Bug: Incorrect results from test for homoscedasticity as part of nonlinear regression
Prism 7 added the ability to test the residuals from nonlinear regression for homoscedasticity (equal variance). This is also called a test for appropriate weighting. This test computes the nonparametric correlation between the absolute value of the weighted residuals and the predicted Y values (the Y values of the fit curve). If you picked an appropriate weighting, there should be low correlation and a large P value.
Prism computes this test differently when the sample size (n) is 16 or less, and when n is 17 or more. There are bugs in both. These bugs are in Prism 7 Windows up to 7.03 an Prism Mac up to 7.0c. They are fixed in 7.04 and 7.0d.
If n<= 16
- The reported P value is correct.
- Regardless of P value, it always concludes that the data passed the test (reports "yes")
If n > 16
- The P value is always reported as P>0.9999
- The conclusion about statistical significance, whether the data passed the test, is correct