Package 'mvPred'

Datasets Included in mvPred

Description

Datasets distributed with mvPred for examples, experiments, and package demonstrations.

Usage

data("auto-mpg")
data(english)
data(NHkids)

Format

auto_mpg: a data frame based on the Auto MPG data, including fuel-efficiency and vehicle-characteristic variables.

english: a data frame used for predictive modeling examples involving vocabulary outcomes and demographic predictors.

NHkids: a data frame used for predictive modeling examples involving child weight, height, and age in months.

Details

These datasets are included to support examples and methodological comparisons within the package.

Examples

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

---

Cross-Validation for Predictive Modeling with Missing Data

Description

Run k-fold cross-validation for regression while applying one of the package's missing-data handling strategies. Depending on the selected method, the function can fit complete-case, available-case, PREFILL, or TOWER-based models and return fold-level summaries.

Usage

bootstrap(
  data,
  yName,
  k = 5,
  task = c("regression", "classification"),
  method = c("AC", "CC", "PREFILL", "TOWER"),
  threshold = 0.5,
  seed = 42,
  test_split = 0.2,
  impute_method = c("mice", "amelia", "missforest", "complete"),
  mice_method = NULL,
  m = 5,
  use_dummies = FALSE,
  tower_regFtnName = "lm",
  tower_opts = list(),
  tower_scaling = NULL,
  tower_yesYVal = NULL,
  ...
)

Arguments

data	A data frame containing predictors and the response variable.
yName	The name of the response column in data.
k	Number of folds used for cross-validation.
task	Whether to run "regression" or "classification" metrics.
method	Missing-data handling strategy: "AC", "CC", "PREFILL", or "TOWER".
threshold	Classification threshold used when task = "classification".
seed	Random seed used for split reproducibility.
test_split	Proportion of the full dataset reserved for the final holdout test set.
impute_method	Imputation method used when method = "PREFILL".
mice_method	Optional mice method specification for PREFILL runs.
m	Number of multiple imputations used by applicable PREFILL methods.
use_dummies	Logical indicating whether categorical variables should be converted to dummies for Amelia-based PREFILL runs.
tower_regFtnName	Regression routine passed to the TOWER method.
tower_opts	Optional list of control settings passed to TOWER.
tower_scaling	Optional scaling object passed to TOWER.
tower_yesYVal	Optional flag passed to TOWER.
...	Additional arguments forwarded to the selected fitting routine.

Details

The function first optionally creates a holdout split controlled by test_split. Cross-validation is then run on the training portion using the requested missing-data handling method. For PREFILL runs, the supported imputation methods are mice, Amelia, missForest, and complete-case preprocessing.

Value

A list containing fold-level performance summaries and fitted model objects. Depending on the selected task, the returned metrics correspond to regression or classification performance.

Common returned components include:

• fold_models: list of fitted models, one per cross-validation fold.

• final_model: fitted model from the final training split when test_split > 0.

• test_results: performance on the final holdout test set when test_split > 0.

• training_missingness_by_fold: missingness tables for each training fold.

• training_missingness_average: average missingness across training folds.

Examples

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 = 3,
  task = "regression",
  method = "CC"
)

res$RMSE_mean

---

Extract Complete Cases from a Data Frame

Description

Remove rows containing missing values and report which rows were retained or discarded.

Usage

compCases(data)

Arguments

data	A data frame or matrix that may contain missing values.

Value

A list with components:

• intactData: rows of data with no missing values.

• intactRows: logical vector indicating complete rows.

• nonintactRows: integer indices of rows containing at least one missing value.

Examples

df <- data.frame(x = c(1, NA, 3), y = c(2, 4, 6))
compCases(df)

---

Linear Regression Using Available Cases Methods

Description

This suite of functions implements linear regression using the Available Cases (pairwise deletion) approach. It makes use of the avaliable cases in the dataset to fit the model and make predictions rather than deleting any rows with missing values. The main function lm_ac fits the model even when the predictors or response contain missing values. The helper functions compute matrix products using only complete pairs of observations. S3 methods are included for summary and prediction.

