HI Sarah,

I really have difficulty understanding your questions and terminology.

> I have a question regarding the way statistics are calculated when
> using 3dttest++ versus using GLT (in the context of running a GLM
> where I am specifying a contrast between two conditions at a single
> time point per trial using a tent function).

3dttest++ is typically used for group analysis on whole brain. Is it still the case here, or no? When referring to GLM, do you mean individual subject analysis? Although GLM is very popular in FMRI, it is really a sloppy usage: I don't really like it being specifically referred to individual subject analysis because it is a generic term for modeling that also includes those models typically used at group level (e.g., t-tests, AN(C)OVA, regression, etc.) and it does not accurately characterize the specifics involved in FMRI individual subject analysis either.

Also, what do you mean by "a single time point per trial"? Maybe pasting a few relevant lines of your analysis scripts would be more helpful.

> If I extract the coefficients for one time point per trial (using stim_times_IM)
> for each condition and compute a ttest using 3dttest++, the resulting t
> statistic is larger than the resulting t-statistic when using GLT.

What HRF model did you use that's combined with stim_times_IM? Tents? 3dttest++ was run across those trials?

> I'm curious about potential differences in the way the statistics are computed
> in both cases because I would have expected the resulting t statistic to be the
> same for each. In other words, how is the t-statistic for the GLT calculated,
> because it does not seem to be by contrasting beta estimates from each trial
> (otherwise it should have led to the same t-statistic as 3dttest)?

You're comparing two different modeling strategies here.

1) With the typical individual subject analysis, a condition (with multiple trials) is assumed to have the SAME BOLD response magnitude across trials. And the significane of the condition effects (regression coefficients) or their linear combinations (GLTs) is assessed by the effect sizes relative to their precision (standard error).

2) What you did with stim_times_IM (my guess) seems to have two steps. (a) You assume that the BOLD response of a condition can have DIFFERENT BOLD response magnitude across trials, and estimate the effect size per trial in a regression model; (b) Gather the effect sizes from those trials, assume they have the SAME precision/reliability, and assess the significance for the overall effect size across trials with a one-sample t-test.

The above two approaches have different models and carry different assumptions. So it's not surprising that they ended up with different results.

HTH,

Gang Chen
[afni.nimh.nih.gov]



Edited 1 time(s). Last edit at 08/15/2012 01:15PM by Gang.