variancemodel-fittingglmmtmb

Strange values for r2_nakagawa() with easystats

I have been evaluating the quality of modelfit with the easystats package (0.7.1.3). Since I have mixed models I use the r2_nakagawa() function and since the last update I did (easystats::install_latest(force = TRUE)) this function gives me strange values.

For example for this model:

m2<-glmmTMB(FvFm ~ hora*temperatura + (1|Experimento), data=d, family=beta_family(link="logit"))

I used to get these values:

r2_nakagawa(m2)
# R2 for Mixed Models
Conditional R2: 1.000
Marginal R2: 0.973

and now (with the latest update) it gives me these values:

# R2 for Mixed Models
Conditional R2: 1.341
Marginal R2: 1.306

I don't know what's wrong, but I have the latest version of easystats and performance and insight.

Someone is having these problems, I think it is not code. Any help is welcome and appreciated!!

I tried to install previous versions of easystats, but the values for R2 did not change.

Solution

  • Please see this issue on GitHub for the insight package, which is internally used to compute the r-squared values: https://github.com/easystats/insight/issues/889

    We have revisited the function and intensively tested against all examples from the Nakagawa et al. paper and against results from other R packages that compute r-squared for mixed models.

    Thereby, we finally could improve the accuracy (or correct the computation) for some families, including the Beta-family. You can check your results by installing the latest development versions from GitHub, which you can do in two ways:

    1. installing packages separately from r-universe:
    install.packages("insight", repos = "https://easystats.r-universe.dev")
install.packages("performance", repos = "https://easystats.r-universe.dev")
    1. or run easystats::install_latest(force = TRUE).

    Then try r2_nakagawa(m2) again. The update documentation for r2_nakagawa() is also much clearer about the validated and working-but-not-fully-validated model families.

    Furthermore, it is now possible to calculate r-squared values based on different approximations for the distribution-specific variance component, using log-normal, delta or trigamma approximations (as proposed by Nakagawa et al., and also implemented in the MuMIn package); see the new approximation argument.

    References