Have an idea?

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

Would this be an acceptable method for presenting conjoint results?


I've read the article on presenting sensitivity analysis, which (if I understand correctly) consists of presenting two concepts that are identical, and then changing one feature of a concept while holding all others constant, and running the simulator to observe the resulting impact of that change.

While this makes sense for attributes in which preference is a priori established (such as people preferring a lower than a higher price), I find that some clients have trouble interpreting the results otherwise. I was wondering if it makes more sense to create as many products as there are levels within an attribute, and have each product on a different level, while all other attributes are held to a constant. For instance, if an attribute has 4 levels, I create 4 products and assign a different level to each product for that attribute, apply the same levels across the 4 products for all the other attributes, then observe the share of preference.

Would this make sense, or are there red flags with this method?
asked Apr 13, 2015 by anonymous

1 Answer

+1 vote
Sorry, these approaches you mention don't make sense to me.

I think you may have misread the article you mention about sensitivity analysis.  This is the way I believe is most correct:

Create a realistic scenario in which there are multiple products on the market as similar as can be made to the real world.

Change just the "test product" one level at a time and observe the change in share of preference due to changing that one level.  When you change an attribute level of an attribute, leave all the other levels of the other attributes at the base case specification (including all the competitive products).

It is wrong to create multiple versions of the test product within the same scenario where the test product's replicates are given different levels of an attribute.

I know it looks like more work to do things the way I've outlined above, but our SMRT market simulator does this all in an automated way using the Sensitivity Analysis mode approach.  It runs the sensitivity analysis as I've described and gives you a summary report at the bottom of all the iterations of the sensitivity analysis in a handy table.
answered Apr 14, 2015 by Bryan Orme Platinum Sawtooth Software, Inc. (128,365 points)
Hi Bryan,

I've used the sensitivity analysis approach in a few projects, and I think it works great for attributes in which preference is a priori established and I can truly replicate a given competitive market scenario. But I find it doesn't work as well when the market cannot be exactly replicated, and I then have to arbitrarily select levels as the base.

Real estate is  a prime example. There are thousands of different options out there that enter and leave the market on a daily basis, so it is not achievable to replicate a competitive market scenario.

It's particularly problematic if we're looking at subgroup differences and the arbitrarily chosen base level is naturally static across all subgroups. While I understand the rationale for this, the client hasn't and I find they have a lot of difficulty interpreting the findings.

Regarding the method I attempted to describe (and perhaps I have not explained it well enough), let's take my earlier example of real estate and one of the attributes is the number of bedrooms. There is no a priori preference for the number of bedrooms (since it is dependent on the buyers' lifestage and family status). Some living single prefer 1 bedroom, while others who have a family would look for 3+ bedrooms.

My rationale is to say "We have 5 levels that we measured in that attribute (1 bedroom to 5 bedrooms), so what would be the popularity be for each of those 5 levels? If I present a scenario where there are 5 houses out there and they all share the same characteristics (let's assume there are no interactions with other attributes), let's see how respondents would break out across those 5 levels." I've tried it both ways, and I'm getting exactly the same conclusions, except that now I'm able to look at how subgroups differ for each of the 5 levels.
Ah, if your goal is more toward segmentation: who (and how many) are the people who most prefer different levels of a given attribute from a conjoint analysis, then putting 5 different products in the simulator (one associated with each level of the same attribute--holding all other attributes constant) is an appropriate way to go.

However, if you are trying to estimate the impact that changing an attribute level has on the relative preference for a product concept, then having a good base case scenario involving competitive products and changing the single "test product" one attribute level at a time (sensitivity analysis) is preferred.
Thank you for your thoughtful replies!