Clean and prepare your data

How to build a data dictionary that prevents data cleaning errors

Define every variable, its allowed values, and validation rules before you collect, so cleaning does not become your analysis.

Every hour spent cleaning data after collection is usually an hour that could have been prevented before collection. A data dictionary defines every variable, its allowed values, and its validation rules before the first participant is entered. The software then rejects many errors automatically, instead of leaving them for analysis day.

By Heidi Andersén, MD, PhD · Updated 18 July 2026 · 6 min read

The problemCleaning becomes the analysis

You open the spreadsheet to run your first test and find ages typed as words, a sex column with nine spellings, dates in three formats, and blank cells you cannot tell from real zeros. None of it was caught when it was entered, so all of it lands on you now. A data dictionary moves that work to the start, where it is cheap, and lets the data-entry tool enforce it.

What it isOne row per variable

A data dictionary (REDCap calls it the instrument; other tools call it a codebook or variable dictionary) is a table with one row for every variable you will collect. It is the specification your research database, REDCap and other clinical research databases, enforces at the point of entry. The columns that carry the work:

Variable name. Stable, machine-readable name (for example age_years).
Field type. Integer, text, date, radio button, checkbox, calculated field.
Label. The question shown to the person entering data.
Choices. Allowed coded values.
Validation. Data type and acceptable range.
Identifier. Whether the field contains personal identifiers.

A minimal dictionary (REDCap-style)

variable       type    label                choices / validation    identifier
record_id      text                         (auto)                  no
age_years      text    Age in years         integer, 18-120         no
sex            radio   Sex at birth         1, Male | 2, Female     no
consent_date   text    Date of consent      date_dmy                no
initials       text    Participant initials -                       yes

The age_years rule (integer, 18-120) refuses “forty” and “220” at entry. The coded sex choices refuse a tenth spelling. initials is flagged as an identifier, so you know to remove it before sharing.

ValidationRules that catch the error at entry

Set the type and range once and the tool refuses anything that fails: an integer field will not accept text, a date_dmy field will not accept 31/02/2026, a minimum and maximum fence out impossible values. Use coded choices instead of a free-text box wherever the answer is a category, and mark fields that must not be blank as required.

But validation reduces errors; it does not guarantee correctness. It cannot tell whether a value is true, only whether it is possible. A participant aged 52 can still be entered as 25, and validation will accept it. It catches impossible values, not incorrect ones. That distinction is easy to forget and worth keeping in front of you.

The same four entries, with and without a dictionary

                without a dictionary          with a dictionary
sex             Male, male, M, m, Female, F   1, Male | 2, Female
age “forty”     accepted as text              rejected
age 220         accepted                      rejected
date 31/02/2026 accepted                      rejected

ThenRead it into R or Python

Because every variable already has a defined type, R and Python import the data with far fewer surprises. Most of the cleaning work has already happened during collection.

library(REDCapR)
d <- redcap_read(redcap_uri = uri, token = token)$data
# or from a CSV export:
d <- readr::read_csv("data.csv")

Python

import pandas as pd
d = pd.read_csv("data.csv")

Why this mattersMost analysis problems start at collection

Most analysis problems are data-collection problems discovered too late. A statistician cannot recover information that was never collected consistently. An hour spent on the data dictionary usually saves many hours of cleaning, recoding, and checking before analysis, and it protects results you could not otherwise defend.

ChecklistBefore you open recruitment

Five checks

Every variable has a name, a type, and a label.
Every category field uses coded choices, not free text.
Numbers and dates have a validation type and a plausible min and max.
Required fields are marked required.
Every field that could identify a person is flagged as an identifier.

Related guidesRead next

Want it built for you? DictVahti builds a REDCap-ready data dictionary with variable names, validation rules, coded values, and identifier flags. You review the structure before collecting data.