This is a fantastic read about the challenges of observational studies that don’t account for selection bias. Wow.
The best of these observational studies try to control for these variables. Even so, we can never be sure that there aren’t unmeasured factors, known as confounders, that are changing the results. Confounding is a classic problem of selection bias, so perhaps the best way to be sure that the two groups aren’t different is to remove selection from their hands by randomly assigning participation.