hi,
I would first delete the two (maybe even 3?) columns with practically no answers - it is just inflating the number of degrees if freedom, and then use fishets exact test.
regards
You can see in the below table that I have multiple cells that have expected counts <5 and cells with counts equal to zero + zero expected counts. As I want to analyze the relationship between my two variables, one being about attitude and the other about behavior, would fishers exact test be good to use in this situation? I'm wondering this as I've read that your variables need to have a dichotomous relationship to use the test. I'm not sure if its still okay to use the test between my two variables as I find it hard to discern if they are dichotomous. Can someone advise on this?
Note: And although the variables are supposed to be ordinal, they have a non-monotonic relationship so I have to use nominal measures of association as suggested to me in other forums such as the fishers exact test (when you have low expected count freq).
hi,
I would first delete the two (maybe even 3?) columns with practically no answers - it is just inflating the number of degrees if freedom, and then use fishets exact test.
regards
Hmm, I'm a bit reluctant to "delete" 2 columns with some data in them let alone three. I think what would be better would be to combine the levels of disagree (D) and strong disagree (SD). Also, people have recommended the same solution that you just did, but because I'm still new to stats, I don't understand this: It does kind of make sense for me to at least combine lets say D and SD and then do calculations because of the very low freq counts. But by changing the original number of levels, aren't the results going to be different had I not changed the number of levels? And if so (as it likely will be), will my results still be accurate after combining levels?
Thanks
As I saw your data I thought, why not add one more colum, with the opinion of those who believe in unicorns, then one more for the dwarfs and so on... , you see the point.
Seriously, the point is, what is your research question? There is no way empty columns will help you answer it.
regards
Umm...lol? I wanted to test for a relationship between the two variables that I showed above. Yes I do see your point. Its just that I thought that if there is at least 2 or 3 counts in the cells, it can still be computed and analyzed for a relationship.
I would say that to usefully apply a statistical hypothesis test, you need to actually have an hypothesis to test. Just wanting to test for a relationship between variables is too vague a specification. Does your study have a specific hypothesis? If it doesn't, forget the hypothesis testing and just describe your data using graphical methods.
Matt aka CB | twitter.com/matthewmatix
ondansetron (05-16-2017)
The hypothesis I want to test is that there is a relationship between recycling attitude and behaviour. Basically the more positive one's attitude towards recycling is, the more they will recycle.
Given your directional hypothesis, I would suggest a Spearman's rho correlation. Fisher's exact test would not really be testing whether "the more positive one's attitude towards recycling is, the more they will recycle". It would just be looking at whether the frequency with which individuals fall in the different recycling categories differs (in some way) depending on which attitude category they're in.
Matt aka CB | twitter.com/matthewmatix
Sorry I didn't mention earlier that my data have a non-monotonic relationship. So Spearmans isn't the best for this case. Heres what I think I need to do but please correct me if I'm wrong: I initially use Fishers test to see if there is a relationship between the variables. Then if I find that they do have a relationship, I further test their relationship by using a nominal-ordinal/nominal-nominal measure of association to test the strength of the relationship. However, I'm not sure whether to do nominal-ordinal or nominal-nominal. What test would you recommend me to do for my case?
the below image is where i graphed the data from the two variables
Last edited by slamer2000; 05-16-2017 at 11:51 PM.
You did mention this, but it isn't actually that relevant imo. Your question relates to the strength of a monotonic relationship: Changing to a Fisher's exact test based on a pattern observed in a sample of data means changing the question your analysis is asking. The data also really don't show very strong evidence of a non-monotonic relationship (maybe it is, and maybe it isn't). I'd stick with the analysis that actually answers your question. Ideally in future you should select and pre-register your analysis plan before collecting data.
Matt aka CB | twitter.com/matthewmatix
Tweet |