Macro models not converging properly

[nickbailey2]

I'm running models from macros. I set 'batch 1' so models should run to convergence. The models appear to converge (I can check that CONV = 1, manually). But if I open up the equations window and click 'Start' (without changing anything else), the models converge to slightly different results. For a couple of models I've looked at, some parameters change by 1 per cent - not huge, but not necessarily trivial either.

Has anyone else found this? Why should I get different results? And is there a way of ensuring that models converge to the same estimate regardless of whether they are run via a macro or not?

Thanks

Nick Bailey
University of Glasgow

[ChrisCharlton]

We have done some investigation of this behaviour and it appears that if you use robust standard errors, set the estimation method to RIGLS and run the model with the Start button then you will get slightly incorrect results. This should be fixed in the next release of MLwiN.