walker (version 1.0.10)

predict_counterfactual: Predictions for walker object

Description

Given the new covariate data and output from walker, obtain samples from posterior predictive distribution for counterfactual case, i.e. for past time points with different covariate values.

Usage

predict_counterfactual(
  object,
  newdata,
  u,
  summary = TRUE,
  type = ifelse(object$distribution == "gaussian", "response", "mean")
)

Value

If summary=TRUE, time series containing summary statistics of predicted values. Otherwise a matrix of samples from predictive distribution.

Arguments

object

An output from walker() or walker_glm().

newdata

A data.frame containing covariates used for prediction. Should have equal number of rows as the original data

u

For Poisson model, a vector of exposures i.e. E(y) = uexp(xbeta). For binomial, a vector containing the number of trials. Defaults 1.

summary

If TRUE (default), return summary statistics. Otherwise returns samples.

type

If "response" (default for Gaussian model), predictions are on the response level (e.g., number of successes for Binomial case, and for Gaussian case the observational level noise is added to the mean predictions). If "mean" (default for non-Gaussian case), predict means (e.g., success probabilities in Binomial case). If "link", predictions for non-Gaussian models are returned before applying the inverse of the link-function.

Examples

if (FALSE) {
set.seed(1)
n <- 50
x1 <- rnorm(n, 0, 1)
x2 <- rnorm(n, 1, 0.5)
x3 <- rnorm(n)
beta1 <- cumsum(c(1, rnorm(n - 1, sd = 0.1)))
beta2 <- cumsum(c(0, rnorm(n - 1, sd = 0.1)))
beta3 <- -1
u <- sample(1:10, size = n, replace = TRUE)
y <- rbinom(n, u, plogis(beta3 * x3 + beta1 * x1 + beta2 * x2))

d <- data.frame(y, x1, x2, x3)
out <- walker_glm(y ~ x3 + rw1(~ -1 + x1 + x2, beta = c(0, 2), 
  sigma = c(2, 10)), distribution = "binomial", beta = c(0, 2), 
  u = u, data = d,
  iter = 2000, chains = 1, refresh = 0)

# what if our covariates were constant?
newdata <- data.frame(x1 = rep(0.4, n), x2 = 1, x3 = -0.1)

fitted <- fitted(out)
pred <- predict_counterfactual(out, newdata, type = "mean")

ts.plot(cbind(fitted[, c(1, 3, 5)], pred[, c(1, 3, 5)]), 
  col = rep(1:2, each = 3), lty = c(1, 2, 2))
}