What statistics test should I use?

Question

I have survey results from two sets of respondents who differ by a single variable. I want to know whether there is a significant difference in their responses to any of the survey questions. Each question has a 1-10 scale. The two groups are unequal in size.

My first thought was to use a simple t-test. But then I realized the answers were not normally distributed in either group. In both groups, more than half the responses were "10". So then I thought I should use a Mann-Whitley U-test. But since so many responses have identical values, ranking them doesn't seem very logical.

Any advice?

Paul · Accepted Answer

Without seeing how this data is arranged I'm having trouble giving the best possible answer!

You could use a t-test because normality is not a requirement (it helps, but isn't absolutely critical). Randomness is an assumption that can't be relaxed, but normality can be. 

A simple binomial test could be used for each question, but wouldn't answer the question for the entire population of whether they differ for the entire survey.  It would just answer for each question taken separately. 

However, you might consider a chi-square test, which wouldn't rely on population parameters, but just on the number of answers on your survey.  The null hypothesis is that the two groups don't differ.  The number of degrees of freedom is the number of questions on the survey minus one.

The chi-square statistic is computed as (observed - expected)^2 / expected for each cell.

Let me know what you think of this approach.

trodriguez2011 · Answer

It depends on what you are looking to do. A two mean t test would certainly not yield any results, as the data is not normal. That is the test to use if you are trying to determine that the two sets are different to a certain significance. If you are merely trying to see if they are independent of one another, chi sq is the way to go.

Trending News

What statistics test should I use?

2 Answers