merTools predictInterval() for model with nested random effect

问题

Does predictInterval() from the merTools package not like nested random effects? For example, using the msleep dataset from the ggplot2 package:

library("lme4")
library("merTools")
library("ggplot2")
mod <- lmer(sleep_total ~ bodywt + (1|vore/order), data=msleep)
predInt <- predictInterval(merMod=mod, newdata=msleep)

Returns an error:

Error in '[.data.frame'(newdata, , j) : undefined columns selected

This runs fine no problem:

mod <- lmer(sleep_total ~ bodywt + (1|vore) + (1|order), data=msleep)
predInt <- predictInterval(merMod=mod, newdata=msleep)

(Well actually it gives a warning about NA levels in the random effect variables, but I'm not concerned about that)

UPDATE

As discussed in the comments of Ben Bolker's answer below, a new version of merTools accounts for nested random effects. However, when I try to predict for data that contains new levels of the nested random effect, I get errors.

This works:

mod <- lmer(sleep_total ~ bodywt + (1|vore/order), data=msleep)
predInt <- predictInterval(merMod=mod, newdata=msleep)

And this works, albeit with a couple of warnings (see below for additional Q about the warnings*):

mod <- lmer(sleep_total ~ bodywt + (1|vore) + (1|order), data=msleep)
msleep2 <- msleep %>% mutate(vore = "omni")
predInt <- predictInterval(merMod=mod, newdata=msleep2)

But this does not work:

mod <- lmer(sleep_total ~ bodywt + (1|vore/order), data=msleep)
msleep2 <- msleep %>% mutate(vore = "omni")
predInt <- predictInterval(merMod=mod, newdata=msleep2)

With the following error:

Error in `[.data.frame`(tmp, alllvl) : undefined columns selected
In addition: Warning message:
In predictInterval(merMod = mod, newdata = msleep3) :
  newdata is tbl_df or tbl object from dplyr package and has been
              coerced to a data.frame

And here, "omni" isn't actually a new level of vore, but when combined with order, it creates new nested combinations of the variables.

If I use "new" or anything else that isn't an observed level of vore, I get similar results: It works for the non-nested version of the model, but not for the nested version.

*Also, should I be concerned about the warning given by the second model chunk above:

> mod <- lmer(sleep_total ~ bodywt + (1|vore) + (1|order), data=msleep)
> msleep2 <- msleep %>% mutate(vore = "omni")
> predInt <- predictInterval(merMod=mod, newdata=msleep2)
Warning messages:
  1: In predictInterval(merMod = mod, newdata = msleep2) :
     newdata is tbl_df or tbl object from dplyr package and has been
       coerced to a data.frame
  2: In chol.default(sigma, pivot = TRUE) :
     the matrix is either rank-deficient or indefinite

I'm going to guess the second one is a result of vore taking on the same value for each observation, but that shouldn't be an issue for prediction, should it? I could see it being an issue if the variable took on the same value when I was fitting the model, but don't think it should be an issue when predicting new observations?

回答1:

One can (apparently) work around this by writing out the interaction term explicitly. Warning: I haven't actually checked to make sure the resulting predictions are correct, just seen that no error is produced and the resulting object is approximately sensible ...

msleep <- transform(msleep,voreOrder=interaction(vore,order,drop=TRUE))
mod2 <- lmer(sleep_total ~ bodywt + (1|vore)+(1|voreOrder), data=msleep)
predInt <- predictInterval(merMod=mod2, newdata=msleep)

This does generate warning messages, but apparently they're due to <NA> values in the vore variable (I don't know this data set ...)

来源：https://stackoverflow.com/questions/35166173/mertools-predictinterval-for-model-with-nested-random-effect

标签

lme4