Have an idea?

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

Likelihood of purchase or RFC with none utility?

Hello,

We used ACBC to run a real estate study, and the simulator is set up to determine whether respondents would purchase a single concept or not. What would be the more appropriate method: using Purchase likelihood or using RFC  and comparing the concept vs. the none utility?
asked Jun 17, 2014 by anonymous

2 Answers

0 votes
First, respondents are notoriously bad at self-reporting their future purchase likelihood of products within market research surveys.  Conjoint analysis is not immune to the over-exaggeration.  So, you cannot use conjoint results to “determine” the purchase likelihood of products.  You can estimate it, given what respondents tell you in the survey, however.  But, exercise caution in interpreting the results, as it is only as good as respondents self-reporting and will probably demonstrate too high of purchase intent vs. reality.

ACBC’s “None” share of preference can come from two sources, depending on how you do it.  1) If you do not include the optional “Calibration” section at the end, then the None threshold is taken from the “Consideration” phase of the survey, where respondents are asked for each of many product concepts if they would consider it or not (binary logit specification).  The “or not” option is the None threshold.  2) If you include Calibration concepts at the end of the survey, then the None threshold can be replaced by the threshold the researcher specifies as associated with the 5-point purchase intent scale used in that optional series of questions.  If you say that a “4” on the 5-point scale marks the boundary between non-purchase and purchase, then the part-worth utility associated with a 4 on the 5-point scale is used as the new None threshold, etc.

The Purchase Likelihood market simulation method was designed for use with ACA and CVA ratings-based conjoint results (if you use single-concept presentation in CVA with the logit transformation on the dependent variable).  It uses the formula:

100*e^Ui / (1 + e^Ui)

Where Ui is the total utility for the product concept in question.  This math is not suited for CBC or ACBC to predict any sort of stated purchase intent.  If using the Purchase Likelihood simulation transformation as above with CBC or ACBC, the interpretation should be: “what’s the likelihood of picking this product concept as specified within the market simulator vs. a product concept with a utility of zero (since e^0 = 1).  Also, note that because we typically zero-center the utilities within CBC and ACBC, the zero-utility concept represents a product concept of “average” utility based on the attributes and levels you specified in your model.

Next, if you use either Share of Preference (logit rule) or RFC in our simulators vs. the None concept (one product vs. the None concept), then you need to interpret the results based on how you estimated the None parameter.  If it comes naturally out of the Consideration phase of ACBC, then the interpretation of the simulation output is “the likelihood that this product would be viewed as ‘a possibility’ by respondents”.  If you use the None calibration from the optional Calibration section in ACBC, then the interpretation is “the likelihood that this product would exceed the purchase threshold as specified by the researcher”.  

The math for the Share of Preference vs. the None would be:

100 * e^Ui / (e^Ui + e^None)

Where Ui is the utility for the product placed in the market simulator and None is the utility for the None threshold.

For example, if the researcher specifies that the purchase threshold is a “4” on the 5-point scale, then a product that achieves a utility equal to the utility associated with a 4 on the purchase intent scale would get a 50% RFC or Share of Preference prediction vs. the None concept.  Once the product achieves higher utility than that associated with the “4” on the purchase intent score would the probability of selection instead of the None exceed 50%.
answered Jun 17, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (128,365 points)
0 votes
Hello,

I'd like to expand upon Bryan's response and offer some general thoughts about forecasting.  I think forecasting is one of the most difficult things researchers are asked to do.  Moreover, forecasting involves a slightly different skill set than many marketing researchers receive during their formal educations.  

First, as Bryan notes about using "none" for forecasting - by itself - doesn't get you to purchase intentions (though it would be closer if you're forecasting a one-of-a-kind product into a brand new category).  Otherwise you're failing to take competition into account (upon which of course you can improve by including multiple competing products in your simulation scenario).  

But I think even that falls short.  Any forecasting method has its biases and inaccuracies, things that it captures and things that it misses and so on.  So my preference would be to use multiple forecasting methods, to get a parallax view.  I might well run a conjoint experiment (after which I'd simulate a realistic scenario in terms of the competitive landscape) but I might also ask a traditional purchase intention question as is common in volumetric forecasting (realizing, as Bryan notes, that purchase intentions are themselves subject to over-statement, about which you can find a good number of papers, mostly from back in the 1970s).  Often enough I'll even fuse the conjoint results and the purchase intentions together to produce a forecast (the calibration questions built into ACBC are one way to do this but there are others).  

And, assuming it's important to do your forecast well, I'd go even a step further.  I'd recognize that there are many other variables that go into the forecast besides purchase intent and conjoint simulations:  How big is the market?  How many buyers are buying how often?  How much advertising/distribution/promotional effort will I be able to get relative to my competition and how much awareness will I be able to generate.  Will my product depend on word-of-mouth and if so what share of that WOM will my product get?  So there are potentially several other inputs a forecast model will need.  Some of these may come from surveys and some from secondary research.  Some may even come from managerial judgment.  Most of them will themselves have some amount of uncertainty around them.   In this case I like to build a Monte Carlo forecast (either in Excel or in some specialized Monte Carlo software) wherein we run a large number (1,000, 10,000) simulations, each using the point estimate for each input measure plus an appropriate random draw from the error distribution of each input measure.  The result is a distribution of forecasts, one for which we can describe a mean, a median, forecast intervals and so on.  If there's a particular break even volume/share target we need to hit we can even identify the portion of the forecast distribution that exceeds the target.  

In summary, there are several different levels to which we might take forecasting, depending on the importance of the forecast (and the willingness of the manager to pay for forecasting).  I think conjoint analysis is a great place to start and is one of the more important inputs.  But one can take forecasting a lot further if one uses conjoint analysis as a starting point rather than as the only forecast input.
answered Jun 17, 2014 by Keith Chrzan Gold Sawtooth Software, Inc. (48,525 points)
...