Sure Independence Screening for High-Dimensional Mediation Analysis

Description

This function performs dimension reduction for high-dimensional mediation analysis using Sure Independence Screening (SIS). Mediators are ranked based on the product of their marginal associations with the exposure and the outcome, and the top-ranked mediators are retained for downstream analysis.

Usage

SIS(
  target_data,
  source_data = NULL,
  Y,
  D,
  M,
  X,
  topN = NULL,
  transfer = FALSE,
  verbose = TRUE,
  ncore = 1,
  dblasso_method = FALSE
)

Arguments

Details

The function supports transfer learning, allowing information from a source dataset to be leveraged to improve screening stability and robustness in the target dataset.

Value

A list with the following components:

• target_SIS: A data frame containing the outcome, exposure, selected mediators, and covariates from the target dataset.

• source_SIS: A data frame containing the same variables from the source dataset if transfer = TRUE; otherwise NULL.

• M_ID_name_SIS: A character vector of selected mediator names.

Examples

set.seed(123)

# Target data
M_target <- matrix(rnorm(200 * 50), nrow = 200)
colnames(M_target) <- paste0("M", 1:50)

target_data <- data.frame(
  Y = rnorm(200),
  D = rnorm(200),
  M_target,
  X1 = rnorm(200)
)

# Source data
M_source <- matrix(rnorm(300 * 50), nrow = 300)
colnames(M_source) <- paste0("M", 1:50)

source_data <- data.frame(
  Y = rnorm(300),
  D = rnorm(300),
  M_source,
  X1 = rnorm(300)
)

# Run SIS
result <- SIS(
  target_data = target_data,
  source_data = source_data,
  Y = "Y",
  D = "D",
  M = paste0("M", 1:50),
  X = "X1",
  transfer = TRUE,
  topN = 10
)

result$M_ID_name_SIS

TransHDM: High-Dimensional Mediation Analysis with Transfer Learning

Description

The TransHDM function performs high-dimensional mediation analysis under a transfer learning framework. It identifies and estimates indirect (mediation) effects of a high-dimensional set of mediators between an exposure and an outcome by integrating a target dataset and a source datasets.

Usage

TransHDM(
  target_data,
  source_data = NULL,
  Y,
  D,
  M,
  X,
  transfer = FALSE,
  verbose = TRUE,
  ncore = 1,
  topN = NULL,
  dblasso_SIS = FALSE
)

Arguments

Value

A list with the following components:

• contributions: A data frame of identified mediators containing:

  • mediator: Mediator name

  • alpha: Estimated exposure mediator effect

  • alpha_pv: P-value for the exposure mediator effect

  • beta: Estimated mediator outcome effect

  • beta_pv: P-value for the mediator outcome effect

  • alpha_beta: Estimated indirect (mediation) effect

  • ab_pv: Joint p-value for the indirect effect

  • pa: Proportion of the total effect mediated

• effects: A data frame summarizing the total indirect effect, direct effect, total effect, and proportion mediated.

• IDE_est: A numeric vector of indirect effect estimates for all specified mediators, with non-selected mediators set to zero.

• DE_est: The estimated direct effect of the exposure on the outcome.

References

Pan L, Liu Y, Huang C, Lin R, Yu Y, Qin G. Transfer learning reveals the mediating mechanisms of cross-ethnic lipid metabolic pathways in the association between APOE gene and Alzheimer's disease. Brief Bioinform. 2025;26(5):bbaf460. doi:10.1093/bib/bbaf460

Examples

set.seed(123)

# Target data
target_data <- gen_simData_homo(n = 50, p_x = 3, p_m = 20, rho = 0.1)$data

# Source data
source_data <- gen_simData_homo(n = 100, p_x = 3, p_m = 20, rho = 0.1, source = TRUE,
transferable = TRUE)$data

# Run TransHDM
result <- TransHDM(
  target_data = target_data,
  source_data = source_data,
  Y = "Y",
  D = "D",
  M = paste0("M", 1:20),
  X = paste0("X", 1:3),
  transfer = TRUE,
  ncore = 1,
  topN = 10
)
summary(result)

Fit Debiased LASSO with Transfer Learning

Description

Fits a debiased LASSO regression model under transfer learning framework, supporting feature selection and coefficient estimation by combining target and source data.

