[R-sig-ME] MCMCglmm predict failing with more than one fixed effect
Tom Houslay
hou@|@y @end|ng |rom gm@||@com
Wed Sep 4 13:02:27 CEST 2019
Hi all,
Hoping someone can help me with this issue, even (especially?) if I'm
just being an idiot. I had been using predict.MCMCglmm with new data
successfully until recently (I think coinciding with updating R and
various packages), and now if I have more than one fixed effect in a
model then it fails when supplying any newdata. The error message is:
Error in predict.MCMCglmm(mcmc_ss_2, newdata = as.data.frame(df_ss_pred_2)) :
model for newdata has fixed effects absent from the original model
I've tried a load of different ways of building the data to predict
from, but nothing seems to work.
Code below for a reproducible example (requires MCMCglmm and the
'sleepstudy' data from lme4). I'm running R 3.6.1 and MCMCglmm 2.29.
Thanks!
Tom
# Basic model of sleepstudy with 1 fixed effect and simple random intercept
mcmc_ss_1 <- MCMCglmm(Reaction ~ Days,
random =~ Subject,
data = sleepstudy)
summary(mcmc_ss_1)
# Predicts fine on original data
predict(mcmc_ss_1)
# Create data on which to predict
df_ss_pred <- data.frame(Days = 0:9,
Subject = 350,
Reaction = 0)
# This prediction works
predict(mcmc_ss_1,
newdata = df_ss_pred)
# Create new data frame with fake additional column
sleepstudy_2 <- cbind(sleepstudy,
extra = rnorm(nrow(sleepstudy)))
# Run model with additional fixed effect
mcmc_ss_2 <- MCMCglmm(Reaction ~ Days + extra,
random =~ Subject,
data = sleepstudy_2)
summary(mcmc_ss_2)
# Predict works fine from original data
predict(mcmc_ss_2)
# Create data frame on which to predict,
# as before but now with new column added
df_ss_pred_2 <- as.data.frame(expand.grid(Reaction = 0,
Days = 0:9,
Subject = 350,
extra = mean(sleepstudy_2$extra)))
# Prediction fails
predict(mcmc_ss_2,
newdata = df_ss_pred_2)
# Check model formula
mcmc_ss_2$Fixed$formula
# Check variable names in data frame
str(sleepstudy_2)
# Check variable names in new data frame
str(df_ss_pred_2)
More information about the R-sig-mixed-models
mailing list