Have an idea?

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

Only one attribute level has a significant t-value

Dear all,

I've got two questions concerning the output of my CBCA Logit Estimation.

1) Do all or at least two of the attribute levels (in case of an attribute with 3 levels) have to have significant t-values? Sometimes, just one of my attribute levels has a significant t-value (+/- 1,96).

2) Why do the values of ChiĀ² etc. change, if I integrate interaction into the Logit Analyse?

I hope you can help me!

Best regards
asked May 6, 2015 by Lena

1 Answer

+1 vote
I assume you're looking at t-values from an aggregate logit report.  An attribute with ALL levels with t=0 could still be extremely important to include within the model!

Let me tell you how: let's imagine you had an attribute called brand with just two levels: Coke and Pepsi.  Of your respondents, 50% strongly love Coke and 50% strongly love Pepsi (to an equal degree).  When you look at pooled aggregate logit, the utilities for Coke and Pepsi are both 0.0 (since the respondents' preferences cancelled each other out), leading to a t=0.   However, CBC/HB or Latent Class modeling would correctly capture that there are different people with strongly differing opinions that need to be accounted for within the market simulator.

Values of the fit statistic (like Chi Squared and log-likelihood) change because adding interaction terms changes the overall fit of the model.

Aggregate logit is useful to gain a quick view of what's happening on average with your data and examining the precision of your design (e.g. the standard errors of utilities)...but you should hardly ever use it for your final analysis.  (Main exceptions in my book are when working with really sparse data or for some MBC applications.)
answered May 6, 2015 by Bryan Orme Platinum Sawtooth Software, Inc. (132,290 points)