PDA

So Lost
08-01-2008, 06:09 PM
Hi, I'm new to the forum, and worried that I have screwed up my thesis project. I thought I designed a standard one group pretest/postest design. But I was working with kids, and the advice I was given was they do better with Likert scale (1-False 2-Somewhat False 3-Neither 4-True 5-Very True) as opposed to a 10 point scale. (Rate how true XYZ is to you- 1 being Not at all true and 10 being very true) Now I have this data, and I'm not sure it is kosher to do a paired t-test on it, because a Lickert scale is technical ordinal data, not interval data. I have taken statistics but it was 10 years ago, and I'm very rusty.

Here is my study in a nutshell:

I wanted to study the effect of a curriculum I designed on the students anxiety, confidence, and motivation for learning English. I administered the pretest, taught the curriculum, and administered the posttest. The pretest/posttest is a lickert scale survey with 5 questions each in 3 categories which are: Anxiety, Confidence, and Motivation toward learning English. Pretest and Posttest were exactly the same survey.

Can I add up the reponses in each section and get a mean and compare the means of the pretest and posttest using a paired t-test? Or should I treat each question as a seperate variable (each question covers a different aspect of the main category) and do a paired t-test for each question? Or is there a completely different test I should be doing?

This is not my only data. This is a mixed methods study with both qualitative and quantitative data. I've triangulated the data with my observations and interviews. I don't know if that helps any with not doing the perfect test.

TheAnalysisFactor
08-27-2008, 03:34 PM
Hi SoLost,

You should be fine. While it's technically true that Likert Scale data is ordinal categories, it is often assumed to be interval as long as:

1. The intervals between the values are generally equal: is the difference between a 1 and a 2 the same as between a 2 and a 3?

2. The underlying concept you are measuring is itself continuous.

The sum of ordinal variables is still ordinal, so running them separately instead of summing them doesn't help anything.

You have a few options. Run the paired t-test as planned, and check your assumptions well. If they look good, stick with the t-test.

If the assumptions don't turn out well, there is a non-parametric equivalent to the paired t-test called the Wilcoxan sign-rank test. It is available in SPSS under nonparametrics.

Good luck,
Karen