Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Why is the ASC in my HB run fluctuating all over the board while the others are stable?

Hi,
can anybody provide a brief interpretation of why the phenomenon above occurs?
I only have 4 parameters to estimate (one of which is an ASC), over 4000 respondents with 11 tasks each and 40.000 burn-ins and still the ASC will not converge but jumps and drops 10times the magnitude of the other "flat-lining" converged values.
Should I remove the ASC even if it was significant in the Logit runs and contributed to (LL) model fit?
Thanks for any thoughts on this.
Alex
asked Oct 23, 2012 by alex.wendland Bronze (2,005 points)
retagged Oct 23, 2012 by Walter Williams
There are many reasons for nonconvergence in HB estimation. You must tell more about your model specification. BTW, what is "ASC"?
ASC is the alternative specific constant.
I just thought that it is odd that this parameter in particular is jumpy while the others are perfectly normal. I have not seen this before. Usually all parameters are more or less stable in my experience...

1 Answer

0 votes
Whenever I see a particular parameter do crazy things, I step back and submit the same run to aggregate logit to make sure that I don't have a fundamentally inefficient (or confounded) design.  Our latent class software (standalone latent class) with a 1-group solution is the same as aggregate logit, so that tool may be used.  Specifically look at the standard errors of the parameter estimates.

--Bryan
answered Oct 23, 2012 by Bryan Orme Platinum Sawtooth Software, Inc. (144,240 points)
Hi Bryan,
I did the aggr logit on the model specification with the ASC (using MBC). The standard error suggests that the ASC effect is significant. When I run the logit without ASC the 3 other (non-ASC) parameters change somewhat (e.g one changes from slight positiv to slight negativ, one doubles in magnitude) but remain significant.
I also looked at the point estimates and the corresponding standard deviations for several HB runs (with and without ASC). Here, when I leave out the ASC, the mean parameter estimates stay much more similar than with logit and either way have significant t-ratios. Only the ASC when included gets a 0.2 t-ratio.
Can you think of an explanation for the ASCs behavior?
It seems I can exclude the ASC from my model without loosing much fit but I wonder if the erratic fluctation is a matter concern and indicator of something bad (although I think I can rule out inefficient or confounded design).
Thanks for the reply,
Alex
...