MetaJAGS

This package provides functions to easily implement custom hierarchical Bayesian meta-regressions using RJAGS. These models are introduced and described in detail in Zapparrata (2024). This package is currently in early beta, and only three main functions. Additional functions and features, as well as refinements will be added frequently. Please download the most recent GitHub version to have the most up-to-date version.

Installation

Instructions on how to install the package from GitHub:

# Install the development version from GitHub:
devtools::install_github("Donnangravelle/MetaJAGS")

JAGSmeta

JAGSmeta() is the main function provided by this package. It fits the hierarchical Bayesian meta-regression, as described in Zapparrata (2024).

At the moment, the function has 8 parameters.

formula: The model formula. This follows standard model formula syntax in R. E.g., yi ~ 1 + predictor_1 + predictor_2

v: The column that contains the variances of the effect sizes to be estimated.

studyID: The column that contains the unique study identifiers that effects will be nested within.

data: The data frame object that contains the effect sizes, predictors, variances, and study identifers. At the moment, the function should only work with data frame objects, not with seperate vectors containing this information.

n_chains: The number of separate sampling chains that will be used to simulate the distribution of effects.

n_updates: The number of times the sampling model will update and refit.

n_iterations: The number of iterations used when sampling.

prob_conditions: This allows users to estimate the probability of a model parameter being of a certain value. Takes as an arugment a list of model parameters and a list of values per parameter. Currently, there are three options for estimation: “greater”: Probability that the model parameter is greater than some value. “lesser”: Probability that the model parameter is less than some value. “Between”: Probability that the model parameter is between two values.

Example

 result <- JAGSmeta(
 formula = yi ~ 1 + moderator_1 + moderator_2, # Model formula
 v = "vi", # Column name for variance
 studyID = "study_id", # Column name for study identifier
 data = example_data, # The dataset
 n_chains = 5, # Number of MCMC chains
 n_updates = 100000, # Number of updates for burn-in
 n_iterations = 100000 # Number of iterations for sampling
  prob_conditions = list(
 "mu.alpha" = list(
   "greater" =c(0,.2, .5, .8)),
 "sd.alpha" = list(
   "lesser" = .1,
   "between" = list(c(.1,.2), c(.2,.5)),
   "greater" = .5)
   )
)

Features in Development

These are not in any order.

1. The ability to specify priors for each model parameter.

Model parameters are currently implemented by uniformative priors. Users will be able to specfiy prior distributions, such as Cauchy, half-Cauchy, and t-distributions to fully make use of the power of Bayesian analyses.

1. Improved model output.

At the moment, model parameters are outputted in hard to interpret verbiage and layout. Model output will be adjusted to become easier to understand.

1. More flexible modeling options

Currently, the model is unable to handle interactions.

JAGS.summary

This is a supplementary function for JAGSmeta. It simplifies the model output to ignore parameters estimated at the study level. Takes as an arugment the JAGSmeta model.

Example

JAGS.summary(JAGSmeta_model)

BayesForest

This allows for the visualization of the distribution of effect size estimates at the study level. The function uses the ggplot2 and ggridges functions as a foundation. It takes several aesthetic arguments. The main arguements are the model that was fitted, the original data, and the column containing the study information to be displayed on the y-axis.

Example

BayesForest(model, study_info = "Study_Info",
data = example_data, dens_color = "skyblue",
point_color = "black", overlap = 1,
CI_offset = .3, y_text_size = 14, CI_size = 5, x_text_size = 14,
x_title_size = 14, y_title_size = 14)

Features in Development

At the moment, the function uses fixed sizes and positions for much of the text. This means that users need to play with the sizes and positions to work with their display. Future versions will make positions and text sizes relative to the native R display by default.

Functions and Features in the Pipeline

There are several features currently in development. Some examples of include the ability to detect and correct for publication bias within a Bayesian framework, additional options for visualizing the analysis, and the ability to easily conduct a sensitivity analysis.