Have an idea?

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

How to interpret "None" part-worth Utility in Latent Class, CBC-HB estimates?

Hello Everyone,

I was wondering if anyone could help me to understand the meaning of None Part-worth utility for latent class clustering and CBC-HB individual partworth utilities.

For example, I did Latent Class clustering into 4 groups.

"None" has the part-worth utility as follows:
1. Group 1: 363.78328           
2. Group 2: -46.69162
3. Group 3: 229.23112
4. Group 4: 401.70326

Suppose Group 1 has a high part-worth utility of 223.70555 for low price $30 and its None utility is 363.78328 as shown above.

Does this mean, even if they prefer $30 more (hence price sensitive), their likelihood of not purchasing is also very high?

For the HB estimates, I have None as well against each respondent's response. I guess, this None is used to estimate the share of preference and thereby purchase likelihood. Am I correct in this reasoning or is there a catch to it?

And is there any way to use HB estimates to use in latent class clustering in Sawtooth?

Kindly advise.

Regards,
Sujit
asked Apr 13 by sacharya (490 points)
added to the section above

1 Answer

0 votes
 
Best answer
First, these utilities you are showing use "zero-centered diffs" rescaling, also known in our latent class procedure as "utilities rescaled for comparability"  which is a transformation that blows the utilities up to a larger scale and makes the scale very similar between respondents or groups of respondents to be able to make better comparisons.

The None utility is a threshold utility that gets estimated and scaled to be compared to the sum of the utilities across all the other part-worths.  For example, if the None utility is higher than the sum of part-worth utilities across all the attributes that define a product alternative (take one level from each attribute and sum their part-worths), then we would think that this person or that this segment of respondents has a higher likelihood of choosing None than choosing the product alternative.

Latent Class Clustering in my mind is when the analyst uses an array of basis variables to cluster respondents into groups, where there is no dependent variable.

Latent Class MNL in my mind is what Sawtooth Software's Latent Class procedure does, where the algorithm fits multiple group (class) vectors of utilities that provide better fit to the data (where the fit is the likelihood of choice, so there is a dependent variable involved), and for which each respondent has a continuous probability of belonging to each group (class).

You cannot use HB utilities in Latent Class MNL automatically in Sawtooth Software's programs, though a researcher could try to do that on their own using a more flexible HB routine (though I would think this would be a strange thing to do).  If you used normalized HB utilities (such as from "zero-centered diffs" rescaling), you could submit those normalized utilities to a Latent Class clustering routine.  However, this approach is viewed as not as good from a statistical standpoint as directly using Latent Class MNL (assuming we're referring to a choice experiment, such as CBC or MaxDiff).
answered Apr 13 by Bryan Orme Platinum Sawtooth Software, Inc. (152,030 points)
edited Apr 13 by Bryan Orme
Thanks a lot Bryan for this kind help and explanations for both the sections.

If I understand you correctly from the statement above:
".....For example, if the None utility is higher than the sum of part-worth utilities across all the attributes that define a product alternative (take one level from each attribute and sum their part-worths), then we would think that this person or that this segment of respondents has a higher likelihood of choosing None than choosing the product alternative.....", I don't have to sum up all the part-worth utilities to see if it is > or < None?

For the below estimates for the various attributes and levels, do I have sum up part-worth utilities of ($30,$60,$90), (1m, 3m, 6m), (yes), (yes), (yes), (10SP, 20SP, 30SP), (yes), (yes), (yes) excluding the last level of each attribute as it is not estimated?

Or just sum up part-worth utilities of ($30, 1 month, Yes, yes, yes, 10 SP, yes, yes, yes ).

Sorry, may be I am asking stupid questions, but kindly bear with me. I just want to be sure I understand properly your explanations.

Extract of Part-Worth Utilities from Latent Class Clustering:

$ 30/Month      223.70555
$ 60/Month     -60.63883
$ 90/Month     -93.65318
$ 110/Month    -69.41354
   
1 Month    140.86286
3 Months    -2.37091
6 Months    -69.63914
12 Months    -68.85280
   
Yes             24.83714
Null    -24.83714
   
Yes    11.14106
Null    -11.14106
   
Yes              42.02101
Null    -42.02101
   
10 SP    38.18847
20 SP    -7.72937
30 SP    -14.38432
40 SP    -16.07478
   
Yes    52.23599
Null    -52.23599
   
Yes    10.87364
Null    -10.87364
   
Yes    17.82916
Null    -17.82916
   
NONE    363.78328
Sum up nine part-worth utilities (one level from each of your nine attributes) and compare that sum to the None utility.
thank you soooo sooo much again, Bryan.. I really appreciate all your patience and help for me.
...