Dear Lauren,
Yes, that is a problem.
In this case I would advocate using an empty EV for the choice (difference between buttons) as you cannot estimate any effect of choice in those conditions, but it is important not to have an EV that would remove all the results from the overall EV (and having an identical EV to the overall EV would do exactly that, and should be avoided). With an empty EV you wouldn't be able to get any results for that choice, but it would still let you estimate something for the overall mean.
All the best,
Mark
On 26 Oct 2013, at 00:30, Lauren S <[log in to unmask]> wrote:
> Dear Mark,
>
> Thanks again for your response to this. It has been a few weeks, but I'm analyzing a few new participants and running into an issue (I've pasted your original response to my question below this one, for reference).
>
> As you suggested, I created 6 EVs to model the mean overall of both button presses (1 EV per stimulus type) and 6 EVs that model the difference between button presses (1 EV per stimulus type).
>
> Here's the issue I am running into: a few of my participants are always choosing the same button for an entire stimulus type. Obviously this means we don't have the ability to differentiate between button choice in these conditions-- for my study, this isn't the end of the world. However, what DOES seem to be a problem is that this is causing my matrix to be rank deficient.
>
> What will be the outcome of this rank deficiency: Will the overall mean EV for this stimulus type be affected? If so, do you have a recommendation for dealing with participants who only chose one response for certain conditions? I don't want to simply delete the choice EV altogether because I am eventually planning to run group analyses- although I could make it Empty (all zeros).
>
> Thanks so much for your help.
> Best,
> Lauren
>
> Original message:
>
> Dear Lauren,
>
> To start with, you should model any baseline condition - so if a null trial is equivalent to what is happening during baseline (which I think it is from your description) then you shouldn't have an EV for this.
>
> You also need to be careful not to over-model things, which is what you are currently doing. For instance, if participants always press either 1 or 2, then the sum of the EVs associated with 1 and 2 for a particular condition (i.e. type of stimulus) includes all of the presentations of that condition and would be exactly the same as the original EV that modelled the mean of that condition over both responses. That is, you can't have separate models for (within one condition) the mean of button 1, the mean of button 2, _and_ the mean overall. This is because the mean overall is simply equal to the average of the other two (weighted appropriately). If you try to model it as a separate thing then this redundancy ends up as a rank deficient design matrix.
>
> So instead I would do the following:
> - have 6 EVs which model the mean overall both button presses for each stimulus type
> - have another 6 EVs that model the difference between button press 1 and 2 for each stimulus type
> The second 6 EVs are easily set up with the 3-column format, using the third column as +1 for each instance of button press 1, and -1 for each instance of button press 2.
>
> In this way you can look look at contrasts associated with any of the first set of 6 EVs, to see the average response (irrespective of button press), or a contrast associated with the second set of EVs to see the difference in activity between instances of button press 1 and 2 (use a +1 contrast to see when activity is higher for button 1, and a -1 contrast to see when it is higher for button 2).
>
> There are other ways that you could also set up a design which is equivalent and correct, so it is possible that some of the different answers from your colleagues are effectively the same as this. However, the initial issues of over-modelling (including null events, and having EVs for mean, button 1 and button 2) are definitely things that you have to avoid, as they will create rank deficient designs and must not be done.
>
> I hope this makes things clearer.
> All the best,
> Mark
|