[Richard Williams]

Going back to post #20 -- the clogit manual explains how to use clogit to replicate asclogit, and that seems to be what you have done. But again, I don't see any reason you would want to do this. As far as I can tell, the coding of the alternatives is arbitrary and inconsistent. Within a case, alternative 3 could have just as easily been coded as 5, right? And across cases, alternative 3 can have a totally different meaning for the first case than it does for any other case. If the numbers assigned to alternatives are arbitrary and inconsistent, there is no reason to think they will somehow be related to the outcome, either as direct effects or interactive effects. I think the numbers assigned to the alternatives are just a red herring and you shouldn't pay any attention to them. Indeed, maybe you just want to drop the variable entirely if it is distracting you. Also, as you see, that line of analysis didn't get you anywhere.

The analysis in 22 looks like it is more likely to be fruitful. On the final model, you should be including the lower level 2 way interactions as well, i.e. don't just skip to the 3 way interaction. I would also start with a model that had no interactions and build from there. You should also estimate the model that has both 2 way interactions. At least you should if you are just trying everything out; if there is some theory behind this you may be able to decide that some models make theoretical sense while others do not.

[William Lisowski]

To Richard's excellent recommendations let me add just a few observations.

I would consider expressing income in thousands, so that the coefficients on the income interactions would be on the same order of magnitude as other coefficients. This will make your eventual reports read more easily.

From the two-way interaction with cost_, we see that the contribution of cost_ and income_ to the scoring equation is

-0.7332874*cost_ + 0.0000618*income*cost_

From that we see that when income is around 12000, the model is cost-neutral, and above that, increasing cost (possibly a proxy for other unmeasured characteristics of the route, perhaps correlated with time_) increases the score.

Finally, I'd recommend visiting the Stata Base Reference Manual PDF and reading not only the documentation on clogit, but the following section on clogit postestimation to see what tools are available to assist in interpretation of the results. I think informed use of the margins command will help you understand the story your models are trying to tell you. (Especially if you follow Richard's advice on formulating your various models.)