3dClustSim Thresholds

Hi,

We noticed that running 3dClustSim on two different masks resulted in higher cluster thresholds for the smaller mask (237709 voxels) than the larger one (246932 voxels). The acf values were very similar (they came from the same sample), but just to test it I ran the 3dClustSim command for the smaller mask with the acf values from the larger one, and this lowered the thresholds only a little bit, so does not seem to be accounting for the discrepancy.

Shouldn’t the smaller mask have lower cluster-size thresholds? Any ideas about what might be accounting for this?

Thanks,
Kara

Hi, Kara-

Out of curiosity, what were the cluster size threshold values?

Also, what were the group-mean ACF parameters that you were entering in each case?

–pt

I’ve pasted the tables from the three different 3dClustSims we ran below (with the command lines).

Original 3dClustSim with larger mask:

3dClustSim -both -mask mask_overlap_youth_27subjs.7+tlrc -acf 0.520152 5.1862 12.4322 -prefix 27youth_ACF_ClustSim.mask_overlap_youth_27subjs.7

2-sided thresholding

Grid: 92x109x86 1.75x1.75x1.75 mm^3 (246932 voxels in mask)

CLUSTER SIZE THRESHOLD(pthr,alpha) in Voxels

-NN 1 | alpha = Prob(Cluster >= given size)

pthr | .10000 .09000 .08000 .07000 .06000 .05000 .04000 .03000 .02000 .01000

------ | ------ ------ ------ ------ ------ ------ ------ ------ ------ ------

0.100000 5592.0 5771.0 5972.5 6209.0 6430.0 6715.5 7063.0 7548.0 8218.0 9349.0
0.090000 4665.5 4846.0 5040.3 5211.0 5452.0 5652.0 5966.0 6341.0 6966.0 8013.0
0.080000 3863.0 3983.0 4110.5 4276.0 4505.0 4720.0 4996.0 5308.0 5769.0 6563.0
0.070000 3145.0 3248.0 3354.3 3492.5 3669.5 3858.0 4145.0 4432.5 4844.0 5569.0
0.060000 2537.0 2608.0 2696.0 2804.0 2930.0 3099.0 3289.0 3559.0 3925.0 4488.0
0.050000 2016.0 2082.0 2151.5 2235.5 2331.0 2460.0 2602.0 2792.0 3054.0 3640.0
0.040000 1555.0 1613.0 1670.0 1737.0 1816.0 1902.0 2001.5 2124.0 2351.0 2677.0
0.030000 1151.0 1180.3 1221.8 1268.0 1330.0 1402.0 1486.5 1584.0 1701.0 1971.0
0.020000 773.8 800.6 832.4 869.0 905.3 945.7 999.7 1070.5 1178.5 1343.0
0.015000 608.7 629.0 650.0 675.3 704.5 737.0 780.7 828.6 908.3 1056.0
0.010000 435.7 452.4 469.8 490.2 509.2 534.3 568.0 611.0 665.0 763.2
0.007000 333.0 347.8 360.6 375.6 389.0 412.3 435.0 469.0 520.0 597.0
0.005000 262.9 272.3 282.8 294.2 307.0 323.3 343.4 373.5 412.5 482.5
0.003000 190.4 197.1 204.0 212.4 223.0 233.7 249.0 269.0 296.6 354.0
0.002000 148.4 153.6 160.1 166.7 175.7 185.8 195.7 211.5 234.7 278.5
0.001500 125.1 129.2 134.4 140.8 149.1 156.7 167.3 179.7 200.5 236.6
0.001000 97.9 103.0 107.0 112.5 117.6 124.8 132.8 143.5 161.0 194.5
0.000700 80.6 83.6 87.3 91.4 97.0 102.6 110.1 118.7 132.5 162.2
0.000500 66.7 69.6 72.6 76.2 80.6 85.3 91.4 100.1 112.0 136.7
0.000300 49.6 51.8 54.4 57.2 60.3 65.0 70.6 76.1 87.7 106.7
0.000200 38.5 40.5 42.6 45.3 48.1 51.9 56.8 62.3 72.1 89.4
0.000150 32.2 34.0 36.0 38.3 40.6 44.2 48.7 53.9 62.6 77.7
0.000100 24.9 26.3 28.1 29.9 32.2 35.1 38.8 43.9 49.6 64.8
7.000e-5 19.5 20.7 22.1 24.0 26.0 28.3 31.7 35.5 41.3 55.2
5.000e-5 15.5 16.4 17.6 19.2 20.8 22.9 25.7 29.2 34.8 47.0
3.000e-5 10.5 11.3 12.4 13.4 14.8 16.6 18.8 22.0 26.5 36.2
2.000e-5 7.8 8.4 9.2 10.1 11.3 12.6 14.7 17.0 21.3 28.9
1.500e-5 6.3 6.8 7.4 8.1 9.1 10.4 12.1 14.1 17.5 24.7
1.000e-5 4.6 5.0 5.5 6.0 6.8 7.7 9.1 11.0 13.6 20.3

