Avoiding immortal time bias in observational studies