I used RStudio to perform Shapiro-Wilk test for normality. In the following data set calling it $D$, I have performed Shapiro-Wilk test for normality with p-value=0.7698.

34.681 26.291 33.280 36.169 41.471 31.528 25.502 43.211 35.330 30.447

If I consider log transform of data set $D$(i.e. take log on each element of $D$), I perform Shapiro-Wilk test on the transformed data set $log(D)$ with p-value=0.7707.

In particular, I have both $D$ and $log(D)$ data set following normal distribution as Shapiro-Wilk test did not reject either one of them.

$Q$: How should I resolve such contradiction? Note data $D$ and data $log(D)$ cannot both be normal.

  • $\begingroup$ There's no contradiction in these results: they merely show that the logarithm does little to change the distribution of a set of data with a small CV. $\endgroup$ – whuber Jan 27 at 14:26

This is an excellent example of a non-rejection not implying that the null hypothesis is true, and I will use this the next time someone asserts that.

You resolve the contradiction by noting that you only have ten observations, so the test has limited power to detect non-normality. Neither the data nor the log of the data look distinctly non-normal, but this does not mean that they are normal.

In each case, the test is shrugging its shoulders: “I can’t tell, folks; I can’t tell.”

  • $\begingroup$ So is there anyway out of this dilemma? $\endgroup$ – user45765 Jan 27 at 14:09
  • $\begingroup$ There is no dilemma. We should back up: what you are doing is similar to procedures that are known to be invalid: that is what we should be concerned about. What statistical problem do you hope your testing will solve? $\endgroup$ – whuber Jan 27 at 14:27
  • $\begingroup$ @whuber I mean do I use a different test to do hypothesis testing? If so which one? That is basically my question. I only know Shapiro-Wilk testing for normality. According to wiki, it is best for overall performance. $\endgroup$ – user45765 Jan 27 at 15:22
  • $\begingroup$ @whuber Essentially, I want to test whether the data is following normal distribution which will tell me what estimators I should use. $\endgroup$ – user45765 Jan 27 at 15:36
  • $\begingroup$ That's usually not a good idea: instead, you should choose your estimators based on what you want to estimate and on your model of the data. $\endgroup$ – whuber Jan 27 at 18:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.