate model. It does not perform model selection over marginal distributions or dependence structures.

Let the two observed counts be

Y_j = B_j N_j,\qquad j=1,2,

where N_j follows a Mixed Poisson model and B_j is an optional Bernoulli activation indicator. For an MP margin, B_j = 1 identically.

Conditional on the latent intensity,

N_j \mid \Lambda_j \sim \mathrm{Poisson}(\Lambda_j),

with marginal mean

\mu_{ij} = \exp(x_{ij}^{\top}\beta_j).

For a zero-inflated margin,

\Pr(B_{ij}=1)=\pi_{ij},\qquad \mathrm{logit}(\pi_{ij})=z_{ij}^{\top}\gamma_j.

Thus, B_{ij}=1 indicates that observation i belongs to the count-generating component of margin j.

Dependence between \Lambda_1 and \Lambda_2 is represented by the copula specified by mixing.copula. The copula is evaluated through a checkerboard approximation with grid size d0.

When both margins are zero-inflated, dependence between B_1 and B_2 is specified separately through zero.dependence.

Value

An object of class "bizimp_fit" containing the fitted marginal models, dependence model, likelihood, and information criteria. Important components include:

fit_margin_1, fit_margin_2

Fitted univariate marginal models returned by fit_zimp().

marginal_1, marginal_2

Internal representations of the fitted marginal models used in the dependence likelihood.

best_model

Fitted dependence model. It contains the requested mixing-copula family and estimated copula parameter(s), the zero-dependence parameter when applicable, the maximized bivariate log-likelihood, convergence information, and the full-model AIC and BIC.

mixing.distribution.1, mixing.distribution.2

Mixing distributions used for margins 1 and 2.

var.power

User-supplied marginal variance-power specification, or NULL when native/default parameterizations were used.

mixing.copula

VineCopula family code used for latent mixing dependence.

zero.dependence

Zero-dependence specification when both margins are zero-inflated; otherwise NA.

zero.copula.rho

Estimated Gaussian zero-copula correlation when both margins are zero-inflated and zero.dependence = "gaussian"; otherwise NA.

zero_inflation

Logical vector of length two indicating whether each margin is zero-inflated.

npar

Total number of parameters used to calculate the full-model information criteria.

AIC, BIC

Akaike and Bayesian information criteria for the complete bivariate model.

interface

Character string identifying whether the formula or direct matrix interface was used.

formula1, formula2

Marginal formulas when the formula interface was used.

call

Matched call to fit_bizimp().

Estimation

Estimation follows the IFM principle:

1. Fit margin 1 using fit_zimp().

2. Fit margin 2 independently using fit_zimp().

3. Conditional on the fitted marginal parameters, estimate the dependence parameter(s) from the bivariate likelihood.

Consequently, the marginal estimates returned in fit_margin_1 and fit_margin_2 are the corresponding independent fit_zimp() estimates and are not re-estimated during dependence fitting.

fit_bizimp() fits exactly one requested mixing-copula family. When the requested family is not Gaussian, a Gaussian dependence model may be fitted internally as a pilot for initialization and dependence-sign information; it is not an additional candidate model.

Marginal configurations

Each margin may independently be an MP or ZIMP model. The four supported configurations are therefore:

• MP–MP;

• ZIMP–MP;

• MP–ZIMP;

• ZIMP–ZIMP.

For direct input, Z = NULL specifies MP–MP. A list such as Z = list(Z1, NULL) specifies ZIMP–MP, while Z = list(NULL, Z2) specifies MP–ZIMP. Supplying a shared matrix Z applies the zero-component design to both margins.

Dependence between the Bernoulli activation indicators is estimated only for the ZIMP–ZIMP configuration.

Model specification

Formula interface

Two formulas are supplied, one for each margin. For example,

formula1 = y1 ~ 1 + x1 + x2
formula2 = y2 ~ 1 + x1

specifies two MP margins. Zero-inflated margins are specified using two-part formulas:

formula1 = y1 ~ 1 + x1 + x2 | z1 + z2
formula2 = y2 ~ 1 + x1      | z1

The expression before | specifies the count component and the expression after | specifies the zero-inflation component.

As in fit_zimp(), intercepts are included only when explicitly requested with 1. This convention applies separately to the count and zero-inflation components.

Matrix interface

Y may be an n \times 2 matrix/data frame or a list containing two response vectors.

X may be a common design matrix for both margins or a list list(X1, X2) containing margin-specific count designs. Similarly, Z may be a common zero-component design or a list containing margin-specific designs. Intercept columns must be included explicitly when required.

Mixing distributions

The mixing distribution is specified separately for each margin through mixing.distribution.1 and mixing.distribution.2. Supported distributions are:

"gamma"

Gamma mixing distribution.

"lnorm"

Lognormal mixing distribution.

"ig"

Inverse Gaussian mixing distribution.

"weibull"

Weibull mixing distribution.

"loglogistic"

Log-logistic mixing distribution.

"betaprime"

Beta-prime mixing distribution.

The two margins may use different mixing distributions.

Dependence structure

mixing.copula is a single integer family code supported by VineCopula. It specifies the copula connecting the two latent mixing variables. The default mixing.copula = 1 corresponds to the Gaussian copula.

For mixing.copula = 2, corresponding to the Student-t copula, the degrees of freedom are selected over the values supplied in t.df.grid.

When both margins are zero-inflated, zero.dependence specifies the dependence structure between their Bernoulli activation indicators:

"gaussian"

Gaussian copula dependence with an estimated correlation parameter.

"comonotone"

Upper Fréchet-bound dependence.

"countermonotone"

Lower Fréchet-bound dependence.

See Also

fit_zimp() for fitting a univariate marginal model and select_bizimp() for selecting among multiple bivariate model specifications.

