www.cmm.bristol.ac.uk/forum

Dear all,

I have a question for clarification and maybe the answer is very obvious. I am currently trying to estimate a random-intercept multinomial model in 2 steps: an initial model following the IGLS-procedure and an MCMC-variant of the same model. To obtain the MCMC-model I draw on the initsb() option to define the starting values input for all parameters. This matrix consists of FP's (fixed parameters for both equations), RP's (random parameters) and OD's. Latter category refers to bcons_1 and bcons_2 (two constants of the multinomial equations I guess) and this estimate is typically about 1.00 but does anyone really know what 'OD' stands for and what its exact meaning is?

In the IGLS-model I obtain these two OD's (bcons_1 and bcons_2) as I only estimate the diagonal of the variance-covariance matrix, while the MCMC-model yields only 1 OD (bcons_1). However, this might be a consequence of estimating the whole variance-covariance matrix in MCMC. Because my guess is that both matrix sizes have to correspond?

Best,
Fon

These parameters are due to the way that unordered multinomial models are estimated with the IGLS algorithm (see Goldstein 2010, section 4.4 or Goldstein 1995 -http://www.bris.ac.uk/cmm/team/hg/multbook1995.pdf section 7.47 for more details). These normally have a value of one, however if you request extra-multinomial variation they can have a different value (although the are constrained to be equal). The "OD" in their name stands to over-dispersion to reflect this.

These values are not used for MCMC estimation (and extra-multinomial variation cannot be specified), although it will return a parameter chain for one of them (with a constant value of 1). Whatever you put in as the starting value for this parameter will be returned as the estimate when the chain isn't used.

In terms of what to use as starting values for MCMC estimation I would suggest that you provide estimates for bcons_1 and bcons_2, both with a value of one.