Categories → Statistics, Variables, Causal Inference
Confounding is a situation where an extraneous variable is related to both the treatment and the outcome, making it difficult to determine the true causal effect of the treatment on the outcome.
Confounding can arise when there are variables that are associated with both the treatment variable and the outcome variable, but are not themselves caused by the treatment. These variables are known as Confounders and can lead to biased (misleading, or inaccurate) estimates of the causal effect.
Confounding is often referred to as a "mixing of effects", which means the true effect of the treatment is "mixed" or confused with the effects of other, "confounding" variables.
Possible confounders should be drawn as ordinary variables in your Causal Diagram. Causal Wizard will work out if they are confounders, or not, and colour them as appropriate, during the Check process.
As an example, let's say we are interested in studying the effect of smoking on lung cancer risk. We collect data on a group of people who smoke and a group of people who do not smoke, and we find that the smoking group has a higher rate of lung cancer. However, if we do not account for other factors that may be associated with both smoking and lung cancer risk, such as age, gender, and exposure to environmental toxins, then we cannot be sure that the observed association between smoking and lung cancer is causal. These other factors could be confounders, and if we do not control for them, they could lead to biased estimates of the true effect of smoking on lung cancer risk.
To address this issue, we need to include these variables in our models, and use appropriate statistical methods that can adjust for the confounding variables, such as stratification or regression analysis. By doing so, we can obtain more accurate estimates of the causal effect of smoking on lung cancer risk while accounting for the potential confounding factors.
For further examples, see this article.
A common intuition is that with enough data, noise cancels out and you'll get the right answer. Confounding doesn't behave like noise - it's a systematic distortion that gets more precisely wrong as you add data, not less wrong.
Suppose, in the smoking example above, that smokers in your data happen to be older on average than non-smokers, and older people have higher lung cancer risk regardless of smoking. If you compare the raw smoking vs. non-smoking lung cancer rates without accounting for age, part of the difference you measure is really the effect of age sneaking in through the smokers group being older - not the effect of smoking itself. Collecting a million more rows, with the same age imbalance between smokers and non-smokers, doesn't fix this: the age effect is baked into every single row, so it doesn't cancel out on average. You'd get a very precise estimate - of the wrong number.
This is the key difference between confounding and ordinary statistical noise: more data shrinks your confidence interval, but the interval keeps missing the true effect if a confounder is unaccounted for.
One of the discoveries of Causal Inference is that unnecessarily controlling for too many variables (known as over-controlling) can also bias the result, so it's important to control for only the right set of variables - see Controlling for and Conditioning on a Variable for the full picture, including colliders and mediators, which should not be controlled for. These are automatically identified during the Check process in Causal Wizard, and displayed in your Causal Diagram.