Usage

lm_ac(data, yName, holdout = NULL, ...)

ac_mul(A)

ac_vec(X, y)

## S3 method for class 'lm_ac'
summary(object, ...)

## S3 method for class 'lm_ac'
predict(object, newdata, ...)

Arguments

data	A data.frame containing predictors and the response variable.
yName	The name of the response column in data.
holdout	Optional holdout specification used to define testing rows. This may be a logical vector of length nrow(data) or a numeric vector of row indices. If NULL, all rows are used for fitting.
...	Additional arguments passed to model.frame in lm_ac.
A	A numeric matrix or data frame to be used in ac_mul to compute ATA.
X	A numeric matrix of predictors for ac_vec.
y	A numeric vector representing the response variable for ac_vec.
object	An object of class lm_ac for S3 methods which is returned from calling lm_ac.
newdata	A data frame with predictor variables for prediction.

Details

The function lm_ac fits a linear regression model using pairwise deletion for missing values. Instead of removing entire rows with missing values (as in complete-case analysis), this approach computes each component of the normal equations (X'X and X'y) using the average of all available pairs.

ac_mul returns a symmetric matrix of average products between all pairs of columns.

ac_vec returns a vector of average products between each column of X and y.

summary.lm_ac prints the estimated coefficients.

predict.lm_ac generates predictions using the fitted model on new data.

Value

lm_ac returns an object of class lm_ac, a list containing:

data	Cleaned input data used in modeling.
yName	Response variable name.
formula	The model formula.
fit_obj	A list with components coef (estimated coefficients) and colnames (predictor names).

ac_mul returns a numeric matrix.

ac_vec returns a numeric vector.

summary.lm_ac returns the coefficient vector (invisible).

predict.lm_ac returns a numeric vector of predictions.

Examples

data(airquality)

# Fit model
mod <- lm_ac(airquality, "Ozone")

# Print coefficients
summary(mod)

# Predict on new data
predict(mod, newdata = airquality)

---

Compare Missing-Data Methods with Bootstrap Summaries

Description

Run bootstrap() across one or more missing-data handling methods and collect the resulting cross-validation summary metrics in a single data frame.

Usage

lm_compare(
  data,
  predictors = NULL,
  yName,
  k = 5,
  task = c("regression", "classification"),
  methods = c("CC", "AC", "TOWER", "PREFILL"),
  seed = 42,
  impute_method = c("mice", "amelia", "missforest", "complete"),
  mice_method = NULL,
  m = 5,
  use_dummies = FALSE,
  tower_regFtnName = "lm",
  tower_opts = list(),
  tower_scaling = NULL,
  tower_yesYVal = NULL,
  ...
)

Arguments

data	A data frame containing predictors and the response variable.
predictors	Optional character vector of predictor names to keep. If NULL, all non-response columns are used.
yName	The name of the response column in data.
k	Number of folds used for cross-validation.
task	Whether to summarize regression or classification performance.
methods	Character vector of methods to evaluate. Supported values are "CC", "AC", "TOWER", and "PREFILL".
seed	Random seed used for reproducibility.
impute_method	Imputation method used when "PREFILL" is included in methods.
mice_method	Optional mice method specification passed to PREFILL runs.
m	Number of multiple imputations used by applicable PREFILL methods.
use_dummies	Logical indicating whether categorical variables should be converted to dummies for Amelia-based PREFILL runs.
tower_regFtnName	Regression routine passed to TOWER.
tower_opts	Optional list of TOWER options.
tower_scaling	Optional scaling object passed to TOWER.
tower_yesYVal	Optional flag passed to TOWER.
...	Additional arguments forwarded to bootstrap().

Details

lm_compare() is intended for method comparison under cross-validation. Internally it calls bootstrap() with test_split = 0, so the reported summaries correspond to cross-validation results rather than final holdout-test performance.

Value

A data frame containing one row per requested method and the corresponding average performance metrics.

