choice-metrics.com

Dear all,

For a labeled DCE based on three screening tests (called blood, stool, and combi) and an opt-out option I used a survey that was based on 3 blocks of 12 choice sets. Based on 3 utility functions with alternative specific parameters I could only estimate 11 parameters for each utility function (not 13 in the first and 9 in the second). My first question is: canthis be explained by the use of 12 choices per repsondents? In other words, increasing the number of blocks doesn't help to increase the number of parameters that can be estimated in a specific utility function, does it? But increasing the choice sets does?

Second, suppose that I have the option to use 600 respondents in the DCE. In this case, I would like to know if it would be useful to use 6 blocks of 12 choice tasks (instead of 3) given that the model doesn't change. Would Nlogit be better able to estimate random parameters with such a design?

Kind regards,

Tim Benning

Hi Tim

Based on your post, it is difficult to reply as to what is happening. Could you perhaps post the syntax you are using? In theory, the number of parameters you should be able to estimate for any given logit model design is constrained by the equation

S*(J-1) >= K

where S is the number of choice tasks, J the number of alternatives and K the number of parameters. Note that the issue of blocking is not a consideration in this calculation. It assumes that the choice tasks of the design are replicated an equal number of times, but it doesn't matter who actually answers the S questions, or how many - S is the magic number. It does however have implications on sampling and sample sizes. In blocking, rather than assume all S are given to N respondent,s you are assuming a subset of b tasks are given to different respondents, but provided all S tasks are used in the design equally, S remains fixed. Your sample size increases by S/b!

In your case you have S = 36 and J = 3, hence, you should be able to at least in theory estimate up to 72 parameters, so there is something else happening. By blocking, in terms of sample size, you have b = 12 (you are giving each respondent 12 of the 36 tasks) so your sample size will be 36/12 or 3 times the number than if you gave all 36 tasks to one respondent.

Please note that Ngene is also for designing SP experiments, not estimation. You can use Nlogit to estimate the models once you have data. Our syntax is somewhat similar to that of Nlogit, so the transition to Nlogit should not be difficult.

John

Dear John,

Thank you very much for your detailed answer. I think this has answered my question.

Kind regards,

Tim