Which statistics to report and covarying site location

AFNI Message Board
1

Hi all,

As a start, I am interested in running a general linear model in 3dMVM.

I think I would say that this is a 3 x 2 ANCOVA - although age is a main effect of interest rather than a covariate. I am also controlling for the site in my model (there are two separate sites). Also, I am not interested in the effect of age ONLY on brain response.

Within subject facts: Conditions (equal, overinclude and exclude)

Between subject factors: age, site, and group (risk, HC)

My primary effect of interest: group and group x age differences in brain response during exclusion (exclusion vs. equal) and over-inclusion (over vs. equal contrasts) when controlling for site location.

Questions:

  1. Do I have to look at it anyway as I am looking at the interaction term (my understanding is yes based on other statistical tests).

  2. Reporting: When reporting the F statistics in my glf output and my glt output that better explain the directionality of findings - do these outputs correct for site location automatically? Or should I be reporting the group:site:age: condition F value and then the directionality coefficients and t-values? Do these t-values/coefficients correct for site? I guess I am still struggling with which values make sense to report in paper.

Also, does this model design in 3dMVM make sense given my research interest? Want to triple-check that I am doing this all correctly.

Thanks again


#!/bin/bash

3dMVM -prefix Feb_2023_rerun_2  \
    -bsVars "grp*age*site"  \
    -mask /scratch/06953/jes6785/SOBP/CYB_UPDATE_JAN_2023/ATLAS/bin_social_uniformity-test_z_FDR_0.01.nii.gz \
    -wsVars "condition" \
    -qVars "age" \
    -GES \
    -num_glt 15 \
    -gltLabel 1 RISKvHCexc_v_equ -gltCode 1 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal' \
    -gltLabel 2 RISKvHC_over_v_equ -gltCode 2 'grp : 1*RISK -1*HC condition : 1*over -1*equal' \
    -gltLabel 3 RISKvHC_over_v_exc -gltCode 3 'grp : 1*RISK -1*HC condition : 1*over -1*exclude' \
    -gltLabel 4 RISK_over_v_exc -gltCode 4 'grp : 1*RISK condition : 1*over -1*exclude' \
    -gltLabel 5 HC_over_v_exc -gltCode 5 'grp : 1*HC condition : 1*over -1*exclude' \
    -gltLabel 6 RISK_exc_v_equ -gltCode 6 'grp : 1*RISK condition : 1*exclude -1*equal' \
    -gltLabel 7 HCexc_v_equ -gltCode 7 'grp : 1*HC condition : 1*exclude -1*equal' \
    -gltLabel 8 RISKover_v_equ -gltCode 8 'grp : 1*RISK condition : 1*over -1*equal' \
    -gltLabel 9 HCexc_v_equ -gltCode 9 'grp : 1*HC condition : 1*over -1*equal' \
    -gltLabel 10 RISKvHC_exc_v_equ_age -gltCode 10 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal age :' \
    -gltLabel 11 RISKvHC_over_v_equ_age -gltCode 11 'grp : 1*RISK -1*HC condition : 1*over -1*equal age :' \
    -gltLabel 12 HC_exc_v_equ_age -gltCode 12 'grp : 1*HC condition : 1*exclude -1*equal age :' \
    -gltLabel 13 RISK_exc_v_equ_age -gltCode 13 'grp : 1*RISK condition : 1*exclude -1*equal age :' \
    -gltLabel 14 HC_over_v_equ_age -gltCode 14 'grp : 1*HC condition : 1*over -1*equal age :' \
    -gltLabel 15 RISK_over_v_equ_age -gltCode 15 'grp : 1*RISK condition : 1*over -1*equal age :' \
    -num_glf 4  \
    -glfLabel 1 EXC_RISKvHCexc_v_equ -glfCode 1 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal'  \
    -glfLabel 2 RISKvHC_OVER_v_equ -glfCode 2 'grp : 1*RISK -1*HC condition : 1*over -1*equal'  \
    -glfLabel 3 RISKvHC_exc_v_equ_AGE -glfCode 3 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal age :'  \
    -glfLabel 4 RISKvHC_exc_v_equ_AGE -glfCode 4 'grp : 1*RISK -1*HC condition : 1*over -1*equal age : '  \
2

Before delving into your questions, a fundamental aspect that needs clarification pertains to the assignment of participants between the two sites concerning both their group allocation and age distribution. In addition, is the age distribution roughly similar between the two groups?

Gang Chen

3

The end goal is to have the age distribution to be similar (we are still recruiting). This is preliminary data. At the moment, it looks similar distribution (maybe a statistician would disagree?), but the Kolmogorov--Smirnov < .05
image

Each site should have an equal number of HC and RISK participants (at least with the final data analysis, again preliminary still).

Does that answer your questions?

4

Do I have to look at it anyway as I am looking at the interaction term (my understanding is yes based on other statistical tests).

I'm seeking clarification on your question, as its precise nature is unclear to me. Could you provide additional details or rephrase your inquiry?

Reporting: When reporting the F statistics in my glf output and my glt output that better explain the directionality of findings - do these outputs correct for site location automatically? Or should I be reporting the group:site:age: condition F value and then the directionality coefficients and t-values? Do these t-values/coefficients correct for site? I guess I am still struggling with which values make sense to report in paper.

Consider combining the following four -glf specifications with other -glt specifications, as they possess only one numerator degree of freedom.

...
    -num_glf 4  \
    -glfLabel 1 EXC_RISKvHCexc_v_equ -glfCode 1 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal'  \
    -glfLabel 2 RISKvHC_OVER_v_equ -glfCode 2 'grp : 1*RISK -1*HC condition : 1*over -1*equal'  \
    -glfLabel 3 RISKvHC_exc_v_equ_AGE -glfCode 3 'grp : 1*RISK -1*HC condition : 1*exclude -1*equal age :'  \
    -glfLabel 4 RISKvHC_exc_v_equ_AGE -glfCode 4 'grp : 1*RISK -1*HC condition : 1*over -1*equal age : '  \

All effects in your -glt specifications have been adjusted for the site effect.

Also, does this model design in 3dMVM make sense given my research interest? Want to triple-check that I am doing this all correctly.

Yes, the model specification appears reasonable.

Gang Chen