The package is on CRAN, so if you have R to install from CRAN by default, you can just:

install.packages("sjstats")

Dear Khelifi,

as Salvatore already mentioned, it is on CRAN. Once you have installed the package and all its dependencies, here is the documentation for all the functions of this powerful and comprehensive add-on:

https://cran.r-project.org/web/packages/sjstats/sjstats.pdf

There is also a lot of info and help available online...

BW

Immo

Hi Salvatore and Immo, thank you for your answer.

In fact, "sjstats" does not work with Rstudio (which i use habituel) but only with R (i don't know why...)

However, i am disappointed because "sjstats::p_value()" gives results similar to "summary()" function (used with glmer), that is p-value associated to the different levels of explanatory variables and not for example to factor1 x factor 2 x Factor 3 interaction (as whole).

Using Anova(m0, ddf="kenw") does not change anything. I'm still looking for an effective solution…

The sjstats package seems to work fine for me in RStudio.

Have you tried the lmerTest package? Does that not give an anova-like table with glmer models?

If all of your independent variables are factor variables, you might try emmeans::joint_tests . The package supports glmer model objects (https://cran.r-project.org/web/packages/emmeans/vignettes/models.html)

Finally, are you familiar with the idea of using anova() with two nested models? That is, if want to get p-value for X, you can compare Y ~ X + Z + (1|Subj) to Y ~ Z + (1|Subj) with anova(). I think in lme4, that will use likelihood ratio tests.

Also, see the lme4 documentation on p-values: https://www.rdocumentation.org/packages/lme4/versions/1.1-18-1/topics/pvalues

I really don't know why it doesn't work for me with RStudio.

I tried emmeans::joint_tests, but it gives only one ddl insted of the two (the other is reported as "inf").

i have used "likelihood ratio tests"in a submited article but one reviewer tell us to use GLMM because it is possible to skewed data without long transformation.

Ìt is strange that the use of anova with "lmer" give p-values while glmer does not?

It's probably correct that emmeans reports inf for the denominator degrees of freedom. It's the way it calculates p-values for some model types.

Hmm... for the other issues, I probably have to play with it and see what output the different approaches produce.

That is very kind of you!

• It appears that lmerTest doesn't return p-values for glmer models, but only for lmer models. (https://cran.r-project.org/web/packages/lmerTest/lmerTest.pdf).
• The documentation for car::Anova mentions glm and lmer, but not glmer. It will return p-values for glmer, though. In this case it appears the only option for the test statistic is Wald ("Chisq"). (?car::Anova).
• emmeans supports glmer objects. (https://cran.r-project.org/web/packages/emmeans/vignettes/models.html). emmeans::joint_tests returns p-values only for factor terms, via F-test. Inf for the denominator degrees of freedom is normal. (?emmeans::joint_tests).
• The most usual course to determine p-values for terms is to use anova(model1, model2) for nested models. This will return likelihood ratio tests. When models are complex, it can be a little tricky to figure out how to specify all the models to compare.
• I don't have much experience with the afex package, but it looks promising. By default, it returns Type III, likelihood ratio tests for effects. (https://cran.r-project.org/web/packages/afex/afex.pdf).

Sample R code:

set.seed(42)

n = 24

X = factor(c(rep("X1",n/2), rep("X2",n/2)))

Z = factor(rep(c("Z1", "Z2"), n/2))

m = n/4

Subj = factor(rep(letters[1:m], 2, each=2))

Y = as.numeric(X) + as.numeric(Z) + as.numeric(Subj) + rnorm(n,0,0.5)

Data = data.frame(Subj,X,Z,Y)

library(lme4)

model = glmer(Y ~ X + Z + (1|Subj), family= Gamma)

### Log-likelihood tests on nested models

model.noX = glmer(Y ~ Z + (1|Subj), family= Gamma)

model.noZ = glmer(Y ~ X + (1|Subj), family= Gamma)

anova(model, model.noX)

anova(model, model.noZ)

### glmer models are not listed as supported by car::Anova

### Wald tests

library(car)

Anova(model)

### glmer models are supported by emmeans

### See ?joint_tests for discussion

### F tests, type III

library(emmeans)

joint_tests(model)

### afex package

### I believe it refits the model with the lme4 package

library(afex)

model.afex = mixed(Y ~ X + Z + (1|Subj), data = Data, method = "LRT", family = Gamma)

anova(model.afex)

Thank you very much for all those information!

I think that the most appropriate solution for me is to use joint_tests() function.

I tried to use method = "LRT", it returns indeed a p-value. I think I need some time to understand really all those possibilities.

Thank you again Salvatore