3dLMEr Triple-Interaction gltCode clarification

I'm looking to assess a very particular contrast involving a categorical x categorical x quantitative triple interaction.

Quick background

My project assesses whether a particular TMEM gene allele offers a protective effect against gray matter atrophy in individuals with genetic frontotemporal dementia.

In R formula syntax the model I am looking at is as follows:

pGM ~ FTDGene_CarrierStatus * TMEM_IsProtective * TimeFromBaseline + AgeAtBaseline + Sex + TIV + (1 | Site / Family / Subj)

(spaces added for readability)

Where:

Variable              Detected_Type Details                                     
Subj                  Subjects      Num Subjects=304                            
ID                    Categorical   Unique levels=1013                          
TIV                   Quantitative  Min=1037.11 | Max=1917.92 | Num outliers=9  
TMEM_IsProtective     Categorical   Counts: NonProtective=380 | Protective=633  
FTDGene_CarrierStatus Categorical   Unique levels=5                             
TimeFromBaseline      Quantitative  Min=0 | Max=7.200546                        
AgeAtBaseline         Quantitative  Min=19.35387 | Max=85.67283                 
Sex                   Categorical   Counts: Female=592 | Male=421               
Site                  Categorical   Unique levels=20                            
Family                Categorical   Unique levels=132    

To clarify, FTDGene_CarrierStatus is 5 levels (Control and GRN/C9orf72 Presymptomatic/Symptomatic)

A picture is 1000 words, so I'd summarize my hypothesis as follows:

My Questions

  1. After doing a trial run with a more parsimonious model (no covariates), the results look promising, but I'm second-guessing myself at the formulation of my gltCode and would like to request a clarification:
	-gltCode DeltaSlope_C9Sympint	'TMEM_IsProtective : 1*Protective -1*NonProtective FTDGene_CarrierStatus : 1*C9orf72_Symptomatic -1*Control TimeFromBaseline :'			\
-gltCode DeltaSlope_GRNSympint	'TMEM_IsProtective : 1*Protective -1*NonProtective FTDGene_CarrierStatus : 1*GRN_Symptomatic -1*Control TimeFromBaseline :'			\
-gltCode DeltaSlope_C9PreSymp	'TMEM_IsProtective : 1*Protective -1*NonProtective FTDGene_CarrierStatus : 1*C9orf72_Presymptomatic -1*Control TimeFromBaseline :'			\
-gltCode DeltaSlope_GRNPreSymp	'TMEM_IsProtective : 1*Protective -1*NonProtective FTDGene_CarrierStatus : 1*GRN_Presymptomatic -1*Control TimeFromBaseline :'			\

These test the interaction effect in differenes of slope changes as per diagram, yes?

  1. I noticed that there appear to be two volumes output for each of the gltCode above. I assume that each pair of volumes corresponds to one of the one-tailed directions, as there aren't any negative values therein, right? Anecdotally, one of the two volumes is near-zero and the other has clusters, so I'm 99.9% sure this is the case, but I would just appreciate a sanity check. Forgive my ignorance, as I was more used to working with FSL/SPM's outputs where t-contrasts give both positive and negative direction outcomes in the same volume.

  2. I noticed that specifying qVars results in each specified variable being automatically centered. Since I'm looking at slopes, this is irrelevant. For future studies however, I would like the interpretation of the intercept to refer to the x-axis variable being zero, not the mean. Is there a way to specify that a variable is quantitative but have it NOT be centered automatically?

Thanks in advance and I will be more than pleased to give reference to AFNI on anything that comes of this! Fingers crossed!
Best,
Maurice

Maurice,

This project is very interesting. Due to the lack of specific domain knowledge, I don't fully understand the nature of (and the causal relationships among) those variables. It might be informative to consider the suggestions regarding covariate selection in this preprint.

These test the interaction effect in differences of slope changes as per the diagram, yes?

Yes, those specifications explore the slope differences among various interactions.

I noticed that there appear to be two volumes output for each of the gltCode above. I assume that each pair of volumes corresponds to one of the one-tailed directions, as there aren't any negative values therein, right?

The sign of the first volume (or sub-brick) indicates whether the effect is positive or negative. The issue of one versus two tails only applies to the second volume when assigning false positives under the conventional null hypothesis significance testing framework.

I was more used to working with FSL/SPM's outputs where t-contrasts give both positive and negative direction outcomes in the same volume.

FYI, this reminds me of an old paper about the topic. On the other hand, excessive emphasis of strict p-values and clusterization might cause other problems as discussed here.

  1. I noticed that specifying qVars results in each specified variable being automatically centered. Since I'm looking at slopes, this is irrelevant. For future studies, however, I would like the interpretation of the intercept to refer to the x-axis variable being zero, not the mean. Is there a way to specify that a variable is quantitative but have it NOT be centered automatically?

Use the option -qVarCenters "0,0,2.3,...".

Gang Chen