Screen Size vs. Interface vs. Double-Entry: Combinations that Yield Low Error Rates When Entering Dates on Touchscreen Devices Dale Rhoda1 Andrea Guillot2, Temitope Adeyoju2, Mary Prier1, Nikhil Patil2, Mary Kay Trimner1, Thomas Albani1 1 Biostat Global Consulting, Worthington, Ohio, USA 2 eHealth Africa, Kano, Nigeria Three Commonsense Assertions • We want survey responses to be recorded accurately – with very low data entry error rates • In some surveys, dates are an important category of response • Error rates for dates are likely higher than rates for some other response types 2 Photo credit: Andrea Guillot Our Experiment 3 Vaccination Records • We did an experiment (the first of several) looking at screen size, interface and confirmation protocol • We made up vaccination records for 30 imaginary children and wrote them on vaccination record cards • Up to 19 dates per card (mean = 10) • All dates ordered D-M-Y & handwritten • We hired 24 data collectors to spend a day entering all 30 cards 2X each • 12 participants on each of two days • At eHA offices in Kano, Nigeria • We obtained ethical approval from the Kano State Ministry of Health 4 Devices • Tablet-- Asus Xplore iX101T1 • Screen dimensions • Length - 23.5 cm (9.25 inches) • Width - 14 cm (5.5 inches) • Phone-- HUAWEI Ascend Y320 U30 • Screen dimensions • Length - 9 cm (3.54 inches) • Width - 5.6 cm (2.20 inches) http://ruggedpcreview.com/images2/xplore_rangerx_angle_400.jpg http://compareindia.news18.com/media/gallery/images/2013/oct/y320_2_311331322688.jpg 5 Interfaces Open Data Kit (ODK) is a popular, free, mobile survey tool for Android devices that is commonly used for data collection in lowresource settings Our Date Interfaces included: • Calendar widget (ODK default) • Wheel widget (‘nocalendar’ option) • A homegrown radio button interface https://opendatakit.org/ 6 Wheel Interface • Default to today’s date • Spin 3 wheels to select month, day & year • Left swipe to complete entry • Constraint: order is M-D-Y instead of D-M-Y • Note: Month uses abbreviation instead of ## 7 Calendar Interface Tablet Phone • • • • Default to today’s date Spin wheels to select month & year Tap calendar grid to select day Left swipe to complete entry • Constraint: order is M-Y-D instead of D-M-Y • Note: Month uses abbreviation instead of ## • Narrow phone screen required left/right scrolling which is easily mistaken for left/right swiping 8 Radio-button Interface (not an ODK date widget) • No default date • Three consecutive screens • Select one and swipe left to proceed • Ordered D-M-Y (like card dates) • Month lists both ## and abbreviation • Phone screen required user to scroll upward to select day > 15 • No check for nonsensical dates (e.g. 31 February) 9 No Confirmation Double-entry Is there a date for X dose? Is there a date for X dose? No No Yes Enter date Yes Enter date again Enter date Up to 4X No Dates agree or 5X disagree? Yes Next dose Next dose 10 Hypotheses (a priori) • Unconfirmed error rates are likely to be > 1% • We hoped that confirmation would bring error rates down < 1% for all combinations of interface & device • Tablet will likely yield lower error rates than phones due to little or no need to scroll screen • No strong prediction about difference between interfaces 11 Balanced Design for Strong Inference • Design is meant to eliminate possibility of confounding factor-related differences in error rates (which we hope to see) with a learning effect or with fatigue • All participants • • • • Enter data from 5 cards at each of 12 stations (60 cards total per person) Enter all 30 cards using confirmation; enter all 30 cards unconfirmed Enter 30 using phone; 30 using tablet Enter 20 cards using each of the three interfaces • Each block of 5 cards holds 55 dates in total • Order of blocks of cards was randomized and balanced 12 Latin Squares This slide shows the order of factor combinations for 24 participants 13 Protocol for Experiment Day • Participants were oriented & gave consent • Filled out questionnaire re: demographics & technology use • Received individual packet of 60 vaccination cards in prescribed order • Visited 12 data entry stations in prescribed order (order customized for each participant according to design) • Could take as long as they needed at each station • Checked with experiment proctor before moving to next station 14 Results Photo credit: Andrea Guillot 15 Demographics & Technology Use • All 24 identified as right handed or ambidextrous • Use