Examples

## Generate a small bivariate Mixed Poisson data set
set.seed(12345)
n <- 50
x <- rnorm(n)
X <- cbind(1, x)

mu1 <- exp(X %*% c(0.4, 0.3))
mu2 <- exp(X %*% c(0.2, -0.2))

## Gamma-mixed Poisson margins
y1 <- rnbinom(n, size = 1 / 0.2, mu = mu1)
y2 <- rnbinom(n, size = 1 / 0.3, mu = mu2)
Y <- cbind(y1, y2)

## MP--MP model with a Gaussian mixing copula
fit <- fit_bizimp(
  Y = Y,
  X = X,
  mixing.distribution.1 = "gamma",
  mixing.distribution.2 = "gamma",
  var.power = 2,
  mixing.copula = 1L,
  progress = FALSE
)

fit$best_model
fit$AIC
fit$BIC

## Margin-specific designs and a ZIMP--MP configuration
z1 <- rnorm(n)
Z1 <- cbind(z1)

fit_zm <- fit_bizimp(
  Y = Y,
  X = list(X, X),
  Z = list(Z1, NULL),
  mixing.distribution.1 = "gamma",
  mixing.distribution.2 = "gamma",
  var.power = c(2, 2),
  mixing.copula = 1L,
  progress = FALSE
)

fit_zm$zero_inflation

Fit a Multivariate Zero-Inflated Mixed Poisson Model

Description

Fits a multivariate Mixed Poisson (MMP) or Zero-Inflated Mixed Poisson (MZIMP) model with factor-copula dependence.

Marginal models may either be fitted directly from Y, X, and optionally Z, or supplied through an object returned by select_multi_margin(). After the marginal models have been fitted or supplied, fit_multizimp() estimates dependence among the latent mixed-Poisson components and, when requested, dependence among structural-zero indicators.

The marginal and dependence parameters are estimated using an inference-functions-for-margins (IFM) strategy.

Usage

fit_multizimp(
  Y = NULL,
  X = NULL,
  Z = NULL,
  margin = NULL,
  mixing.distribution = "gamma",
  var.power = NULL,
  margin.ini.par = NULL,
  parallel = NULL,
  margin.block_size = NULL,
  copula.block_size = 500L,
  d0 = 10,
  J = 1L,
  L = 0L,
  latent.dependence = c("gaussian", "biexp"),
  latent.factor.skewness = 0.5,
  latent.loglike = NULL,
  gh_order = 16L,
  approx.hermite.joint.order = 4L,
  hermite.polar.cycles = 6L,
  hermite.polar.v.iter = 8L,
  hermite.polar.r.iter = 4L,
  hermite.polar.tol = 1e-05,
  hermite.polar.r.max = 0.995,
  genz.conditional.Q = 256L,
  genz.conditional.antithetic = TRUE,
  qmc.n_rep = 4L,
  qmc.seed = 1L,
  zero.dependence = c("gaussian", "biexp"),
  zero.factor.skewness = 0.5,
  biexp.root_tol = 1e-10,
  psi0_start = NULL,
  rho0_start = NULL,
  gh.progressive = TRUE,
  use.gradient = TRUE,
  progress = TRUE,
  trace = 0,
  eval.max = 1000L,
  iter.max = 100L,
  tiny = 1e-300,
  big_penalty = 1e+12,
  rel.tol = 1e-04
)

Arguments

Details

For margin j=1,\ldots,k, the observed response is represented as

Y_j = B_j N_j,

where N_j follows a mixed Poisson model and B_j is an activation indicator when zero inflation is present. For an ordinary mixed Poisson margin, B_j=1 with probability one.

The model separates two sources of cross-margin dependence:

1. dependence among the latent mixed-Poisson components underlying N_1,\ldots,N_k; and

2. dependence among the activation indicators for zero-inflated margins.

These two dependence structures are modeled separately through the count layer and the structural-zero layer.

Value

An object of class "multi_zipm_fit" containing the fitted marginal models and dependence structure. Important components include:

method

Estimation method. Currently "IFM".

n, k

Number of observations and response margins.

marginal

List of length k containing the fitted marginal information. Each element contains the margin index, mixing distribution, variance power, zero-inflation status, count coefficients (be), dispersion parameter (phi), zero-component coefficients (gam) when present, and marginal log-likelihood.

marginal.dist

Mixing distribution used for each response margin.

zero_inflation

Logical vector indicating which margins contain a zero-inflation component.

count_layer

Fitted dependence structure for the mixed-Poisson count layer. Important components include dependence, J, psi, A, implied_correlation, and information on the likelihood method. The loading matrix psi has dimension k by J.

zero_layer

Structural-zero dependence information. When zero-layer dependence is estimated, this includes the fitted loading matrix rho and its implied correlation matrix. The component also records cases in which no zero-layer dependence is estimated.

copula_fit

Dependence-fitting results, including the fitted log-likelihood, optimization objective, convergence information, elapsed time, optimizer output, and available final likelihood-evaluation diagnostics.

summary

Ordinary list containing the principal user-facing results, including dimensions, marginal-model summary, count- and zero-layer summaries, log-likelihood, convergence information, elapsed time, and final likelihood diagnostics.

call

Matched function call.

Marginal models

Marginal models can be supplied in two ways.

Direct specification.

If margin = NULL, Y and X must be supplied. The function fits exactly one marginal model for each response. No marginal model selection is performed.

mixing.distribution specifies the mixing distribution and var.power specifies its variance-power parameterization.

Z = NULL gives ordinary mixed Poisson margins. Supplying Z gives zero-inflated mixed Poisson margins.

X and Z may be common matrices/data frames shared by all margins or lists containing margin-specific designs. When lists are supplied, their lengths must equal the number of response margins.

Previously selected margins.

Alternatively, margin may be an object returned by select_multi_margin(). In this case the selected marginal distributions, variance powers, regression coefficients, dispersion parameters, and zero-inflation specifications are reused without refitting or reselecting the marginal models.

If compatible checkerboard probability matrices were precomputed by select_multi_margin(), they are reused when possible.

A margin object may contain a mixture of ordinary MP and ZIMP margins.

Count dependence

Dependence among the latent mixed-Poisson components is represented by a J-factor copula. The number of factors must satisfy

1 \leq J < k.

The factor family is controlled by latent.dependence:

"gaussian"

Gaussian factor copula.

