`afex_plot` not fully compatible with glmm

[mattansb]

When fitting a glmm model (either with lme4::glmer or with afex::mixed), I've found that I cannot use afex_plot to plot the following:

• Original data is not plotted (I suspect due to the link-transformed scale?)
• When trying to plot the response scale (by passing emmeans_arg = list(type = "response")), an error is produced.

> library(afex)
> library(mlmRev)
>
> gm1 <- mixed(use ~ age + I(age^2) + urban + livch + (1 | district), method = 'LRT',
+              family = binomial, data = Contraception)
>
> afex_plot(gm1,~age,emmeans_arg = list(type = "response"),data_plot = TRUE)
Aggregating data over: district
Error in FUN(X[[i]], ...) : object 'y' not found
In addition: There were 50 or more warnings (use warnings() to see the first 50)

I suspect that I am pushing afex_plot to its limit, perhaps beyond its intended use.

Thanks!

[singmann]

This should be solved in the development version: devtools::install_github("singmann/afex@master")

You do not even need to pass the extra emmeans_arg argument. The following should produce a plot:

afex_plot(gm1, "urban", data_plot = TRUE)

However, it still does not print the raw data in the background as the original dv is a factor.

[mattansb]

Great, thank!

I also managed to plot the data:

library(afex)
library(mlmRev)

Contraception <- within(Contraception,use <- as.numeric(use)-1) # for plotting the data

gm1 <- mixed(use ~ age + I(age^2) + urban + livch + (1 | district), method = 'LRT',
             family = binomial, data = Contraception)

afex_plot(gm1,~age,data_plot = TRUE)

[mattansb]

Any chance that the x-axis will not automatically be treated as a factor? Would allow for plotting continuous variables on it?

[singmann]

Unfortunately this is not easily possible. This really requires thinking about and constructing the plot in a very different manner. The reason is that this requires plotting a linear slope or other function, thus completely different geoms, and the usage of emtrends instead of emmeans. So it really would be quite a different function then it is today.

I think your best bet here is to use the effects package, which was developed exactly for that (but is not so convenient for factorial designs for which afex_plot was developed).

library(afex)
library(mlmRev)
gm1 <- mixed(use ~ age + I(age^2) + urban + livch + (1 | district), method = 'LRT',
             family = binomial, data = Contraception, set_data_arg = TRUE)
library(effects)
plot(Effect("age", gm1$full_model))

[mattansb]

Thanks for the suggestion!
I really like how afex_plot also has the original data, and it seems that effects is having some difficulties with this with this a glmm model (I can only guess do to the residuals being put through the link function):

library(afex)
library(mlmRev)
Contraception <- within(Contraception,use <- as.numeric(use)-1) # for plotting the data
gm1 <- mixed(use ~ age + I(age^2) + urban + livch + (1 | district), method = 'LRT',
             family = binomial, data = Contraception, set_data_arg = TRUE)
library(effects)
plot(Effect("age", gm1$full_model,residuals = TRUE))

I guess I didn't think about the need to change geoms etc. Though I think it might still be enough to just coerce the x back to continuous via as.numeric(as.charachter(x)) and keep the discrete geoms (doesn't look too bad):

afex_plot(gm1,~age,data_plot = TRUE, emmeans_arg = list(cov.red = FALSE))

So for the time being... If I ever need a continuous x-axis, I will use with afex_plot with return = "data", and reconstruct my plot (in a very very very roundabout way):

library(ggplot2)
X <- afex_plot(gm1,~age,data_plot = TRUE, emmeans_arg = list(cov.red = FALSE),return = 'data')
ggplot(mapping = aes(as.numeric(as.character(x)),y)) + 
  geom_point(data = X$data, alpha = .1) + 
  geom_ribbon(aes(ymin = lower, ymax = upper),data = X$means, color = NA, alpha = .4) + 
  geom_line(data = X$means)