Wilcoxon Rank Sum Test between countries

[Marius Bauer]

Hello everyone,

I was talking to my supervisor and he asked me to do a Wilxoxon ranksum test for my dataset. He wants to see that my datasets are not identical and show differences.

My dataset looks as following: I do have 3 countries which are actually run one by one and are not combined into one dataset. Now I did combine dataset one with dataset two and run the ranksum test by using the following code:

ranksum zfund, by(country)

Two-sample Wilcoxon rank-sum (Mann-Whitney) test

country       obs    rank sum    expected

1      4291    13414650    13261336
2      1889     5684640   5837954.5

combined      6180    19099290    19099290

unadjusted variance   4.175e+09
adjustment for ties  -463574.07

adjusted variance     4.175e+09

Ho: zfund(country==1) = zfund(country==2)
z =   2.373
Prob > z =   0.0177

My zfund is my independent variable which indicates my investment volume which is different for each country and time. However, I am confused about the Prob > z = 0.0177 does this not mean that my datasets are similar?

thanks

[Clyde Schechter]

No. "Prob > z = 0.0177" means "p-value = 0.0177," which, if you are going to do conventional null hypothesis significance testing, means you have a "statistically significant" difference in the value of zfund between the two countries.

That said, if your advisor really said and meant that he wants to see that your data sets are not diagonal, one would not do any statistical test for that. It would require only finding two observations, one from each data set, that are different. I'll assume that you (or he) were just speaking loosely.

[Marius Bauer]

I am confussed. I thought the Wilcoxon ranksum test is doint exactly this what you said?

For what else would I use the ranksum test?

Thank you for the explination with the p-value. I guess I missread something there

[Nick Cox]

It seems artificial to me to combine two countries when there are three. So that detail alone implies Kruskal-Wallis, not Mann-Whitney. But, but, but:


1. With thousands of data points, even trivial differences will lead to a report of significance at conventional levels.

2. Is independence of observations within countries satisfied, because that's an underlying assumption here?

3. I don't follow your supervisor's argument as you explain it.

[Carlo Lazzaro]

Marius:
as Others highlighted in their replies, there something strange in the way your data were appended and analyzed (as requested by your supervisor; what strikes me is the 2:1 countries comparison).
At the risk of tasting plain vanilla (with no chocolate variations), I would have gone -regress- (or -qreg-):

regress zfund i,country