"biexp"

Factor copula based on a standardized bilateral-exponential distribution.

The fitted count-layer loading matrix is denoted by \Psi and has dimension k \times J. The parameterization guarantees

\sum_{r=1}^{J}\psi_{jr}^2 < 1

for every margin j, leaving a positive idiosyncratic variance.

For Gaussian factors, latent.loglike controls likelihood evaluation. "GH" uses Gauss–Hermite quadrature and "hermite" uses the Hermite-based approximation implemented by the package. If latent.loglike = NULL, the method is selected automatically.

Structural-zero dependence

Zero-inflated margins may additionally share a dependence structure through their activation indicators.

L = 0 specifies no factor dependence in the structural-zero layer. L = 1 requests a one-factor zero-layer dependence model when at least two zero-inflated margins are available.

The zero-layer factor family is specified by zero.dependence, with supported values "gaussian" and "biexp".

If no margin is zero inflated, or if only one margin is zero inflated, no cross-margin zero-layer dependence can be estimated. Ordinary MP margins do not participate in the structural-zero dependence model.

Thus a multivariate model may contain any combination of MP and ZIMP margins.

Checkerboard probabilities

Dependence estimation uses checkerboard interval probabilities derived from the fitted marginal distributions. The checkerboard resolution is controlled by d0.

These marginal quantities are computed before dependence optimization and do not depend on the factor-loading parameters. When a compatible cache is available in an object returned by select_multi_margin(), it is reused where possible.

Estimation

The model is fitted using an IFM strategy. Marginal parameters are estimated first, or taken from margin, and are subsequently held fixed while the dependence parameters are estimated.

For Gaussian count factors, gh.progressive = TRUE allows lower-dimensional fitted factor models to be used as starting values for higher-dimensional models.

parallel controls supported parallel computations. NULL or 1 requests serial computation; an integer greater than one requests the corresponding level of parallelism where supported.

See Also

select_multi_margin() for selecting marginal models before dependence fitting; fit_zimp() and select_zimp() for univariate models; fit_bizimp() for the specialized bivariate model.

Examples

set.seed(12345)

n <- 15
k <- 3
x <- rnorm(n)
X <- cbind(Intercept = 1, x = x)

beta <- c(0.5, 0.3)
mu <- exp(X %*% beta)

Y <- cbind(
  y1 = rpois(n, mu),
  y2 = rpois(n, mu),
  y3 = rpois(n, mu)
)

## Directly specified mixed-Poisson margins
fit <- fit_multizimp(
  Y = Y,
  X = X,
  Z = NULL,
  mixing.distribution = "gamma",
  var.power = 2,
  J = 1,
  L = 0,
  d0 = 3,
  parallel = 1
)

fit$summary
fit$count_layer$psi

## Alternatively, select the marginal models first
margins <- select_multi_margin(
  Y = Y,
  X = X,
  Z = NULL,
  mixing.distribution = c("gamma", "lnorm"),
  var.power = 2,
  compute.I = TRUE,
  I.d0 = 3,
  parallel = 1
)

fit2 <- fit_multizimp(
  margin = margins,
  J = 1,
  L = 0,
  d0 = 3,
  parallel = 1
)
fit2$summary

Fit a Univariate Mixed Poisson or Zero-Inflated Mixed Poisson Model

Description

Fits a single univariate Mixed Poisson (MP) or Zero-Inflated Mixed Poisson (ZIMP) regression model for a specified mixing distribution and variance-power parameterization.

Both formula and matrix interfaces are supported. The count component models the mean of the Poisson mixing distribution, while an optional zero-inflation component models the probability that the count-generating component is active.

Usage

fit_zimp(
  formula = NULL,
  data = NULL,
  Y = NULL,
  X = NULL,
  Z = NULL,
  offset.count = NULL,
  mixing.distribution = "gamma",
  var.power = NULL,
  n_gh = 20,
  gh = NULL,
  ini.par = NULL,
  progress = TRUE,
  trace = 0,
  eval.max = 100,
  rel.tol = 1e-05,
  x.tol = 1e-05,
  iter.max = 100
)

Arguments

Details

fit_zimp() fits one specified model. It does not perform model selection over mixing distributions or variance-power specifications. For model selection, see select_zimp().

The conditional count model is

Y \mid \Lambda \sim \mathrm{Poisson}(\Lambda),

where the distribution of the latent intensity \Lambda is determined by mixing.distribution. The marginal mean is linked to the count covariates through

\mu_i = \exp(x_i^\top \beta + o_i),

where o_i is an optional count offset.

For a zero-inflated model, let B_i denote the indicator that the count-generating component is active. The model uses

Y_i = B_i N_i,

with

\Pr(B_i = 1) = \pi_i,\qquad \mathrm{logit}(\pi_i) = z_i^\top \gamma.

Thus, Z models the probability of belonging to the count-generating component rather than the structural-zero component.

The dispersion parameter phi is constrained to be positive and is optimized internally on the log scale.

Value

An object of class "zimp_fit" containing the fitted model. Important components include:

be

Estimated regression coefficients \hat\beta for the count component.

phi

Estimated dispersion parameter on its natural scale.

gam

Estimated zero-component coefficients \hat\gamma for a ZIMP model.

log_lik_value

Maximized log-likelihood.

neg_log_lik_value

Negative log-likelihood evaluated at the fitted parameters.

AIC

Akaike information criterion.

BIC

Bayesian information criterion.

par

Estimated parameter vector on the internal optimization scale.

convergence

Optimizer convergence code.

message

Optimizer convergence message, when available.

hessian

Estimated Hessian matrix, when available.

vcov

Estimated covariance matrix, when available.

coef_table

Coefficient and inference table produced by the fitting engine.

count_coef_table

Coefficient table for the count component.

zero_coef_table

Coefficient table for the zero component for a ZIMP model.

phi_se

Standard error of the reported dispersion parameter, when available.

y, X, Z

Response and design matrices used for fitting.

offset.count

Count offset used in the fitted model.

mixing.distribution

Mixing distribution used in the model.

var.power

Variance-power specification used in the model.

zero_inflation

Logical indicating whether the fitted model contains a zero-inflation component.

nobs

Number of observations used in the fit.

npar

Number of estimated parameters.

call

Matched function call.

summary

Printable summary containing coefficient tables, dispersion, log-likelihood, AIC, BIC, and convergence information.

