Key considerations for choosing a statistical method to deal with incomplete treatment adherence in pragmatic trials