Original 3dClustSim with smaller mask:

3dClustSim -nodec -both -mask outside_rois_27youth+tlrc -acf 0.513804 5.16813 12.352 -prefix 27youth.ACF_ClustSim.outside_rois_27youth

2-sided thresholding

Grid: 92x109x86 1.75x1.75x1.75 mm^3 (237709 voxels in mask)

CLUSTER SIZE THRESHOLD(pthr,alpha) in Voxels

-NN 1 | alpha = Prob(Cluster >= given size)

pthr | .10000 .09000 .08000 .07000 .06000 .05000 .04000 .03000 .02000 .01000

------ | ------ ------ ------ ------ ------ ------ ------ ------ ------ ------

0.100000 5770 5965 6152 6367 6603 6909 7310 7883 8658 9914
0.090000 4918 5067 5246 5427 5622 5935 6247 6666 7383 8457
0.080000 4116 4254 4390 4569 4769 4992 5269 5688 6177 7114
0.070000 3411 3518 3650 3813 3983 4198 4444 4796 5213 6015
0.060000 2789 2901 2994 3119 3246 3415 3620 3930 4291 4854
0.050000 2236 2305 2381 2480 2586 2749 2912 3110 3417 3906
0.040000 1739 1793 1863 1927 2019 2122 2240 2416 2676 3040
0.030000 1284 1325 1370 1429 1497 1581 1672 1816 2000 2306
0.020000 884 918 951 987 1030 1071 1137 1224 1361 1560
0.015000 687 709 734 764 801 843 884 953 1056 1215
0.010000 501 519 536 557 585 620 656 706 774 902
0.007000 389 402 416 432 450 473 504 546 605 687
0.005000 312 321 333 345 359 381 402 433 485 562
0.003000 227 234 244 253 262 278 296 314 352 408
0.002000 177 184 192 200 209 219 233 250 280 321
0.001500 149 155 162 169 178 186 198 212 237 276
0.001000 118 123 129 135 142 149 159 172 193 230
0.000700 98 101 106 111 118 123 133 146 161 192
0.000500 82 84 88 92 97 104 112 122 139 164
0.000300 62 64 67 70 74 80 87 96 111 130
0.000200 49 51 54 57 60 65 71 79 91 109
0.000150 42 44 46 48 51 55 61 68 79 95
0.000100 33 35 37 39 41 45 49 56 67 81
7.000e-5 27 28 30 32 34 37 41 46 55 68
5.000e-5 22 23 24 26 28 31 34 39 46 59
3.000e-5 15 16 18 19 21 23 26 30 36 47
2.000e-5 11 12 13 14 16 18 20 24 29 39
1.500e-5 9 10 11 12 13 15 17 20 25 33
1.000e-5 7 7 8 9 10 11 13 15 20 28

Testing the smaller mask with the other ACF values:

3dClustSim -both -mask outside_rois_27youth+tlrc -acf 0.520152 5.1862 12.4322 -prefix 27youth.test.outside

