Is it appropriate to identify and remove outliers because they cause problems? This all pertains to my Psychology honours thesis.
I have two groups (Autism and control) and all participants completed four tasks. It is very important to my study that the groups do not differ on reaction time in each of the tasks. However, they do. The autism group responded faster than the control group. This confounds the results for the construct we actually want to investigate.
I thought I might correct the difference by excluding outliers from the study. I tried to identify outliers at both the univariate (Boxplots, SD = +/- 2.5, for each of the four tasks) and multivariate level (Mahalanobis Dsq). No participant comes up as an outlier. Then I thought I would exclude participants that have low average reaction times ('low' being an arbitrary value), but even so the difference between the two groups was significant. 


*

*Is there anything else I can do? 

*And how would I report such a process in my thesis?

 A: You should not exclude outliers just because they cause problems, nor should you use a subset of your data because the full data causes problems. Neither of these solved the "problem" in your case, but even if they did, it wouldn't be right.
You haven't given a lot of detail about what you are trying to do or how you are doing it, but can you add reaction time as a covariate? 
A: It is very important that you consider the possibility that the categories of subject have a real difference in reaction times. If that is the case then anything that makes the difference go away will lead to potentially artifactual results. Don't assume that an inconvenient effect is a result of the presence of outliers.
Perhaps you could look for a relationship between reaction time and another outcome measure. The form of the relationship may differ between autistic subjects and normal subjects.
A: It sounds like you need to explore your data a little more.  Why don't you try some unsupervised techniques like clustering.  Outliers would show up in their own groups.  And you would think there'd be some kind of grouping of your controls.
Regardless, you can still have a thesis about not seeing an effect you expected to see.  You'd have to explain how your data/method was not flawed.  And add a section about what variables you might add to explain why your test subjects and controls are grouping together.  This work still helps future researchers.
