logLik_lnorm: Information criteria for log-normal models

In mattansb/MSBMisc: Some functions I wrote that I find useful

Information criteria for log-normal models

Description

Log-likelihood (and by extension AIC and BIC) for log-normal models fit with stats::lm(log(y) ~ ...) are computed with stats::dnorm(log(y), ...) instead of with stats::dlnorm(y, ...), which makes comparing different families difficult. This function is aimed at rectifying this. See examples.

Usage

logLik_lnorm(object, REML = FALSE)

AIC_lnorm(object, k = 2, REML = FALSE)

BIC_lnorm(object, REML = FALSE)

Arguments

object	A fitted model object. The model must meet all of the following (will throw an error if not met): A Gaussian likelihood with an identity link function. The LHS of the model's formula must use the log() function. No weights (not yet supported).
REML	Only FALSE supported.
k	numeric, the penalty per parameter to be used; the default k = 2 is the classical AIC.

Note

REML is not (yet) supported. Make sure you are comparing correct LL/AIC/BIC values.

Examples

data("mtcars")
mtcars$mpg <- floor(mtcars$mpg)

model_lnorm <- lm(log(mpg) ~ factor(cyl), mtcars)
model_norm <- lm(mpg ~ factor(cyl), mtcars)
model_pois <- glm(mpg ~ factor(cyl), mtcars, family = poisson())

# No, that first one is wrong...
(aics <- AIC(model_lnorm, model_norm, model_pois))

aics[1, "AIC"] <- AIC_lnorm(model_lnorm)
aics # better!


# Should support any model really... =====================
model_lnorm <- lme4::lmer(log(mpg) ~ factor(cyl) + (1 | gear), mtcars, REML = FALSE)
model_norm <- lme4::lmer(mpg ~ factor(cyl) + (1 | gear), mtcars, REML = FALSE)
model_pois <- lme4::glmer(mpg ~ factor(cyl) + (1 | gear), mtcars, family = poisson())

# No, that first one is wrong...
(aics <- AIC(model_lnorm, model_norm, model_pois))

aics[1, "AIC"] <- AIC_lnorm(model_lnorm)
aics # better!

mattansb/MSBMisc documentation built on March 22, 2023, 6:02 p.m.