2-sided thresholding

Grid: 92x109x86 1.75x1.75x1.75 mm^3 (237709 voxels in mask)

CLUSTER SIZE THRESHOLD(pthr,alpha) in Voxels

-NN 1 | alpha = Prob(Cluster >= given size)

pthr | .10000 .09000 .08000 .07000 .06000 .05000 .04000 .03000 .02000 .01000

------ | ------ ------ ------ ------ ------ ------ ------ ------ ------ ------

0.100000 5752.0 5942.0 6128.7 6346.0 6578.0 6903.0 7288.0 7766.0 8428.0 9678.0
0.090000 4905.0 5049.0 5228.0 5401.0 5633.5 5879.0 6225.0 6653.0 7311.0 8291.0
0.080000 4089.7 4228.0 4381.5 4554.5 4786.0 4995.5 5277.0 5630.5 6145.5 7010.0
0.070000 3393.0 3509.0 3650.0 3791.5 3932.5 4132.0 4400.0 4741.0 5153.0 5921.0
0.060000 2761.3 2863.0 2962.0 3083.0 3217.5 3399.0 3583.0 3834.5 4234.0 4816.0
0.050000 2206.0 2269.0 2343.0 2448.0 2564.0 2707.5 2871.0 3106.0 3400.0 3897.0
0.040000 1701.0 1751.0 1820.0 1899.0 1989.2 2099.7 2250.0 2410.0 2641.0 3031.0
0.030000 1262.7 1301.7 1349.8 1399.0 1466.0 1554.5 1642.0 1784.0 1953.5 2247.0
0.020000 858.0 884.6 915.0 953.0 999.0 1048.0 1111.5 1197.0 1321.5 1532.0
0.015000 674.4 696.2 723.0 754.0 784.7 822.6 877.0 947.0 1035.0 1192.0
0.010000 494.0 510.2 531.3 553.7 576.8 609.0 651.3 700.8 763.7 871.0
0.007000 383.4 396.2 408.8 428.3 446.3 468.0 505.0 545.7 598.5 679.0
0.005000 305.2 315.1 326.3 340.0 354.1 371.6 397.2 432.0 473.7 548.7
0.003000 220.8 227.5 236.1 244.4 255.8 269.2 286.7 308.3 340.7 394.7
0.002000 172.5 178.1 184.1 191.5 199.8 211.9 223.8 243.2 272.0 314.0
0.001500 145.7 150.6 155.9 162.8 170.6 179.5 192.2 208.0 233.2 272.0
0.001000 115.2 119.8 125.4 131.1 137.4 144.9 153.2 167.2 187.6 222.0
0.000700 94.7 98.5 102.6 107.4 113.5 120.0 126.9 135.9 156.0 187.0
0.000500 78.5 81.8 85.2 89.2 94.3 100.8 107.6 115.1 130.0 158.7
0.000300 58.7 61.0 64.4 68.6 72.4 77.3 82.1 90.2 102.4 126.3
0.000200 46.9 49.2 51.9 54.8 58.4 62.2 67.3 74.6 84.5 104.7
0.000150 39.6 42.0 44.3 46.6 49.5 53.6 58.1 64.5 74.5 91.7
0.000100 31.1 32.8 34.8 37.1 39.6 43.3 47.0 52.7 61.7 76.5
7.000e-5 24.9 26.5 28.1 30.1 32.3 35.3 39.0 44.5 51.8 63.8
5.000e-5 20.0 21.2 22.8 24.5 26.2 29.1 32.7 37.5 44.2 55.2
3.000e-5 14.0 14.9 16.3 17.7 19.3 21.4 24.4 28.8 34.1 43.5
2.000e-5 10.2 11.1 12.2 13.3 14.8 16.6 19.4 23.0 27.3 36.3
1.500e-5 8.1 8.9 9.8 10.9 12.2 14.1 16.2 19.2 23.5 31.5
1.000e-5 5.8 6.4 7.2 8.0 9.0 10.3 12.0 15.0 18.2 26.0

Hi Kara,

What is the difference in cluster sizes?

