BICBF: Use delta-BIC to approximate Bayes Factors for model...

In akcochrane/ACmisc: Aaron Cochrane's miscellaneous functions

Use delta-BIC to approximate Bayes Factors for model comparison

Description

Calculates BIC for the entire model, and for the entire set of models dropping each fixed effect one at a time. Uses the change in BIC between models to approximate Bayes Factors (essentially penalized Likelihood Ratios, or evidence for one model over another). This approach is promoted by, for example, Kass and Raftery (1995; 10.1080/01621459.1995.10476572) and Wagenmakers (2007; 10.3758/BF03194105).

Usage

BICBF(modIn, logBase = 3)

Arguments

modIn	lm, glm, rlm, lmer or glmer model.
logBase	Base of the logBF

Details

NOT intended for relations between two variables (in that case, use robustCor or similar).

Could use further testing to ensure that formula extraction is correct.

Follows the recommendations of Wagenmakers (2007; PBR), e.g., for LMEMs calculates BIC using -2*logLik(modIn) + attr(logLik(modIn),'df')*log(sum(ngrps(modIn)))

Note

BIC_dropped of the Intercept term is the full model BIC. BIC_dropped of all other coefficients are the BIC of the model without that coefficient. See comment(output) for more information.

Known Bug preventing use with non-multi-level models. In other words, it is only currently functional for multi-level models (i.e., lmer() or glmer())

Examples

d <- data.frame(y = rnorm(60),grup = sample(c('a','b','c','d'),60,replace = T),arb = sample(c('e','f','g','h'),60,replace = T))
d$x  <- d$y + rnorm(60,0,.2)

m_lmer <- lme4::lmer(y~x*arb+(1|grup),d,REML = F)
BICBF(m_lmer)

m_glmer <- lme4::glmer(y~x*arb+(1|grup),d,family=gaussian())
BICBF(m_glmer)

m_lm <- lm(ravens ~ alert * corsi + isAdult, dat_cochraneEtAl_2019_PLOSOne)
BICBF(m_lm)

m_rlm <- MASS::rlm(ravens ~ alert * corsi + isAdult, dat_cochraneEtAl_2019_PLOSOne)
BICBF(m_rlm)

m_glm <- glm(y ~ x*arb,d,family=gaussian())
BICBF(m_glm)

akcochrane/ACmisc documentation built on Nov. 24, 2024, 11:22 a.m.