calc.ppci: Calculate Predicted Probabilities and 95% Confidence...

In jenniferthompson/ClusterBootMultinom: Multinomial Logistic Regression Using Clustered Bootstrapping

Description

Given a matrix of bootstrapped coefficient estimates from multinomial regression using vglm(), calculates linear predictors, predicted probabilities, and SEs and confidence limits for a continuous exposure variable at all non-reference levels of the outcome. Has the capabilitiy to include restricted cubic splines using rcs().

Usage

calc.ppci(design.matrix, model.obj, use.coefs, use.vcov)

Arguments

design.matrix	Design matrix with covariate values, stacked by outcome level.
model.obj	vglm() model object, used to get number of outcome levels and label linear predictors.
use.coefs	Numeric vector of coefficients, if not taken from original model (eg, colMeans(matrix of bootstrapped coefficients).
use.vcov	Numeric matrix to use as variance-covariance matrix (eg, var(matrix of bootstrapped coefficients)).

Value

List of linear predictors (LinearPredictors); variance-covariance matrix of linear predictors (VarianceLP), predicted probabilities (PredictedProbs) and their variance-covariance matrix (VariancePPs); SEs (ProbsSes), lower and upper confidence limits of predicted probabilities (ProbsSEs, ProbsLCLs and ProbsUCLs, respectively). Each has (number of outcome levels - 1) columns, representing the quantities for all outcome levels except the reference.

See Also

vglm, which this function assumes you are using; multi.plot.probs, which calls this function; rcs().

Examples

## Create data frame
df <- data.frame(id = sample(1:20, size = 100, replace = TRUE),
                 x1 = rnorm(n = 100),
                 x2 = rbinom(p = 0.75, n = 100, size = 1),
                 y = sample(LETTERS[1:3], size = 100, replace = TRUE))
df <- df[order(df$id),]
df$time <- unlist(lapply(1:length(unique(df$id)),
                         FUN = function(idnum){ 1:nrow(df[df$id == unique(df$id)[idnum],]) }))

## Using create.sampdata(), generate list of cluster bootstrapped data sets
bootdata.list <- create.sampdata(org.data = df,
                                 id.var = 'id',
                                 n.sets = 25)

## Fit model to original and bootstrapped data frame, saving errors and warnings to .txt file
boot.fits.a <- multi.bootstrap(org.data = df,
                               data.sets = bootdata.list,
                               ref.outcome = grep('A', levels(df$y)),
                               multi.form = as.formula('y ~ x1 + x2'))

## Create matrices of coefficients for all bootstrap fits
boot.matrix.a <- do.call(rbind,
                         lapply(boot.fits.a$boot.models,
                                FUN = function(x){ x@coefficients }))

## Calculate predicted probs and CIs for x1 at outcomes B, C
## Design matrix: first two columns = intercepts, second two set X2 = 1
design.tmp <- matrix(c(1, 0, 0, 1, 1, 0, 0, 1), nrow = 2)
x1.vals <- sort(unique(df$x1))

## Add all unique x1 values to get complete design matrix
x1.design <- do.call(rbind,
                     lapply(1:nrow(design.tmp), FUN = function(r){
                       tmp <- matrix(rep(c(design.tmp[r,],
                                           rep(0, (length(unique(df$y)) - 1))),
                                         length(x1.vals)),
                                     nrow = length(x1.vals), byrow = TRUE)
                       tmp[,(ncol(design.tmp) + r)] <- x1.vals
                       tmp
                     }))

## Calculate linear predictors, predicted probabilities, etc
x1probs <- calc.ppci(design.matrix = x1.design,
                     model.obj = boot.fits.a$org.model,
                     use.coefs = colMeans(boot.matrix.a),
                     use.vcov = var(boot.matrix.a))

jenniferthompson/ClusterBootMultinom documentation built on May 19, 2019, 4:03 a.m.