this is for holding javascript data
Anisha Keshavan edited In_a_multisite_model_we__.tex
over 8 years ago
Commit id: 02e397f159ada4c001781aa9699bc956e19a0bc6
deletions | additions
diff --git a/In_a_multisite_model_we__.tex b/In_a_multisite_model_we__.tex
index 94a86b4..53c8d96 100644
--- a/In_a_multisite_model_we__.tex
+++ b/In_a_multisite_model_we__.tex
...
In a multisite model, we are sampling effect sizes from our set of sites, and then taking an average of those samples and testing whether or not this average effect is significantly different from 0. Becuase of this, its really important to have enough site-level samples to estimate a mean effect. The plots don't go down to the single site case because the power curves
wouldn't would not apply there - the model would be different. In a single site case, one simply needs to power a two sample T-test, given an effect size, number of subjects, false positive rate. If
we take similar parameters
are taken to a single site case (effect size=0.2, alpha=0.002, power = 80\%),
you one would need 1550 subjects, all acquired at one site, to power this.
The reason we don't do this is because However, it takes a really long time to acquire that many subjects for one site, and it is likely the scanner will go through upgrades or protocols will change in the meantime. The n cutoff (number subjects per site for our 20 sites) that
we chose was chosen for this plot
is was 150 subjects per site, which
we feel is the maximum amount
we'd we would ask our consortium to
collect, though ideally collect. Ideally this would be even lower, especially if researchers wanted to study very rare diseases. At a certain point, even with 0 variability from MRI, there
simply aren't are not enough sites for an effect sizes that are so small (which is the case with genetics) and this is why the # of sites do not go below 10 for this particular effect size ($<10$ samples is not enough for 80\% power, even with no bias).