Usage

dblasso(
  target,
  source = NULL,
  transfer = FALSE,
  level = 0.95,
  lambda = "lambda.1se"
)

Arguments

Value

A list containing:

• dbcoef.hat: Debiased LASSO coefficient vector (including intercept)

• coef.hat: Original LASSO coefficient vector

• CI: Data frame with confidence intervals (lb = lower bound, ub = upper bound)

• var.est: Variance estimates for debiased coefficients

• se.est: Standard errors for debiased coefficients

• P.value: Vector of p-values for coefficients

Examples

# Prepare target and source data
target <- list(x = matrix(rnorm(100 * 20), 100, 20), y = rnorm(100))
source <- list(x = matrix(rnorm(200 * 20), 200, 20), y = rnorm(200))

# Non-transfer mode
result_no_transfer <- dblasso(target = target, transfer = FALSE,
                             level = 0.95, lambda = 'lambda.min')
summary(result_no_transfer)

# Transfer learning mode
result_transfer <- dblasso(target = target, source = source, transfer = TRUE,
                          level = 0.95, lambda = 'lambda.min')
summary(result_transfer)

Simulated Dataset Generation for High-Dimensional Mediation Analysis

Description

Generates synthetic datasets mimicking high-dimensional mediation structures, optionally incorporating transferable source data under varying covariate correlation and heterogeneity levels. This function supports heterogeneous settings for data generation.

Usage

gen_simData_hetero(
  n = 100,
  p_x = 5,
  rho = 0,
  p_m = 100,
  h = 0,
  source = FALSE,
  transferable = TRUE,
  seed = NULL
)

Arguments

Details

This function generates data according to a structural equation model (SEM):

• Treatment D depends linearly on covariates X.

• Mediators M depend on D and X, with residual correlation controlled by rho.

• Outcome Y depends on D, M, and X.

• If source = TRUE, then a source dataset is simulated with potentially transferable mechanisms.

Value

A list with the following components:

• data: A data.frame of dimension n × (2 + p_m + p_x) containing outcome (Y), treatment (D), mediators (M1–Mp_m), and covariates (X1–Xp_x).

• coef: A named list of true model coefficients (alpha1, alpha2, beta1, beta2, beta4, etc.).

Examples

source_data <- gen_simData_hetero(
  n = 100, p_x = 5, rho = 0, p_m = 100, h = 0,
  source = TRUE, transferable = TRUE, seed = 123
)
source_data <- gen_simData_hetero(
  n = 100, p_x = 5, rho = 0, p_m = 100, h = 0,
  source = FALSE, transferable = TRUE, seed = 123
)

Simulated Dataset Generation for High-Dimensional Mediation Analysis

Description

Generates synthetic datasets mimicking high-dimensional mediation structures, optionally incorporating transferable source data under varying covariate correlation and heterogeneity levels. This function supports homogeneous settings for data generation.

Usage

gen_simData_homo(
  n = 100,
  p_x = 5,
  rho = 0,
  p_m = 100,
  h = 0,
  source = FALSE,
  transferable = TRUE,
  seed = NULL
)

Arguments

Details

This function generates data according to a structural equation model (SEM):

• Treatment D depends linearly on covariates X.

• Mediators M depend on D and X, with residual correlation controlled by rho.

• Outcome Y depends on D, M, and X.

• If source = TRUE, then a source dataset is simulated with potentially transferable mechanisms.

Value

A list with the following components:

• data: A data.frame of dimension n * (2 + p_m + p_x) containing outcome (Y), treatment (D), mediators (M1–Mp_m), and covariates (X1–Xp_x).

• coef: A named list of true model coefficients (alpha1, alpha2, beta1, beta2, beta3, beta4, etc.).

Examples

source_data <- gen_simData_homo(
  n = 100, p_x = 5, rho = 0, p_m = 100, h = 0,
  source = TRUE, transferable = TRUE, seed = 123
)
target_data <- gen_simData_homo(
  n = 100, p_x = 5, rho = 0, p_m = 100, h = 0,
  source = FALSE, transferable = TRUE, seed = 123
)

K-Fold Cross-Validation Data Splitting

Description

Splits input data into k folds for cross-validation, generating training and test sets for each fold. Particularly useful for mediator selection stability assessment in high-dimensional mediation analysis.

