Have an idea?

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

Should I add interactions if it makes many of my attributes insignificant according to their T value?

If I run a latent class analysis without adding interactions only two of my attributes show insignificant t-values. When I run the same analysis with the interaction then 7 individual attributes show an insignificant t value.

Note: 2 interactions were found to be significant, and they increase the log likelihood from -947.85 to -928.43 and the chi-square from 1023.80 to 1062.63.
asked Jul 13, 2018 by Paul van Empel

1 Answer

0 votes
Paul, do you mean that two and seven of the attribute LEVELS are non-significant without and with the interaction?  That would be less concerning than entire attributes becoming non-significant, which would be odd.

Because the utilities of the levels are centered at zero, it's not a surprise that some of them have non-significant t-tests (which would change, by they way, under different kinds of coding) so that part really doesn't concern me.

The improvement in LL is pretty substantial, but of course I don't see how many d.f. you're adding with those interactions.
answered Jul 13, 2018 by Keith Chrzan Platinum Sawtooth Software, Inc. (92,075 points)
Entire attributes become insignificant when adding the interaction effects. So I have four classes with 4 attributes, each attribute has either 2 or 3 levels. When adding the interaction effects, the entire attributes (all levels) become insignificant for certain classes (7 in total).
Hmmm, are those variables that become non-significant the ones involved with the interactions or are they different ones?  I'm just trying to visualize your data.  Perhaps you'd like to send me the utility summary at keith@sawtoothsoftware.com?