Dear all,
I am using linear mixed effect models in my experiment (using r package). I would like to show a measurement of the fitting quality of the model. I see that there are lot of methods but two of the most frequently are:
1) Xu, R. (2003). Measuring explained variation in linear mixed effects models. Statistics in medicine, 22(22), 3527-3541.
2) Nakagawa, S., & Schielzeth, H. (2013). A general and simple method for obtaining R2 from generalized linear mixed‐effects models. Methods in Ecology and Evolution, 4(2), 133-142.
I have tried to use those aproach to my data obtaining different results. Please look:
Approach 1. Method from Nakagawa et al.(2012) basing on this post: https://ecologyforacrowdedplanet.wordpress.com/2013/08/27/r-squared-in-mixed-models-the-easy-way/
> r.squaredGLMM(reves_m1)
R2m R2c
[1,] 0.02128424 0.5280393
Approach 2. Method from Xu(2013) basing on this post:https://stats.stackexchange.com/questions/95054/how-to-get-an-overall-p-value-and-effect-size-for-a-categorical-factor-in-a-mi
> 1-var(residuals(reves_m1))/(var(model.response(model.frame(reves_m1))))
[1] 0.5338557
As you can see R^2 marginal from the Nakagawa formula is very similar to the R^2 obtained with the formula of Ronghui Xu but in other post I have read that it is better to use the R^2 marginal from the Nagawaka formula. Could someone give me some idea of the differences between R^2 marginal and R^2 conditional? Which one I should comment in my paper if I would like to talk about how the model fits? In my case the R^2 conditional is moreless high (0.52) but the R^2 marginal is very low (0.02)
Thanks very much,
Gabriel Delgado
You may want to look at the function "get_variance()" from the "insight"-package (see https://easystats.github.io/insight/ resp. https://easystats.github.io/insight/reference/get_variance.html).
This function extracts all variance-components of mixed models and supports different packages (like lme4, glmmTMB, GLMMadaptive, lme, rstanarm, afex, ...). The code is based on the latest proposals from Nakagawa et al 2017 and Johnson 2014.
There is currently a package in development that will be based on this functions to compute R2 and ICC (intraclass correlation coefficient) for mixed models, but you can easily compute the values by yourself for now, which will also help you understand what the R2 values mean:
m
Take a look at page 53 of this url: https://cran.r-project.org/web/packages/MuMIn/MuMIn.pdf
Very interesting. I have just read, but I need to study it more deeply. Why I always have a r conditional very near to 0.5? It must be an error because r marginal fluctuates a lot...
You may want to look at the function "get_variance()" from the "insight"-package (see https://easystats.github.io/insight/ resp. https://easystats.github.io/insight/reference/get_variance.html).
This function extracts all variance-components of mixed models and supports different packages (like lme4, glmmTMB, GLMMadaptive, lme, rstanarm, afex, ...). The code is based on the latest proposals from Nakagawa et al 2017 and Johnson 2014.
There is currently a package in development that will be based on this functions to compute R2 and ICC (intraclass correlation coefficient) for mixed models, but you can easily compute the values by yourself for now, which will also help you understand what the R2 values mean:
m
Hello Daniel,
Thanks very much... Very clear reply. I will try the code that you bring me this morning... I should obtain the same results than with the code r.squaredGLMM(reves_m1) ?
I am a little scared that I am doing something wrong because I obtain very similar r^2 conditional for every variable I instert in the model (a total of 9 variables)...
I will try your code to see if I obtain more logical results...
Thanks very much again
> I should obtain the same results than with the code r.squaredGLMM(reves_m1) ?
Yes, maybe they differ after some decimals, but they should roughly be the same.
A short update: the performance-package was released recently, which wraps the above mentioned functions and provides an "r2()" function for regression models, including mixed models. For mixed models, "r2()" internally calls "r2_nakagawa()" (https://easystats.github.io/performance/reference/r2_nakagawa.html). Due to the latest package updates, "r2_nakagawa()" (or "r2()") now also computes the best approximated r-squared for mixed models with zero-inflation - something, which other packages currently do not (afaik).
Oh that is very interesting to know. Thanks very much. I will try this package
- Badges
- Science topic
- Similar topics
- Papers