Fixing mlogit real parameters
Posted: Tue Apr 29, 2014 3:28 pmI want to summarize an offlist discussion about fixing mlogit real parameters that I had with Torbjorn Ergon and Gary White.
First of all, you can go wrong by deleting design data records for Psi to fix the values to 0. In particular if you delete all of the records in a group to set all transitions to 0, it may not work. In a recent release I added a preferred way to fix real parameter values that are set for all models. In the design data for a parameter, add a field with the name fix. Assign NA to values that should be estimated and for fixed values assign it the fixed real value. This avoids any problems with deleting design data. Deleting design data will work fine in most cases but using fix is safer.
Now, most likely you won't be fixing real parameters for transitions and other mlogit parameters to anything other than 0. But if you are tempted to do so, do not use the mlogit link. Gary said there is no guarantee it will maintain the constraint that the probabilities will sum to 1 and I was able to verify that with an example. Also, as presently written the code in compute.real in RMark will not produce the same real estimates as MARK if the fixed mlogit value is non-zero. In the next release RMark will throw an error if any non-zero real fixed value is used with an mlogit link.
regards --jeff
First of all, you can go wrong by deleting design data records for Psi to fix the values to 0. In particular if you delete all of the records in a group to set all transitions to 0, it may not work. In a recent release I added a preferred way to fix real parameter values that are set for all models. In the design data for a parameter, add a field with the name fix. Assign NA to values that should be estimated and for fixed values assign it the fixed real value. This avoids any problems with deleting design data. Deleting design data will work fine in most cases but using fix is safer.
Now, most likely you won't be fixing real parameters for transitions and other mlogit parameters to anything other than 0. But if you are tempted to do so, do not use the mlogit link. Gary said there is no guarantee it will maintain the constraint that the probabilities will sum to 1 and I was able to verify that with an example. Also, as presently written the code in compute.real in RMark will not produce the same real estimates as MARK if the fixed mlogit value is non-zero. In the next release RMark will throw an error if any non-zero real fixed value is used with an mlogit link.
regards --jeff