Adjusting for bias and unmeasured confounding in Mendelian randomization studies with binary responses.