Getting my ordinal data ranked

#1
Hello,

I have the following issue.
I have data where 9 characteristics have been rated using an ordinal scale ("very important", "important", "neutral", "unimportant", "very unimportant").

I now want to bring the data in one rank so that I could e.g., say characteristic 5 is most important, characteristic 9 is second etc...

What methodology could I use best in this case?

Thanks for the help.
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
Can you just ascribe a numeric value (integer) to the groups then sum and find mean or use median for each characteristic.

How many respondents did you have?

And did the exact same group rank all of the characteristics?
 
#3
I have 120 respondents. And the data is "dependent" as they all rated all of the characteristics.

Giving integers to the groups and calculating "arithmetic mean" would mean I assume my data is not ordinal but interval scaled. This is however not true as I cannot tell that there is always the same difference between my values, i.e. difference between "very important" and "important" might be different than the difference between "important" and "neutral".

I have read something about Friedmann test and post-hoc Wilcoxon ranked pair test. I am not sure though how this should work. Any idea?
 

hlsmith

Less is more. Stay pure. Stay poor.
#4
Friedmann and Wilcoxon tests are used with non-parametric data (not normally distributed), so if you had reservations in finding mean or median - these reservations will probably carry over to these tests as well.
 
#5
Hallo,

According to what I have read, Friedman only tests if my characteristics are different in their individual ranks. But I cannot use the calculated ranks. Hence in some places it is recommended to use a Wilcoxon ranked pair test as "post hoc".

Wilcoxon test would have to be run on EVERY characteristic combination. I have nine characteristics. This makes 36 combinations in total. To compensate for multiple comparisons (Type I error) the significance value has to be adjusted. With 36 tests this adjustment is another challenge which method to choose to correct the alpha level.

BTW. I have no issues with the median, but with the mean. the wilcoxon was explicitly developed using the median and is hence working for ordinal data.

https://statistics.laerd.com/spss-tutorials/wilcoxon-signed-rank-test-using-spss-statistics.php

Thanks for support.
 

hlsmith

Less is more. Stay pure. Stay poor.
#6
Ideally, most of these decisions are made apriori. The reason is, so you do not run into these issues. Yes, pairwise post-hoc comparisons will definitely present some problems since you do not want to come off as fishing for significance.