I have found opposing views about outlier analysis for Likert scale data. Since the lowest and highest values are bound, does it make sense to apply the times standard deviation rule to identify outlier for likert data? Sometimes, a respondent can consistently rank his/her answer on a higher scale, that is a valid response, and why such responses should be eliminated just to increase the model fit? Any idea or comment with references would be appreciated. Thanks.
Dear Khandoker,
Outlier decisions are always based on judgement; the statistics themselves inform us about the influence of certain points and the consequences of our judgements. Keeping extreme points because they represent "real" data is perfectly reasonable! Using methods that assume normal distributions when some of your points belie this assumption is not always as reasonable...
I recommend that you consider models that either accommodate the extreme values, or examine the consequences of reducing their weight just to be able to respond to reviewers about the sensitivity of your conclusion to your choice. For what it is worth, I have almost never removed any extreme values - but I have always run my analysis with and without them to make sure that I understand the consequences of my decision.
Dear scholars, thanks for your feedback and answer. Much appreciated.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis