Hi Spencer

You should read articles about directed acyclic graphs (DAGs), which is an analytical method to determine what to adjust for.

Adrian

http://sfi.nr.no/sfi/images/d/d8/Greenland.pdf

In your example, you are apparently testing to see if the observed association between smoking and lung cancer holds even after adjusting for potential confounders. The choice of which factor to enter into the model is very important. This is not statistics but pure methodology. You need to clearly identify how each factor relates to both smoking and lung cancer.

For a factor to be a confounder, these three conditions have to be met:

1. The confounder needs to be a known cause of lung cancer.

2. The confounder needs to be associated to smoking.

3. The confounder must not be on the causal pathway between smoking and lung cancer.

In your model, you therefore presume that socio-economic status predefines smoking status, and is a known cause of lung cancer. You also ruled out that smoking could affect our live prior to having lung cancer. This could however be the case smoking has some direct and indirect costs (ex: the cost for paying for cigarettes, the loss of income due to other diseases caused by smoking).  Adjusting for socio-economic index could therefore over-adjust the true association between smoking and lung cancer. By including the socio-economic index measured after the person has started smoking, we would be removing the socio-economic burden's effect on cancer due to smoking.

So you need to clearly think out your model at the planning phase to privilege co-factors that are clearly present prior to the moment the person has started smoking. Once your model's rationals' are sound, you then need to make sure you do not have factors that are two highly correlated between themselves. Finally,  you need to keep in mind that there will always be some residual confounding left from unknown factors.

In the end, all we have is a statistical association.  Cardiologists assumed for decades based on "overwhelming evidence" that eating too much cholesterol or salt causes heart disease, but those theories are now discredited.  Smoking could ultimately be a proxy for people who choose to live risky lifestyles and work around more toxic fumes for a living.  I don't doubt that inhaling a lungful of smoke can cause health problems, but smoking variables are picking up on more than just smoking.

Thank you all for answering. Adrian, I will read this article. I learned about DAGs very briefly but never would have thought about them for this. Paul, just to be clear, this isn't my model but one from an already published paper. It essentially adjusts for SES multiple times by using multiple proxies for SES in a single model. I guess, as you say, these factors would be highly correlated since they are all SES measures? Also I think it may be established that smoking is more prevalent with lower social classes, but any relatedness will do for SES to be a confounder.