Have an idea?

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

Are zero-centered diffs in ACA calculated from calibrated or un-calibrated final utilities?


I'm chewing on the interpretation of some ACA results and was just wondering how the zero-centered diffs come about. That is, I understand the maths, but what I can't see is whether the diffs are calculated from calibrated or un-calibrated final utilities - or whether that would make a difference at all, as the rescaling could kind of "undo" the calibration.

Any advice on this is apreciated!

asked May 30, 2014 by DMay (120 points)

1 Answer

0 votes
Good question.

If you are speaking of Calibrated vs. Uncalibrated ACA utilities, then that means you are referring to the output of ACA/HB estimation.  If using the output of ACA/HB estimation and then requesting the additional step of calibrating  the utilities based on the additional information provided in the Calibration Concepts, then we fit a slope and an intercept to the Calibration ratings.  The intercept shifts all the HB-estimated utilities left or right.  The slope is a single multiplier on the HB-estimated utilities.  This happens within individual.

So, you are right that zero-centered diffs leads to the same results whether starting with raw ACA/HB utilities or calibrated ACA/HB utilities.  Zero-centered Diffs undoes the calibration step (removes the slope differential per person and removes any shifting).

Please note that it has been found many times that the calibration step in ACA typically leads to utilities that are too flat (have too small of "scale factor"...also known as "exponent") for predicting choices (getting the share of preference ratios right between products within a simulation scenario), relative to scaling we typically see in CBC or ACBC data.  So, the calibration step really is only helpful if you want to be able to use the Purchase Likelihood simulation model and to predict the stated purchase intent for individual products at a time in market simulations.  If you are planning on using the RFC or Share of Preference models for competitive market simulations involving multiple products, then the calibration step doesn't help you...it potentially hurts matters.  ACA's utilities somehow need to be adjusted in terms of scale factor to best predict probability outcomes (such as choice or "market share") and thus the researcher will need to adjust the "Exponent" in the market simulator to obtain good results.

If you have holdout choice scenarios in addition to your ACA data, then the holdout choice probabilities give you a target to adjust to.  But, once we start talking in these terms, one starts to wonder why we are using a ratings-based conjoint method at all and why don't we move to something like CBC or ACBC.
answered May 30, 2014 by Bryan Orme Platinum Sawtooth Software, Inc. (128,365 points)
Dear Brian,

thanks for getting back to me so quickly! Actually I was referring to the standard OLS regression estimation of utilities - which, if I get it right, at least technically produces both calibrated and uncalibrated final utilities (i.e., combined priors and paired utilities only). That's at least how I understood the estimation procedure in the ACA Technical Paper. Or am I on a wrong track here completely?