Additional components describing the formula, model frames, coefficient names, interface, and missing-data handling may be included when relevant.

Model specification

Two input interfaces are available.

Formula interface

A Mixed Poisson model can be specified as

y ~ x1 + x2

and a Zero-Inflated Mixed Poisson model as

y ~ x1 + x2 | z1 + z2

The expression before | specifies the count component and the expression after | specifies the zero-inflation component.

In the mzimp formula interface, intercepts are included only when explicitly requested with 1. For example,

y ~ x1 + x2

specifies a count model without an intercept, whereas

y ~ 1 + x1 + x2

includes an intercept. The same convention applies independently to the zero-inflation component.

Matrix interface

The response and design matrices can instead be supplied directly through Y, X, and optionally Z. If Z = NULL, an MP model is fitted. If Z is supplied, a ZIMP model is fitted. Intercept columns, when required, must be included explicitly in X and Z.

Mixing distributions

The following mixing distributions are supported:

"gamma"

Gamma mixing distribution.

"lnorm"

Lognormal mixing distribution.

"ig"

Inverse Gaussian mixing distribution.

"weibull"

Weibull mixing distribution.

"loglogistic"

Log-logistic mixing distribution.

"betaprime"

Beta-prime mixing distribution.

The optional var.power argument controls the variance-power parameterization. If var.power = NULL, the native/default parameterization of the selected mixing distribution is used. Otherwise, var.power must be one of 1, 2, or 3.

Initial values

Initial parameter values may be supplied through ini.par. Parameters are ordered internally as

c(beta, log(phi), gamma)

where beta contains the count-component coefficients, log(phi) is the log-dispersion parameter, and gamma contains the zero-component coefficients. The gamma block is omitted for an MP model.

See Also

select_zimp() for selecting among multiple univariate MP or ZIMP model specifications.

Examples

## Generate a small Mixed Poisson data set
set.seed(12345)
n <- 50
x1 <- rnorm(n)
x2 <- rnorm(n)
X <- cbind(1, x1, x2)
mu <- exp(X %*% c(0.5, 0.3, 0.7))

## Gamma-mixed Poisson data are Negative Binomial marginally.
## This provides a short reproducible example without relying on
## non-exported simulation functions.
phi <- 0.2
y <- rnbinom(n, size = 1 / phi, mu = mu)

dat <- data.frame(y = y, x1 = x1, x2 = x2)

## Formula interface
fit_formula <- fit_zimp(
  y ~ 1 + x1 + x2,
  data = dat,
  mixing.distribution = "gamma",
  var.power = 2,
  progress = FALSE
)

fit_formula

## Equivalent matrix interface
fit_matrix <- fit_zimp(
  Y = y,
  X = X,
  mixing.distribution = "gamma",
  var.power = 2,
  progress = FALSE
)

Print a select_bizimp summary

Description

Print a select_bizimp summary

Usage

## S3 method for class 'bizimp_selection_summary'
print(x, ...)

Arguments

Value

The summary object, invisibly.

Print a select_zimp summary

Description

Print a select_zimp summary

Usage

## S3 method for class 'zimp_selection_summary'
print(x, ...)

Arguments

Value

The summary object, invisibly.

Select a Bivariate Mixed Poisson or Zero-Inflated Mixed Poisson Model

Description

Performs model selection for bivariate Mixed Poisson (MP) and Zero-Inflated Mixed Poisson (ZIMP) regression models.

Marginal models are first selected independently using select_zimp(). The best margin.top models from each margin are retained, and dependence structures are then evaluated for every retained marginal-model pair. The final bivariate model is selected by the largest maximized observed-data log-likelihood.

Usage

select_bizimp(
  formula1 = NULL,
  formula2 = NULL,
  data = NULL,
  Y = NULL,
  X = NULL,
  Z = NULL,
  mixing.distribution = c("gamma", "lnorm", "ig", "weibull", "loglogistic", "betaprime"),
  var.power = NULL,
  margin.top = 1L,
  mixing.copula = c(1L),
  d0 = 10L,
  n_gh = 20,
  gh = NULL,
  ini.par = NULL,
  parallel = NULL,
  progress = FALSE,
  show.result = FALSE,
  trace = 0,
  eval.max = 100,
  rel.tol = 1e-05,
  x.tol = 1e-05,
  iter.max = 100,
  gaussian.sign.tol = 1e-06,
  t.df.grid = c(3, 5, 10, 30)
)

Arguments

Details

select_bizimp() extends select_zimp() from univariate marginal selection to bivariate model selection. In contrast to fit_bizimp(), which fits one specified bivariate model, select_bizimp() searches over marginal specifications and mixing-copula families.

The selection procedure consists of three stages:

1. Select candidate models independently for each margin using select_zimp().

2. Retain the best margin.top models from each margin and perform a Gaussian dependence prescreen for every retained marginal pair.

3. Fit the compatible user-requested mixing copulas, together with the admissible zero-dependence structures, and select the candidate with the largest maximized bivariate log-likelihood.

