External Email - Use Caution
Dear Freesurfer Mailing list,
I have run a mass-univariate spatiotemporal model using 'lme_mass_fit_EMinit' and then applied an FDR correction across both hemispheres (syntax provided below). On completion I realised that m ost of the corrected p-values for my main and interaction effects were well above the alpha of 0.05. The corrected p-value thresholds in log base 10 format are included in the attached figure and summarised below.
Summary of p-value thresholds in attached images:
Figure
Key
Log10P
P
Figure 1
A1
-1.193
0.064
A2
-0.695
0.202
B1
-0.969
0.107
B2
-0.243
0.571
C1
-0.195
0.638
C2
-0.001
0.998
Figure 2
A
-2.663
0.002
B
-1.295
0.051
I have no idea why this has happened and would greatly appreciate any insight you can provide. I would also note that the same issue occurred in a model where we exclude all subjects with only a single time-point. However, it did not occur when we included only Caucasians in our analysis and excluded the two dummy coded ethnicity covariates. This leads me to think its something to do with the complexity of the model, but I cannot understand why the error is only occurring when we correct for multiple comparisons. I have provided the details of sample and methods below.
Sample
Our sample includes 112 controls subjects, and 106 at-risk subjects (individuals with a first degree relative of BD but not personal history of BD). All subjects are aged between 12 and 30 years. 153 or these subjects have 2 time-points (77 control, 76 at-risk), while the remaining 65 individuals have only 1 MRI time-point. We also have mixed ethnicites - 165 Caucasians, 23 Asians, and 30 mixed (Asians-Caucasians).
Analysis Design
The QDEC file contains the following 6 variables:
1. Y (years between scans) 2. A (baseline age) 3. G (group, 1=case, 0=control) 4. S (sex, 1=female, 0=male) 5. E1 (Ethnicity 1, 1=asian, 0=other) 6. E2 (Ethnicity 1, 1=mixed-asian-caucasian, 0=other)
Design matrix: [ones(length(M),1) M M(:,1).*M(:,3)] i.e. main effects for each of the qdec variables + an interaction term for years X group
DVs: Cortical thickness, area and volume
Model: Mass-univariate spatiotemporal model using 'lme_mass_fit_EMinit'
CODE:
% Read in surface files
[Y,mri] = fs_read_Y(mgh);
% Read in qdec file
Qdec = fReadQdec(qdec); % Remove fsid from qdec Qdec = rmQdecCol(Qdec,1); % Store col 1 (fsid-base) in sID variable sID = Qdec(2:end,1); % Remove col 1 (fsid-base) from Qdec array Qdec = rmQdecCol(Qdec,1); % Convert Qdec to numeric matrix M M = Qdec2num(Qdec);
% Sort data and evaluate design matrix
[M,Y,ni] = sortData(M,1,Y,sID); X = eval([ones(length(M),1) M M(:,1).*M(:,3)])
% Compute vertex-wise temporal covariance estimates.
[Th0, Re] = lme_mass_fit_EMinit(X,[1],Y,ni,cortex,3);
%Segmentation and model fitting.
[Rgs, RgMeans, stats] = fit(Th0, Re, [1], sphere, cortex, X, Y, ni);
%Check surfaces.
surfcomp(Th0, RgMeans, sphere, fig1, fig2)
Correction for multiple comparisons: FDR across both hemispheres
CODE:
P = [ F_lhstats.pval(lhcortex) F_rhstats.pval(rhcortex) ]; G = [ F_lhstats.sgn(lhcortex) F_rhstats.sgn(rhcortex) ]; [detvtx, sided_pval, pth] = lme_mass_FDR2(P,G,[],0.05,0); pcor = -log10(pth); thrlh(r,2) = {pcor}; thrrh(r,2) = {pcor}; [~,dc] = size(detvtx); dvtx(r,2) = {dc};
Kind regards,
Bronwyn Overs
Research Assistant
Neuroscience Research Australia Margarete Ainsworth Building Barker Street Randwick Sydney NSW 2031 Australia M 0411 308 769 T +61 2 9399 1725
[ http://neura.edu.au/ | neura.edu.au ]
[ https://twitter.com/neuraustralia | Twitter ] | [ https://www.facebook.com/NeuroscienceResearchAustralia | Facebook ] | [ http://www.neura.edu.au/help-research/subscribe | Subscribe ]
