I'm using NLOGIT to analyze data from a choice experiment. I have 6 attributes with 2 levels and 1 attribute with 4 levels. I have 3 alternatives. My alternatives are 2 contract types and a status quo.
So basically I had three utility functions. Utility function for the two contracts use all the attributes but utility function for status quo uses only one attribute (income). I was able to estimate the model using both conditional logit and mixed logit. Using mixed logit i found preference heterogeneity. So I tried to interact some socio-economic variables with the attributes. But when I do that I'm getting error 1085. I created the interaction terms before running the new model, but I'm still getting the error. Looks like I'm getting the error if I use more that 8 covariates on the right hand side. Could this be a problem with degrees of freedom? I had only 59 surveys back and each survey had 4 choice sets. So basically in NLOGIT i had 59*4*3 number of rows.
How can I generate the confidence interval for resulting marginal effects using NLOGIT?
Also, I included ASC and interacted age with ASC. But the output gives the coefficient for ASC only not for ASC*age. Is there a way to estimate them separately?
Is it possible to estimate ASC using STATA as well?
With any maximum likelihood model, it is a good idea to start small and simple and then to generalize the specification to accommodate heterogeneity one dimension (i.e. one new parameter) at a time. This is tedious but the best way to see just where you introduce a parameter that cannot be identified by the data you have available. This is a rather small set of choices (236). Also, it is not clear whether you are trying to add systematic heterogeneity in parameters on top of random variation in the same parameters. I have had trouble in the past with a dataset where age was distributed across the same according to a distribution that could be described as "roughly normal." I could put a normal distribution on some of the key marginal utility parameters in my model OR I could let the same parameters vary systematically with age. But it was not possible to do both because that was essentially redundant. I opted to rely on the systematic variation because that gave me a story about the heterogeneity, rather than just confirming that there WAS heterogeneity (via a statistically significant variance coefficient on the relevant random parameters. To compare these non-nested models, of course, you'd need to use some sort of information criterion.
Error 1085 is a conflict in the notation. Be more creative in the naming of these interactions to make sure that they all have very different names. I believe you
have only 6 characters to work with on these, but you should be able to come up with unique names. It is not an issue of degrees of freedom. At the point at which the 1085 error occurs, it is not yet known how many "degrees of freedom" there are.
Also, I agree with Trudy's advice. Start small. Build up the model a bit at a time. The error is likely to show up when you cross a boundary of some sort.
Her other point is also pertinent. You would find out about at estimation time, however, not when the model specification is being parsed. Again, starting small will help.
I think the problem should stem from how the model is constructed or specified. Probably you could provide details on how the left hand variables were constructed, or simply take a step at a time as proposed William.
ThanksTrudy, Bill and Dobdinga for your suggestions.
I finally figured out the problem. I just had to shorten the names of the variables and I could use all the interaction terms that I meant to study.
- Badges
- Science topic
- Similar topics
- Engineering
- Computer Engineering
- Computer