There’s a lot of confusion on confounding.
We need to understand confounding when we try to use non-randomised studies to see if doing a Thing is going to produce more good than harm, or looking for risk factors we then will hope to influence to produce goodness.
If we look at observational data on time-to-diagnosis in some brain tumours, we find shorter times between symptoms and diagnosis lead to higher risks of death / disability. Ergo, we should lengthen scan waiting times, and neurosurgical review accessibility, in order to save lives / prevent disability… “No!” you all cry (at least, I hope you do; if not, please self-refer to your governing body for immediate review of fitness to practice). This finding is confounded by the nature of those tumours which present quickly, and lead to worse outcomes. The underlying biology is the cause of worse outcomes in the early group, compared to the later ones. The counfounder is the feature which is the actual cause, compared to the observed / presumed cause.
Whenever you’re looking at observational studies, you need to very carefully explore what else might be going on. Even with statistical techniques to adjust for the measured and observed differences between groups which may be affecting the outcome, there is always residual confounding (the bits left unseen). The GRADE group suggest if the observed effect, after all those adjustments and considerations, is an odds ratio of less than 2 (conversely more than 0.5) you should be wary. “Sketchy stuff” as the young people say.
When you’re away from randomised trials, don’t get confused. Grasp confounding.