Screen Size vs. Interface vs. Double-Entry: Combinations That Yield Low Error Rates When Entering Dates on Touchscreen Devices (303374)

Temitope Adeyoju, eHealth Africa 
Thomas Albani, Biostat Global Consulting 
Andrea Guillot, eHealth Africa 
Nikhil Patil, eHealth Africa 
Mary Prier, Biostat Global Consulting 
*Dale Rhoda, Biostat Global Consulting 
Mary Kay Trimner, Biostat Global Consulting 

Keywords: data entry errors, information bias, touchscreen, accuracy

Double-data entry with cross-comparison is an effective method of minimizing data entry errors when keying survey responses recorded on paper forms. When survey data are recorded on digital devices rather than paper, it may often not be practical to have two persons independently and simultaneously enter data, for example, during household interviews. A single-device double-entry with confirmation, however, might be a more practical approach---especially for data fields where errors are likely and accuracy is extremely important.

Questions that ask about dates (month & day & year) are important components of some surveys, e.g., when assessing whether children were vaccinated at the right age. But data entry error rates tend to be higher for dates than for single-stroke responses because it takes more than a single click, tap or keystroke to enter a date.

This talk describes a full factorial experiment that compares date entry error rates for a) two Android touchscreen devices: smartphone & tablet, b) three common date entry interfaces: radio button, graphical calendar & scrollable dials, and c) two confirmation conditions: single-entry (no confirmation) vs. double-entry repeated until the two entries agree---similar to common password confirmation checks. We hired 24 interviewers from a commercial survey contractor in Nigeria to enter a set of vaccination dates from 30 child health cards. The cards contained an average of ten dates each. Each experimenter entered 50 dates in each of 12 factorial combinations of factors. The order of factor combination and of individual child health cards was balanced across participants to minimize bias from order effects or learning effects and to provide adequate power to detect a moderate difference between error rates in confirmed and unconfirmed conditions. This talk will summarize error rates across the 12 combinations and identify those that yielded the lowest rates.