lmerMod vs lmerModLmerTest - what are the differences and which is correct?

Question

I have three trials, want to know if the order of trial affected the outcome. I was checking my work with ChatGPT and noticed a discrepancy between their results and mine. This is the dummy code from ChatGPT:

# Load necessary libraries
library(lme4)
library(lmerTest)
library(emmeans)

# Define synthetic dataset with explicit scores for each trial
set.seed(123)  # Set seed for reproducibility

# Scores for 20 subjects, each completing 3 trials
scores <- c(80, 85, 78,  # Subject 1
            75, 80, 82,  # Subject 2
            84, 79, 81,  # Subject 3
            77, 82, 79,  # Subject 4
            80, 83, 85,  # Subject 5
            79, 78, 76,  # Subject 6
            82, 84, 79,  # Subject 7
            81, 80, 82,  # Subject 8
            78, 80, 84,  # Subject 9
            82, 81, 83,  # Subject 10
            79, 78, 77,  # Subject 11
            84, 87, 89,  # Subject 12
            83, 85, 84,  # Subject 13
            82, 81, 83,  # Subject 14
            80, 79, 82,  # Subject 15
            78, 80, 81,  # Subject 16
            85, 88, 87,  # Subject 17
            79, 82, 80,  # Subject 18
            81, 82, 85,  # Subject 19
            84, 85, 86)  # Subject 20

# Create the dataset
data <- data.frame(
  Subject = rep(1:20, each = 3),   # 20 subjects
  Trial = rep(1:3, times = 20),    # Three trials
  Score = scores                    # Defined scores
)

# Treat Trial as a numeric variable
data$Trial <- as.numeric(data$Trial)

# Fit the linear mixed-effects model treating Trial as numeric
model_continuous <- lmer(Score ~ Trial + (1 | Subject), data = data)

# Model summary
summary(model_continuous)

I am getting a different output. I think it has something to do with lmerMod and lmerTest, and I don't know how to replicate their results.

Mine:

Linear mixed model fit by REML. t-tests use Satterthwaites method ['lmerModLmerTest']
Formula: Score ~ Trial + (1 | Subject)
   Data: data

REML criterion at convergence: 282.2

Scaled residuals: 
     Min       1Q   Median       3Q      Max 
-1.93939 -0.48327  0.08381  0.51236  1.94184 
......
Fixed effects:
            Estimate Std. Error      df t value Pr(>|t|)    
(Intercept)  80.0833     0.8456 54.7267  94.707   <2e-16 ***
Trial         0.7500     0.3157 39.0000   2.376   0.0225 *  

ChatGPT:

Linear mixed model fit by REML ['lmerMod']
Formula: Score ~ Trial + (1 | Subject)
   Data: data

REML criterion at convergence: 104.7192

Scaled residuals:
    Min      1Q  Median      3Q     Max 
-2.1934 -0.5690 -0.0162  0.5091  2.3361
.....
Fixed effects:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  80.1000     0.7123  112.48   <2e-16 ***
Trial         2.0000     0.6232    3.21    0.0056 ** 

From R Documentation:

In cases when a valid lmer-object (lmerMod) is produced, but when the computations needed for Satterthwaite df fails, the lmerMod object is returned - not an lmerModLmerTest object.

What does this mean? Which is correct? How do I know which to use? How do I switch between lmerMod and lmerModLmerTest if I need to?

Answer 1

ChatGPT is simply making up numbers.

lmerTest provides a wrapper around lme4 which provides denominator degrees-of-freedom computations and a few other helper functions. The actual fitting is done with exactly the same machinery in lme4 and lmerTest.

The only reason I can think of that you would need to revert from lmerTest to lme4 is that occasionally when something goes wrong with the fitting it can be a little harder to figure out what's going on with lmerTest (it hides some warning messages). In this case you could detach("package:lmerTest"); library("lme4"), but this isn't guaranteed to work perfectly — it would be safer to start a clean R session where you load lme4 instead of lmerTest.