Value

An object of class "bizimp_selection". Important components include:

margin_selection_1, margin_selection_2

Results from the independent select_zimp() selections for margins 1 and 2.

margin_top

Requested maximum number of marginal models retained from each margin.

retained_margin_1, retained_margin_2

Lists of fitted marginal models retained for the dependence search.

pair_results

Detailed dependence-search results for every retained marginal pair, including the Gaussian prescreen, compatible mixing-copula families, zero-dependence candidates, and fitted dependence models.

candidate_table

Data frame containing all fitted bivariate candidates, with columns describing margin 1, margin 2, zero dependence, mixing dependence, maximized log-likelihood, AIC, and BIC. Rows are ordered by decreasing log-likelihood.

pair_summary

Data frame containing the best dependence model obtained for each retained marginal pair, ordered by decreasing bivariate log-likelihood.

best_pair

Detailed result for the retained marginal pair containing the globally selected bivariate model.

best_model

Selected dependence model. It contains the selected mixing-copula family and parameter estimates, zero-dependence parameter when applicable, maximized bivariate log-likelihood, convergence information, total parameter count, AIC, and BIC.

best_dependence

Complete dependence-engine result associated with best_model.

best_margin_1, best_margin_2

Fitted marginal models belonging to the globally selected bivariate model.

interface

Character string identifying whether the formula or direct matrix interface was used.

formula1, formula2

Marginal formulas when the formula interface was used.

mixing.copula.candidates

User-requested candidate mixing-copula family codes.

summary

Object of class "bizimp_selection_summary" containing the candidate table, selected candidate, pair summary, and matched call for formatted display.

call

Matched call to select_bizimp().

Model specification

Two input interfaces are supported.

Formula interface

Supply formula1, formula2, and data. Each formula follows the syntax used by select_zimp(). For example,

formula1 = y1 ~ 1 + x1 + x2 | z1 + z2
formula2 = y2 ~ 1 + x1 + x2

specifies a ZIMP first margin and an MP second margin.

The expression before | specifies the count component and the expression after | specifies the zero-inflation component. Intercepts are included only when explicitly requested with 1.

Both formulas must refer to the same observations. If missing values cause different observations to be retained for the two margins, bivariate model selection is stopped.

Matrix interface

Alternatively, supply Y, X, and optionally Z.

Y may be an n \times 2 matrix/data frame or a list containing two count vectors. X may be a design matrix shared by both margins or a list containing two margin-specific design matrices.

Z = NULL specifies two MP margins. A shared matrix/data frame Z specifies two ZIMP margins. A list may be used for mixed configurations; for example, Z = list(Z1, NULL) specifies ZIMP–MP and Z = list(NULL, Z2) specifies MP–ZIMP.

Marginal model selection

Each margin is selected independently using select_zimp() with the candidate mixing distributions supplied in mixing.distribution and the variance-power specifications supplied in var.power.

Supported mixing distributions are:

"gamma"

Gamma mixing distribution.

"lnorm"

Lognormal mixing distribution.

"ig"

Inverse Gaussian mixing distribution.

"weibull"

Weibull mixing distribution.

"loglogistic"

Log-logistic mixing distribution.

"betaprime"

Beta-prime mixing distribution.

If var.power = NULL, each distribution uses its native/default variance-power specification. Otherwise, every requested mixing distribution is crossed with the requested variance powers, following the candidate construction used by select_zimp().

Marginal candidates are ranked by decreasing maximized univariate log-likelihood. At most margin.top models are retained from each margin. If m_1 and m_2 models are retained from margins 1 and 2, respectively, dependence selection is performed for

m_1 m_2

marginal-model pairs.

Thus, margin.top = 1 restricts the dependence search to the independently best marginal model from each margin.

Mixing-copula candidates

Dependence between the two latent mixing variables is specified using bivariate VineCopula family codes. mixing.copula may contain one or more candidate families.

Supported family codes are:

0L

Independence copula.

1L

Gaussian copula.

2L

Student-t copula.

3L

Clayton copula.

4L

Gumbel copula.

5L

Frank copula.

6L

Joe copula.

7L

BB1 copula.

8L

BB6 copula.

9L

BB7 copula.

10L

BB8 copula.

13L, 14L, 16L–20L

180-degree rotations of the corresponding Archimedean and BB families.

23L, 24L, 26L–30L

90-degree rotations of the corresponding Archimedean and BB families.

33L, 34L, 36L–40L

270-degree rotations of the corresponding Archimedean and BB families.

104L, 114L, 124L, 134L, 204L, 214L, 224L, 234L

Supported Tawn families and their rotations.

The default is mixing.copula = 1L, corresponding to the Gaussian copula.

For example,

mixing.copula = 1L             # Gaussian only
mixing.copula = c(1L, 3L, 5L) # Gaussian, Clayton, and Frank

For the Student-t copula (2L), the correlation parameter is optimized for each degrees-of-freedom value supplied in t.df.grid, and the likelihood-maximizing specification is retained.

Gaussian dependence prescreen

For every retained marginal pair, a Gaussian checkerboard copula is fitted first to the latent mixing variables. This fit is used only as a prescreen for the sign of the mixing dependence.

If the absolute fitted Gaussian correlation exceeds gaussian.sign.tol, user-requested copula families incompatible with that dependence direction are removed before the full dependence search. If the fitted correlation is not finite or does not exceed the tolerance, no sign restriction is imposed.

If sign filtering would remove all user-requested families, all requested families are restored. Consequently, the prescreen never introduces a mixing-copula family that was not requested by the user.

The Gaussian family is fitted internally for this prescreen even when 1L is not included in mixing.copula, but it is eligible for final selection only when explicitly requested.

Zero dependence

When both retained margins are ZIMP models, dependence between their Bernoulli activation indicators is also selected. For every compatible mixing copula, the following three zero-dependence specifications are evaluated:

Gaussian

Gaussian copula dependence with an estimated correlation parameter.

Frechet upper

Comonotone dependence at the upper Frechet bound.

Frechet lower

Countermonotone dependence at the lower Frechet bound.

Thus, for a ZIMP–ZIMP marginal pair, the full dependence search crosses each compatible mixing-copula family with all three zero-dependence specifications.

If either margin is an MP model, the zero-dependence layer is inactive.

Final model selection

The complete candidate set consists of the fitted combinations of

\text{retained margin 1} \times \text{retained margin 2} \times \text{compatible mixing copula} \times \text{admissible zero dependence}.

Candidates are evaluated using the full bivariate observed-data likelihood. candidate_table is sorted by decreasing maximized log-likelihood, and its first row corresponds to the selected model.

AIC and BIC are also calculated for each candidate. Their parameter counts include both marginal models, the mixing-copula parameters, and the Gaussian zero-copula correlation parameter when applicable. AIC and BIC are reported for comparison but are not used for model selection.

