(SPSS- Logistic Regression) Interpretation of moderation effect with one non-significant main effect

[MH_Thesis]

Hello everyone,

Currently, I am in the final stages of my thesis on political instability and firm innovation propensity. However, I have some trouble interpreting an interaction effect and I've been looking for an answer for a long time. Unfortunately, I have not found a convincing answer yet.

I am researching the effects of political instability (scale 0-5) on firm innovation (binary, not innovative / Innovative) using a logistic regression model. This effect is non-significant, however, I also hypothesize moderating effects. One of these is Internationalization (binary, not international (0) or international (1)). The main effect of internationalization on firm innovation is significant, the interaction effect of internationalization*politicalinstability on firm innovation is also significant. However, like I mentioned, the main effect of political instability on firm innovation is non-significant.

SPSS output (DV=Firm Innovation)

Political instability:
(WALD=1.97, df=1, p=.161) B=.157 Exp(B)=1.170
Internationalization(1):
(WALD=28.12, df=1, p<.001) B=1.228 Exp(B)=3.415
Internationalization(1)*Political Instability:
(WALD=13.84, df=1, p<.001) B=-.322 Exp(B)=.725

I've only followed two statistics courses and I am far from properly understanding SPSS, therefore I hope I have added the necessary information. If not, please let me know what is missing and I will provide whatever else is needed. After days of searching and not finding anything helpful, I hope someone on this forum can help me out with this. How would I interpret these results? Any help at all would be extremely appreciated!

Regards,

Mees

 

[Karabiner]

There is no problem with a non-signficant main effect if you want to interpret a statistically signficant interaction.

With kind regards

Karabiner

 

[hlsmith]

Agreed and I will recommend always plotting interactions.

 

[MH_Thesis]

Thank you both for the quick reply. I have a few questions on this.

1. I've also seen sources that say ''do not interpret the interaction if the main effect is not significant''. Do you have a known source that supports the argument of there being no problem in interpreting this interaction result? I've found many papers on interaction in logistic regression, but from what I see none of them explicitly state that it is no problem to have insignificant main effects and why this is the case.

2. If I were to interpret the results I have, I usually could say something like ''international firms have a reduced impact from Political Instability on the likelihood to be innovative compared to domestic firms'' however, as the main effect of political instability is insignificant, this seems kind of strange as Internationalization then would moderate an effect that is not there? Or how am I seeing this wrong?

 

[Karabiner]

1. I've also seen sources that say ''do not interpret the interaction if the main effect is not significant''.

Could you name some of these sources? I have never read such statement, since obviously it does not make sense.
Consider an "extreme" interaction such as that displayed here. None of both main effects is present, neverthesless
there is a clearly interpretable interaction effect.

You will instead often find the observation that it is difficult to interpret a main effect if an interaction is present, since
the effect of a predictor is now conditional on the interaction.

2. If I were to interpret the results I have, I usually could say something like ''international firms have a reduced impact from Political Instability on the likelihood to be innovative compared to domestic firms'' however, as the main effect of political instability is insignificant, this seems kind of strange as Internationalization then would moderate an effect that is not there? Or how am I seeing this wrong?

As hlsmith suggested, you should maybe create plots. And/or you could calculate the values of the dependent variable
for different configurations of your predictors (what is predicted if A has a low value and B has a low value, if A is low and
B is medium etc.).

With kind regards

Karabiner

 

[hlsmith]

Think of this, you have two exposures - smoking and asbestos exposure, both smoking and asbestos exposure cause lung cancer. Exposure to both result in a multiplicative increase in cancer risk. In the small sample you fail to find an effect beyond chance with the single exposures and find an effect with the combination. are you going to ignore these results. Now imagine the same scenario where the single exposures are not related to outcome but the combination is. are you going to ignore these results. In both scenarios, ignoring the interaction effect would not be prudent. It is important information.