Alphasim and GRF correction

Submitted by yaosir on

Dear experts,

I am now doing some rsFC analyses among individuals with Internet addiction and controls and trying to exmaine the differences between these two groups. Unfortunately, no clusters remained after FDR corrected. Although there remains some significant results if I used Alphasim correction instead, I found that Alphasim correction might be doubted by some reviewers and was rarely used in papers published in top journal. Epecially, If I set a voxel threshold of P < 0.001 (with FWHM=4mm and brainmask_05_61*73*61), the threshold of the number of clusters was only 6! I am not sure if it was right, but if so, the results corrected by Alphasim were not even as strict as that with a voxel threshold of P < 0.001 and cluters > 10 (which was refered to as uncorrection in some papers). I wonder if my understanding about Alphasim correction is right?

In addition, I found there was another correction available in REST: GRF correction, which was used in some papers published in top journal (eg., Di Martino et al., 2009). In these papers, authors stated that they used GRF correction with voxel-wise: minimum z score  >2.3 and cluster significance: P < 0.05, which was a little defferent from that in REST. According to http://restfmri.net/forum/node/1434, is it equal to voxel level P < 0.017 and cluster level P < 0.05? And if the steps are as follows: selet underlay and overlay---Misc---GRF correction---select brainmask_05_61*73*61---fill out Ps of voxel level and cluster level--- select one tailed---click show it?

Finally, I want to know the order of these three corrections (FDR, Alphasim, and GRF) according to strictness.

Thank you for reading so long post!

Best,

Vincent

YAN Chao-Gan

Mon, 08/18/2014 - 16:51

Hi Vincent,

Please have a look on http://rfmri.org/Course section V or section 八. DPABI is specifically dealing with this multiple comparison issue.

It's not correct to use AlphaSim in such a way (using preprocessing smooth kernel), including my first paper (Yan et al., 2009).

The correct way is using the estimated smoothness. "Smoothness estimation based on the 4D residual is built in regression function – smoothness is written to the NIfTI headers automatically. For AlphaSim and GRF multiple comparison correction, only using smooth kernel applied in preprocessing is NOT sufficient, please use the estimated smoothness instead."

That's correct, one tailed P: P < 0.017 and cluster level P < 0.05. And corresponds to two tailed P: P < 0.0214 and cluster level P < 0.1.

Strictness is a tricky term, it should depend on data. You can write a post on how these appears on your data.

Best,

Chao-Gan

 

 

Dear Dr. Yan,

Thank you for your helpful reply. I have estimated smoothness instead through REST utilities---REST Alphasim, and the AlphaSimtext was as follows:

Mask filename = BrainMask_05_61x73x61

 
Voxels in mask = 70831
 
Gaussian filter width (FWHM, in mm) = 11.712
 
Gaussian filter width (FWHM, in mm) = 12.134
 
Gaussian filter width (FWHM, in mm) = 10.992
 
Cluster connection radius: rmm = 5.00
 
Individual voxel threshold probability = 0.001
 
Number of Monte Carlo simulations = 1000
 
Output filename = AlphaSimtext2
 
 
 
