rvlenth
commented
1 year ago First, emmeans() and its relatives need the factors to identify what is to be averaged. The model matrix is not the same thing, and it can't read your mind.
Second, if you want a common baseline, parameterize the model (again, in terms of factors) so that that criterion is met:
> x$Trt <- factor(rep(c("Bas","A","Bas","B"), each = 5))
> library(nlme)
> mod <- gls(val ~ Trt, data = x,
+ correlation = corSymm(form = ~1 | id),
+ weights = varIdent(form = ~1 | tim))
> library(emmeans)
> emmeans(mod, ~ Trt)
Analytical Satterthwaite method not available; using appx-satterthwaite
Trt emmean SE df lower.CL upper.CL
A 1.008 0.392 8.09 0.105 1.91074
B 3.089 0.392 8.09 2.186 3.99122
Bas -0.463 0.201 8.89 -0.918 -0.00818
Degrees-of-freedom method: appx-satterthwaite
Confidence level used: 0.95 Note that these are exactly the values you obtained with predict().
I had wanted to fit the model val ~ Trt + tim, but that is singular and gls does not have that capability. But you can add it later using add_grouping():
> RG <- add_grouping(ref_grid(mod), "Time", "Trt", c("W1", "W1", "Bas"))
> confint(RG)
Trt Time prediction SE df lower.CL upper.CL
Bas Bas -0.463 0.201 8.96 -0.917 -0.0087
A W1 1.008 0.392 7.98 0.103 1.9128
B W1 3.089 0.392 7.98 2.184 3.9933
Degrees-of-freedom method: appx-satterthwaite
Confidence level used: 0.95
> contrast(RG, "trt.vs.ctrl1")
contrast estimate SE df t.ratio p.value
A W1 - Bas Bas 1.47 0.436 9.64 3.372 0.0141
B W1 - Bas Bas 3.55 0.436 9.64 8.142 <.0001
Degrees-of-freedom method: appx-satterthwaite
P value adjustment: dunnettx method for 2 tests 
Generalized
Dear @rvlenth,
The cLDA (proposed many years ago) is now gaining more and more popularity (actually, it's one of the standards today) in randomized clinical trials under the pharmaceutical regulators' acceptance. Briefly - when comparing treatment arms, it forces the baseline predictions to be equal. This is justified for RCTs (and invalid elsewhere), increasing the power of the analysis (at the cost of slightly increased type-1 error). The classic longitudinal analysis (LDA) puts the baseline values at the right side of the formula (removing it from the response vector) and includes the full interaction of treatment * time. The cLDA does it oppositely: it puts the baseline values to the response vector and forces exact baseline by removing from the model matrix the intercept and the treatment. This is the exception, where we have the interaction without one of the main effects - but that's fine here.
Of interest is exclusively the interaction term and the appropriate contrasts comparing treatment arms at selected time points.
A few examples: https://datascienceplus.com/taking-the-baseline-measurement-into-account-constrained-lda-in-r/ https://rdoodles.rbind.io/2020/03/analyzing-longitudinal-data-a-simple-pre-post-design/ https://www.middleprofessor.com/files/applied-biostatistics_bookdown/_book/pre-post.html
Some literature: (free) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5223669/ (paywalled) https://www.jstor.org/stable/25053123 (paywalled) https://pubmed.ncbi.nlm.nih.gov/19764951/ https://arxiv.org/ftp/arxiv/papers/2203/2203.15555.pdf
To the point - it's quite simple, when doing it by hands:
The data:
The model matrix:
The model (can be done with nlme::gls, but Frank Harrell's rms::Gls gives me also the CIs)
Result:
The predictions - here you can nicely see what happened to the baseline at both study arms:
Now let's do something with emmeans (same result with rms::Gls or nlme::gls) Whatever I provide to the "specs" gives me the same error.
Tell me, please, how to do it with emmeans? I need it to do all the usual work (testing contrasts, adjusting p-values and CIs) in a consistent manner, rather than employing lmTest or contrasts packages.
Is this possible?