Usage

kfold_split(data, kfold = 3)

Arguments

Value

A list containing two elements:

• train_set: List of length kfold, each element is a training subset

• test_set: List of length kfold, each element is the corresponding test subset

Fit LASSO Regression with Transfer Learning

Description

Fits a LASSO (Least Absolute Shrinkage and Selection Operator) regression model under a transfer learning framework. Supports feature selection and coefficient estimation by combining target data and source data.

Usage

lasso(target, source = NULL, transfer = FALSE, lambda = "lambda.1se")

Arguments

Value

A numeric vector coef containing LASSO coefficient estimates (including intercept).

Examples

# Prepare target and source data
target <- list(x = matrix(rnorm(100 * 20), 100, 20), y = rnorm(100))
source <- list(x = matrix(rnorm(200 * 20), 200, 20), y = rnorm(200))

# Non-transfer mode
coef_no_transfer <- lasso(target = target, transfer = FALSE, lambda = 'lambda.min')
print(coef_no_transfer)
summary(coef_no_transfer)

# Transfer learning mode
coef_transfer <- lasso(target = target, source = source, transfer = TRUE, lambda = 'lambda.min')
print(coef_transfer)
summary(coef_transfer)

Estimate Null Proportion in Joint Mediation Analysis

Description

Estimates the proportion of null hypotheses in dual-stage mediation analysis using empirical p-value distributions. This implements the statistical framework for estimating the proportion of non-mediating variables.

Usage

null_estimation(input_pvalues, lambda = 0.5)

Arguments

Value

A list containing estimated null proportions:

• alpha00: Proportion of mediators with null effects in both exposure-mediator and mediator-outcome paths

• alpha10: Proportion with non-null exposure-mediator but null mediator-outcome effects

• alpha01: Proportion with null exposure-mediator but non-null mediator-outcome effects

• alpha1: Marginal null proportion in exposure-mediator path

• alpha2: Marginal null proportion in mediator-outcome path

Detect Transferable Source Data via Cross-Validation

Description

Determines whether external source datasets can be effectively transferred to the target data by comparing predictive performance using LASSO regression under a transfer learning framework.

Usage

source_detection(
  target_data,
  source_data = NULL,
  Y,
  D,
  M,
  X,
  kfold = 5,
  C0 = 0.05,
  verbose = TRUE
)

Arguments

Value

A list containing:

• transfer.source.id: Indices of source datasets deemed transferable

• source.loss: Mean validation loss for each source dataset

• target.valid.loss: Mean validation loss using target-only model

• T_index: Difference between source loss and target-only loss for each source

• threshold: Calculated transferability threshold

• loss.cv: Full k-fold cross-validation loss matrix

Examples

## Reproducible example
set.seed(123)

# Generate synthetic target data
target_data <- data.frame(
  Y  = rnorm(200),
  D  = rnorm(200),
  M1 = rnorm(200),
  M2 = rnorm(200),
  X1 = rnorm(200)
)

# Generate synthetic source data
source1 <- data.frame(
  Y  = rnorm(300),
  D  = rnorm(300),
  M1 = rnorm(300),
  M2 = rnorm(300),
  X1 = rnorm(300)
)

source2 <- data.frame(
  Y  = rnorm(250),
  D  = rnorm(250),
  M1 = rnorm(250),
  M2 = rnorm(250),
  X1 = rnorm(250)
)

# Run source detection
result <- source_detection(
  target_data = target_data,
  source_data = list(source1, source2),
  Y = "Y",
  D = "D",
  M = c("M1", "M2"),
  X = "X1",
  kfold = 5,
  C0 = 0.05,
  verbose = FALSE
)

# Get Summary
summary(result)

# Transferable source indices
result$transfer.source.id

# Compare validation losses
data.frame(
  Source = c(paste0("Source", seq_along(result$source.loss)), "Target"),
  Loss   = c(result$source.loss, result$target.valid.loss)
)

Summary of TransHDM Mediation Analysis

Description

Summary of TransHDM Mediation Analysis

Usage

## S3 method for class 'TransHDM'
summary(object, top = 10, digits = 4, ...)

Arguments

Value

An object of class "summary.TransHDM".

Summary of Debiased Lasso Inference

Description

Summary of Debiased Lasso Inference

Usage

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

Arguments

Value

An object of class "summary.dblasso".

Summary of Lasso Regression

Description

Summary of Lasso Regression

Usage

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

Arguments

Value

An object of class "summary.lasso".

Summary of Source Detection Results

Description

Summary of Source Detection Results

Usage

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

Arguments

Value

An object of class "summary.source_detection".