Different Shapiro-Wilk test results in R, JASP and SPSS?

eniseg2

Hi everyone,
I know this isn't strictly speaking a Bayes question but it's a JASP question.

I've run the Shapiro-Wilk test on the same data in all three platforms and SPSS gives me W=.789, p=.027; R gives me W=.766, p=.008 and JASP gives me W=.804, p=.022.

What should I do here?

Is there any objective reason as to why this should be the case?

EJ

JASP uses an R package for this, must be the settings that make the difference. I'll ask Johnny to look into this.

eniseg2

Thank you. I was mildly confused by SPSS and R differing but I expected JASP to be the exact same as R because I knew it 'runs on R'.

This is for my thesis. How do I know which one of them to report? I've got some cases where one of them shows S-W as significant and in others it doesn't..

EJ

We'll look into it. If you are really eager, you can inspect the JASP code on GitHub and see what arguments to the R function we call. I'm inclined to trust R over SPSS, but that's a general prior value judgement.

eniseg2

I was using R to begin with and my analyses currently are based on the R Shapiro-Wilk tests. Was using them to determine which inferential stats to use. So I'll report the R ones for now but please do let me know what happens with the difference between JASP and R.

JohnnyB

Hi,
I just looked into this, and JASP uses the shapiro.test() function in R. For the two-sample t-test, JASP does this separately for each group, so you get two Shapiro statistics and p-values. It might be the case that you calculated the aggregated Shapiro statistic across the two groups, leading to a different value. So for instance, for the kitchen rolls example dataset in JASP:

shapiro.test(x = kitchenDat$mean_NEO) # to get the overall Shapiro statistic
shapiro.test(x = kitchenDat$mean_NEO[kitchenDat$Rotation == "clock"]) # separate per condition
shapiro.test(x = kitchenDat$mean_NEO[kitchenDat$Rotation == "counter"])

If you repeat this with the one sample t-test in JASP, you get the first result, and if you run the Shapiro test with the two sample t-test in JASP, you get the latter two results. I do not have access to SPSS, so I'm not sure what they do with the test.
In case this is not the reason for the differing results, could you please include a data example and R-code used to calculate the Shaprio statistics?
Kind regards,
Johnny