# Interpretation of multivariate linear regression significance with interaction term

I am working through an obesity dataset for a project I'm involved in, and I wanted to run a fairly simple multivariate linear regression on the data. What I was interested in is the correlation between income and education on obesity. In R, I ran two version of the model:

(I) obesity ~ income + education

and

(II) obesity ~ income * education

The results for (I) showed that income was significantly correlated with obesity while education was not, while the results for (II) showed that both income and education were significantly correlated with obesity (while the interaction term was not).

I am wondering what an intuitive explanation/interpretation of these results is; that is, why is education not significant in (I) but is significant when the interaction term is added in (II)? I have attached a screenshot of the results from the two runs. • The results show that there are nonlinear relationships that you haven't modelled. That's not surprising given the nature of the variables -- it would hardly make sense for someone on 10 times the income to be 10 times as obese (or 10 times less obese). I suggest you transform income to log-scale and look at some plots! Aug 28 '17 at 0:34
• @bogdanknez See this stats.stackexchange.com/questions/52067/…. Aug 28 '17 at 2:38
• @GordonSmyth That is not the interpretation of the parameters of an ordinary linear regression... Aug 28 '17 at 7:38
• @GordonSmyth I'll have to agree with Knarpie above- the coefficient estimates here are not indicating any 10-fold difference. Rather, let's take the first example (obesity~income + edu)- the coefficient for income is -0.28. I already have income reported as thousands of dollars, meaning that a single unit change in income ($1000) correlates with a 0.28 unit change in obesity (measured in % here). That's not that unbelievable at all, given what we already know about relationships between income and obesity. I have looked at plots for all of this and the data is quite linear. Original Q stands Aug 28 '17 at 11:08 • @tintinthong thanks for that thread! This is along the lines that I was thinking. The reason why I posted my question is that, in my example, the variable I'm adding is actually an interaction term of the first two variables, so I'm trying to intuitively explain why this is causing the changes it's causing. My own explanation would be that since income and education are related to an extent, we should include the interaction term in order to avoid OVB (which causes an underestimation of education if we only do obesity~ income + edu Aug 28 '17 at 11:16 ## 1 Answer I will assume that you already understand the basic principles of multiple regression. In particular, I assume you already understand that all the regression coefficients will change when you add a new predictor variable to a regression, unless the new variable is uncorrelated with the old variables. So I won't spent time on that and will concentrate specifically on the type of interaction model that you have fitted. The basic explanation here is that interaction effects are non-linear in the original variables and it is always possible for non-linear effects to average out to zero on the scale of the original linear variables. For this reason, it is quite common for an interaction to be significant even when neither (or only one) of the interacting variables were significant without the interaction. Any variation on this scenario is possible. Even if the interaction isn't quite strong enough to be statistically significant itself, it nevertheless can easily soak up enough variability to push a previously borderline main effect to become significant. In your case the interaction term is simply a new covariate equal to the product of the original variables. In other words, you added a third predictor variable equal to income_thousands$\times$edu_uni. This new covariate is obviously nonlinear in the original variables but nevertheless also correlated with the original variables. The new product covariate just fails to make 0.05 significance (p = 0.06), but is nevertheless strong enough to push edu_uni from borderline significant (p = 0.10) to significance (p = 0.018). One possible explanation in your case is that the original linear regression might be underestimating slightly the obesity for cities that are simultaneously above average for both income and education. The original linear regression predicts a reduction in obesity with both income and education, but this trend might perhaps taper off for cities with both variables high, and that would lead to this sort of underestimate. It might be that either income or education need to be high to get the obesity reduction but that having both high doesn't give a further reduction. In this scenario, the interaction term would get a positive coefficient to pick up this effect at high levels of both variables, but then the coefficients of the original variables would have to become more negative to compensate for the collinear interaction term in the main body of the regression. The original variables would therefore become more significant. The results you show are consistent with this scenario. Note that this is an essentially nonlinear phenomenon. If obesity depended on income and education in a purely linear fashion, then these effects are not possible. It is not true that you need to include the interaction term in the model simply because income and education are themselves correlated. The multiple regression arithmetic already accounts for correlations. In fact the type of model you have fitted: $$E(y) = \beta_0 + \beta_1x_1 + \beta_2x_2 + \beta_3 x_1x_2$$ is one that generally should be avoided when$x_1$and$x_2$are continuous covariates. It makes sense to try nonlinear terms, but including the product term without either of the square terms very seldom makes sense. To check whether$x_1$and$x_2$have a nonlinear relationship with$y\$, it would be reasonable to try a quadratic response surface. But to fit that, you need the quadratic terms as well: $$E(y) = \beta_0 + \beta_1x_1 + \beta_2x_2 + \beta_3 x_1x_2+\beta_4x_1^2 +\beta_5x_2^2$$ You could then use an anova approach to test whether the quadratic regression is a worthwhile improvement over the linear regression.

The 3-variable regression that you've fitted is a sort of half-way house between the linear and quadratic regressions without any clear interpretation. To see what is wrong with the 3-variable regression, suppose that you redefined your income and education variables by subtracting off the country-wide averages. The new income and education variables would be the same as the old, just shifted by a constant value. Doing so would make no difference to the linear regression -- only the intercept would change. It would also make no difference to the quadratic regression in that fitted values should be unchanged and the p-values for the quadratic terms would stay the same. However your 3-variable regression would change completely, giving different fitted values from before. The trouble is that the 3-variable regression is too special: it isn't invariant under sensible rescalings of the predictor variables and so doesn't have any general interpretation.

When you fit in interaction model ~A*B in R with factors A and B, the interaction model takes into account all possible combinations of the factors A and B. Unfortunately the A*B formula doesn't do this when A and B are both continuous covariates. In that case A*B just adds a single product term, which accounts for only a small part of the possible bivariate effects. Despite the terminology in R, this should not truly be considered to fully represent possible interactions'' between the covariates.