Is one way anova with ranking appropriate test?

canrunner

New Member
I have two independent variables that I have data in which they have ranked 9 choices, 1-9. One of the groups has 30 respondents while the other is 120. A chi- square will not work because the expected response rate is below five on one group. I have read a one way anova has issues in this situation. I have calculated percents of each group response rate to each ranking etc but really need a better test. What is my best option?

katxt

Active Member
What are you actually wanting to find out? Give us an example of what your conclusion may look like.

canrunner

New Member
The questions,(short version) "Is there a significant difference in the factors that influence a person to leave their position." two groups ranked thier predicted reason for leaving one through nine. So can I run a test to determine if there is a difference. I have calculated percent of each group ranking each item. Is this the best I can do?

Karabiner

TS Contributor
You could create 9 pairs, one for each reason. For each reason, you indicate its median rank in group 1 and the median rank in group 2.
Then you calculate Spearman's rank correlation for the n=9 pairs.

You could perform up to 9 Mann-Whitney U-tests, one for each reason (grouping variable "group", dependent variable "rank of reason A [or B, C... I, respectively]")

With kind regards

Karabiner

katxt

Active Member
Here's an another fairly direct approach which could (should?) indicate a difference if there is one. For each subject in one group, count the total number of agreements with the subjects in other group. Total all these agreements. This is your measure of inter-group agreement.
Now do a permutation test. Mix up all the subjects separating them into two groups at random and get a distribution for the inter-group agreement assuming no difference. See where your original measure fits and get a p value. Low p means a difference.

(It might be useful to first establish that there is agreement within each group before you start anything else.)
kat

katxt

Active Member
You could perform up to 9 Mann-Whitney U-tests, one for each reason (grouping variable "group", dependent variable "rank of reason A [or B, C... I, respectively]")
Or, possibly even better, do a repeated measures MANOVA comparing the two groups. The data isn't normal but it isn't too badly behaved, and, as Karabiner has often quite rightly pointed out, with the amount of data you have, the results can be relied on.