choice-metrics.com

Dear Experts,

I have a choice experiment data including the population of cancer and non-cancer. I want to compare whether preferences for attributes differ between cancer and non-cancer groups. My first thought is to generate interaction terms between cancer indicator and attributes. According to Greene, W.H. (2010) and Ai, C. and Norton, E. (2003) , it seems interaction terms in non-linear model may not provide correct information. However, I also noticed some papers created interaction terms between social characteristics (such as age) and attributes. I am confused with this paradox.
My questions are: 1. may I analyze the data as a whole and generate interaction terms between cancer indicator and attributes as a way to compare preferences between two groups ? 2. if not, do I need to run the model on two groups separately? Any other suggestions?

Many thanks,

Connie

HI Connie,

I would create an interaction term between a dummy coded cancer variable and attributes, exactly as you propose. I am not sure why this would lead to wrong results as this has been the way people have done it for a long time. I am not aware of issues mentioned by Greene or Ai & Norton, what do they say specifically?

Michiel

Thanks for the confirmation, Michiel. Ai, C. and Norton, E. (2003) said the sign of interaction term and its marginal effect are not easy to calculate. May be I got the wrong understanding.

I am not sure what article you are referring to, but I do not see why computing interaction terms and their marginal effects would be an issue, so I would suggest that you add interaction terms as you suggested as this is commonly done.

Michiel