2

I have a set of data which contains answers to a questionnaire (questions about satisfaction rated 1 to 5) and some demographic information (sex, age). I can divide the data in two groups and I want to see if there is any statistical difference between those groups. I tried the Wilcoxon test, but I have many ties, so I'm not sure this is the test that I need. What would be the best test in this case?

StasK
  • 29,235
  • 2
  • 80
  • 165
rslite
  • 185
  • 1
  • 8
  • Are you saying that the response scale for each item is 1 to 5? Or are you saying that respondents rank five questions 1 to 5? Or did you form an overall scale scores by taking the mean or sum of items? How are the two groups formed? – Jeromy Anglim Apr 25 '13 at 08:04
  • Yes, the response scale is 1 to 5 (how satisfied a patient was with different aspects of his/her admission to a hospital). – rslite Apr 25 '13 at 20:22
  • As a somehow related question - if you jiggle the results so there are no ties (i.e. instead of 1, 1, 2, 2 you use 1.01, 1.02, 2.01, 2.02) would Wilcoxon test be applicable to show a significant difference between the two groups? I have about 300-400 patients that answered the questions. – rslite Apr 25 '13 at 20:27
  • So participants are not "ranking", they are "rating"? How many items do you have? Are you combining these items into a composite score? or if not how are you dealing with multiple items? – Jeromy Anglim Apr 26 '13 at 09:10
  • Yes, they are rating. I want to check the differences for each question separately, not a composite score. So for the purpose of my question you can assume there is only one set of answers – rslite Apr 26 '13 at 12:24
  • 1
    Is your question any different from this one? http://stats.stackexchange.com/questions/203/group-differences-on-a-five-point-likert-item – Jeromy Anglim Apr 26 '13 at 12:40
  • @Jeromy - looks like it's the same question, however I don't know the terms, so I didn't find it in my search. What did you end up using in the end? Do you think t-test would be good for my purpose? – rslite Apr 26 '13 at 15:20
  • 2
    There are at least five questions on how to compare two groups or two ordinal variables: http://stats.stackexchange.com/q/35382/5739, http://stats.stackexchange.com/q/49318/5739, http://stats.stackexchange.com/q/203/5739, http://stats.stackexchange.com/q/59625/5739, http://stats.stackexchange.com/q/43113/5739, http://stats.stackexchange.com/q/56949/5739. So this is one of the most FAQed things on the site. – StasK Jun 25 '13 at 16:48

1 Answers1

-2

two-sample Kolmogorov-Smirnov Test

Glen_b
  • 257,508
  • 32
  • 553
  • 939
sav
  • 187
  • 1
  • 1
  • 10
  • The two-sample Kolmogorov-Smirnov does not deal with ties. If you ignore heavy ties you end up with a much lower significance level than desired, with corresponding impact on power. You would need to explain how to deal with ties. – Glen_b Apr 25 '13 at 07:55
  • I'm interested in the issue with the ties, since by design there are going to be plenty of ties. There are 300-400 respondents and there are 5 answers to choose from. – rslite Apr 25 '13 at 20:25
  • K-S test only works for continuous data, assuming that your CDF does not have jumps that mess up the asymptotic distribution. – StasK Jun 25 '13 at 16:49