What test should I use?

[Chan Sutherland]

Hi All,
I’m an SPSS statistics novice and I was hoping for some help on a question which is;

Are people more likely to help strangers who ask for a cigarette than those who ask for money?
Both are ordinal variables with both labels being;
Definitely not
Probably not
Possibly yes
Definitely yes
I’m guessing it would be a non parametric test either Chi, Mann-Whitney, Wilcoxon, Kruskal-Wallis? Or maybe I’m completely wrong. Any help would be appreciated. Thanks

 

[obh]

Hi Chan,

Since it is an ordinal variable you should use the non-parametric test.
You can use Mann-Whitney (one group will be a constant)
Paired Wilcoxon is for a dependent test like paired -t (no this case)
Kruskal Walis is a generalization of the Mann-Whitney, when using two groups it is the same test.
For the Chi-squared test, you need to have an expected model, for example, the expected model may be the following model:
50% Definitely not or Probably not and 50% Possibly yes or Definitely yes.
But in this case, you will lose some knowledge as you merge the two "no" groups and two "yes" groups.

 

[Karabiner]

How did your study design look like? Did you ask each participant 2 questions? Or did you asknthe cigarette question in one group, and the money question in another group?

With kind regards

Karabiner

 

[obh]

I assumed it is one question:
Are people more likely to help strangers who ask for a cigarette than those who ask for money?

 

[Chan Sutherland]

How did your study design look like? Did you ask each participant 2 questions? Or did you asknthe cigarette question in one group, and the money question in another group?

With kind regards

Karabiner[/QUOTE
Hi there, each participant was asked the two separate questions and was wanting to compare who would be more likely to help one group from the other. Wilcoxon maybe? Thanks for your reply

 

[Karabiner]

The Wilcoxon signed rank test requires interval scaled variables, yours are ordinal.
You can use the sign test for dependent ordinal variables.

With kind regards

Karabiner

 

[Chan Sutherland]

Thanks for your reply, in regards to the question I posted I am only allowed to choose a suitable test from the above that I posted. Sign test not included unfortunately. I was under the impression that the Wilcoxon test would allow for at the very least ordinal data?

 

[Karabiner]

No. It compares differences, and that of course requires interval scale level.
But since many people think that it is for ordinal scaled variables, and the
Wilcoxon signed rank test is the only one for dependent samples on your
list, maybe you will try it.

With kind regards

Karabiner

 

[obh]

The Wilcoxon signed rank test requires interval scaled variables, yours are ordinal.
You can use the sign test for dependent ordinal variables.

With kind regards

Karabiner

Hi Karabiner,

If you assume that the null assumption of the Wilcoxon signed-rank test is:

H0: The probability of the difference is greater than zero equals the probability of the difference is less than zero

Then I assume ordinal data is okay.
I don't think that only the ratio scale is required, I assume any continuous data will do the job.
The only problem with the ordinal data is the ties (equal values), the traditional exact test assumes there are no ties.
For this problem, you may use the correction for the ties, but with z approximation.
I also read that there are exact calculations with ties, not sure if yet in any of the main software.

Clearly, only 4 options for a question has many ties.

When you use the sign test you loose knowledge since there is no difference between "Definitely not" and "Probably not"
I would only guess that the Wilcoxon signed-rank will have a better test power.

If the labels would be 1,2,3,4 instead of the ordinal "Definitely not", "Probably not",...
we may say it is an interval scale?
My personal opinion is that there is no significant "psychological" difference between the two options.

 

[Karabiner]

If you assume that the null assumption of the Wilcoxon signed-rank test is:
H0: The probability of the difference is greater than zero equals the probability of the difference is less than zero
Then I assume ordinal data is okay.

The Wilcoxon signed rank test compares the magnitude of positive differences
with the magnitude of negative differences. To compare differences, you need
an interval scale.

With kind regards

Karabiner

 

[obh]

Yes, you are correct.
I assume for Mann U Witney continuous variable is sufficient.
but definitely, when comparing differences it must be an interval scale.