1

I want to know the minimum number of data required to check normality of my data, I found in Minitab it should >= 20 , other say should be 30 data, on what basis this number is based, and in this case what simulation to be done to verify this number?

Gamal MOHAMED
  • 11
  • 1
  • See https://www.google.no/search?q=minimum+number+of+observations+to+test+for+normality+site:stats.stackexchange.com&safe=off&nfpr=1&sa=X&ved=2ahUKEwjCrJ_np8rhAhWp5aYKHWmJAO8QrQIoBDAEegQIBhAO&biw=1280&bih=583 – kjetil b halvorsen Apr 12 '19 at 10:10
  • 4
    Why do you want to test for normality? – mdewey Apr 12 '19 at 13:22
  • [This post](https://stats.stackexchange.com/questions/2492/is-normality-testing-essentially-useless) on the question of normality testing is relevant and worth reading, imo. – COOLSerdash Apr 12 '19 at 13:56

1 Answers1

2

The number needed is going to depend on how non-normal your data are and what those violations mean in your specific situation.

  1. If you are testing the normality of your data for an OLS regression, stop. Regression does not require normal data, it requires normal residuals.

  2. If you are testing the normality of residual in an OLS regression I think graphical methods are much better than any "test" (also, normality of residuals is only important for some aspects of regression).

  3. More generally, I'd still say that graphs (such as density plots, quantile normal plots and so on) are going to be more useful than any test.

  4. When you simulate, you should try simulating the effect of your non-normality on the method you are using. Or, if it's regression and you are worried about non-normality, you can use a regression method that does not depend on normality, such as quantile regression or robust regression.

Peter Flom
  • 94,055
  • 35
  • 143
  • 276
  • The number I need for testing process capability and control chart, yet control charts doesn't assume normality but capability does. – Gamal MOHAMED Apr 15 '19 at 12:12