Parallel computation

parallel is passed to select_zimp() and controls parallel fitting of candidate marginal models within each margin.

The two marginal selections themselves are performed sequentially. Dependence searches across retained marginal pairs, mixing copulas, and zero-dependence specifications are also performed sequentially by select_bizimp().

See Also

fit_bizimp() for fitting one specified bivariate model, select_zimp() for univariate marginal model selection, and fit_zimp() for fitting one specified univariate marginal model.

Examples

## Generate a small bivariate Mixed Poisson data set
set.seed(12345)
n <- 30
x1 <- rnorm(n)
x2 <- rnorm(n)
X <- cbind(1, x1, x2)

mu1 <- exp(X %*% c(0.5, 0.3, 0.7))
mu2 <- exp(X %*% c(0.3, 0.5, 0.4))

## Gamma-mixed Poisson margins can be generated from negative-binomial
## distributions for this short reproducible example.
y1 <- rnbinom(n, size = 1 / 0.2, mu = mu1)
y2 <- rnbinom(n, size = 1 / 0.3, mu = mu2)
Y <- cbind(y1, y2)

## Compare several mixing-copula families while retaining only the
## independently best marginal model from each margin.
sel <- select_bizimp(
  Y = Y,
  X = X,
  Z = NULL,
  mixing.distribution = "gamma",
  var.power = 2,
  margin.top = 1L,
  mixing.copula = c(1L, 3L, 5L)
)

## Candidate models are ordered by decreasing log-likelihood
sel$candidate_table

## Selected marginal and dependence models
sel$best_margin_1
sel$best_margin_2
sel$best_model

Select Marginal Models for Multivariate Count Data

Description

Fits and selects a univariate Mixed Poisson (MP) or Zero-Inflated Mixed Poisson (ZIMP) model independently for each margin of a multivariate count response.

Candidate models may differ in mixing distribution, variance-power specification, and, when requested, the presence or absence of zero inflation. The best model for each margin is selected using AIC or BIC.

select_multi_margin() is designed primarily as the marginal-selection stage preceding fit_multizimp(). It supports common or margin-specific count and zero-component designs, serial or parallel computation, and optional precomputation of the checkerboard interval-probability matrices required by the multivariate dependence model.

Usage

select_multi_margin(
  Y,
  X,
  Z = NULL,
  X.columns = NULL,
  Z.columns = NULL,
  check.zero = FALSE,
  selection.criterion = c("BIC", "AIC"),
  mixing.distribution = c("gamma", "lnorm", "ig", "weibull", "loglogistic", "betaprime"),
  var.power = NULL,
  benchmark.distribution = "lnorm",
  ini.par = NULL,
  parallel = NULL,
  block_size = NULL,
  compute.I = FALSE,
  I.d0 = 10L,
  progress = FALSE,
  show.result = FALSE,
  trace = 0,
  eval.max = 100,
  iter.max = 100,
  rel.tol = 1e-05,
  x.tol = 1e-05,
  n_gh = 20,
  gh = NULL,
  seed = NULL
)

Arguments

Details

Let

\mathbf{Y}=(Y_1,\ldots,Y_k)

denote an n \times k multivariate count response. The function fits a collection of candidate univariate models separately to each margin Y_j and selects the best marginal model according to selection.criterion.

Marginal selection does not estimate dependence between response margins. Dependence estimation is performed subsequently by fit_multizimp().

Value

An object of class "multi_margin_fit" containing the marginal-selection results and the information required by fit_multizimp(). The public components are:

Y

Validated n \times k response matrix.

X, Z

Count- and zero-component designs after internal conversion.

X_is_list, Z_is_list

Logical indicators identifying whether the corresponding designs are stored as margin-specific lists.

X_columns, Z_columns

Margin-specific column selections when common designs were supplied.

selected_margins

List of length k containing the selected marginal models. Each element has class "mzimp_margin_fit" and contains, among other quantities, the selected mixing distribution, variance power, zero-inflation status, count coefficients (be), dispersion parameter (phi), zero-component coefficients (gam) when applicable, and maximized marginal log-likelihood.

benchmark.distribution

Mixing distribution used as the benchmark.

benchmark_by_margin

Benchmark-fit information for individual response margins.

benchmark

Summary of the comparison between the selected models and the benchmark, including marginal and aggregate log-likelihood gains.

selected_summary

Data frame with one row per response margin describing the selected distribution, variance power, zero-inflation status, log-likelihood, AIC, BIC, benchmark log-likelihood, and gain relative to the benchmark.

I_precomputed

Logical indicating whether the checkerboard probability cache was computed.

I_d0

Checkerboard grid size used for the cache, or NULL when compute.I = FALSE.

I_y_list, I_0_list

Lists of checkerboard interval-probability matrices evaluated at observed counts and at zero, respectively, when compute.I = TRUE; otherwise NULL.

pb_list

List of Bernoulli activation-probability vectors when compute.I = TRUE; otherwise NULL.

y0_list

List of observed-zero indicators for ZIMP margins when compute.I = TRUE. Elements corresponding to MP margins are NULL.

mixing.distribution, var.power

Candidate marginal specifications requested by the user.

candidate_grid

Data frame describing the candidate marginal specifications participating in model selection.

check.zero

Logical indicating whether MP versus intercept-only ZIMP comparison was requested when Z = NULL.

selection.criterion

Information criterion used for marginal model selection.

n, k

Number of observations and response margins.

response_names

Names of the response margins.

parallel, n_workers

Requested parallel setting and the actual number of workers used.

performance

Timing and, when available, memory-use information for marginal fitting and optional checkerboard-cache computation.

summary

A regular list containing selected_summary, benchmark information, selection criterion, response dimensions and names, and performance information.

Response and design specification

Y must be an n \times k numeric matrix or data frame containing finite, non-negative integer counts. Each column represents one response margin.

The count-component design X can be supplied either as:

• one matrix or data frame shared by all margins; or

• a list of length k containing a separate design for each margin.

When a common X is supplied, X.columns may be used to specify a different subset of columns for each margin.

