Have an idea?

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

Average Raw HB Utilities & Aggregate Logit Utilities

Hi (again),

Should it not be the case that the averages of the individual utilities (raw) of a HB run
- the averages of each of the raw individual utility columns,  for the same data set (a simulated data set of 125 respondents) be roughly the same as the aggregate Logit utilities?

I've run two the two sets of analyses and while they are highly correlated (.95) meaning they are largely directionally correct, the actual values are wildly different, for example:     

Here are the average of the individual raw utilities for three ASCs (from 125 simulated respondents)
Bus                         Train                     Tram                     Car                          Taxi                         Uber
0.26345184     0.49724232     0.26748096     -0.74642424    -0.16523104    -0.11651976

Here are the aggregate logit utilities (for the same experiment with the same simulated responses)
Bus                         Train                     Tram                     Car                          Taxi                         Uber
0.00415            0.17455             0.08132           -0.11076              -0.08145              -0.0678
Both are standard runs (no changes to the default settings. I'm confused as to why the magnitude of the part-worths are so different? is 125 simulated respondents not enough? I would have expected much closer absolute values...

Does anyone have a good line of sight on this?


asked Mar 6, 2017 by Jasha Bowe Bronze (1,745 points)

1 Answer

+1 vote
Best answer

I would not expect the values to be of the same size, no (though I would expect them to be correlated, as you are seeing).  The aggregate model's utilities are calibrated to predict sample-level aggregate shares while the HB utilities are calibrated to fit individual respondent-level choices.

Clearly they should not disagree (hence the high correlation) but changing the scale of either model will decrease its fit to the calibration data.
answered Mar 7, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (84,950 points)
selected Mar 7, 2017 by Jasha Bowe
Thanks Keith! I feel a bit silly.. I completely forgot that (to steal the title of a well known article) that pesky little scale issue! I came from a background which didn't take advantage of the power of HB and individual level estimates. Thanks for taking the time to ask this rudimentary question!