Examples

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

df <- auto_mpg
df$car_name <- NULL

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

lm_compare(df, yName = "mpg", k = 3, task = "regression", methods = c("CC", "AC"))

---

Linear Model Fitting with Pre-Imputation of Missing Data

Description

Fit linear models to data with missing values by first imputing missing data using multiple imputation methods, then fitting linear models to the imputed datasets. Supports multiple imputation methods including mice, Amelia, and missForest, as well as complete-case analysis.

Usage

lm_prefill(
  data,
  yName,
  impute_method = "mice",
  m = 5,
  use_dummies = FALSE,
  holdout = NULL,
  ...
)

## S3 method for class 'lm_prefill'
summary(object, ...)

## S3 method for class 'lm_prefill'
predict(object, newdata, type = "response", ...)

Arguments

data	A data frame containing the variables for modeling, possibly with missing values.
yName	A character string specifying the name of the response variable in data.
impute_method	Character string specifying the imputation method to use. One of "mice" (default), "amelia", "missforest", or "complete".
m	Integer specifying the number of multiple imputations to generate. This is only used for "mice" and "amelia" methods. Default is 5.
use_dummies	Logical indicating whether to convert categorical predictors to dummy variables before imputation. This is only relevant for "amelia".
holdout	Optional holdout specification used to define testing rows. This may be a logical vector of length nrow(data) or a numeric vector of row indices. If NULL, all rows are treated as training data.
...	Additional arguments passed to the imputation function (mice::mice, Amelia::amelia, or missForest::missForest) or to stats::lm for model fitting.
object	An object of class "lm_prefill" returned by lm_prefill().
newdata	A data frame containing new data for prediction. Rows with missing values in newdata are omitted before prediction.
type	Type of prediction; passed to stats::predict.lm. Default is "response".

Details

lm_prefill is a wrapper function that first imputes missing values in the dataset using one of several supported methods, then fits linear models to the imputed datasets.

Supported imputation methods:

• "mice": Multiple Imputation by Chained Equations via the mice package.

• "amelia": Multiple imputation using the Amelia package.

• "missforest": Nonparametric imputation using random forests via the missForest package.

• "complete": Complete-case analysis by omitting rows with missing values.

The use_dummies argument controls whether categorical variables are converted to dummy variables before imputation. If use_dummies = TRUE, the package regtools is required for dummy variable creation.

Additional arguments passed via ... are forwarded to the underlying imputation or modeling functions as appropriate.

The returned object is of class "lm_prefill" and contains the imputed data, fitted models, and metadata.

summary.lm_prefill returns model summaries for each imputed dataset or a single summary for complete-case or missForest methods.

predict.lm_prefill returns predictions averaged across imputations for multiple imputation methods, or direct predictions for single imputation or complete-case models.

If newdata contains missing values, those rows are omitted with a warning before prediction.

Value

An object of class "lm_prefill" containing at least the following components:

data	Original input data.
yName	Response variable name.
formula	Model formula used for fitting.
imputed_data	Imputed dataset(s) or imputation object.
impute_method	Imputation method used.
fit_obj	Fitted linear model(s).
missforest_OOBerror	If impute_method = "missforest", the out-of-bag error estimate from missForest.

Author(s)

Cynthia Mascarenhas cynmascarenhas@ucdavis.edu

See Also

mice, amelia, missForest, lm, predict.lm, na.omit, factorsToDummies

Examples

library(mice)
library(Amelia)
library(missForest)

# Example dataset with missing values
data(airquality)
airquality$Ozone[1:10] <- NA

# Fit linear model with mice imputation
lm_obj <- lm_prefill(
  airquality,
  yName = "Ozone",
  impute_method = "mice",
  m = 5
)

# Summarize fitted models
summary(lm_obj)

# Predict on new data (complete cases only)
newdata <- airquality[11:20, ]
predict(lm_obj, newdata)

# Fit with Amelia and use dummy variables for categorical predictors
lm_obj2 <- lm_prefill(
  airquality,
  yName = "Ozone",
  impute_method = "amelia",
  m = 5,
  use_dummies = TRUE
)
summary(lm_obj2)

---

Fit Linear Models with Missing Data Using the Tower Method

Description

lm_tower fits a linear regression model using the Tower Method implemented in the toweranNA package. This method handles missing data in both training and new datasets without explicit imputation by leveraging regression averaging.

