Calculating preference at the individual level

Question

Hello all-
I have study with 9 attributes and the max number of levels is 5.
I want to know if it is possible to calculate an individual's preference for one treatment option over another (total 3 treatment options) by doing this:
1. I can go over each attribute level and mark it if it indicates Treatment Option 1 or Option 2 or Option 3. While some attributes are not so specific to any one treatment option, it is possible to distinguish for some of them.
2. Can I just add up the partworths for all levels that are connected to Option 1 to get the preference for Option 1? Can we then repeat this for Option 2 and get the preference for Option 2 and so on for Option 3?
3. Once I am done adding up the part worths, can I then say, Option 2 has the highest score, so this particular person prefers Option 2 the most, followed by Option 1 and Option 3?

I am a novice in this area, so please bear with me. I might using some terms incorrectly but I am happy to learn.

Answer 1

What you describe is indeed what we do when we conduct a "market simulation" with conjoint part-worths.   You add the part-worths for one level from each attribute to construct the total utility for each product alternative.

But...

If the attributes in your experiment that include levels corresponding to  options 1, 2, or 3 represent all the attributes that in the real world describe options 1, 2, and 3, then this makes sense.

However, if not all the information that makes up the differences among the options is included in your attribute list, then you will be missing important information that would be needed to predict which option each respondent preferred.

Comments

Thank  you for your response Bryan. I was trying to see what might be the best way to understand what socio-demographic factors affect a patient's decision to choose treatment A vs. treatment B vs. Treatment C. However, one of the issues we are facing is that we don't explicitly call these concepts as treatment A or B or C; however, based on your answer I am not sure if our attributes "represent all the attributes that in the real world describe options 1, 2, and 3" - they certainly represent some of the most important, but probably not ALL. Knowing this, would it be possible to run a figure out a way to see which socio-demographic factors were associated with a person choosing a treatment concept that was similar to treatment option  A/B/C.  Can the part-worth utilities be used as the dependent variable in a regression equation as a proxy for their treatment preference? It is ok, if it can't be done, but that would be helpful to know.

---

Yes, it is quite common to try to figure out which other survey variables seem to be associated with respondents' preferences for one product concept or another from a conjoint analysis.  There are multiple approaches to doing this and the choice often depends on who you will be showing the research to and what their preference is for such statistical testing.

One of the easiest approaches that practitioners tend to do is to use the rescaled part-worth utilities (normalized, per the approach called "zero-centered diffs").  Our software automatically provides this within the Lighthouse Studio platform (I don't think you've described how you computed the part-worth utilities).  I'll assume you are using HB within our Lighthouse Studio platform.

Next, sum the normalized part-worths associated with treatments A, B, and C.  This becomes your dependent variable (the sum of the zero-centered diffs part-worth utilities associated with these options).  

Last, use your favorite cross-tabulation software to compute F-tests based on other survey variables, to see which variables are predictive of choice of treatment.

---

Thank you again for your help with this. We are using Sawtooth for our analyses. I will try this approach and follow up if this doesn't work for us for any reason.

---

Following up from the previous posts, can we add part-worths from multiple levels (within the same attribute) to indicate one treatment option? Specifically, I to know, say an attribute has 5 levels but we have only 3 treatment options, how do we reconcile while calculating the share of preferences for each treatment? Must each attribute level be associated with one (or more) treatment options or is it ok if an attribute level remains unassigned to any treatment option?

---

The experimental design showed respondents one and only one level per each attribute in the questionnaire.  Thus, summing multiple levels within the same attribute is not permitted.

One and exactly one level MUST be assigned to each treatment option.  If you are going to use an attribute, you must use one level from that attribute for each treatment option.  

You can skip an attribute altogether for ALL treatment options.  But, it isn't proper to add the utility for one level of an attribute for some of the alternatives but not others.  That would imply that the missing level had a zero utility, which isn't necessarily the correct imputation for what a missing attribute level means.