mvPred: A Central Repo for Handling Missing Values in Predictive Applications

Methods for handling missing values abound, but they are almost all focused in applications in which the goal is effect estimation, e.g. calculation and statistical inference for regression coefficients. Our focus here is instead on prediction.

Note that this should not be confused with packages that use regression methods for inputation. There, the missing values in a variable X_i may be inputed using predictions in which X₂ is regressed on the other X_j. Instead, we are interested in applications in prediction itself is the focus, such as forecasting or disease diagnosis.

Installation

You can install mvPred from CRAN with:

install.packages("mvPred")

Then load the package with:

library(mvPred)

Main Methods

The package currently includes methods for predictive modeling with missing data, including:

• complete-case analysis via bootstrap(..., method = "CC")
• available-case regression via bootstrap(..., method = "AC")
• PREFILL-based approaches via bootstrap(..., method = "PREFILL")
• TOWER-based modeling via bootstrap(..., method = "TOWER")

For PREFILL, the supported imputation methods include:

• mice
• Amelia
• missForest
• complete-case imputation through impute_method = "complete"

Core modeling helpers in the package include:

• bootstrap()
• lm_ac()
• lm_prefill()
• lm_tower()

Example

library(mvPred)

data("auto-mpg", package = "mvPred")

df <- auto_mpg
df$car_name <- NULL

for (nm in names(df)) {
  suppressWarnings(df[[nm]] <- as.numeric(df[[nm]]))
}

res <- bootstrap(
  data = df,
  yName = "mpg",
  k = 5,
  task = "regression",
  method = "CC"
)

res$RMSE_mean

Package Data

The package also includes datasets used in examples and experiments, such as:

• auto-mpg
• english
• NHkids