choice-metrics.com

Dear Michiel and Andrew

I have just completed a pilot survey and about to undertake the design for the main survey. My question relates to the use of generic parameters where the alternatives have different attribute levels.

For my pilot design I assumed alternative specific parameters where attribute levels differed across alternatives and generic where they were the same (It is large design 7 choices with up to 7 attributes each).

However, using the data from the pilot survey I found that the best model has all, non-constant, generic parameters (all but one attribute significant) . Using a design with the mixture of specific and generic parameters as described above resulted in a much poorer performing model with a large number of non significant parameters and overall poorer fit.

My question is that it is likely that after the main survey the final model is likely to have all non constant generic parameters should I when coding NGENE:

1 use generic parameters but have the different attribute levels (is this even possible in NGENE?)
2 keep the AS parameters but set them to the generic values
3 keep the AS parameters using the values from the poorer performing model

kind regards

Doug

It is no problem using different attribute levels for attributes that have a generic parameter, Ngene accepts any levels you specify.

Whether you are better off using generic parameters or alternative-specific, it is the final model that you think you will want to estimate that counts the most. Is there any reason for the parameters to be alternative-specific? I am not sure what experiment you are doing. If you have more parameters, it will be harder to estimate the parameters, so with alternative-specific parameters the standard errors will be poorer and you will probably need larger sample sizes. If you are not entirely sure what to do, you can always use model averaging. In that case, Ngene can evaluate a single design using both models (one with generic and one with alternative-specific parameters), so it will (on average) be efficient for both of them.

Dear Michiel

Thank you for that feedback. I had not thought of model averaging within a logit model type (RPL). That seems to me the way forward.

I am will be using the logit model for forecasting market shares so there is no absolute reason for using AS parameters for the attributes.


regards

Doug