README.md

In JonasMoss/straussR: What the Package Does (One Line, Title Case)

straussR

Note: This project has no pre-release yet! If you're interested in a working package, you will have to wait.

Description

straussR is an R-package for Bayesian meta-analysis that corrects for publication bias and p-hacking. Its main features are:

• Explicit modeling and correction for publication bias and p-hacking. The degree of p-hacking can depend on covariates.
• The random effects distribution can be chosen among possibly non-normal alternatives.
• Support for covariates in all the parameters of the random effects distribution, and a large number of links.
• Flexible choice of priors for all parameters.

The sampling is done in STAN.

Example usage

Here's an artificial example with an imaginary data set. x and y are two covariates, z are the z-values of the studies, and n contains the (scaled) number of participants in each study. The effect size distribution is gumbel, which is a more reasonable choice than the normal if we suspect the effects to be skewed. p is the propensity to p-hack, and lives on the unit interval. I use the term probit(p) ~ 1 + n since the p-hacking propensity should decrease with the study size, due to the intuition that large studies get published no matter what.

formula = z ~ gumbel(mean ~ 1 + x, 
                     log(sd) ~ 1 + y, 
                     probit(p) ~ 1 + n)

priors = list(mean = list((Intercept) ~ gamma(2, 1),
                          x ~ weibull(2, 3),
              sd   = list((Intercept) ~ normal(0, 1)),
                          y ~ gumbel(0, 1),
              p    = list((Intercept) ~ student_t(2, 0, 1)),
                          n ~ skew_normal(0, 1, 10))

straussR(formula = formula, data = data, priors = priors)

JonasMoss/straussR documentation built on May 17, 2019, 7:02 p.m.