A confounding variable should meet the following conditions:
1. It is associated with the outcome
2. It is associated with the outcome, independent of the variable of interest
3. It is associated with the variable of interest
4. It is not in the causal pathway between variable of interest and outcome
(See: http://en.wikipedia.org/wiki/Confounding)
How should 'association' be determined? The outcome of a statistical test or size of the effect? Do we need to provide a rationale for confounding variables in our research? If so, what guidelines should we follow? Can you suggest any good papers to guide our decisions on what should and should not be controlled for in observational studies?
I agree with you Stefan. I'd probably also highlight that this means you need to think about it, draw up a DAG etc., *before* you do the analysis. Specifically addressing one of the questions, it's generally a bad idea to base the decision on statistical tests, and a bad idea to base those on the dataset you're using. That approach leads to problems in almost every department.
Please look at chapter 7 in the book of Miguel Hernan (http://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/) for the definition of a confounder based on the counterfactual-framework. The main idea is, that confounders are "common causes" of both the outcome and the exposure and/or confounders close "backdoor paths" from outcome to exposure in causal diagrams (DAG).
Hope this helps.
I agree with you Stefan. I'd probably also highlight that this means you need to think about it, draw up a DAG etc., *before* you do the analysis. Specifically addressing one of the questions, it's generally a bad idea to base the decision on statistical tests, and a bad idea to base those on the dataset you're using. That approach leads to problems in almost every department.
You are absolutely right, Darren. I always try to convince my colleagues to draw up a DAG in advance before the analysis based on their (mechanistical) knowledge. So they can decide which covariable is a confounder, a collider or maybe a mediator.
I think the Hernan book is very good introduction in the usage of DAGs. I can also only recommend to participate at the Hernan course about causal inference in the Erasmus Rotterdam Summer Programme if possible. Unfortunately I wasn't able to participate myself yet.
You have some good and important suggestions here, and also I would like to add that since you are the researcher and you have the knowledge of the research problem, you can spend some time to think about the possible confounding variables which might affect the outcome of yours. Generally speaking, depending on your subject knowledge of the research problem, you may find that you can even ignore some of these according to your objectives. To this process if you can incorporate the things like DAGs you may come out with a good solution. I am just pointing out that all of the statistical solution are not a panacea for the research problem, there is always room for error. So do not get confused at the beginning, if you can justify your actions in a fair manner ,I think that is enough because we cant always be perfect.
Besides that, there are actions to account for the confounding effects. In the designing phase you can randomize, to allow for all the confounding clusters to have a equally likely chance to be included in the study and then you will minimize the bias. Also you can isolate most important confounders that you think are the most important ones. As an example if you are determining the association of eating habits on lung cancer and you know that heavy smoking is a confounder here you can choose different subjects which all of them are heavy smokers and then check the relation with your variables (just an imaginary example).
Also in the analysis phase you can perform statistical methods like multivariate regression to account for the confounding, still in here you cannot adjust for all the variables, but can choose most important variables.
you can also refer the book " Modern Epidemiology by Kenneth J. Rothman, Sander Greenland, Timothy L. Lash, Lippincott Williams & Wilkins, 2008"
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics