Magnitude and direction of missing confounders had different consequences on treatment effect estimation in propensity score analysis.