Guides · Process data

Decimal commas, units and dates: transcription errors in GMP data

Many transcription errors come from separators, date order, units, lost zeros and wrong rows, often made by the spreadsheet rather than the typist. Agree conventions, keep the value as written beside the number, type text cells as text and check every transcribed value.

When values move from a paper batch record into a spreadsheet, many errors are not slips of the finger. They come from conventions: a decimal comma read as a thousands separator, a date read month first, a unit dropped, a trailing zero lost. Some are made by the person typing. Many are made by the spreadsheet, which decides what a typed or imported string means according to the computer’s regional settings, and nobody sees it happen.

The fix is mostly agreement and discipline: decide the conventions per form, keep the value exactly as written beside the number you derive from it, stop the spreadsheet from converting cells that must stay text, and have every transcribed value checked.

What does the regulation expect of a transcribed value?

The MHRA data integrity guidance (2018) says data should be attributable, legible, contemporaneous, original (or a true copy) and accurate (6.1). On transcription it is direct: “Where manual transcriptions occur, these should be verified by a second person or validated system” (6.11.2). The same section says a complete copy should keep the meaning of the data, and names date formats as an example.

PIC/S PI 041-1 section 9.7 expects data to be entered in a format the software controls, with invalid formats refused, and every manual entry of critical data verified by a second operator or by validated computerised means. Section 9.10, on hybrid systems, asks inspectors to look for secondary checks wherever data is transcribed between paper and computer, and for the original data to be kept afterwards. In the US, 21 CFR 211.68(b) asks that input to and output from a computer system be checked for accuracy.

The source side matters too. EU GMP Chapter 4 asks for handwritten entries that are clear, legible and indelible (4.7), and for any alteration to leave the original readable (4.9). A transcriber copying a corrected entry has to decide which value is the valid one.

Which errors are typical?

The examples below are fictional.

Error Written on the record What can end up in the sheet
Decimal comma read on a point system 7,2 (pH) 72, or text that no chart plots
Thousands separator mistaken for a decimal 1,250 (g) 1.25, a thousand times too small
Spreadsheet turns a number into a date step 3-2 or 12/2 2 March and 2 December on a month-first computer; 3 February and 12 February on a day-first one
Trailing zero dropped step 3.10, weight 50.20 3.1, 50.2
Leading zero dropped lot suffix 007 7
Day and month swapped 03/04/2026 4 March read as 3 April, or the reverse
Time past midnight start 23:40, end 00:15 a negative duration
12-hour clock 2:30 with no am or pm 02:30 or 14:30
Unit lost or converted wrongly 1.5 kg in a column of grams 1.5 among values near 1500
Copied to the wrong row value of row 6 in row 7 plausible numbers, all shifted by one

Several of these are made by the software, not the person. Microsoft’s own support page says Excel is preprogrammed to make dates easy to enter, so 12/2 becomes 2-Dec, and that there is no way to turn this off; the workaround is to format cells as Text first or type an apostrophe. Excel also removes leading zeros and keeps at most 15 significant digits. It takes its decimal and thousands separators from the computer’s regional settings, so the same file can be read differently on a colleague’s computer in another country.

A trailing zero is not cosmetic. 50.20 g says the balance read to two decimals; 50.2 says one. If you round or compare against a limit, the number of decimals written is part of the value.

Why do they matter for CPV and investigations?

A transcription error in a trend does one of two things. It creates a signal that is not there: a value ten times too large on a control chart starts a review of a process that did nothing wrong. Or it hides one that is: a value read inside its limits because a digit was dropped never reaches anyone. The second kind is worse, because nothing about it looks odd.

The FDA process validation guidance expects stage 3 data to be collected and evaluated to judge process stability and capability, with trending that guards against both overreacting to single events and missing unintended variability. Neither is possible on data that does not match the record. In an investigation, swapped dates break the sequence of events and a midnight crossing can turn a hold time into nonsense. Every conclusion then has to be checked again against the paper.

How do you prevent them?

Agree the conventions per form, and write them down.

  • The decimal separator the form’s entries use, and that thousands separators are not used in the dataset.
  • The date order for numeric dates, and the date format in the dataset (year first, such as 2026-04-03, cannot be read two ways).
  • 24-hour times, and how a time after midnight carries the next date.
  • One unit per column, named in the header, with the conversion rule when the record allows several units.

Keep two columns for every value. One holds the value exactly as written, as text: 7,2, 50.20, 1,250. The other holds the number you derived from it. Anyone can check the derivation, and a converted cell can never overwrite the original.

Stop the spreadsheet deciding. Format identifier and as-written columns as Text before typing. When importing a CSV, use the Text Import Wizard or Power Query rather than a double-click: set each column’s type, and tell it which decimal and thousands separators the file uses. Lock the template so that types cannot be changed by accident.

Control the entry format. Data validation on number columns (decimal only, within a plausible range for the unit) is a simple form of the software-controlled format PIC/S describes.

Verify every transcribed value. A second person compares each entry with the record, or a validated system does. PIC/S section 8.8 asks inspectors to check that secondary reviews verify any calculations used and, where possible, to view the original data to confirm it was transcribed correctly. How much is checked, and by whom, is for your quality system to decide on risk, but every critical value, checked against the page, is the plain reading of the guidance above.

How do you find errors already in a dataset?

Run these checks before the data is used, and send each hit back to the page rather than correcting it from the spreadsheet:

  • values outside their printed limits (an apparent deviation is often a misread or a mistyped value; see NOR, PAR and unlabelled ranges);
  • values impossible for their unit (a pH of 72, a negative weight, a minute past 59);
  • values close to ten or a thousand times the column’s usual level;
  • fewer or more decimals than the column normally carries;
  • text in a number column, and dates in a column that should hold numbers;
  • dates outside the batch’s manufacturing window, and times out of order within a log;
  • two neighbouring rows with the same value, a sign of a row shift.

Range checks have a limit worth stating: a wrong value that still falls within its limits passes every one of them. Only a comparison with the record catches it, which is why the second-person check stays.

For how to organise that check, see verifying transcribed data; for the wider path from record to trend, see from batch records to CPV data.

Sources

  1. MHRA, 'GXP' Data Integrity Guidance and Definitions, Revision 1 (March 2018), 6.4 and 6.11, accessed 18 November 2026.
  2. PIC/S PI 041-1 Good Practices for Data Management and Integrity in Regulated GMP/GDP Environments (8.8, 9.7, 9.10), accessed 18 November 2026.
  3. EudraLex Volume 4, Chapter 4: Documentation (4.7, 4.9), accessed 18 November 2026.
  4. 21 CFR 211.68 Automatic, mechanical, and electronic equipment, accessed 18 November 2026.
  5. FDA, Process Validation: General Principles and Practices (2011), accessed 18 November 2026.
  6. Microsoft Support, Stop automatically changing numbers to dates, accessed 18 November 2026.
  7. Microsoft Support, Keeping leading zeros and large numbers, accessed 18 November 2026.
  8. Microsoft Support, Change the character used to separate thousands or decimals, accessed 18 November 2026.
  9. Microsoft Support, Text Import Wizard, accessed 18 November 2026.

General information, not regulatory advice. Check the current text of each regulation; your quality system decides how it applies to you.

Related guides