#StackBounty: #anova #mixed-model #lme4-nlme #repeated-measures #lm Repeated measures anova: lm vs lmer

Bounty: 50

I’m trying to reproduce several interaction test between with both lm and lmer on repeated measures (2x2x2). The reason I want to compare both methods is because SPSS’s GLM for repeated measures yields the exact same results as the lm approach presented here, so at the end I want to compare SPSS vs R-lmer. So far, I’ve only managed to reproduce (closely) some of these interactions.

You’ll find below a script to better illustrate my point:

library(data.table)
library(tidyr)
library(lmerTest)
library(MASS)

set.seed(1)

N     <- 100 # number of subjects
sigma <- 1   # popuplation sd
rho   <- .6  # correlation between variables

# X1:   a  a  a  a  b  b  b  b
# X2:   a  a  b  b  a  a  b  b
# X3:   a  b  a  b  a  b  a  b
mu <- c(5, 3, 3, 5, 3, 5, 5, 3) # means

# Simulate the data
sigma.mat <- rep(sigma, length(mu))
S <- matrix(sigma.mat, ncol = length(sigma.mat), nrow = length(sigma.mat))
Sigma <- t(S) * S * rho  
diag(Sigma) <- sigma**2
X <- data.table( mvrnorm(N, mu, Sigma) )
setnames(X, names(X), c("aaa", "aab", "aba", "abb", "baa", "bab", "bba", "bbb"))
X[, id := 1:.N]

# Long format
XL <- data.table( gather(X, key, Y, aaa:bbb) )
XL[, X1 := substr(key, 1, 1)]
XL[, X2 := substr(key, 2, 2)]
XL[, X3 := substr(key, 3, 3)]

# Recode long format (a = +1; b = -1)
XL[, X1c := ifelse(X1 == "a", +1, -1)]
XL[, X2c := ifelse(X2 == "a", +1, -1)]
XL[, X3c := ifelse(X3 == "a", +1, -1)]


### Composite scores to be used with lm
# X2:X3 2-way interaction (for half the data; i.e. when X1 == "a")
X[, X1a_X2.X3 := (aaa - aab) - (aba - abb)]

# X2:X3 2-way interaction (for all the data)
X[, aa := (aaa + baa) / 2]
X[, ab := (aab + bab) / 2]
X[, ba := (aba + bba) / 2]
X[, bb := (abb + bbb) / 2]
X[, X2.X3 := (aa - ab) - (ba - bb)]

# X1:X2:X3 3-way interaction (for all the data)
X[, X1.X2.X3 := ( (aaa - aab) - (aba - abb) ) - ( (baa - bab) - (bba - bbb) )]


### Fit models
# X2:X3 2-way interaction (for half the data; i.e. when X1 == "a")
summary( lm(X1a_X2.X3 ~ 1, X) ) # t = 34.13303
summary( lmer(Y ~ X2c*X3c + (X2c+X3c|id), XL[X1 == "a"]) ) # t = 34.132846  close match
summary( lmer(Y ~ X2c*X3c + (X2c+X3c||id), XL[X1 == "a"]) ) # t = 34.134624  close match

# X2:X3 2-way interaction (for all the data) 
summary( lm(X2.X3 ~ 1, X) ) # t = 0.3075025
summary( lmer(Y ~ X2c*X3c + (X2c+X3c|id), XL) ) # t = 0.1641932
summary( lmer(Y ~ X2c*X3c + (X2c+X3c||id), XL) ) # t = 0.1640710
summary( lmer(Y ~ X2c*X3c + (X2c*X3c|id), XL) ) # t = 0.1641765
anova(   lmer(Y ~ X2c*X3c + (X2c*X3c|id), XL), ddf = "Kenward-Roger" ) # t = 0.1643168
summary( lmer(Y ~ X2c*X3c + (X2c*X3c|id), XL, REML = FALSE) ) # t = 0.1645303
summary( lmer(Y ~ X2c*X3c + (X2c*X3c||id), XL) ) # t = 0.1640704

# X1:X2:X3 3-way interaction (for all the data)
summary( lm(X1.X2.X3 ~ 1, X) ) # t = 46.50177
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c - X1c:X2c:X3c|id), XL) ) # t = 49.0317599
anova(   lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c - X1c:X2c:X3c|id), XL), ddf = "Kenward-Roger" ) # t = 49.03176
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c - X1c:X2c:X3c|id), XL, REML = FALSE) ) # t = 49.2677606
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c - X1c:X2c:X3c||id), XL) ) # t = 46.5193774 close match
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c|id), XL) ) # unidentifiable
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c|id), XL,
              control = lmerControl(check.nobs.vs.nRE="ignore")) ) # t = 46.5148684 close match

As you can see from above, none of the lm estimations are exactly matched by the lmer ones. Although some of the results are very similar and may differ only due because of numerical/computational reasons. The gap between both estimation method is specially large for the X2:X3 2-way interaction (for all the data).

My question is if there’s a way to obtain the exact same results with both methods, and if there is a correct way to perform the analyses with lmer (although it may not match the lm results).


Bonus:

I’ve noticed that the t value associated with the 3-way interaction is affected by the way factors are coded, which seems very strange to me:

summary( lmer(Y ~ X1*X2*X3 + (X1*X2*X3 - X1:X2:X3||id), XL) ) # t = 48.36
summary( lmer(Y ~ X1c*X2c*X3c + (X1c*X2c*X3c - X1c:X2c:X3c||id), XL) ) # t = 56.52


Get this bounty!!!

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.