Dear Philip. How different are the p-values?

If the data you used are exactly the same, I would guess the difference is due to the precision (number of decimal places) used by the softwares. Maybe, Stata uses a higher or lower precision than Graphpad. By doing so, the SE values might be different and, consequently, the p-values.

If this is not the case, you might have faced some problem on data import. You could try to export the Stata DB into an Excel file and than importing it into Graphpad (or vice-versa).

It depends on the software used for study. there is no problem you interpret your results. there is worry when in one software result is significant and in another not significant.

Philip, I suspect you are getting an exact test from GraphPad and an approximate test from Stata. From https://www.graphpad.com/guides/prism/7/statistics/index.htm?how_the_kruskal-wallis_test_works.htm:

"If your samples are small (even if there are ties), Prism calculates an exact P value. If your samples are large,  it approximates the P value from a Gaussian approximation."

Stata's -kwallis- command (https://www.stata.com/manuals/rkwallis.pdf) appears to have no option for obtaining an exact test. There may be a user-written program for that, but I am not aware of any.

HTH.

I have done the same thing using Ranksum (Mann-Whitney) test and a have a difference between the two softwares: Stata p values was 0.01 and Prism GraphPad < 0.0001.

As Bruce Weaver noted, there are often different methods to calculate these statistical tests, and different software may use different methods for different conditions.

My advice is to read the software documentation carefully so that you can report how the statistics were calculated.

But also, it is helpful to post some toy data giving different results. Others can then check against known methods, for example in R, to confirm (maybe) the method being used.

Gabriel Ferreira – Stata provides methods and formulas for all its procedures, so you can check it. By default, Prism gives exact Wilcoxon Mann-Whitney test output for small sample sizes. Stata, on the other hand, uses the usual calculation. I ran the Prism example data in Stata (ranksum) and R (wilcox.test) and got the same answer, so I don't imagine it's a bug in either package, but due to the exact algorithm used by Prism.

Actually, my first response would be to check that the packages are analysing the same data! However, if you can, please share the data that caused the problem and we can have a look at it.

Ronán Michael Conroy thank you for your answser.

Please find attached the plan.

Gabriel Ferreira I've only got Prism 5, but it gives the P-value as 0.0157, using a normal approximation,

Mann Whitney test P value 0.0157

Exact or approximate P value? Gaussian Approximation

while Stata gives the P-value as 0.0148.

Ho: fv(group==0) = fv(group==1)

z = -2.437

Prob > |z| = 0.0148

Meantime, R gives this :

> wilcox.test(fv~group, data=foo)

Wilcoxon rank sum test

data: fv by group

W = 43, p-value = 0.01368

Good grief! Is this what you are seeing?

Using R:

p = 0.01368, exact test

p = 0.01572, normal approximation, with continuity correction

p =0.01481, normal approximation, without continuity correction

Very good! I have to check what happen in my prism software so. Anyway I always use Stata. Ronán Michael Conroy I am trying also to control three differents groups related with this variable. Age, gender and HIV infection. Therefore, I want to test among those that are HIV negative, into children group, is there difference bewteen males and females? Same thing for adults. In that case a regression is not good, I believe so, because it will include all HIV persons. Does anyone can give me some tip?

Gabriel Ferreira Regression will work. If you want to run the regression separately on adults and children, and your variable that records whether they are adult or child is called 'adult', then you just add a prefix

. by adult, sort : regress yvariable gender

You can check whether the effect is different in adults and children by using an interaction

. regress yvariable adult##gender

I'm not clear what your hypothesis is, or why adults and children might be different. It's hard to be more specific without understanding your data generation process.

Ronán Michael Conroy tthe variable that I am testing with is an immunological marker which has a different biological distribuition between these groups. I did not use regress once it does not fit into a parametric curve. I found the logic command to test it, however, sometimes I dont have a significant p value using for exemple Ranksum, but when I cut the variable and transform it in dicotomous and applyed a pararametric test I do find the difference. For the multivariate the analysis sometimes I have colinearity.