simDAG (version 0.3.0)

node_zeroinfl: Simulate a Node Using a Zero-Inflated Count Model

Description

Data from the parents is used to first simulate data for the regular count model, which may follow either a poisson regression or a negative binomial regression, as implemented in node_poisson and node_negative_binomial respectively. Then, zeros are simulated using a logistic regression model as implemented in node_binomial. Whenever the second binomial part returned a 0, the first part is set to 0, leaving the rest untouched. Supports random effects and random slopes (if possible) in both models. See examples.

Usage

node_zeroinfl(data, parents, parents_count,
              parents_zero, formula_count, formula_zero,
              betas_count, betas_zero,
              intercept_count, intercept_zero,
              family_count="poisson", theta,
              var_corr_count, var_corr_zero)

Value

Returns a numeric vector of length nrow(data).

Arguments

data

A data.table (or something that can be coerced to a data.table) containing all columns specified by parents, parents_count and parents_zero.

parents

A character vector specifying the names of the parents that this particular child node has. Note that this argument does not have to be specified if parents_count and parents_zero are specified. If non-linear combinations or interaction effects should be included, the user should specify the formula_count and/or formula_zero arguments instead.

parents_count

Same as parents but should only contain the parents of the count model part of the node.

parents_zero

Same as parents but should only contain the parents of the zero-inflation model part of the node.

formula_count

An enhanced formula passed to the node_poisson or the node_negative_binomial function, used to generate the count part of the node. If this argument is specified, there is no need to specify the parents_count, betas_count and intercept_count arguments. The syntax is the same as in the usual formula argument as described in node.

formula_zero

An enhanced formula passed to the node_binomial function, used to generate the zero-inflated part of the node. If this argument is specified, there is no need to specify the parents_zero, betas_zero and intercept_zero arguments. The syntax is the same as in the usual formula argument as described in node.

betas_count

A numeric vector with length equal to parents_count, specifying the causal beta coefficients used to generate the node in the count model.

betas_zero

A numeric vector with length equal to parents_zero, specifying the causal beta coefficients used to generate the node in the zero-inflation model.

intercept_count

A single number specifying the intercept that should be used when generating the count model part of the node.

intercept_zero

A single number specifying the intercept that should be used when generating the zero-inflated part of the node.

family_count

Either "poisson" for a zero-inflated poisson regression or "negative_binomial" for a zero-inflated negative binomial regression.

theta

A single number specifying the theta parameter (size argument in rnbinom). Ignore if family_count="poisson".

var_corr_count

If random effects or random slopes are included in formula_count, this argument should be specified to define the variance structure of these effects. It will be passed to the var_corr argument of node_poisson. Random effects or slopes are currently not supported with family_count="negative_binomial".

var_corr_zero

If random effects or random slopes are included in formula_zero, this argument should be specified to define the variance structure of these effects. It will be passed to the var_corr argument of node_binomial.

Author

Robin Denz

Details

It is important to note that data for both underlying models (the count model and the zero-inflation model) are simulated from completely independent of each other. When using random effects in either of the two models, they may therefore use completely different values for each process.

See Also

empty_dag, node, node_td, sim_from_dag, sim_discrete_time

Examples

library(simDAG)

set.seed(5425)

# zero-inflated poisson regression
dag <- empty_dag() +
  node(c("A", "B"), type="rnorm", mean=0, sd=1) +
  node("Y", type="zeroinfl",
       formula_count= ~ -2 + A*0.2 + B*0.1 + A:B*0.4,
       formula_zero= ~ 1 + A*1 + B*2,
       family_count="poisson",
       parents=c("A", "B"))
data <- sim_from_dag(dag, n_sim=100)

# above is functionally the same as:
dag <- empty_dag() +
  node(c("A", "B"), type="rnorm", mean=0, sd=1) +
  node("Y_count", type="poisson", formula= ~ -2 + A*0.2 + B*0.1 + A:B*0.4) +
  node("Y_zero", type="binomial", formula= ~ 1 + A*1 + B*2) +
  node("Y", type="identity", formula= ~ Y_zero * Y_count)
data <- sim_from_dag(dag, n_sim=100)

# same as above, but specifying each individual component instead of formulas
dag <- empty_dag() +
  node(c("A", "B", "C"), type="rnorm", mean=0, sd=1) +
  node("Y", type="zeroinfl",
       parents_count=c("A", "B"),
       betas_count=c(0.2, 0.1),
       intercept_count=-2,
       parents_zero=c("A", "B"),
       betas_zero=c(1, 2),
       intercept_zero=1,
       family_count="poisson",
       parents=c("A", "B"))
data <- sim_from_dag(dag, n_sim=100)

# zero-inflated negative-binomial regression
dag <- empty_dag() +
  node(c("A", "B"), type="rnorm", mean=0, sd=1) +
  node("Y", type="zeroinfl",
       formula_count= ~ -2 + A*0.2 + B*3 + A:B*0.4,
       formula_zero= ~ 3 + A*0.1 + B*0.3,
       family_count="negative_binomial", theta=1,
       parents=c("A", "B"))
data <- sim_from_dag(dag, n_sim=100)