How we found a potential confounder