4

In the testing of normality, how would the 2 compare? Is one significantly better than the other?

data_miner
  • 41
  • 1
  • 2
  • http://stats.stackexchange.com/questions/36212/what-tests-do-i-use-to-test-if-error-residuals-are-normally-distributed – Taylor Sep 13 '12 at 05:09
  • 1
    There was a recent question on the SPSS Nabble group that gave a reference to [a similar question](http://spssx-discussion.1045642.n5.nabble.com/jarque-bera-test-td5714665.html#a5714669). Also some googling for Shapiro Wilks power finds a question [here that is highly rated](http://stats.stackexchange.com/q/1645/1036). Possible duplicate without further clarification. – Andy W Sep 13 '12 at 06:14
  • 4
    Which is more powerful depends on which alternatives you consider - since they measure nonnormality differently, each has good power against slightly different things. As general tests of normality, both are excellent - with really good power against a wide range of alternatives. Neither is uniformly better. Do you need power studies? – Glen_b Sep 13 '12 at 07:18
  • @Taylor The question you linked is relevant here, but this is not an exact duplicate. –  Sep 13 '12 at 10:24
  • Glen_b is right. The recent similar question asked how to test model residuals for normality. One respondent cited a paper that shows Shapiro-Wilk is more powerful than Anderson-Darling. But he does not cite what the assumptions are that lead to the result. I mentioned that it cannot be globally true. You might want to look at the similar questions. As mbq points out, the recent one which I am referring to is definitely not an exact duplicate. Andy W. gives in his second link a question that is very close and has the answers you are looking for but does not address power. – Michael R. Chernick Sep 13 '12 at 10:48
  • An answer on the recent (preceding one) does provide a reference that does. – Michael R. Chernick Sep 13 '12 at 10:53

0 Answers0