Post-Hoc Test with lsmeans error with uncorrelated random effects

[René Sch.]

Dear Henrik,

I have a simple model, a categorical predictor with 3 levels and find a sign. effect, and now want to test, where this effect comes from by using lsmeans::pairs (or ::contrast) post-hoc tests (same happens with the “emmeans” package).

But when I run a mixed model (any dataset) with uncorrelated random intercepts and slopes, e.g.
mod1<-mixed(y~x+(1+x||subjects), method=”LRT”, data=data, expand_re=TRUE)

and afterwards do
lsm.options(lmer.df = “Satterthwaite”)
lsmeans(mod1, ~ x, data=mod1$data)

I get the following error:
Error in eval(predvars, data, env) : Object ‘re1.x1’ not found

Can you please give me a hint, or is there an alternative? I need this post – hoc test desperately 🙂

Best, René

• This topic was modified 6 years, 8 months ago by René Sch..

[henrik]

One surprising solution (surprising in the sense that I only know about it for a few days) is to set set_data_arg = FALSE. In the current CRAN version (but not in the current development and next CRAN version), this throws quite a few errors which can be safely ignored.

library("afex")
data("Machines", package = "MEMSS") 
m1 <- mixed(score ~ Machine + (Machine||Worker), data=Machines, 
            expand_re = TRUE, set_data_arg = FALSE)
emm_options(lmer.df = "Satterthwaite")
emmeans(m1, "Machine")
#  Machine   emmean       SE    df lower.CL upper.CL
#  A       52.35556 2.397324  7.35 46.74094 57.97017
#  B       60.32222 2.633283  9.04 54.36898 66.27546
#  C       66.27222 2.836796 11.56 60.06544 72.47901
# 
# Degrees-of-freedom method: satterthwaite 
# Confidence level used: 0.95 

Alternatively, a different method for calculating the df can also be chosen:

m2 <- mixed(score ~ Machine + (Machine||Worker), data=Machines, expand_re = TRUE)
emm_options(lmer.df = "Satterthwaite")
emmeans(m2, "Machine")
# Error in eval(predvars, data, env) : object 're1.Machine1' not found

emm_options(lmer.df = "asymptotic")
emmeans(m2, "Machine")
#  Machine   emmean       SE  df asymp.LCL asymp.UCL
#  A       52.35556 2.397324 Inf  47.65689  57.05422
#  B       60.32222 2.633283 Inf  55.16108  65.48336
#  C       66.27222 2.836796 Inf  60.71220  71.83224
# 
# Degrees-of-freedom method: asymptotic 
# Confidence level used: 0.95 

emm_options(lmer.df = "kenward-roger")
emmeans(m2, "Machine")
#  Machine   emmean       SE    df lower.CL upper.CL
#  A       52.35556 2.397324  7.35 46.74087 57.97024
#  B       60.32222 2.633283  9.04 54.36897 66.27548
#  C       66.27222 2.836796 11.60 60.06746 72.47698
# 
# Degrees-of-freedom method: kenward-roger 
# Confidence level used: 0.95 

I am currently deliberating whether to use set_data_arg = FALSE as the new default.

[Author]

Posts