Skip to contents

Set default prior distributions

Source: R/priors.R
set_default_priors.Rd

Set default prior distributions for RoBMA models.

Usage

set_default_priors(parameter, null = FALSE, rescale = 1)

Arguments

parameter

a character string specifying the parameter for which the prior distribution should be set. Available options are "effect", "heterogeneity", "bias", "hierarchical", "covariates", "factors".

null

a logical indicating whether the prior distribution should be set for the null hypothesis. Defaults to FALSE.

rescale

a numeric value specifying the re-scaling factor for the default prior distributions. Defaults to 1. Allows convenient re-scaling of prior distributions simultaneously.

Value

A prior distribution object or a list of prior distribution objects.

Details

The default prior distributions corresponds to the specification of RoBMA-PSMA and RoBMA-regression outlined in bartos2021no;textualRoBMA and bartos2023robust;textualRoBMA.

Specifically, the prior distributions are:

For the alternative hypothesis:

  • Effect: Normal distribution with mean 0 and standard deviation 1.

  • Heterogeneity: Inverse gamma distribution with shape 1 and scale 0.15.

  • Bias: A list of 8 prior distributions defining the publication bias adjustments:

    • Two-sided: Weight function with steps 0.05.

    • Two-sided: Weight function with steps 0.05 and 0.1.

    • One-sided: Weight function with steps 0.05.

    • One-sided: Weight function with steps 0.025 and 0.05.

    • One-sided: Weight function with steps 0.05 and 0.5.

    • One-sided: Weight function with steps 0.025, 0.05, and 0.5.

    • PET-type model with regression coefficient: Cauchy distribution with location 0 and scale 1.

    • PEESE-type model with regression coefficient: Cauchy distribution with location 0 and scale 5.

    All weight functions use a unit cumulative Dirichlet prior distribution on relative prior probabilities.

  • Standardized continuous covariates: Normal distribution with mean 0 and standard deviation 0.25.

  • Factors (via by-level differences from the grand mean): Normal distribution with mean 0 and standard deviation 0.25.

For the null hypothesis:

  • Effect: Point distribution at 0.

  • Heterogeneity: Point distribution at 0.

  • Bias: No prior distribution.

  • Standardized continuous covariates: Point distribution at 0.

  • Factors (via by-level differences from the grand mean): Point distribution at 0.

The rescaling factor adjusts the width of the effect, heterogeneity, covariates, factor, and PEESE-style model prior distributions. PET-style and weight function prior distributions are scale-invariant.

Examples


set_default_priors("effect")
#> Normal(0, 1)
set_default_priors("heterogeneity")
#> InvGamma(1, 0.15)
set_default_priors("bias")
#> [[1]]
#> omega[two-sided: .05] ~ CumDirichlet(1, 1)
#> [[2]]
#> omega[two-sided: .1, .05] ~ CumDirichlet(1, 1, 1)
#> [[3]]
#> omega[one-sided: .05] ~ CumDirichlet(1, 1)
#> [[4]]
#> omega[one-sided: .05, .025] ~ CumDirichlet(1, 1, 1)
#> [[5]]
#> omega[one-sided: .5, .05] ~ CumDirichlet(1, 1, 1)
#> [[6]]
#> omega[one-sided: .5, .05, .025] ~ CumDirichlet(1, 1, 1, 1)
#> [[7]]
#> PET ~ Cauchy(0, 1)[0, Inf]
#> [[8]]
#> PEESE ~ Cauchy(0, 5)[0, Inf]