Usage

lm_tower(data, yName, regFtnName = "lm", opts = list(), scaling = NULL, yesYVal = NULL)

## S3 method for class 'lm_tower'
summary(object, ...)

## S3 method for class 'lm_tower'
predict(object, newdata, ...)

Arguments

data	A data frame containing the training data with possible missing values.
yName	A character string specifying the name of the response variable in data.
regFtnName	A character string specifying the regression function to use (default is "lm").
opts	A list of options passed to toweranNA::makeTower().
scaling	Optional scaling parameters passed to toweranNA::makeTower().
yesYVal	Optional parameter passed to toweranNA::makeTower().
object	An object of class "lm_tower" returned by lm_tower().
newdata	A data frame containing new observations for prediction. May contain missing values.
...	Additional arguments (currently unused).

Details

The Tower Method implemented in the toweranNA package fits regression models that can handle missing data without explicit imputation. It uses regression averaging over subsets of complete cases to provide predictions even when new data contain missing values.

This function provides a convenient wrapper around toweranNA::makeTower() and associated prediction methods.

lm_tower is particularly useful when the primary goal is prediction rather than inference.

Value

lm_tower returns an object of class "lm_tower" containing the fitted Tower model and related information.

summary.lm_tower prints a brief summary of the fitted model, including the response variable, regression function, and number of complete cases used.

predict.lm_tower returns a numeric vector of predicted values for newdata. The Tower method handles missing values internally, so predictions are returned for all rows of newdata.

Author(s)

Cynthia Mascarenhas cynmascarenhas@ucdavis.edu

See Also

makeTower, predict.tower

Examples

library(toweranNA)

data(airquality)
airquality$Ozone[1:10] <- NA
airquality$Solar.R[5:15] <- NA

# Fit the tower model
lm_tower_obj <- lm_tower(airquality, yName = "Ozone")

# Summarize the model
summary(lm_tower_obj)

# Prepare new data with missing values
newdata <- airquality[20:30, -which(names(airquality) == "Ozone")]
newdata[1, "Solar.R"] <- NA

# Predict on new data
preds <- predict(lm_tower_obj, newdata = newdata)
print(preds)

---

Interface qeML Models with Missing-Value Preprocessing

Description

Apply a missing-value handling strategy before fitting a predictive model from the qeML package. The helper addListElement() appends or replaces an element in a list.

Usage

qeMLna(
  data,
  yName,
  qeMLftn,
  mvFtn,
  qeMLopts = NULL,
  mvPredOpts = NULL,
  retainMVFtnOut = TRUE,
  seed = 9999,
  holdout = 1000
)

addListElement(l, newEltName, newElt)

Arguments

data	A data frame containing predictors and the response variable.
yName	The name of the response column in data.
qeMLftn	Character string naming the predictive function from qeML to call.
mvFtn	Character string naming the missing-value handling routine to apply first.
qeMLopts	Optional list of arguments passed to the selected qeML function.
mvPredOpts	Optional list of arguments passed to the missing-value handling routine.
retainMVFtnOut	Logical indicating whether to store the missing-value routine output in the returned object.
seed	Random seed passed through to applicable routines.
holdout	Holdout size used by the predictive routine if applicable.
l	A list to modify. If NULL, a new list is created.
newEltName	Name of the element to add or replace.
newElt	Value of the element to add or replace.

Value

qeMLna() returns an object of class "MVqeMLout" containing the qeML fit and, optionally, the missing-value handling output.

addListElement() returns the modified list.

Examples

data("english", package = "mvPred")

eng <- english[, c("age", "birth_order", "ethnicity", "sex", "mom_ed", "vocab")]
qeMLna(eng, "vocab", "qeLin", "compCases", retainMVFtnOut = FALSE)


addListElement(list(a = 1), "b", 2)

---