Bias–variance trade‐off in pharmacoepidemiological studies using physician‐preference‐based instrumental variables: a simulation study