Print

Print
See below.

Best Regards, 
Donald McLaren, PhD


On Tue, Oct 27, 2015 at 10:20 AM, Laurens Van der Cruyssen <[log in to unmask]> wrote:
Dear list, I'm posting this on behalf of a colleague who was unable to post this on the mailinglist.

Hi all, My supervisor asked me to post this on behalf of him as he is not in the mailing list. "It is often recommended to model error responses and fillers in separate vectors. The reasoning seems to be that by doing this, error responses and fillers are excluded from the residual error term, leading to better estimation of the model. But what is SPM actually modelling?

SPM builds regressors for each trial type that you specify. The model uses these to find the best fit of these regressors to the data and the remainder is he residual error term.

Let's assume for a second, that there is no residual error in the model. This means that the model is a linear combination of the data. If we fit the complete model, there is no error in the model. Now, let's remove the error trial regressors. If we refit the data to the reduced model, then you will find that you have non-zero residuals as there is data that can't be explained by the model because you gave it a model that said nothing happened during those periods of time.

 
If we do not specify onsets for error responses or fillers, then it is simply not part of the model because not part of the design.

Correct.

 
Hence, these errors and fillers will not inflate the residual error.

Incorrect. Because the errors and fillers are not in the model, they must go into the error term.
 
If this reasoning is correct, there is no need to model errors and fillers as separate vectors of no interest. Please give advice." Thanks in advance. Cheers,

As the reasoning is incorrect, you need to include the errors and fillers in your model.

You also risk mis-fitting the regressors you do specify to account for the data that you failed to add a known source of variance for in the model. 

Although not a task analysis, but a gPPI analysis, McLaren et al. 2012 demonstrated that the failure to model all trial types could lead to false positive of contrasts for the trial types in the model. This reinforces the notion that you should model everything that the subject experiences.

Hope this helps.
 


--
Laurens Van der Cruyssen

---

Postdoctoral researcher VUB
Experimental & Applied Psychology
Cognitive Psychology

Postdoctoral researcher UGent
Ontwikkelings-, Persoonlijkheids-, & Sociale Psychologie
Sociale Psychologie

Visiting Professor UGent
Data-Analyse
Onderzoeksmethoden I & II, Onderzoeksmethoden in de Psychologie