Sophisticated statistical techniques like two-stage least squares (instrumental variables) are designed to correct for reverse causation and confounding, but they are often applied to data that is simply not good enough for the task, so we can never control for all factors and may falsely attribute correlations to an unobserved variable.