Dear all, I want to clarify myself. I have about 99 items in my questionnaire meaning I have 99 variables in SPSS. I have run the outlier test for all the 99 variables and found that almost all the boxplot to show outliers. Shall I count the number of repetitive outliers and report the present of outliers? What must I do?
Yang Li, exactly I am interested outlier at the respondent level. How should I detect it?
What do you think constitutes an outlier in Likert scale data? Someone answered "5" on all the questions? Why would this be an outlier? What if it is just their answer?
Meaning to say, Likert Scale has its floor and ceiling. Anything fall in between should not be an outlier?
What I am saying is that it doesn't matter how SPSS draws a boxplot. What matters is what would be a meaningful "outlier" to you.
Salvatore is correct- this would depend on type of variable, responses given, and what is natural.
I do agree with @Salvatore S. Mangiafico. It may be just their personal opinion or rating. Perhaps I should take @Chalamalla Srinivas advice, to ignore outliers since it does not really representative outlier behavior. It may just the true opinion from the respondent. What do you guys think?
I think it is (always!) worthwhile to look at the data and see if there is anything impossible or suspicious. (That is, really an "outlier".)
It's difficult to come up with cases with Likert items or Likert scales that would qualify. 1) If someone answered all 5 or all 1, then you might suspect they didn't answer honestly, but just answered as quickly as possible. 2) There may be other suspicious things. Like if someone identified as an atheist and then had a really high score on a religiosity scale. Or a child answering as only an adult could.
If you are analyzing each of the 99 items as a separate variable, you might sum all values for each respondent. (Or sum them in logical groups of the 99 items). If you plot a histogram of the sums, that you give you some sense if there were one or two individuals that were very different than the others. This wouldn't justify excluding their data, but you might investigate these individuals for anything suspicious.
Different software has different rules for labelling a value in a box plot an outlier. This in no sense means it is an outlier in a statistical sense (which itself has no agreed definition).
In most cases a value being extreme on a box plot is insufficient to warrant any action in and of itself (with exceptions being odd situations such as where the value is clearly an error such as getting 7 on a 5 point scale). When you have a statistical model what usually matters is not the raw data but the residuals of the model. In such cases it usually makes sense to look at influence using statistics such as Cook's distance rather than extremeness (e.g., using studentized residuals).
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS