Skip to contents

Bayesian Evidence for Meta-Regression Coefficients

Source: R/meta_reg_helpers.R
coefficient_evidence.Rd

Computes Bayesian measures of evidence for moderator coefficients, including probability of direction, null range (Region of Practical Equivalence), and credible intervals.

Usage

coefficient_evidence(object, null_range = NULL, ci_level = 0.95)

Arguments

object

A bayesma_reg object.

null_range

Numeric vector of length 2 defining the null range, or NULL (default) to skip null range calculation. For standardized coefficients, c(-0.1, 0.1) is a common choice.

ci_level

Numeric. Credible interval width (default: 0.95).

Value

A tibble with columns:

  • term: Coefficient name

  • estimate: Posterior median

  • std_error: Posterior SD

  • ci_lower, ci_upper: Credible interval bounds

  • pd: Probability of direction (max of P(β>0), P(β<0))

  • direction: Most probable direction ("positive" or "negative")

  • p_null: Proportion of posterior inside null range (if specified

  • p_outside_null: Proportion outside null range (if specified)

Details

Probability of Direction (pd)

The probability of direction is the proportion of the posterior on the same side of zero as the median. It ranges from 0.5 (no evidence) to 1 (strong evidence for direction). A pd > 0.95 is often considered meaningful evidence, though this is not a formal threshold.

null range (Region of Practical Equivalence)

The null range defines a range of values considered "practically equivalent to zero". The proportion of the posterior inside the null range indicates whether the effect is practically meaningful:

  • p_null > 0.95: Effect is practically zero (accept null)

  • p_null < 0.05: Effect is practically meaningful (reject null)

  • Otherwise: Inconclusive

Common null range choices:

  • c(-0.1, 0.1) for standardized coefficients (Cohen's d scale)

  • c(-0.05, 0.05) for log-OR (corresponds to OR 0.95-1.05)

Examples

if (FALSE) { # \dontrun{
fit <- meta_reg(data, studyvar = "author", yi = "yi", vi = "vi",
                mods = ~ year + quality)

# Basic evidence summary
coefficient_evidence(fit)

# With null range
coefficient_evidence(fit, null_range = c(-0.1, 0.1))
} # }