The shapes of the masks can have an effect, too. The more spherical a
mask boundary, the more likely it is to get larger clusters. Is there any
notable difference in the shapes?

  • rick

Yes, that might be it. One mask is a whole-brain mask, and the other covers most of the brain but does not include ROIs that we are doing small volume corrections for.

EDIT: Sorry, I got confused and stated this backwards.
Correction: since the blur estimates are used for applying
the blur in Monte Carlo simulations, the smaller numbers
will apply smaller blurs and lead to smaller random clusters,
and therefore lead to smaller cluster requirements. In your
case, the mask shape is likely the dominant factor.

It would be reasonable to use the full mask ACF parameters
for both commands.

  • rick

Just to check, took a WB mask “mask+orig” (25379 vox), and then dilated by 5 voxels, so there was a much smaller WB-like mask (i.e., similar shape), called “mask_shrunk+orig” (4130 vox-- prob I shrank it too much, but for this comparison it probably shows what is relevant). I then ran 4 test cases of 3dClustSim using your two sets of ACF parameters:


A) mask+orig with ACF = 0.513804 5.16813 12.352
B) mask+orig with ACF =  0.520152 5.1862 12.4322
C) mask_shrunk+orig with ACF = 0.513804 5.16813 12.352
D) mask_shrunk+orig with ACF =  0.520152 5.1862 12.4322

In each case, the command I ran was like the following (substituting mask names and ACF param names):


3dClustSim -nodec -both -mask mask+orig. -acf 0.513804 5.16813 12.352 -prefix mask_0.513804_5.16813_12.352

For space, I am just copy-pasting results from some of the middle-range, most relevant cases of pthr:

A


0.010000      43     44     45     47     49     52     55     60     65     76
 0.007000      33     35     36     37     39     41     44     47     52     60
 0.005000      27     28     29     30     32     33     35     38     43     50
 0.003000      20     21     22     23     24     25     27     29     32     36
 0.002000      16     17     17     18     19     20     21     23     25     30
 0.001500      14     14     15     16     16     17     18     20     22     26
 0.001000      11     12     12     13     13     14     15     16     18     21

B) same mask as A, but other ACF pars


0.010000      43     44     46     47     49     52     54     58     63     76
 0.007000      33     34     36     37     39     41     43     46     51     59
 0.005000      27     28     29     30     31     33     35     38     41     49
 0.003000      20     21     21     22     23     24     26     28     32     38
 0.002000      16     17     17     18     19     20     21     23     25     30
 0.001500      14     14     15     16     16     17     18     20     21     26
 0.001000      11     12     12     13     13     14     15     16     18     21

C) shrunken mask, same ACF pars as A


0.010000      22     23     24     25     26     28     30     33     37     47
 0.007000      17     18     19     20     21     22     24     26     30     38
 0.005000      14     15     15     16     17     18     20     21     24     30
 0.003000      10     11     11     12     12     13     15     16     18     22
 0.002000       8      9      9     10     10     11     12     13     15     18
 0.001500       7      7      8      8      9      9     10     11     13     15
 0.001000       6      6      6      7      7      8      8      9     10     13

C) shrunken mask, different ACF pars from A


0.010000      22     22     24     25     26     28     30     33     36     43
 0.007000      17     18     19     20     21     22     24     26     29     35
 0.005000      14     15     15     16     17     18     20     21     24     28
 0.003000      10     11     11     12     12     14     15     16     18     21
 0.002000       8      9      9      9     10     11     12     13     14     17
 0.001500       7      7      8      8      8      9     10     11     12     15
 0.001000       6      6      6      7      7      7      8      9     10     12

Summary:

  • for the same mask but different ACF pars, the results are quite similar, esp. as smaller p
  • for same pars but smaller mask, the cluster sizes are uniformly much smaller (as expected).

–pt

Sounds good. Thank you so much!

Hi Kara,

You might now care anymore, but my interpretation of the
blur effects was backward in the previous post, it should
be correct now. Sorry.

  • rick