Immortal time bias for life-long conditions in retrospective observational studies using electronic health records