# Stepwise regression, moderation effects, main effects

I have a simple model:

$A$ is hypothesized to be a predictor / regressor / explanatory / input variable

$B$ is hypothesized to be the response / regressand / explained / outcome variable

So, the relationship looks something like:

$A\longrightarrow B$

Additionally, $C$ is hypothesized to be a moderator of the relationship between $A$ and $B$.

When I run the regression by including all the variables ("enter" procedure in SPSS), none of the relationships are significant.

When I use "step-wise" regression, and let SPSS choose the variables to include, $C\times A$ has a statistically significant effect on $B$, but SPSS stops the "step-wise" regression procedure before including $A$. I suppose one can assume that $A$ doesn't have a statistically significant effect on $B$ (after the inclusion of $C\times A$ in the model).

Thus, I have a statistically significant moderation term $(C\times A)$ , but the main term $(A)$ has not been included in the model.

What can I do with such a result? I was taught that moderation effects are not valid if main effects were not included in the model. Is there a way around that admonition? Is there some way I could still employ this result profitably?

• Thus, I have a statistically significant moderation term (C×A), but the main term (A) has not been included in the model. What SPSS command you use? Linear regression command does not create interactions internally. Did you compute AC youself first and include three variables A, C, AC? – ttnphns Jan 14 '14 at 10:14
• @ttnphns Yes, I computed AC and included A, C and AC. – thanks_in_advance Jan 15 '14 at 5:41
• Use entering in blocks. Block1= A C (method enter); Block2= AC (method stepwise or forward). Hereby, you force main effects first. Then, if the prediction allows, the interaction will be added. – ttnphns Jan 15 '14 at 6:01

• Did you make sure that the value 0 on both $A$ and $C$ are meaningfull and near the range of the data? For example, if $A$ is year of birth, than the value 0 would be the year 0, which in most cases is way outside the range of the data. In that you would need to center the offending variable(s) before computing $A \times C$. – Maarten Buis Jan 15 '14 at 8:47