> In your example, the positive correlation between the dose of the antidepressant and the probability of suicide still exists in the data. Of course it does b

> In your example, the positive correlation between the dose of the antidepressant and the probability of suicide still exists in the data. 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.