Have an idea?

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

Summed pricing vs. conditional pricing in CBC

Hey there,

I just read "Becoming an expert in conjoint analysis" (I really recommend it!) and still have a few questions about summed and conditional pricing.
My professor wants me to show as realistic prices as possible. In a first approach, I defined 61 different price levels showing all possible sums of attributes that define the price. Now we know that we will not have sufficent respondents for so many levels and, if I understood right, due to the correlation it would not be possible to calculate the part-worth of the price attribute.
Now we were thinking about conditional pricing in CBC. As all other attributes (5) define the price, I was thinking about including all of them in a conditional relationship and defining all possible sums, each for "low", "medium" and "high" price levels. Did I understand right that for this approach, there are e.g. 3 different price levels and "low" would be the summed price for the alternative in the choice task -30%, "medium" the exact one and "high" the sum +30%, all rounded? And if so, how many respondents do I need? Does it depend on the 3 defined price levels or on the 41 actual ones?  There are 8 choice task with 3 sets each.
We're not sure if we understood correctly that summed pricing could also be possible for CBC, not only for ACBC. I just tried it in excel as described in the book. In that case, if I implemented it in CBC as conditional relationship for attributes, how many price levels should I define and how should I discriminate between them? If I just used random shock for 3 price attributes, after rounding there is no price difference, but maybe I just didn't get the concept right.
For both cases: Does the number-of-levels-effect apply but would it be manageable? And am I right that I could just import a lookup-table in the simulator for analysis in both cases?
Sorry for so many questions! Many thanks in advance!
asked Mar 5 by lisach (200 points)

1 Answer

0 votes
Remember, if you create summed pricing without the random shock to the sum total each time, you will correlate the features with the prices shown, and you will not be able to estimate a price slope (price sensitivity).   Sounds like you're understanding that, though.

Also, if you do the same thing through conditional pricing without having a separate price attribute that varies the prices, say, -20%, +0%, +20%, you won't be able to estimate price sensitivity.

You understand correctly that conditional pricing takes the different premiums or discounts for, say, premium brands and discount brands and then varies those by, say, -30%, +0%, and +30% (if you had 3 levels for the price attribute).

But, I'd be cautious about trying to make conditional pricing based on more than one attribute.  If there are significant interaction effects in the utility parameters, then you'll not be able to untangle them only with main effects + first order interactions.   That said, many researchers still do use conditional pricing based on more than just one other attribute.  Just make sure the conditional pricing elements are additive in a logical formula rather than haphazard.  If the conditional pricing elements are combined haphazard, you'll confound the utilities with the experimental design.

Summed pricing is possible in CBC, but only via a "power trick".  I just did one last year like this.  It involves exporting the CBC design file (assuming
a standard price attribute with, say, 7 to 10 levels) to the .CSV file, then modifying that file.   It's somewhat involved, but it involves calculating the summed price plus a random shock for each concept in the design.  Then, it involves mapping those summed prices into the 7 or 10 discrete levels for price...assigning those concept prices discretely into the nearest 7 to 10 discrete price buckets.  So, essentially, you're creating a summed pricing design where you are rounding the end prices to 7 or 10 discrete price values.  Once you've done this, you import the design and test the design efficiency again.  In the model building phase, you can estimate price as either the standard part-worth functional form or also as a linear form.

The number of levels effect is not as big of a deal as the papers from the 1980s suggested.  Those articles from the 1980s pointed out the problem if you had 2 levels for the full range of a quantitative attribute rather than 4 levels (but both cases covering the same range).  It turns out that if you are comparing 3 levels to 6 levels; or 4 levels to 8 levels, you just don't see nearly the same number of levels effect as when comparing 2 to 4 levels.  So, it seems mostly a corner case involving such a low number of levels for a quantitative attribute (2) that it is unrealistic to real world practice.
answered Mar 5 by Bryan Orme Platinum Sawtooth Software, Inc. (184,340 points)