Ethnicity discrepancies

Ethinicity is a key variable of my dataset and there are discrepancies across instances. 

Is there a particular instance which may be more accurate eg touchscreen responses are verified by research staff? 

Thank you!