Negative WAIC

[Leon-Alph]

Dear all,
I'm using the WAIC measure to compare the goodness of fit of some model. For some of them, the WAIC is negative. I don't understand how I can interpret such a result.
Any idea ?

[Richard McElreath]

It is not necessarily a problem for WAIC or any other criterion to be negative. The same thing happens with AIC. Smaller is still better.

WAIC is scaled like deviance, and deviance is scaled as -2 log-probability. For a continuous probability density, like Gaussian, it is easy for deviance to be negative. Consider for example:

-2*dnorm(0.1,0,0.1,TRUE)

[Jonah]

Like Richard says, it's fine for it to be negative. There's nothing good or bad about a negative value, it just happens to be negative, like in Richard's example. The particular values you get for different models are not what's of interest but rather how those values relate to each other. 

[Leon-Alph]

Thank you for this clear answer.