News
Typed twice: the measured error rate of re-entering a fact
November 15, 2026 · PenDraco
When clinic staff retyped a lab value the instrument had already sent, 3.7% of 6,930 entries did not match. A second study found 2.8% per field and an error in 24% of records. The count that matters is how often a fact is typed twice.
Staff at 60 clinics retyped a glucose value that the meter had already sent to the health record. A 2019 study in the Journal of the American Medical Informatics Association compared the two copies. Of 6,930 hand-typed values, 260 did not match the instrument. That is 3.7%.
One number, typed twice
The authors are Mays and Mathias. The meter result reached the record through an interface. A billing documentation problem then changed the routine. The clinics chose to "manually re-enter the exact same POC glucose measurement into the EHR", the authors write. POC means point of care. EHR means electronic health record. So each result existed twice: one copy from the instrument and one from a keyboard.
The entries came from 506 clinical staff. Leave out entries with non-numeric characters, and 223 still held a wrong number. Of the 260 errors, 37 differed from the true value by more than 20%. Only 1 provider noted a discrepancy in a chart: a glucose of 13 where the true value was 132.
The authors draw one conclusion. The results "underline the importance of interfacing instruments when feasible," they write.
The rate holds outside one clinic
A second study checked a clinical database in Australia. Hong and colleagues published it in BMJ Open in 2013. They compared 5,148 hand-copied data points with an electronic import of the same reports. Four surgical residents who knew the terms did the typing. The error rate was 2.8% across all fields. It ran from 0.5% to 6.4% by field, and text fields did worse than numbers.
A review of research data entry gives the same band. "Error rates typically range from 0.55% to 3.6%", Barchard and colleagues wrote in Behavior Research Methods in 2020. Raymond Panko of the University of Hawaii found the pattern in spreadsheets. There, "errors seem to occur in a few percent of all cells", he wrote. Of 113 spreadsheets audited in 7 studies since 1995, 88% contained errors.
The federal government reports a higher figure for one kind of entry. The National Taxpayer Advocate reported it to Congress for 2022. The IRS's "manual transcription error rate for paper returns" was 22 percent. The report does not define the unit. IRS staff spent about 2.4 million hours that year preparing and keying paper returns.
A small rate becomes a large one
A record has many fields, and each field is a chance. Panko gives the formula for at least one error in n steps: E = 1 - (1-e)^n. Here e is the error rate for one step.
Hong's team measured the effect. At 2.8% per field and a median of 12 fields, 24% of records held at least one error. Only 320 of 421 records matched in full. Our arithmetic from the formula gives 29% for 12 fields. The measured share is lower because "the errors appear not to occur independently across the fields", the authors write. Some records hold several errors.
Now apply the formula to an order with 10 fields that a person types into a second system. At 0.55% per field, the low end of the research range, 5% of orders carry an error. At 3.6%, the high end, 31% do. These two shares are our arithmetic, not a measurement. Each added system that takes the order by keyboard adds 10 more chances.
What the studies do not show
We found no public study that measures the hours a small company loses to re-entry. The Federal Reserve Banks' 2026 survey report lists technology use as a challenge for 29% of small employers. Its footnote defines that category, and re-entry between systems is not in it. The U.S. Chamber of Commerce reports that 41% of small businesses use 6 or more technology platforms. Its checklist counts a smartphone camera as one.
App counts do not show it either. Okta counted an average of 93 apps per customer in its 2024 report and 101 a year later. Those figures describe the customers of an identity vendor. The Census Bureau's 2023 Annual Business Survey covers 2020 to 2022. In it, 36.0% of reporting firms with fewer than 500 employees used specialized software.
The measurements above come from clinics and research labs, where a second copy existed to check against. The entry screen matters too. In a 2012 study in PLOS ONE, single-key entry into a data entry program erred in 0.370 of 1,000 fields. So the rate depends on the screen and the field as much as on the person.
Four steps before the year closes
The 1099 season reads back what staff typed all year. Two rules meet in January covers what changes in that season.
- List the facts that a person types more than once. Start with customer details and prices. Then check hours and payments.
- For each fact, count the fields and the systems that take it by keyboard.
- Remove the second entry where an error costs the most. Let one system hold the fact, and let the other read it.
- Where a second entry must stay, key it twice. Then compare the two. In the 2020 study, double entry gave an error-free data set 83% of the time. A visual check gave 35%.
If two of your systems take the same fact by keyboard, start there. The embedded sprint puts one workflow into production in 30 to 60 days. Get in touch.
Sources
- Mays and Mathias, Measuring the rate of manual transcription error in outpatient point-of-care testing, Journal of the American Medical Informatics Association 26(3), 2019: the 6,930 paired results, the 3.7% and the errors above 20%.
- Hong and colleagues, Error rates in a clinical data repository: lessons from the transition to electronic data transfer, BMJ Open, 2013: the 2.8% per field and the 24% of records.
- Barchard, Freeman, Ochoa and Stephens, Comparing the accuracy and speed of four data-checking methods, Behavior Research Methods 52, 2020: the 0.55% to 3.6% range and the 83% against 35%.
- Paulsen, Overgaard and Lauritsen, Quality of Data Entry Using Single Entry, Double Entry and Automated Forms Processing, PLOS ONE 7(4), 2012: the 0.370 per 1,000 fields.
- Panko, What We Know About Spreadsheet Errors, Journal of End User Computing 10(2), 1998, revised May 2008, archived copy of December 29, 2019: the formula and the 88% of 113 spreadsheets.
- National Taxpayer Advocate, 2022 Annual Report to Congress, Most Serious Problem 1: Processing Delays: the 22 percent and the 2.4 million hours.
- Federal Reserve Banks, 2026 Report on Employer Firms: Findings from the 2025 Small Business Credit Survey: the 29% and its footnote.
- U.S. Chamber of Commerce, Empowering Small Business: The Impact of Technology on U.S. Small Business, Fifth Edition, October 2026: the 41% and the checklist.
- Okta, Businesses at Work 2024 and Businesses at Work 2025: 93 and 101 apps per customer, Okta's own customers.
- U.S. Census Bureau, Annual Business Survey, table AB2200MCB04, released December 19, 2024: the 36.0% of reporting firms.
- The 29%, 5% and 31% in the third section are PenDraco's arithmetic from the published rates.