I am trying to replicate the Sawtooth results from the aggregate logit maxdif model, with a direct binary anchor. Exporting the data into Stata I can replicate results exactly for the best/worst questions alone.
But when I add the 2 additional tasks to reflect the anchor (as per Lattery 2011) I cannot replicate the results. Looking at the LL for the null model estimated by Sawtooth, I am wondering if the implementation is actually "..the binary version where each attribute was compared with a zero vector." (in my case that would add 12 additional tasks, not 2). Is it possible to confirm which implementation is used?
Update: worked out more quickly than I expected how to generate the data in that form, and that does appear to be the case: when I use the binary version I exactly replicate the results.