# How to handle different important variables, including overlapping information, in regression?

I am interested in smoking_ads effect on smoking_rate. I also have two categorical confounders in the data: town and year.

Possible DAG shows two backdoors that need to be closed in conditioning. Each town may have different characteristics that possibly have an effect on ads (laws) or smoking rate (habitational factors).

However, my data is a bit tricky as its detailedness is limited. Smoking_ads is calculated for each town by study year: the values of "smoking_ads" are very similar for each town. For example, town "C" values range between 22-24 and town D between 233-257.

To illustrate, a stupid analogy of this is would be the following model:

weight ~ sex + breast_size


1) Should I adjust for town or not? It seems, et adding town kills the effect. How you handle such situations?

2) How does it differ if "town" is added as a another level into model + (1 |town)?

1. Should I adjust for town or not? It seems, et adding town kills the effect. How you handle such situations?

Yes, you will want to include town in your model because it is a confounder according to the DAG.

1. How does it differ if "town" is added as a another level into model + (1 |town)?

Not only is it a confounder according to your DAG but you also have repeated measures. Provided you have sufficient of them, random intercepts are a good way to do this. Fitting fixed effects will also achieve the same thing. You might want to check the answer here, to a related question: How do random effects adjust for confounding in a model?

• Thank you so much Robert! The link you provided was very enlightening! – st4co4 Aug 7 at 17:03
• You're welcome. Happy to help:) – Robert Long Aug 7 at 17:20