boolean3 (version 3.1.6)

boolprep: Specify Boolean Model

Description

Construct a boolean object that can then be fit with boolean.

Usage

boolprep(..., data, subset, weights, na.action, offset, family = list(binomial(link = "logit")))

Arguments

...

formula specification for boolean model. See the details and examples sections below.

data

data.frame containing the data to be used in the model.

subset

select subset of data. See lm for details.

weights

specify model weights (not implemented).

na.action

set na.action. See lm for details.

offset

specify an offset. Offsets are not implemented and this parameter is simply ignored.

family

a model family to use for estimation. This can be comprised of a list of link functions when the desire is to have model components with different links. binomial is the only family supported at this time. See family for more details.

Value

boolprep returns a boolean-class object containing the model components needed to estimate a boolean model.

Details

boolprep sets up a boolean model object that can then be fit with boolean. A properly specified model (contained in ...) will contain at least three components. The first component must specify the boolean logic to be employed. For instance, the y ~ (a | b) formula would indicate a logical or between the a and b submodels, while y ~ (a & b) would indicate a logical and. y is the name of the response variable of interest. Logical operators can be nested; e.g., y ~ (a | (b & c)) is valid. The second and third components are submodels and are specified as usual: a ~ x1 + x2 and b ~ x3 + x4 + x5, where are the x-variables are covariates. a, b, and c are labels indicating the submodel position in the boolean specification.

References

Braumoeller, Bear F. (2003) ``Causal Complexity and the Study of Politics.'' Political Analysis 11(3): 209--233.

See Also

See boolean for model estimation.

Examples

## Generate some fake data
require(mvtnorm)
set.seed(12345)
N  <- 2000
Df <- cbind(1, rmvnorm(N, mean=rep(0, 5)))

## Set coefficients
beta.a <- c(-2.00, 0.33, 0.66, 1.00)
beta.b <- c(0.00, 1.50, -0.25)

## Generate path probabilities following a normal model.
y.a <- as.vector(pnorm(tcrossprod(beta.a, Df[, 1:4])))
y.b <- as.vector(pnorm(tcrossprod(beta.b, Df[, c(1, 5, 6)])))

## AND and OR-model
or <- function(x, y) { x + y - x * y }
and <- function(x, y) { x * y }
y.star.OR  <- or(y.a, y.b)
y.star.AND <- and(y.a, y.b)

## Observed responses
y.OR <- rbinom(N, 1, y.star.OR)
y.AND <- rbinom(N, 1, y.star.AND)

## Set up data.frame for estimation
Df <- cbind(1, Df)
Df <- as.data.frame(Df)
Df[,1] <- y.OR
Df[,2] <- y.AND
names(Df) <- c("y.OR", "y.AND", "x1", "x2", "x3", "x4", "x5")

## Before estimating, boolean models need to be specified using the
## boolprep function.

## OR model, specified to use a probit link for each causal path. This
## matches the data generating process above.
mod.OR <- boolprep(y.OR ~ (a | b), a ~ x1 + x2 + x3, b ~ x4 + x5,
                   data = Df, family=list(binomial("probit")))