Have an idea?

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

Standard error for price in ACBC

Hi everyone,

I am currently setting up an ACBC with summed pricing. I tested the design with 500 test respondents (which is approximately what I also expect for the actual data collection) to check level counts and standard errors.

The relation of my base price (150€) to the average price (287,5€) is 52,2%. I vary prices from 30% above to 30% below summed price.

During the first test run, level counts and standard errors of all attribute levels were fine, while the standard error for price was above 0,05 (0,0755). Subsequently, I tried some adaptions to the design and was able to reduce the standard error for price, but it is still above 0,05 (0,0684).
Is this acceptable or should I reduce it to below 0,05 before starting the data collection?

Also, the second test showed me that for three test respondents of 500 one of the summed pricing levels appeared only once. Is that fine or a problem?

Thanks in advance!
asked Jun 7 by MichaelFu (140 points)

1 Answer

0 votes
Wow, Michael, it looks like you've done your homework on norms for design testing using robotic respondents for our aggregate logit-based tests in ACBC!  Nice work.

It also seems you're paying attention to our recommendations for degree of random shock coming from our white paper about summed pricing:  https://sawtoothsoftware.com/resources/technical-papers/three-ways-to-treat-overall-price-in-conjoint-analysis

...because you're calling attention to the fact that the base price makes up ~50% of the total price of the product, such that only half of the price summation involves correlation with the attribute features.  The greater the base price is as a contribution to the product whole price, the less correlated summed price is with the feature attributes in the experimental design--leading to even more precision in the price slope beta.

Also great that you're remembering that we recommend in design testing and sample size planning (for CBC and ACBC) that robotic respondents (random ones) should lead to standard errors for attribute levels of about 0.05 or less.

However, what I don't think we've ever written in our help documentation or the white papers is that the 0.05 standard error recommendation is for standard attributes that are part-worth (effects-coded).  I've also found, as you have, that the standard errors for the summed price attribute will tend to be a bit higher for the summed price attribute than the part-worth coded attributes.  This is a function of the fact that the summed price attribute (even after shocking with random error) still has a modest correlation with other attributes in the design matrix.  (But, not enough to foul up your experiment, in our experience.)  So, you'll expect the standard error to float a bit higher than for the non-correlated attributes.   As long as you follow the +/-30% or so rule for random shock to summed price, in our experience you'll be fine.
answered Jun 7 by Bryan Orme Platinum Sawtooth Software, Inc. (186,865 points)
edited Jun 7 by Bryan Orme
Ah, I just noticed that on page 7 of the white paper I reference above (three ways to treat overall price in conjoint analysis) we do report that the relative precision of summed price coefficient is about 70% as precise as a 3-level standard attribute when base price makes up half the total price of a summed pricing concept.  So, you can see that's what is expected (that standard errors are a bit higher than for normal attributes) for summed price coefficients.
Dear Bryan,

thank you for the quick and helpful response!