# How to understand the significance and effect of each term in a non-linear model with interaction terms?

Let's say we have a model as given below:

$$Y_1 = \beta_0 + \beta_1 X_1 + \beta_2 X_2+ \beta_3 \frac{X_2}{X_1} + \beta_4\frac{X_2^2}{X_1}$$, $$R^2= 0.98$$

Here, $$X_1$$ & $$X_2$$ are positive integers.

The value of $$\beta_3$$ and $$\beta_4$$ are high compared to $$\beta_1$$ and $$\beta_2$$.

If we fit multiple regression model for $$Y_1$$ with only the linear terms, the R-squared is very high.

$$Y_1 = \gamma_0 + \gamma_1 X_1 + \gamma_2 X_2$$, $$R^2= 0.9$$

So, can we say the interaction terms or ratio terms here are not significant? Also how can we find the significance/ contribution of each term to the model prediction?

Following are my observations when fitting only the interaction terms

$$Y_1 = \alpha_0 + \alpha_1 \frac{X_2}{X_1}$$, $$R^2= 0.4$$

$$Y_1 = \rho_0 + \rho_1 \frac{X_2^2}{X_1}$$, $$R^2= 0.7$$

• "The value of β3 and β4 are high compared to β1 and β2." their units are different, so their relative size is a case of apples and oranges Aug 26, 2022 at 1:40

Technically, this is still a linear model as it is linear in the coefficients, the $$\beta$$ values. As @Glen_b notes in a comment, the magnitudes of the coefficients aren't to be compared when the associated predictors are on different scales. The sub-models on their own also aren't reliable guides when the outcome-associated predictors are correlated, due to omitted-variable bias.