Extending conditional likelihood in models for stratified binary data