The zero-component design Z follows the same convention. It may be NULL, a common matrix/data frame, or a list of length k.

No intercept is added automatically to either X or Z. Include an intercept column explicitly when required.

Data-frame designs may contain numeric, integer, logical, factor, and character variables. Character variables are converted to factors and categorical variables are expanded internally into numeric design columns.

MP and ZIMP margins

The interpretation of Z determines whether a margin is fitted as an ordinary MP model or as a ZIMP model.

Z = NULL, check.zero = FALSE

All margins are fitted as ordinary MP models.

Common matrix/data frame Z

All margins are fitted as ZIMP models using the supplied zero-component design.

Z supplied as a list

Each margin is specified independently. Z[[j]] = NULL gives an MP margin, whereas a non-NULL Z[[j]] gives a ZIMP margin.

Z = NULL, check.zero = TRUE

For every marginal distribution and variance-power specification, both an MP model and an intercept-only ZIMP model are considered. The presence or absence of zero inflation is therefore selected independently for each margin.

If Z is explicitly supplied, check.zero is ignored.

For a ZIMP margin, the zero-component regression models the Bernoulli activation probability. If B_i denotes the activation indicator,

\Pr(B_i=1)=\operatorname{logit}^{-1}(z_i^\top\gamma).

Thus, B_i=1 corresponds to the count-generating component and B_i=0 to a structural zero.

Candidate marginal models

The following mixing distributions are supported:

"gamma"

Gamma mixing distribution.

"lnorm"

Lognormal mixing distribution.

"ig"

Inverse-Gaussian mixing distribution.

"weibull"

Weibull mixing distribution.

"loglogistic"

Log-logistic mixing distribution.

"betaprime"

Beta-prime mixing distribution.

If var.power is supplied, every requested mixing distribution is crossed with every requested variance power. For example,

mixing.distribution = c("gamma", "lnorm")
var.power = c(1, 2, 3)

defines six distribution–variance-power specifications before any MP/ZIMP expansion.

If var.power = NULL, each distribution uses its native/default variance-power parameterization:

Gamma

2

Lognormal

2

Inverse Gaussian

3

Weibull

2

Log-logistic

2

Beta-prime

2

The complete candidate specifications used for marginal selection are returned in candidate_grid.

Marginal model selection

Candidate models are fitted independently for each response margin. The selected model minimizes the information criterion specified by selection.criterion.

For selection.criterion = "BIC",

\mathrm{BIC}=-2\ell+\log(n)q,

whereas for selection.criterion = "AIC",

\mathrm{AIC}=-2\ell+2q,

where \ell is the maximized marginal log-likelihood and q is the number of estimated marginal parameters.

Selection is performed separately for every margin. Consequently, different response margins may select different mixing distributions, variance powers, and, when check.zero = TRUE, different MP/ZIMP specifications.

Candidate fitting uses the same initialization strategy as select_zimp(). Gamma, lognormal, and inverse-Gaussian models use distribution-specific initialization. Weibull, log-logistic, and beta-prime models use fitted Gamma models with the same effective variance power to initialize their regression coefficients, while retaining distribution-specific initialization for dispersion.

Gamma models required only for initialization are internal fits and do not participate in final model selection unless the corresponding Gamma specification was explicitly requested.

Benchmark model

benchmark.distribution specifies a reference marginal distribution used to compare the selected marginal models. Supported benchmark distributions are "lnorm", "gamma", and "ig".

The benchmark is fitted using its native/default variance-power parameterization under the corresponding zero-inflation specification. It is used only for reporting and does not affect marginal model selection.

The returned object contains the benchmark log-likelihood for each margin and the log-likelihood gain of the selected model relative to the benchmark.

Parallel computation

Candidate marginal models can be fitted in parallel.

If parallel = NULL or parallel = 1, fitting is serial. If parallel is an integer greater than one, a PSOCK cluster is created and marginal-model fitting jobs are distributed across workers.

Parallelization is across independent margin–candidate fitting jobs. The number of workers actually used is bounded by the number of available fitting tasks.

block_size optionally controls how fitting jobs are grouped for load-balanced parallel execution. If block_size = NULL, an appropriate block size is selected automatically.

Serial and parallel execution use the same model-selection procedure and return the same public result structure.

Checkerboard probability cache

If compute.I = TRUE, checkerboard interval probabilities are precomputed for each selected marginal model for subsequent use by fit_multizimp().

For each margin j, the cache contains:

I_y_list[[j]]

An n \times d matrix of interval probabilities evaluated at the observed counts, where d is I.d0.

I_0_list[[j]]

An n \times d matrix of interval probabilities evaluated at zero.

pb_list[[j]]

Bernoulli activation probabilities. For an MP margin this is a vector of ones. For a ZIMP margin it is \operatorname{logit}^{-1}(Z_j\gamma_j).

y0_list[[j]]

Logical indicator of observed zeros for a ZIMP margin. It is NULL for an MP margin.

The checkerboard resolution d is controlled by I.d0.

If compute.I = FALSE, no cache is created and I_d0, I_y_list, I_0_list, pb_list, and y0_list are returned as NULL.

See Also

fit_multizimp() for fitting the multivariate dependence model after marginal selection, select_zimp() for selecting a univariate marginal model, and fit_zimp() for fitting one specified univariate model.

Examples

set.seed(12345)

n <- 30
k <- 3
x1 <- rnorm(n)
x2 <- rnorm(n)

X <- cbind(Intercept = 1, x1 = x1, x2 = x2)

Y <- cbind(
  y1 = rpois(n, exp(0.5 + 0.3 * x1 + 0.2 * x2)),
  y2 = rpois(n, exp(0.3 - 0.2 * x1 + 0.3 * x2)),
  y3 = rpois(n, exp(0.4 + 0.1 * x1 - 0.2 * x2))
)

## Select among Gamma and lognormal marginal models
fit <- select_multi_margin(
  Y = Y,
  X = X,
  Z = NULL,
  mixing.distribution = c("gamma", "lnorm"),
  var.power = 2,
  selection.criterion = "BIC",
  parallel = 1
)

## One selected model per response margin
fit$selected_summary

## Selected marginal model objects
fit$selected_margins

Select a Univariate Mixed Poisson or Zero-Inflated Mixed Poisson Model

Description

Fits and compares multiple univariate Mixed Poisson (MP) or Zero-Inflated Mixed Poisson (ZIMP) regression models across candidate mixing distributions and variance-power specifications. Candidate models are ranked by maximized log-likelihood, and the model with the largest log-likelihood is selected.