right hand to sign their name & right hand only or both to type on phone • Over 90% of participants between 18-34 years old • 75% female • Device ownership • 87.5% (n=21) own smartphone; 19 used it 7 days in prior week • 20.8% (n= 5) own tablets; all used it 5+ days in prior week 16 Analyze Entries that Should Have Dates • In this analysis we consider only entries that should have had dates: 10,830 entries • Respondents did sometimes lose their place on the card and enter dates where they shouldn’t have… and that is an interesting type of data entry error… but not our primary concern here 17 We Use Inchworm Plots to Represent 95% Confidence Intervals (CIs) Inchworm No Bar Bar with 95% CI 50 55 60 65 70 75 80 85 90 18 Mean error rate across all 12 combinations: 13.2% Error Rate by Person For the remainder of this presentation I exclude the five participants with the highest error rates – we could either fire them or train them to do better… 19 Mean after excluding bad faith and high error rates: 5.1% Error Rate by Person 20 Error Rates for 12 Combos = No Confirmation = Double-entry 21 Sobering Conclusions & Implications 1. Error rates vary considerably by participant Train & practice & test before hiring workers and sending to the field 2. People cheat Hire based on test results & make it clear that you will check their work 3. The default ODK interface yields the worst error rates here More work to recommend optimal solution, but radio buttons are promising 4. Even the best rates here are too high Use onboard camera to photograph original document for later verification Consider making accuracy a factor in compensation 22 Next Steps Additional Analyses More Experiments • Understand double-entry errors on wheel interface with tablets • How many errors could be caught with logic checks? • How many errors give misleading results in vaccination quality indicators? • Better training & practice with feedback • Cognitive & usability testing with double-entry • Try wheel with D-M-Y order and month in numbers • Lots more ideas… 23 Acknowledgements & Feedback We are grateful for partial funding from the Bill & Melinda Gates Foundation and for in-kind services from eHealth Africa, Kano, Nigeria Send suggestions or request for forthcoming manuscript to [email protected] 24 Backup Slides 25 Accurate Data Entry is Important • There is a rich history of research on data entry error rates, first by keyboard and more recently using mobile devices • Strategies to reduce errors include: • • • • Double entry with forced matching Visual checking Read-aloud checking Onboard or post-hoc logic checks for plausibility 26 Themes in Literature Search • Mobile technologies are popular for data collection • Skip patterns are well-executed • Data entry is complete as soon as field work is complete • Onboard GPS & camera are useful • Entering dates is more error prone than other field types, but not well explored with mobile devices • Open fields versus widgets • Higher rates of error when a default value present • Double data entry confirmation seems to help 27 Incomplete Responses & Bad Faith Data Entry Q1 Q2 Q3 Q4 Missing Bad Faith Entry (repeatedly entered default date or repeatedly entered numbers off-by-one) 28 Incomplete Responses, Bad Faith Entry & Very High Error Rates Q1 Q2 Missing Q3 Q4 High Error Rate Bad Faith Entry (repeatedly entered default date or repeatedly entered numbers off-by-one) 29 Number of Cards Data are Reasonably Balanced Even After Omissions TW Tablet Wheel TR Tablet Radio TC Tablet Calendar PW Phone Wheel PR Phone Radio PC Phone Calendar 30 No Evidence of Aggregate Learning Effects 31 Results if we exclude only the bad faith entries…(so include cards from 23 of 24 respondents) 32 Error Rates for 12 Combos = No Confirmation = Double-entry 33 Error Rates for 12 Combos = No Confirmation = Double-entry 34 Learning Effects Error rates dropped slightly from Q2-Q4 (This slide includes 23 of 24 respondents) 35 Future Experiments • Incorporate usability testing!!! • Reverse radio button interface to use M-D-Y order and see if it gets worse • Modify wheel and calendar widgets to show date in D-M-Y order • Revise double-entry instructions: goal is to enter the correct date twice (not just the same date twice) • Detect bad faith entries in real-time & do not pay for them • Experiment with read-aloud protocol • Same task with double-entry using a computer keyboard • Conduct orientation & training by video (for consistency) 36
© Copyright 2025 Paperzz