Cl Size Frequency Cum Prop p/Voxel Max Freq Alpha
1
1487 0.203587 0.000810 2 1.000000
2
958 0.334748 0.000789 7 0.998000
3
657 0.424699 0.000762 7 0.991000
4
556 0.500821 0.000734 7 0.984000
5
451 0.562568 0.000703 11 0.977000
6
392 0.616238 0.000671 21 0.966000
7
309 0.658543 0.000637 17 0.945000
8
282 0.697152 0.000607 17 0.928000
9
260 0.732749 0.000575 44 0.911000
10
212 0.761774 0.000542 33 0.867000
11
181 0.786555 0.000512 25 0.834000
12
165 0.809146 0.000484 36 0.809000
13
138 0.828039 0.000456 41 0.773000
14
118 0.844195 0.000431 39 0.732000
15
114 0.859803 0.000407 47 0.693000
16
94 0.872673 0.000383 36 0.646000
17
104 0.886911 0.000362 51 0.610000
18
69 0.896358 0.000337 29 0.559000
19
68 0.905668 0.000320 35 0.530000
20
47 0.912103 0.000301 19 0.495000
21
56 0.919770 0.000288 32 0.476000
22
52 0.926889 0.000271 33 0.444000
23
59 0.934967 0.000255 35 0.411000
24
38 0.940170 0.000236 28 0.376000
25
32 0.944551 0.000223 22 0.348000
26
27 0.948248 0.000212 14 0.326000
27
33 0.952766 0.000202 21 0.312000
28
33 0.957284 0.000189 27 0.291000
29
28 0.961117 0.000176 20 0.264000
30
26 0.964677 0.000165 21 0.244000
31
20 0.967415 0.000154 15 0.223000
32
20 0.970153 0.000145 18 0.208000
33
20 0.972892 0.000136 14 0.190000
34
14 0.974808 0.000127 9 0.176000
35
19 0.977410 0.000120 17 0.167000
36
17 0.979737 0.000111 15 0.150000
37
12 0.981380 0.000102 11 0.135000
38
16 0.983571 0.000096 13 0.124000
39
9 0.984803 0.000087 8 0.111000
40
8 0.985898 0.000082 7 0.103000
41
10 0.987267 0.000078 10 0.096000
42
3 0.987678 0.000072 2 0.086000
43
4 0.988226 0.000070 4 0.084000
44
5 0.988910 0.000068 5 0.080000
45
4 0.989458 0.000065 3 0.075000
46
4 0.990005 0.000062 4 0.072000
47
7 0.990964 0.000060 6 0.068000
48
3 0.991375 0.000055 3 0.062000
49
5 0.992059 0.000053 5 0.059000
50
3 0.992470 0.000049 3 0.054000
51
7 0.993428 0.000047 6 0.051000
52
5 0.994113 0.000042 5 0.045000
53
8 0.995208 0.000039 8 0.040000
54
0 0.995208 0.000033 0 0.032000
55
4 0.995756 0.000033 4 0.032000
56
1 0.995893 0.000029 0 0.028000
57
8 0.996988 0.000029 7 0.028000
58
0 0.996988 0.000022 0 0.021000
59
0 0.996988 0.000022 0 0.021000
60
1 0.997125 0.000022 1 0.021000
61
2 0.997399 0.000021 1 0.020000
62
0 0.997399 0.000020 0 0.019000
63
2 0.997673 0.000020 2 0.019000
64
2 0.997946 0.000018 2 0.017000
65
1 0.998083 0.000016 1 0.015000
66
2 0.998357 0.000015 2 0.014000
67
0 0.998357 0.000013 0 0.012000
68
0 0.998357 0.000013 0 0.012000
69
1 0.998494 0.000013 1 0.012000
70
0 0.998494 0.000012 0 0.011000
71
0 0.998494 0.000012 0 0.011000
72
0 0.998494 0.000012 0 0.011000
73
1 0.998631 0.000012 1 0.011000
74
1 0.998768 0.000011 1 0.010000
75
1 0.998905 0.000010 1 0.009000
76
2 0.999179 0.000009 2 0.008000
77
1 0.999315 0.000007 1 0.006000
78
0 0.999315 0.000006 0 0.005000
79
0 0.999315 0.000006 0 0.005000
80
0 0.999315 0.000006 0 0.005000
81
1 0.999452 0.000006 1 0.005000
82
0 0.999452 0.000005 0 0.004000
83
1 0.999589 0.000005 1 0.004000
84
0 0.999589 0.000004 0 0.003000
85
1 0.999726 0.000004 1 0.003000
86
2 1.000000 0.000002 2 0.002000
 
The first Alpha < 0.05 corresponded to Cl size = 5. Thus, do I only need 5 as the cluster size threshold? If so, I think it is too small, can I use Cl size = 10 or 20 instead but still saying that it is corrected by means of Monte Carlo simulation?
 
In addition, I have also tried GRF correction following the steps I mentioned above, but the results were different from that corrcted by Alphasim. The most important difference I found was that the remaining clusters corrected by GRF had much larger cluters size than that corrcted by Alphasim, but the results corrcted by Alphasim has higher peak value. The location of the clusters were also different. What means of correction do you recommand?
 
Thank you in advance.
 
Best,
Vincent