Tahir,

The Shapiro-Wilk test rejects the null hypothesis (of errors being distributed normally) with large sample sizes - incorrectly in most of these instances. Your QQ-Plot seems reasonably and sufficiently Normal to me.

Hope this helps

Tahir,

The Shapiro-Wilk test rejects the null hypothesis (of errors being distributed normally) with large sample sizes - incorrectly in most of these instances. Your QQ-Plot seems reasonably and sufficiently Normal to me.

Hope this helps

One of the variables? -- The assumption is that the conditional distribution of the response is normal, what can be rephrased as that the model residuals should be normal distributed. Thus, actually, the normal-QQ-plot should show 99 points.

Formal tests on assumptions don't make sense. There are hundreds of posts even just here in RG explaining this. The point is not whether of not your residuals are normally distributed. The point is whether the assumption is reasonable for your kind of data. This means, there should be no obvious patterns or severe deviations from a straight line in the normal-QQ-plot. How "obvious" and "severe" should be defined is context-specific and, eventually, your expert decision (as nobody knows the context better than you do - hopefully...).

The normality condition is for the residuals from the fitted model. Check the residuals not individual variables. IMO it is OK to do something like a K-S or S-W test here but again, IMO the interpretation should be like that of the normal-Q-Q plot as suggested by Prof Wilhelm. I would not require a 0.05 condition absolutely , for example.

Thanks all for such a quality answers!