I'm looking for an analysis (I assume a book or website, but feel free to put it all in comments here) that provides an in-depth discussion of the power/accuracy of normality assessments such as Shapiro-Wilk. In short, I understand that Shapiro-Wilk "is good for small data sets" and with large data sets it is overly sensitive to small deviations from normality - but what does this mean quantitatively?
Overall I'd like to get a feel for something like this: If I have n data points, I'm mm% confident that the Shapiro-Wilk (and/or other tests) test is providing me a reasonable conclusion. How does mm% change when n is 15 ... 10 ... 7? How about as it goes from 100 to 500 ... 1000? Please note that I do believe in the power of graphical assessments as well, but I'm keeping those out of scope for this question and I'm looking for numeric, quantitative assessments in this case.