Adding New Methods

This vignette explains how to add new publication bias correction methods to the PublicationBiasBenchmark package. In the following, we will use the PET (Precision-Effect Test) method as an example.

Overview

Each method in the package consists of three key components:

  1. Main method function: Implements the statistical method
  2. Settings function: Defines available method configurations
  3. Extra columns function: Specifies additional result columns

All three functions must be implemented in a single file named method-{METHOD_NAME}.R in the R/ directory. Implementation of these three functions allows users to apply the method via the run_method() function.

File Structure and Naming

For a method called “PET”, you need to create a file named R/method-PET.R containing three functions:

The naming pattern is crucial for the package’s S3 method dispatch system to work correctly.

1. Main Method Function: method.{METHOD_NAME}()

This is the core function that implements your statistical method. Here is the PET implementation as an example:

#' @title PET (Precision-Effect Test) Method
#'
#' @description
#' Implements the Precision-Effect Test for publication bias correction.
#' PET regresses effect sizes against standard errors to test for and correct
#' publication bias. The intercept represents the bias-corrected effect size
#' estimate.
#'
#' @param method_name Method name (automatically passed)
#' @param data Data frame with yi (effect sizes) and sei (standard errors)
#' @param settings List of method settings
#'
#' @return Data frame with PET results
#'
#' @export
method.PET <- function(method_name, data, settings = NULL) {
  
  # Extract data
  effect_sizes    <- data$yi
  standard_errors <- data$sei
  
  # Input validation and error handling
  if (length(effect_sizes) < 3)
    stop("At least 3 estimates required for PET analysis", call. = FALSE)
  
  if (stats::var(standard_errors) <= 0)
    stop("No variance in standard errors", call. = FALSE)
  
  # Implement the statistical method
  pet_model <- stats::lm(effect_sizes ~ standard_errors, 
                        weights = 1/standard_errors^2)
  
  # Extract and process results
  coefficients    <- stats::coef(pet_model)
  se_coefficients <- summary(pet_model)$coefficients[, "Std. Error"]
  p_values        <- summary(pet_model)$coefficients[, "Pr(>|t|)"]
  
  # Main estimates
  estimate    <- coefficients[1]  # Intercept = bias-corrected effect
  estimate_se <- se_coefficients[1]
  estimate_p  <- p_values[1]
  
  # Additional method-specific results
  bias_coefficient <- coefficients[2]
  bias_p_value     <- p_values[2]
  
  # Calculate confidence intervals
  estimate_lci <- estimate - 1.96 * estimate_se
  estimate_uci <- estimate + 1.96 * estimate_se
  
  # Return standardized results
  return(data.frame(
    method         = method_name,
    estimate       = estimate,
    standard_error = estimate_se,
    ci_lower       = estimate_lci,
    ci_upper       = estimate_uci,
    p_value        = estimate_p,
    BF             = NA,
    convergence    = TRUE,
    note           = NA,
    # Method-specific columns
    bias_coefficient = bias_coefficient,
    bias_p_value     = bias_p_value
  ))
}

Key Requirements for the Main Function:

Input Parameters:

Output: Must return a data frame with these required columns:

If your method does not provide certain values (e.g., Bayes factor), use NA.

Error Handling:

2. Settings Function: method_settings.{METHOD_NAME}()

This function defines the available configurations for your method:

#' @export
method_settings.PET <- function(method_name) {
  
  settings <- list(
    "default" = list() # PET has no configurable settings
  )
  
  return(settings)
}

The selected settings are passed to your main function as the settings parameter. There, the main function can use these settings to adjust its behavior. Once defined, these settings cannot be changed retrospectively to ensure reproducibility and continuity of the benchmark.

For an example of a method that defines several settings we can examine the random effects meta-analysis (RMA) method:

# Example with multiple settings (from RMA method)
method_settings.RMA <- function(method_name) {
  
  settings <- list(
    "default" = list(
      method = "REML", 
      test.uni = "knha", 
      test.mv = "t", 
      control = list(stepadj = 0.5, maxiter = 500)
    )
  )
  
  return(settings)
}

3. Extra Columns Function: method_extra_columns.{METHOD_NAME}()

This function specifies additional columns that your method returns beyond the required ones:

#' @export
method_extra_columns.PET <- function(method_name) {
  c("bias_coefficient", "bias_p_value")
}

The column names must match exactly the additional columns that your main function returns. These columns will be included in the final output data frame alongside the required columns and guarantee that the results can be merged in case the method fails with an error.

Use character(0) if your method has no extra columns

Using Your New Method

Once implemented, your method can be used through a unified interface:

# Create example data
data <- data.frame(
  yi  = c(0.2, 0.3, 0.1, 0.4, 0.25),
  sei = c(0.1, 0.15, 0.08, 0.12, 0.09)
)

# Run your method
result <- run_method("PET", data)
print(result)

# Use specific settings (if available)
result <- run_method("PET", data, "default")