Tukey test after LMM keeping contrasts

I want to test a 2x3 factorial design and contrasted the variables like this

    library(lme4)
    library(emmeans)
    my.helmert = matrix(c(2, -1, -1, 0, -1, 1), ncol = 2)

    contrasts(Target3$mask) = my.helmert
    contrasts(Target3$length)

So for mask I want to compare the first group with the average of the two other groups and in a second step the second with the third group.

This works fine in my LMM

Target3.2_TT.lmer = lmer(logTotalTime ~ mask*length+ (1+length|Subject) +(1|Trialnum), data = Target3)

There is a significant interaction between mask and length, that´s why I want to take a look at this effect and calculate a post hoc test (Turkey) like this:

emmeans(Target3.2_TT.lmer, pairwise ~ mask : length)

This also works pretty fine with one problem: now my contrasts are gone. The text calculates the differences for all masks and not just 1 vs. 2 and 3 and 2 vs. 3. Is there a possibility to keep my contrasts in the Post hoc test?

This is what the data looks like:

> dput(Target3)
structure(list(mask = structure(c(2L, 1L, 2L, 3L, 1L, 2L, 3L, 
2L, 1L, 3L, 3L, 3L, 3L, 2L, 2L, 3L, 3L, 1L, 3L, 1L, 1L, 2L, 3L, 
3L, 2L, 1L, 3L, 2L, 3L, 2L), contrasts = structure(c(2, -1, -1, 
0, -1, 1), .Dim = c(3L, 2L), .Dimnames = list(c("keine Maske", 
"syntaktisch\n korrekt", "syntaktisch \n inkorrekt"), NULL)), .Label = c("keine Maske", 
"syntaktisch\n korrekt", "syntaktisch \n inkorrekt"), class = "factor"), 
    length = structure(c(1L, 2L, 1L, 2L, 1L, 2L, 1L, 2L, 1L, 
    2L, 1L, 2L, 2L, 1L, 2L, 1L, 2L, 2L, 1L, 2L, 1L, 2L, 2L, 1L, 
    2L, 2L, 1L, 1L, 2L, 1L), .Label = c("kurzes \n N+1", "langes\n N+1"
    ), class = "factor"), logTotalTime = c(4.969813299576, 5.37989735354046, 
    5.14166355650266, 5.40717177146012, 5.27299955856375, 5.72358510195238, 
    5.4249500174814, 6.18001665365257, 5.67675380226828, 5.44241771052179, 
    5.66988092298052, 5.04985600724954, 5.78996017089725, 5.03043792139244, 
    5.92958914338989, 5.15329159449778, 6.11146733950268, 5.26269018890489, 
    5.17614973257383, 6.18001665365257, 6.03068526026126, 5.68697535633982, 
    5.17614973257383, 5.19849703126583, 5.29330482472449, 5.89989735358249, 
    5.73979291217923, 5.65599181081985, 5.94017125272043, 5.72031177660741
    )), .Names = c("mask", "length", "logTotalTime"), row.names = c(2L, 
4L, 6L, 8L, 9L, 11L, 13L, 15L, 16L, 18L, 20L, 22L, 27L, 29L, 
31L, 33L, 35L, 37L, 39L, 41L, 42L, 44L, 47L, 49L, 51L, 54L, 55L, 
57L, 59L, 61L), class = "data.frame")

Solution

Well, if you ask for pairwise comparisons, that’s what you get, and Helmert contrasts are not the same as pairwise comparisons. Further, the Tukey (not Turkey) method applies only to pairwise comparisons, not to other types of contrasts.

Here’s something to try that may give you what you want.

emm = emmeans(Target3.2_TT.lmer, 
    ~ mask | length)
contrast(emm, list(
    c1 = c(2, -1, -1)/2,
    c2 = c(0, 1, -1)),
  adjust = “mvt”)

This will work independently of whatever parameterization (i.e., contrasts settings) were used when fitting the model. The model parametrrization affects how the model matrix is set up and the interpretation of the coefficients, but does not affect the results from emmeans or its relatives.