Have an idea?

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

Is it possible to use HB Mean Importance to show heterogeinity in the logit estimation?

Hello everyone,

I ran a labelled DCE with 6 attributes. I got 159 usable responses. I calculated the utilities and WTP in another statistical program, to include base utilities of the options (different product categories).

So far, I have used the logit utilities for my report because they align with what I calculated in WTP AND because I am certain about the way it is calculated. To give a full picture, I want to include importance weights, which I would like to take from the HB analysis because I learned it accounts for heterogeneity.

Would that be a cardinal error?

I am not sure, how HB estimation is derived. Therefore, this may be a stupid question. Thank you so much in advance!
related to an answer for: What's the difference of logit and HB?
asked Jan 26 by Cora

1 Answer

0 votes
Pooled (aggregate) logit can give a distorted view of "attribute importances", because if people disagree about an unordered attribute (like brand or color), their disagreement can cancel out in the aggregate, making it look like the population doesn't care much about brand or color.  (It sounds like you understand this, but I'm stating it for others on the Forum who might be reading.)

Importance scores quantify the range of utilities and put them on a scale summing to 100%.  It's more reflective of what respondents think about the impact of attributes to compute this at the individual level, or with latent class segments if you use a handful of segments or more.

I don't think it would be a big mistake to report importances from a latent class MNL or an HB-MNL, though importances can be misleading to interpret.  The range of levels for attributes the analyst chooses to include in the experiment has a direct bearing on the importance one gets for attributes.  Also, the importance score capitalizes on any difference between best and worst levels for a respondent, even if that difference is due to random noise.

I think the bigger potential flaw in your approach is to rely on aggregate logit models to estimate WTP.  Such an approach often exaggerates WTP.  If interested, see this white paper we recently wrote that outlines what we think is a better approach to WTP: https://sawtoothsoftware.com/resources/technical-papers/estimating-willingness-to-pay-in-conjoint-analysis
answered Jan 26 by Bryan Orme Platinum Sawtooth Software, Inc. (195,815 points)
...