I'm trying to get a deeper understanding of the loglinear variance models. Therefore I used the InjectionMolding-example from the JMP-documentation. Working with JMP I am able to reproduce the results from the website, of course.
Additionally I implemented an ML-estimator (based on the Harvey-paper) in R. Further I tried the remlscore function from the statmod-package in R (REML estimation like Smyth 2002 describes it). The funny result is, that I get quite similar results with both functions in R for the estimated residual variance (~1.88, ~1.9) but that number is pretty different in JMP: ~6.69.
Now I was curious if this is just some difference in the parametrization? All other numbers match quite nicely.