Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/namakala/inappropriate-acid-suppressor-agent-use/llms.txt

Use this file to discover all available pages before exploring further.

Meta-regression extends the random-effects model by adding study-level covariates as predictors of the effect size, allowing you to quantify how much of the between-study heterogeneity each variable explains. The pipeline fits five univariable models (one per moderator) and a single multivariable model, all using metafor::rma() on the logit-transformed prevalence. Targets for the univariable models are generated automatically via tar_map(); the multivariable model is a single named target.

​
Univariable meta-regression

fitMetareg() in src/R/meta-analysis.R handles both the univariable and multivariable cases through a single function: 
fitMetareg <- function(tbl, varname, ...) {
  require("metafor")
  require("meta")

  sub_tbl <- tbl |>
    subset(
      subset = var_logit_prevalence != Inf,
      select = c("logit_prevalence", "var_logit_prevalence", varname)
    ) |>
    na.omit()

  is_single <- length(varname) == 1

  if (is_single) {
    form <- paste("~", varname)
  } else {
    form <- paste("~", paste(varname, collapse = " + "))
  }

  form %<>% as.formula()

  mod <- metafor::rma(
    yi     = logit_prevalence,
    vi     = var_logit_prevalence,
    mods   = form,
    data   = sub_tbl,
    method = "REML",
    ...
  )

  return(mod)
}
The function:
  1. Drops rows with infinite sampling variance (var_logit_prevalence == Inf) — these arise when a study reports a prevalence of exactly 0 or 1.
  2. Removes any remaining rows with NA in the outcome or the moderator columns via na.omit().
  3. Builds a one-sided formula from varname; a single string produces ~ Age, while a vector produces ~ Year + JBI_Classification + ....
  4. Passes logit prevalence as the outcome (yi) and its sampling variance (vi) to metafor::rma() with method = "REML".
The five univariable models correspond to the same uni_vars used for subgroup analysis: 
uni_vars <- c(
  "Age", "Continent", "Setting", "JBI_Classification", "use_guideline"
)
_targets.R expands these into individual targets with tar_map(): 
tar_map(
  values = list("varname" = uni_vars),
  unlist = FALSE,
  tar_target(mod_metareg, fitMetareg(tbl_clean, varname))
)
This creates mod_metareg_Age, mod_metareg_Continent, mod_metareg_Setting, mod_metareg_JBI_Classification, and mod_metareg_use_guideline.

​
Multivariable meta-regression

The multivariable model includes all predictors believed to drive heterogeneity, specified in mv_vars: 
mv_vars <- c(
  "Year", "JBI_Classification", "use_guideline", "Setting", "Continent",
  "Sample_size"
)
A single target fits this model: 
tar_target(mod_metareg_mv, fitMetareg(tbl_clean, mv_vars))
Because fitMetareg() detects length(varname) > 1, it builds the formula ~ Year + JBI_Classification + use_guideline + Setting + Continent + Sample_size automatically — no separate function is needed. The multivariable model jointly estimates the contribution of each predictor while adjusting for the others, and the residual I² indicates how much heterogeneity remains unexplained.

​
Accessing results

targets::tar_load(mod_metareg_mv)
summary(mod_metareg_mv)
For a univariable model: 
targets::tar_load(mod_metareg_Age)
summary(mod_metareg_Age)

​
Interpreting meta-regression output

metafor::rma() returns a rma object. The summary() output reports:
FieldInterpretation
b (intercept)Expected logit prevalence when all moderators are 0 / reference level
b (slope)Change in logit prevalence per unit increase in the moderator
se, zval, pvalStandard error, z-statistic, and two-sided p-value for each coefficient
R2Proportion of τ² explained by the moderator(s)
I2Residual heterogeneity after accounting for moderators
tau2Residual between-study variance
The outcome variable is logit-transformed prevalence. Regression coefficients are therefore on the logit scale. To convert a predicted value back to a probability, apply plogis():
# Predicted logit for a reference-category study
predicted_logit <- coef(mod_metareg_mv)[["intrcpt"]]
plogis(predicted_logit)  # back-transforms to [0, 1]
Slopes represent additive shifts on the logit scale; for small effects you can approximate the change in probability, but for large coefficients use plogis() evaluated at specific covariate values.

Build docs developers (and LLMs) love

Get started for freeTalk to us