Using Precomputed Measures"
In PublicationBiasBenchmark: Benchmark for Publication Bias Correction Methods

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(PublicationBiasBenchmark)

This vignette explains how to access and use the precomputed performance measures from the PublicationBiasBenchmark package. The package provides comprehensive benchmark results for various publication bias correction methods across different data-generating mechanisms (DGMs), allowing researchers to evaluate and compare method performance without running computationally intensive simulations themselves.

For the sake of not re-downloading the performance measures every time you re-knit this vignette, we disable evaluation of code chunks below. (To examine the output, please copy to your local R session.)

Overview

The package provides precomputed performance measures for multiple publication bias correction methods evaluated under different simulation conditions. These measures include:

Bias: Average difference between estimates and true effect sizes
Relative Bias: Bias expressed as a proportion of the true effect size
MSE (Mean Square Error): Average squared difference between estimates and true values
RMSE (Root Mean Square Error): Square root of MSE, measuring overall estimation accuracy
Empirical Variance: Variability of estimates across repetitions
Empirical SE: Standard deviation of estimates across repetitions
Coverage: Proportion of confidence intervals containing the true effect
Mean CI Width: Average width of confidence intervals
Interval Score: Proper scoring rule for probabilistic interval forecasts, here adapted to evaluate CI quality
Power: Proportion of significant results (in conditions when the null hypothesis is true this corresponds the Type I error rate)
Positive Likelihood Ratio: Indication of how much a significant test result changes the odds of H1 versus H0
Negative Likelihood Ratio: Indication of how much a non-significant test result changes the odds of H1 versus H0
Convergence: Proportion of successful method convergence

The precomputed results are organized by data-generating mechanism (DGM), with each DGM representing different patterns of publication bias and meta-analytic conditions.

Available Data-Generating Mechanisms

The package includes precomputed measures for several DGMs. You can view the specific conditions for each DGM using the dgm_conditions() function:

# View conditions for the Stanley2017 DGM
conditions <- dgm_conditions("Stanley2017")
head(conditions)

Downloading Precomputed Measures

Before accessing the precomputed measures, you need to download them from the package repository. The download_dgm_measures() function downloads the measures for a specified DGM:

# Download precomputed measures for the Stanley2017 DGM
download_dgm_measures("Stanley2017")

The measures are downloaded to a local cache directory and are automatically available for subsequent analysis. You only need to download them once, unless the benchmark measures were updated with a new method (in that case, you need to specify overwrite = TRUE argument).

Retrieving Precomputed Measures

Once downloaded, you can retrieve the precomputed measures using the retrieve_dgm_measures() function. This function offers flexible filtering options to extract exactly the data you need.

Retrieving Specific Measures

You can retrieve measures for a specific method and condition:

# Retrieve bias measures for RMA method in condition 1
retrieve_dgm_measures(
  dgm            = "Stanley2017",
  measure        = "bias",
  method         = "RMA",
  method_setting = "default",
  condition_id   = 1
)

The measure argument can be any of measure function names listed in the measures() documentation.

Retrieving All Measures

To retrieve all measures across all conditions and methods, simply omit the filtering arguments:

# Retrieve all measures across all conditions and methods
df <- retrieve_dgm_measures("Stanley2017")

This returns a comprehensive data frame with columns:

method: Publication bias correction method name
method_setting: Specific method configuration
condition_id: Simulation condition identifier
bias, bias_mcse, rmse, rmse_mcse, ...: Performance measures and their Monte Carlo standard errors

Filtering by Method or Setting

You can also filter by method name:

# Retrieve all measures for PET-PEESE method
pet_peese_results <- retrieve_dgm_measures(
  dgm    = "Stanley2017",
  method = "PETPEESE"
)

Visualizing Precomputed Results

Once you have retrieved the measures, you can create visualizations to compare method performance. Here's an example that creates a multi-panel plot comparing all methods across all conditions:

# Retrieve all measures across all conditions and methods
df <- retrieve_dgm_measures("Stanley2017")

# Retrieve conditions to identify null vs. alternative hypotheses
conditions <- dgm_conditions("Stanley2017")

# Create readable method labels
df$label <- with(df, paste0(method, " (", method_setting, ")"))

# Identify conditions under null hypothesis (H₀: mean effect = 0)
df$H0 <- df$condition_id %in% conditions$condition_id[conditions$mean_effect == 0]

# Create multi-panel visualization
par(mfrow = c(3, 2))
par(mar = c(4, 10, 1, 1))

# Panel 1: Convergence rates
boxplot(convergence * 100 ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(20, 100), 
        data = df, 
        xlab = "Convergence (%)")

# Panel 2: RMSE
boxplot(rmse ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(0, 0.6), 
        data = df, 
        xlab = "RMSE")

# Panel 3: Bias
boxplot(bias ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(-0.25, 0.25), 
        data = df, 
        xlab = "Bias")
abline(v = 0, lty = 3)  # Reference line at zero

# Panel 4: Coverage
boxplot(coverage * 100 ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(30, 100), 
        data = df, 
        xlab = "95% CI Coverage (%)")
abline(v = 95, lty = 3)  # Reference line at nominal level

# Panel 5: Type I Error Rate (H₀ conditions only)
boxplot(power * 100 ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(0, 40), 
        data = df[df$H0, ], 
        xlab = "Type I Error Rate (%)")
abline(v = 5, lty = 3)  # Reference line at α = 0.05

# Panel 6: Power (H₁ conditions only)
boxplot(power * 100 ~ label, 
        horizontal = TRUE, 
        las = 1, 
        ylab = "", 
        ylim = c(10, 100), 
        data = df[!df$H0, ], 
        xlab = "Power (%)")