I will continue to maintain that, given this evidence amidst the existing evidence that smoking is not good for you, you should not start smoking for reasons related to health risks from the Covid.
I will continue to maintain that, given this evidence amidst the existing evidence that smoking is not good for you, you should not start smoking for reasons related to health risks from the Covid.
Sure there is.
As an example, let's say there's a positive correlation between the dose of some antidepressant and the probability of suicide. There could be a confounding variable with respect to the severity of depression in those samples such that people with more severe depression are more likely to take larger doses and more likely to commit suicide. The original correlation may not necessarily be between the dose of the antidepressant and the probability of suicide, but rather the severity of the depression and the probability of suicide. In other words, the correlation could be a correlation of the wrong thing because you didn't control for depression severity.
It does not matter only for causality, but also for correlation. In both cases, it's important to control for confounding variables so that you're measuring and analyzing the actual variable(s) that you're trying to, not some other variable(s) (i.e., confounding) indirectly. Otherwise, it may no longer be a relationship, whether causal or correlative, of the thing you think it is, but of some other thing.
What you're describing is sampling bias: the correlation between the antidepressant and probability of suicide in the sample may not reflect a similar correlation in a larger population.
Of course it does because the confounding factor of depression severity wasn't controlled for. You can't always control for confounding variables after the fact.
> What you're describing is sampling bias: the correlation between the antidepressant and probability of suicide in the sample may not reflect a similar correlation in a larger population.
Yes and no. I am describing sampling bias, but that's because the sampling bias is the source of the confounding variable in my example. The point was that the correlation was supposed to represent the relationship between antidepressant dosage and probability of suicide, not the relationship between people who typically take antidepressants and their probability of suicide. In my example, the dosage ended up being a dependent variable along with the probability of suicide and the severity of depression ended up being the independent variable, when the dosage was supposed to be the independent variable.
Even though looks a lot like a classic example of collider bias, where obesity looks like it protects you from dysglycemia.