Nonrandomized studies using causal-modeling may give different answers than RCTs: a meta-epidemiological study