# Using Mann-Whitney U?

#### picaritos

##### New Member
Hi,

I'm confused about which statistic to use. I want to find out if a law in two different states influences service delivery (get referral/ do not get referral). So I have 2 groups that I want to compare their yes/no response and see if there is a difference. Am I using the Mann-Whitney U? Other? Thanks.

#### picaritos

##### New Member
Thanks for the quick response. The sample size is 290. I surveyed a group of psychologists from 2 different states about their services. My research question is given different state laws, is there a difference in referrals for the two states. Psychologists answered yes to receives referrals and no. In working with dichotomous variables and nonparametric I believe.

#### victorxstc

##### Pirate
I think depending on your subgroups and sample size, you can use a Mann Whitney U, or a chi-square, or a Fisher exact. Please give us more detail on your project, the subgroups, risk factors.

Since you are working with proportions here (proportion of the Yes answer), you should run a chi-square goodness of fit test.

You should use a chi-square goodness of fit test as you don't have a 2x2 table here. This chi-square compares the proportions of your sample with a default 50/50 proportion to see if the proportions in your two groups differ significantly from an even 50/50 proportion. See these links (plz click here and here).

ok, that makes more sense. I ran it last night and was able to do it. However, I have another question... I'm using the same procedure to look at an association between 1 demographic (nominal) variable and a likert scale (1-6). However, I have too many cell counts below 5... I looked into the fisher's exact test but I don't have that add on. My contingency table will be 5x6 I believe.

Thanks for all the help so far!!

I think the best option for your situation might be Rank Biserial correlation coefficient as it deals with correlation between ordinal and nominal variables.

I think it was a misunderstanding as think he mentioned association not necessarily difference; but am not sure, maybe he has said difference in another post.

I'm using the same procedure to look at an association between 1 demographic (nominal) variable and a likert scale (1-6).
Btw, I too agree that chi-square or dichotomizing the 6 levels (for a Fisher) were also practical, good solutions. But once I faced a similar design, and after some trials I thought that correlation test might be the best option because it maintains a higher power (I think) than a chi-square with many cells less than 5, and also doesn't loose data by collapsing the groups.

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Verification, 2 x 2 table consisting of: state A or state B by referral yes or referral no. If this is it, then as previously proposed, the Chi-square should be appropriate.

#### picaritos

##### New Member
ok, I was thinking I'm trying to compare the responses of 2 groups (psychologists from state 1 and from state 2) and see if there is a difference in their self-report of referrals received (yes/no).

just to be sure... I have to weight cases on my referral variable and then plug-in state variable on SPSS chi-square stat (under nonparametric). sorry for such elementary questions but when I do that I don't get that 2x2 table

Last edited:

#### bukharin

##### RoboStataRaptor
Huh? This sounds like a classic 2x2 table of independent binary proportions:

Code:
           |       referral
state |        no        yes |     Total
-----------+----------------------+----------
1 |         a          b |     a + b
2 |         c          d |     c + d
-----------+----------------------+----------
Total |     a + c      b + d |     a + b + c + d
Therefore picaritos should use a simple chi-squared test.

#### picaritos

##### New Member
ok, that makes more sense. I ran it last night and was able to do it. However, I have another question... I'm using the same procedure to look at an association between 1 demographic (nominal) variable and a likert scale (1-6). However, I have too many cell counts below 5... I looked into the fisher's exact test but I don't have that add on. My contingency table will be 5x6 I believe.

Thanks for all the help so far!!

#### hlsmith

##### Less is more. Stay pure. Stay poor.
So the ...
1 demographic (nominal) variable
has six groups, making it a 5x6 tables? With the six coming from the six Likert categories?

Two possible options, perhaps if you have to use the Fisher's test and have small cell values, you can collapse some of your groups if it seems appropriate. So collapse data down into 3 groups.

However, if it seems like Fisher's Exact is your best option, you can do many simple statistics in Microsoft Excel, which available to many people. Just do some searching online on how to do this. Though, you may want to also do the chi square with Excel as well if you go this route. This will make the write-up cleaner and more straightforward to readers and individuals wanting to replicate your work. Instead of saying, I did this test with this and that test with that, etc.

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Clarification, the collapsing would get you closer to using the chi square.

#### picaritos

##### New Member
yes, the 6 would be coming from the likert scale... I'm afraid I'd lose information by collapsing the categories and I don't think they land themselves nicely into 3 groups (particular the 3 & 4 points). I'll look into the excel option. Thank you so much!

Last edited:

#### hlsmith

##### Less is more. Stay pure. Stay poor.
The Rank Biserial may work with these data types, but the person stated that they wanted to see if there was a difference, not sure if detection of a correlation would best answer the question.