I have been wading through the many discussions on outliers on this site but I am still unfortunately having difficulty determining what to do with my data set.

My study consists of a simple pre-post test setup whereby i conduct 6 tests prior to and 6 tests following my treatment. The purpose of the repeated tests is to decrease the effect of anomalous performance on the results and increase the sensitivity. We therefore expect that generally 1 or 2 of these 6 tests to be a bit different to the other 4. At present I've been removing abnormal scores by identifying those results that lie outside +/- 1 standard deviation. This practice has proven somewhat satisfactory (in that it detects most of those scores that upon visible inspection appear anomalous), however I read more and more that the use of standard deviation is not appropriate for determining outliers, and sometimes I've found that the scores deemed as outliers are not always those that appear the most anomalous.

Therefore I'm wondering whether you can offer any suggestions as to how i can improve/better standardise my selection process? Is my current method acceptable? If so, do you know of any published material that has validated this approach (or does this approach have a name that i could google?) Alternatively, would the use of an interquartile range approach using the median value prove more satisfactory?

  • 3
    $\begingroup$ I don't think it's a good idea to standardize the removal of outliers. Rather, you should work on a method for identifying possible outliers. Then look at them more closely. However, even identifying possible outliers is tricky and not easily automated. For example, there can be masking, where the existence of an outlier masks other outliers. Also, depending on sample size, various deviations will be expected to occur. $\endgroup$ – Peter Flom Oct 23 '12 at 20:50
  • $\begingroup$ Thanks very much for the reply. I understand the reluctance to remove outliers but the nature of my testing (human subjects, a technical physical movement, ~1 sec testing period) means that with the 6 tests we perform often 1 or 2 are all over the place whilst the other 4 are very similar. Although I agree that those anomalous movements are of likely significance, we think the 'truer' measure are the 4 values that would be obscured by the other two. We have found instances of masking which is why we wondered whether there was a more suitable approach to identifying outliers. $\endgroup$ – Rob Gathercole Oct 25 '12 at 20:35
  • $\begingroup$ As you say, rather than standardising the removal of outliers does anyone know of any standard ways of identifying possible outliers that could be adopted with my small sample? Thanks again! $\endgroup$ – Rob Gathercole Oct 25 '12 at 20:37
  • $\begingroup$ Look at a box plot and see not only who is an outlier by the default definition, but who is really far from other points. $\endgroup$ – Peter Flom Oct 25 '12 at 20:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.