I estimated a model for which I tested the assumption that the effect of my variable of interest is homogeneous for observations belonging to two different groups (for simplicity assume rural/urban).

Results show that, on the one hand, both the main effects are not statistically different from zero*, on the other the interaction term is statistically significant. This denotes a cross-over interaction where the effect changes direction in sign according to the group being considered.

Specification ONE:

In this particular example I constructed the dummy as 1 if the observation is a rural household, 0 otherwise. As anticipated above the only statistical significant coefficient is that of the interaction term, which is negative (-0.62), thus the overall effect for rural households coincides with the interaction's coefficient.

I thus reject the assumption that the effect of my variable of interest is homogeneous for the two groups. However, I am interested to see the extent of this switch in sign of the coefficient of interest.

Specification TWO:

Conversely, if I construct the dummy in the opposite way (i.e 1 if the observation is urban, 0 otherwise) and estimate the same model. I get the following results:

  • interaction: statistically different from zero (0.62);
  • main effect of variable of interest: statistically different from zero (-0.63);
  • main effect of the dummy defining the groups: not statistically different from zero.

The overall effect of the interest variable for urban households equals -0.01.

However, two questions arise:

  1. Given the detection of a cross-over interaction in the first specification, is it worth computing and showing the results re-computing the dummy for rural/urban as in specification two above? As a reader I would like to see at what extent the effect is positive (something I cannot see from the results of specification one).
  2. Can I still talk of cross-over interaction even if in specification two I get a rather very small but still negative and significant effect?

Note: Statistical significant results are those for which p<0.05; for not statistically significant results p>0.1.


1 Answer 1


When a cross product is in the model the effects of other variables are really simple effects: they are effects when the other variables are 0. You might find your results more interpretable if you use contrast coding (-1,1) and center your other variable.

  • $\begingroup$ I see your point. However, the contrast coding will allow me to explore the differences in mean of the two groups. Moreover, its interpretation is straightforward only for farily balanced designs. I do not see how this is able to provide an answer to my question. What I am interested at is the change in sign and the magnitute of the overall effect by group, not difference in groups' means. $\endgroup$
    – Caserio
    Mar 16, 2017 at 8:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.