Hello! Just a couple of questions about chi squares.

Current situation: students have taken a surve that has a likert scale of 1-5 (with 1= strongly disagree and 5 = strongly agree). I want to test if there is a difference in attitudes between student year (freshmen vs. sophomores vs. juniors). There are 26 questions.

1- Is chi-square the only way to interpret whether there is a significant difference between groups since this is categorical data? I'm assuming I'll do a chi-square for each question.

2- should I do the chi square for all three years at once? (can I just add a third category if I change my degrees of freedom accordingly?)

3- is there a way to detect WHERE the difference lies (post hoc test for chi-squares)? or do I just have to look at the frequency of responses and make an educated guess where the difference is?

I'm using excel but have access to SPSS as well.

Thank you in advance!

There really is no "correct" way to analyze data from Likert scales - just a lot of opinions. Likert scales are at least ordinal, and many would argue that they can be treated as interval as well, without much concern over misleading results.

Chi-square would be one way to go, as well as a test to compare medians (Mann-Whitney U Test) or a simple t-test. I would try all three ways.

You could do post-hoc tests for proportions to see where the differences are if you have a significant overall chi-square statistic.

Yes, you can do a chi-square for all three levels - just adjust the degrees of freedom.

Thanks, John M! I will try them all.

When you say do a post-hoc test for proportions, how do I do that in SPSS? (is it part of a chi-square test, Mann U or t-test?)

Here's a link to a PowerPoint file that does an excellent job of explaining how to do post-hoc tests for chi-square in SPSS:

http://www.utexas.edu/courses/schwab/sw318_spring_2004/SolvingProblems/Class24_ChiSquareTestOfIndependencePostHoc.ppt

That is perfect! John, you're a star. :-)

Thank you!

In one of my stats books, it says that as a rule of thumb, you shouldn't use the Chi Square if more than 20% of your squares have an expected value less than 5.

I had quite a few questions that had percentages higher than 20. I tried combining (Strongly Agree with Agree) and (Strongly Disagree with Disagree) to reduce the number of categories. This got some of the percentages down, but not all of them.

Can I still use the results if I report that I did not satisfy the expected values criteria? Or am I expected to throw these questions out?

Yes, I would still use them - but you need to note the problem in the report....sometimes you can't avoid it, but at least mention it.

Virtually all reports / theses / dissertations have a section entitled "Limitations of this Study" or "Issues With This Study"

Thanks, John. I'm glad I can still use them if I make a note of the problem. Most of them got down to about 26%, so that's not too bad.

hi there,
i happened to stumble onto the post hoc of chi sq powerpoint before seeing it on talk stats... tink its awesome too! very clear. had no idea that a post hoc for chi sq existed before this!

One thing though my chi sq was significant (p=0.047) but the post hoc was not significant for any of the cells. All the std residual was within -1.96 and 1.96.

Would some one be able to tell me how that could occur? i read somewhere else that there is some adjustment to be made to the critical values to prevent type I error for stats? is that the problem here?

thanx

There's nothing wrong - this can happen. Remember, the overall significance test is working with a larger sample size than the individual post-hoc tests, so it has a higher statistical power....