Usage

select_zimp(
  formula = NULL,
  data = NULL,
  Y = NULL,
  X = NULL,
  Z = NULL,
  offset.count = NULL,
  mixing.distribution = c("gamma", "lnorm", "ig", "weibull", "loglogistic", "betaprime"),
  var.power = NULL,
  n_gh = 20,
  gh = NULL,
  ini.par = NULL,
  parallel = NULL,
  progress = TRUE,
  show.result = TRUE,
  trace = 0,
  eval.max = 100,
  rel.tol = 1e-05,
  x.tol = 1e-05,
  iter.max = 100,
  .return.all = FALSE
)

Arguments

Details

select_zimp() is the model-selection counterpart of fit_zimp(). While fit_zimp() fits one fully specified model, select_zimp() fits a collection of candidate models and returns the best-fitting specification.

The function supports the same formula and matrix interfaces as fit_zimp().

Value

An object of class "zimp_selection". By default, the object contains:

summary_table

A data frame containing one row for each requested candidate model, ordered by decreasing maximized log-likelihood. Columns include distribution, var.power, log_lik, neg_log_lik, and phi.

best_model

The selected model, returned as an object of class "zimp_fit". It corresponds to the first row of summary_table and contains full post-estimation inference.

summary

An object of class "zimp_selection_summary" containing the candidate table and selected model for formatted display.

call

The matched call to select_zimp().

If .return.all = TRUE, an additional internal component all_results containing the candidate fit objects is returned.

Model specification

Formula interface

A Mixed Poisson model can be specified as

y ~ x1 + x2

and a Zero-Inflated Mixed Poisson model as

y ~ x1 + x2 | z1 + z2

The expression before | specifies the count component and the expression after | specifies the zero-inflation component.

In the mzimp formula interface, intercepts are included only when explicitly requested with 1. Thus,

y ~ x1 + x2

specifies a model without an intercept, whereas

y ~ 1 + x1 + x2

includes an intercept. The same convention applies independently to the zero-inflation component.

Matrix interface

The response and design matrices may instead be supplied through Y, X, and optionally Z. If Z = NULL, MP candidate models are fitted. If Z is supplied, ZIMP candidate models are fitted. Intercept columns, when required, must be included explicitly in X and Z.

Candidate models

Candidate models are generated from the requested mixing.distribution and var.power values.

The supported mixing distributions are "gamma", "lnorm", "ig", "weibull", "loglogistic", and "betaprime".

If var.power = NULL, one model is fitted for each requested mixing distribution using its native/default variance-power specification:

"gamma"

var.power = 2

"lnorm"

var.power = 2

"ig"

var.power = 3

"weibull"

var.power = 2

"loglogistic"

var.power = 2

"betaprime"

var.power = 2

If one or more values are supplied to var.power, every requested mixing distribution is fitted for each requested variance power. For example,

mixing.distribution = c("gamma", "lnorm")
var.power = c(1, 2, 3)

defines six candidate models.

Duplicate distribution and variance-power specifications are removed before fitting.

Model fitting and selection

Candidate models are fitted by maximum likelihood using the same distribution-specific likelihood and gradient routines as fit_zimp().

Gamma, lognormal, and inverse-Gaussian candidates are fitted directly. Weibull, log-logistic, and beta-prime candidates use matched Gamma models as initialization anchors. For these candidates, the count coefficients and, for ZIMP models, the zero-component coefficients are initialized from a Gamma model with the same effective variance power. The dispersion parameter is initialized using the target distribution's own initialization procedure.

When a required Gamma anchor is not among the requested candidate models, it is fitted internally for initialization only and is not included in the returned candidate table.

Candidate models are initially fitted without post-estimation inference. They are then ordered by decreasing maximized log-likelihood. The first row of summary_table therefore corresponds to the selected model.

Full post-estimation inference is subsequently computed only for the selected model, at its already optimized parameter estimates; the selected model is not re-optimized. The resulting best_model is an object of class "zimp_fit".

Selection is based exclusively on maximized log-likelihood. AIC and BIC are available for the selected model through best_model, but are not used to rank candidate models.

Parallel computation

Candidate models within the same fitting stage may be evaluated in parallel by setting parallel to an integer greater than one.

Parallelization is performed across candidate models; optimization of an individual candidate is not itself parallelized by select_zimp(). Dependencies between fitting stages, including construction of Gamma initialization anchors, are preserved.

On Unix-like systems, fork-based parallel processing may be used. On Windows, PSOCK workers are used.

See Also

fit_zimp() for fitting one specified univariate MP or ZIMP model.

Examples

## Generate a small Mixed Poisson data set
set.seed(12345)
n <- 50
x1 <- rnorm(n)
x2 <- rnorm(n)
X <- cbind(1, x1, x2)
mu <- exp(X %*% c(0.5, 0.3, 0.7))

## Gamma-mixed Poisson data
phi <- 0.2
y <- rnbinom(n, size = 1 / phi, mu = mu)

dat <- data.frame(y = y, x1 = x1, x2 = x2)

## Compare candidate mixing distributions
sel <- select_zimp(
  y ~ 1 + x1 + x2,
  data = dat,
  mixing.distribution = c("gamma", "lnorm", "ig"),
  progress = FALSE
)

## Ranked candidate models
sel$summary_table

## Selected model
sel$best_model

## Compare variance-power specifications
sel_power <- select_zimp(
  Y = y,
  X = X,
  mixing.distribution = c("gamma", "lnorm"),
  var.power = c(1, 2, 3),
  progress = FALSE
)

sel_power$summary_table