Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Measuring multiple coefficients interactions with the levels of the MaxDiff in MNL analysis


I am doing a MaxDiff Case 2 (conjoint BWS) study and already collected the data. I understand that BWS does not allow interactions between levels of attributes but we can include interactions with other coefficients like age, gender ..etc.

My questions are:
1- How can these interactions in the logit analysis of the Lighthouse Studio program?
2- How can I discover the coefficients that produce significant interactions from the once that do not?
3- If I added weight of let's say by gender, how can I interpret the results?

Note: I have conducted the Latent Class analysis and HB for the data but I need to perform the aggregate equivalent too for general comparison as this is an academic study.

asked Mar 2 by AMYN Bronze (2,980 points)

1 Answer

0 votes
First, I should say that crossing respondent characteristics by conjoint or MaxDiff preferences using aggregate logit is an approach that few practitioners would do (I think you recognize this as you indicate this is an academic study).  Most practitioners would think about estimating individual-level utilities via HB and then using the respondent characteristics as filters to summarize and do statistical tests between respondent groups on the individual-level utilities.

So, being a tool aimed quite at practitioners, our software doesn't directly support the idea of interacting respondent characteristics with conjoint or MaxDiff attributes in a utility estimation model.  However, like many things in our software, it's often possible to trick our software into doing such things.  But, it can require manual effort and manual reformatting of data files.

To do the power trick to code respondent characteristics as new attributes in the data file, you need to be using the standalone latent class or standalone HB tools.  (You cannot do this within the Lighthouse Studio interface).

The data file that contains the design matrix and respondent choices is usually a .CSV or a .CHO file.  The .CSV file is easier to work with and modify.  The trick is to add new columns to the file to accommodate attributes representing respondent characteristics.  Because you are doing BW Case 2, I'm thinking you have tricked Lighthouse Studio into fielding a BW Case 2 via MaxDiff (via conjoint-style prohibitions in MaxDiff).  That means, I think, you'll be working with a .CHO file that you exported from Lighthouse Studio.  This is a text-only file with a somewhat difficult format to work with unless you are a scripter and are able to process and modify the data file with something like Python, SPSS scripting, R, Java, or C.  

The format of the .CHO file is described in the following documentation: https://www.sawtoothsoftware.com/download/techpap/lclass_manual.pdf   in Appendix B, section 7.2.

Also, please note that it would seem strange to use HB to estimate a model in which respondent characteristics are included as additional attributes in the data file.  The respondent characteristics don't change within a respondent and HB is an individual-level estimation approach.
answered Mar 3 by Bryan Orme Platinum Sawtooth Software, Inc. (172,790 points)