Hi
when you go through the stats it turns out that it does not matter if
you're interested in significant positive or negative changes - the
interpretation wrt the validity of the stats is the same. The 'room to
decrease' argument I find hard to follow - the estimates are not bound
by any arbitrary limit (0, say) so there is not such 'room' (unless
you were to move to some explicit biophysical model. In any case, your
null assumption is that of no change between t2 and t1, independent of
baseline at t1 or t2. You probably would want to use randomize to make
sure that you're not constrained by the Gaussian assumptions for the
null in the GLM.
hth
Christian
On 8 Jul 2008, at 11:39, Anna Rieckmann wrote:
> Hi Christian
> yes I agree with you and this is also the way we have our analysis
> set up. What I should have maybe said is that we are mainly
> interested in deactivations from t1 to t2. If we then compare the
> groups in the t2-t1 difference we see much greater deactivations in
> the treatment group compared to the control group. Although this is
> according to our predictions, how can we be sure that the greater
> deactivations are due to the treatment rather than to the fact that
> the treatment group had more "room to decrease" from the beginning
> (i.e. both groups show a regression to the mean) ?
>
> Thanks,
> Anna
>
>
>
>
>
> On Jul 8, 2008, at 12:00 PM, Christian F. Beckmann wrote:
>
>> Hi
>>
>> fundamentally this isn't such a problem - if I understand your
>> study correctly you'll be dealing with a paired t-test comparison,
>> i.e. what is relevant is the t2-t1 difference and not the
>> comparison A vs B at t1. For peace of mind you should verify that
>> t2-t1 for the control population is centered at 0 - that's the
>> relevant null assumption to test against.
>> hth
>> Christian
>>
>>
>> On 3 Jul 2008, at 10:49, Anna Rieckmann wrote:
>>
>>> Hi all
>>>
>>> I am looking for fMRI expert opinions on the following situation:
>>>
>>> We have data from an intervention study over two time points. The
>>> aim of the
>>> study was to compare changes in activations during a cognitive
>>> task over
>>> time between the treatment and the control group.
>>> Unfortunately, the two groups are not equal at baseline. Although
>>> allocation
>>> to groups was random, the intervention group happens to show much
>>> stronger/more activations at baseline. This makes it obviously
>>> hard to
>>> attribute the changes in activation over time to the treatment
>>> rather than
>>> the baseline activation.
>>>
>>> I am wondering if there is a good way to statistically control for
>>> baseline
>>> activation when looking at the change over time or whether we
>>> should accept
>>> the situation as it is and not proceed any further with this data.
>>> Any opinions are much appreciated!
>>>
>>> Thanks
>>>
>>> Anna
|