plot a subset for 4-way-interactions

[qwang178]

I am trying to plot a subset of the data for effects from an LME model with 4 way interactions, and found the legend doesn't work as desired. Taking the efc data as an example:

fit model with 4-way-interaction

library(ggeffects)
data(efc)
fit <- lm(neg_c_7 ~ c12hour * barthtot * c161sex * c172code, data = efc)
pr <- ggpredict(fit, c("c12hour[0,80,160]", "barthtot[30,50,70]", "c161sex", "c172code"))
plot(pr) + ggplot2::theme(legend.position = "bottom")

This shows the plots well with three panels. However, when I subset the object and plot them:

pr_subset<-pr[c(1,18,19,36,37,54)]
pr_subset

Predicted values of Negative impact with 7 items

barthtot = 30
c161sex = [1] Male
c172code = 1

c12hour | Predicted | 95% CI
0 | 15.75 | [13.05, 18.45]
80 | 13.71 | [12.13, 15.29]
160 | 11.67 | [ 8.92, 14.43]

barthtot = 70
c161sex = [2] Female
c172code = 3

c12hour | Predicted | 95% CI
0 | 11.46 | [10.72, 12.19]
80 | 13.20 | [12.21, 14.19]
160 | 14.95 | [12.80, 17.10]

plot(pr_subset)

This shows the two panels, but cannot correctly color/label the BARTHEL index. Is there a way to address the problem?

[bbolker]

Could we please have a reproducible example? (Also, your question will be more readable if you edit it to put your code inside a code block, using triple-back-quotes (```) as delimiters ...)

[qwang178]

Thanks for the quick response! The above example is from the efc data used in the package. The code line has been updated for clarity.

[strengejacke]

That's tricky. For four stratification characteristics (i.e. 4 focal terms), predictions are subset by panel:

library(ggeffects)
data(efc)
fit <- lm(neg_c_7 ~ c12hour * barthtot * c161sex * c172code, data = efc)
pr <- ggpredict(fit, c("c12hour[0,80,160]", "barthtot[30,50,70]", "c161sex", "c172code"))
pr_subset <- pr[c(1, 18, 19, 36, 37, 54)]

panels <- unique(pr_subset$panel)
lapply(seq_along(panels), function(.i) {
  as.data.frame(pr_subset[pr_subset$panel == panels[.i], ])
})
#> [[1]]
#>      x predicted std.error  conf.low conf.high group    facet panel
#> 1    0  15.74648 1.3750517 13.047342  18.44562    30 [1] Male     1
#> 19  80  13.70928 0.8032895 12.132472  15.28609    30 [1] Male     1
#> 37 160  11.67208 1.4039684  8.916174  14.42798    30 [1] Male     1
#> 
#> [[2]]
#>      x predicted std.error conf.low conf.high group      facet panel
#> 18   0  11.45581 0.3737412 10.72218  12.18944    70 [2] Female     3
#> 36  80  13.20333 0.5039332 12.21414  14.19252    70 [2] Female     3
#> 54 160  14.95084 1.0933392 12.80469  17.09700    70 [2] Female     3

The color aesthetic is based on the group column, but each plot now has just "one group", which results in the wrong legend.

The good thing is we know when we have a subset data frame of predictions:

attributes(pr_subset)$is_subset
#> [1] TRUE

Based on this information, it might be possible to tweak the color-aes, though I'm not sure right now how to do it. Must look a bit more into ggplot2 to figure this out.