3 values ordinal variable: Mann-Whitney or Chi-squared?


I have an ordinal variable which assumes only 3 values. I need to do a comparison between 2 treatments.

I used Mann-Whitney test, since is an ordinal. My "p-value" is quite high (0.9388) which is strange because the values distribution is quite different for each group.

Then I perform a Fisher a chi-squared test and my "p-value" decreased to 0.0885, which I believe, reflects a little bit better the behavior of the sample.

What should I do? Stick to Mann-Whitney, go to chi-squared or do another type of evaluation?

Thank you very much
It would help if you could post your results (the 3x2 table) but it sounds like there may be a non-linear relationship, or you may have coded something incorrectly
Thank you for answering so quickly.

Here goes my data:

(bad) 1 2 (6.7%) 6 (21.4%)
(ok) 2 15 (50%) 7 (25%)
(great) 3 13 (43.3%) 15 (53.6%)

I also applied a Fisher Exact. I got a "p-value" of 0.0965
I'm sorry if this a dummy question, but what is the difference between a ordinal variable (bad,ok,great) and a ordinal relationship between 2 subgroups (A and B)?
The variable is ordinal because it is "bad", "OK", "great" - they are in order but not equally spaced.

The relationship is not ordinal because the ratio of A to B is 0.31 for bad, 2.00 for OK and 0.86 for great. The ratio goes up and then down. So, you can't say which is "better" only that they are different.
So, imagine that instead of "bad", "ok" and "great" I am working with classes of ages: [18,39] years, [40,59] years and [60,80] years
So, even if my variable is based on a quantitative, since the ratio is not constant I can't work it as an ordinal. I will have to treat the age classes as nominal?
No. Ordinal does NOT require equal spacing. That would be interval. Your variable is ordinal but it has a non-monotonic relationship with your groups.


Less is more. Stay pure. Stay poor.
Was treatment randomized or do you have baseline data to show groups were equal (comparable) to start with?
It was a randomized treatment with a similar baseline. Right now I'm evaluating pain which is measured in a scale between 1 to 5. My values decreased from 4 and 5 to 1 2 3, but the distribution is not identical any more, therefore, there is a difference in the treatments during follow-up. My sample size is very small, however it is important for us to have an idea of this behavior is affecting both groups.

I didn't calculated the ratio of the categories between groups because I wanted a statistics on the overall behavior. Since I am not used to work with ordinal variables my first choice was the Mann-Whitney. However, the "p-value" indicated a similarity which I don't think it reflects the reality.

Since the variable is a qualitative I applied a fisher and a chi-squared test just to have another overview about the categories distribution. Both these tests indicate gave me a calculated probability closer to my opinion of existing differences. Therefore, I need an alternative test for Mann-Whitney since I want to keep on treating the variable as a ordinal.

Peter, I'm sorry if I wasn't understanding your answers but now I understand what you where trying to say. However, my goal since the beginning is to evaluate difference in the behavior and not correlation between variables. Do you think that I can not compare the differences between the two groups? For me, in a clinical point of view, having only 6.7% of subjects with pain against 21.4% is quite significant.