Frank wrote: > When the test on normality revealed non-normality of dep variable, > most often, we do log transformation to improve the normaliy before we > do regression. Why we need to do the > transformation for regression analysis? Any help will be much > appreciated. > Frank In linear regression, it is the residuals (not Y) that should be approximately normal. But even so, statistical tests of normality are not very helpful when they are used to test the assumption of some parametric procedure. The problem is that they have too little power when the sample size is small (which is when normality is most necessary), and too much power when sample sizes are large (and normality is less important due to the central limit theorem). In other words, tests of normality fail to detect important departures from normality when the sample size is small, and they detect unimportant departures from normality when the sample size is larger. -- Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM."

0 |

11/10/2009 12:57:54 AM