Suppressor-Effect? - Should the magnitude of the beta for a suppressor decline when other variables are added to the model?

helLW

New Member
#1
I'm working on a logistic regression in R regarding the association of prenatal care (PC) with later breastfeeding. With the stepAIC function I've find out that the 2 variables "pregnancy-duration" (PD) and "socio-economic status" (SES) are important predictors. But now I have the following problem:
# The base-model is not significant
RR: 1.03
p-value: 0.6
AIC: 91.89
ß-PC: 0.03

#After adding PD and SES the association becomes significant:
RR: 1.31
p-value: 0.02
AIC: 71.3
ß-PC:0.27
ß-PD: 1.65
ß-SES: 1.11

#Additional, here are the zero β-values:
ß-PC: 0.03
ß-PD: 0.93
ß-SES: 0.90

All β-values increased and the non-significance turned into significance. However, in a regular suppressor-situation the β-value of the suppressor should decline or I'm wrong? Additional both predictors are significantly associated with the outcome and are not significantly correlated with each other.
Is it still a suppressor-effect or what could it be?
 

hlsmith

Not a robit
#4
Are you using stepwise selection based on AIC, otherwise I am not familiar with stepAIC. Can you provide documentation on it. I am not overly familiar with suppression, but as I remember there is a decrease - thus the suppression part of the term.

You have a > 10% change in the RR, which is a basic effect size used in regards to confounding. Instead of just looking at results, what you are missing is drawing out the a priori hypothesized relationships between variables (e.g., signed directed acyclic graph). Such an illustration helps to interpret relationships and visualize possible mediation, moderation, mediated moderation, confounding, and controlling for an effect of covariates and the outcome (backdoor path).