Have an idea?

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

LCA Large/ hugh part worths but not sigificant


I have run a Latent class analysis for around 100 respondents. For the smallest segment (14,5%) the price attributes shows very large part worths (rescaled) (1 level = 155.86/ 2 level= 134.35/ 3 level = -290.21).  At the same time the T-value indicates that the price is not significant for this segment. I am aware that the T-values are calculated using the standard errors which are again extremely large (1 level = 273.22/ 2 level= 273.22/ 3 level = -546.43).

How should I interpret this result? Can I say something about the preference regarding price for this segment?
asked Nov 29, 2018 by Aniek

1 Answer

0 votes
This raises some interesting questions.  15% of 100 respondents is around 15 respondents (weighted count).  That is certainly a small sample size for CBC.  Normally, one would not want to conduct a CBC with just 15 respondents and try to interpret the results.  Especially if the number of attributes and levels was relatively large and the number of choice tasks shown per respondent were relatively small.

It would be interesting to examine the raw utilities (on the logit scale) across all attributes and levels for this tiny segment.  Are other attributes even more extreme in utility and do other attributes for this segment have significant T values?

How was your experimental design generated?  Using Sawtooth Software's methods?  Did you you use any prohibitions that might have hindered the quality of the experimental design?
answered Nov 29, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (174,415 points)
Hi Bryan,

I understand that a CBC with 15 people would not make sense in a normal research, but this was given an assignment for a Conjoint class. The number of attributes was 6 with in total 12 parameters to be estimated. Respondents filled in 12 choice sets with three options and a none. No prohibitions were included. Using the balanced overlap method.

Other attributes do have significant results given the T-values. These t-values do seem normal.
Well, I think the most likely answer to your issue is that with just 15 respondents and 12 choice tasks, with six attributes, it's not enough statistical power to find a statistically significant T for the price attribute.  It doesn't mean that segment of respondents ignores price.  It just means we don't have enough data to be 95% confident.