Data quality dimensions for survey data
Data quality dimensions are the headings used to say what good data means. A common list of six, set out by DAMA UK, is accuracy, completeness, consistency, timeliness, validity and uniqueness. In a field survey each one becomes a check you can run on every batch of interviews, and surveys need a seventh, integrity: whether the interview took place as recorded.
By Mehrab Ali, founder of DataFlow · Updated 8 November 2026
Search for data quality dimensions and most of what you find is written for data warehouses. The ideas carry over to a household survey, but the record is an interview and the truth sits in a village you can go back to. This page takes each dimension in turn: what it means in a field survey, how to measure it during fieldwork, and which check covers it.
Which dimensions, and whose list
There is no single official list. The UK government's Data Quality Framework, published in 2020, uses six core dimensions as defined by DAMA UK, the Data Management Association of the UK: completeness, uniqueness, consistency, timeliness, validity and accuracy. ISO/IEC 25012, the international standard for a data quality model, has fifteen characteristics, considered from two points of view: inherent in the data, and dependent on the system that holds it. Other frameworks rename or add to these.
The six describe the values in a record. In fieldwork there is an earlier question: did the record come from a real interview? A form filled in at a tea stall can be complete, valid, consistent and on time, and describe a household nobody visited. So surveys need a seventh dimension, integrity: the interview took place, with the right respondent, when and where the form says.
The dimensions at a glance
Uniqueness comes first because a repeated ID breaks every other measure.
| Dimension | In a household survey | How to measure it | Checks |
|---|---|---|---|
| Uniqueness | Each household is interviewed once, and each form arrives once | IDs that appear more than once; repeated phone numbers | Duplicates |
| Completeness | Every sampled household has an interview, and every interview has the answers it should | Interviews against the target by area; blanks in required questions; “don't know” share | Missing answers, don't know and refusals |
| Validity | Answers fit the question: the type, the allowed range, the choice list | Answers outside the agreed limits; “other” answers that belong to a listed option | Range checks, “other, specify” |
| Consistency | Answers in one interview, one roster or two visits agree with each other | Breaches of each logic rule and of the skip logic, per 100 interviews | Logic checks |
| Accuracy | The answer recorded is the true one | Back-check agreement on questions that shouldn't change; outliers; GPS accuracy | Back-checks, back-check sample size, outliers |
| Timeliness | Data arrives in time to act on, from the current form | Days from interview to upload; interviews on an old form version; days to answer a flag | High-frequency checks |
| Integrity | The interview took place, with the right person, where and when the form says | Length, timing, GPS and answer patterns per enumerator, confirmed by back-checks | Fabricated interviews, interview duration, GPS checks |
Uniqueness
Each household or respondent appears once, and so does each submitted form. Uniqueness fails when a mistyped ID is shared by two households, one household is interviewed twice, a form is uploaded twice after a broken connection, or an interview is copied from another. Count the IDs that appear more than once, across the whole dataset with every batch, and check phone numbers too: a household interviewed under two IDs passes an ID check. Start here because a merge with the sample list, a back-check sample or a link to last year's round goes wrong when one ID points at two interviews. See duplicate survey responses.
Completeness
Completeness works at two levels. Every sampled household has an interview, or a recorded reason why not. And every question that should have been asked has an answer. Many blanks are correct, since a question skipped by its relevance condition should be empty. A blank is a problem when the question is required and the conditions for asking it hold.
Track completed interviews against the target by area and enumerator, daily, so a village falling behind shows in time. Count blanks in questions every respondent is asked, and the share of blanks in each question for each enumerator. “Don't know” and “refused” belong here too: they fill the cell without filling the gap. See missing answers and don't know and refusal rates.
Validity
A valid answer has a form the question allows: a number in its range, a code from the choice list, a date inside fieldwork. A household of 0 people or an age of 150 fails. XLSForm constraints block most of this at the door, but interviews on an older form version, values edited on the server and data entered from paper skip them, so check the data too. Count answers outside the agreed limits, question by question, and read the “other, specify” answers: text that names an option already on the list should be recoded to it. See range checks and “other, specify” answers.
A valid answer can still be wrong. An age of 34 recorded for a respondent who is 43 passes every validity check.
Consistency
Consistent answers don't contradict each other: no more children in school than people in the household, no 15-year-old with 30 years of farming, no mother younger than her child in the roster, no roof that changes between the interview and the back-check. Skip logic is a consistency rule too. A question answered when it shouldn't have been asked, or blank when it should have been, breaks it. Write each rule as a condition that describes something wrong, run it on every interview, and count breaches per 100 interviews by rule and by enumerator. When a rule flags many interviews, check the rule before you blame the team. See logic and consistency checks.
Accuracy
An accurate answer matches the truth. The data alone can't show the truth, so there are two ways in. Indirectly, look for answers that are possible but unlikely. Outlier checks flag values far from the others, best compared within the same area: an income many times the next highest, or a plot given in square metres when the question asked for acres. A GPS point carries its own accuracy in metres. These flags find candidates, and many will turn out to be true.
Directly, back-check: a separate team revisits or calls a random sample of respondents and asks a few questions again. On questions that shouldn't change, such as roof material, the share of answers that disagree, per enumerator, is the closest thing a survey has to an accuracy rate. See back-checks, how many interviews to back-check and outliers.
Timeliness
The UK framework describes timeliness as how well data reflects the period it represents. Fieldwork adds a practical side: a problem found after the enumerators have left the village costs a day of travel to fix. Measure the days from the end of an interview to its upload; interviews that end after they were sent, which points to a wrong clock on the device; interviews on an old form version after a new one went out; completed interviews against the schedule; and the days from a flag to its answer. Checks run once a week leave every problem a week old, so running the high-frequency checks daily is part of timeliness too.
Integrity: was the interview real?
Integrity, or authenticity, asks whether the interview took place as recorded: with the sampled respondent, at the place and time on the form, with the questions asked as written. Made-up interviews (sometimes called curbstoning) and shortcuts, such as answering “no” so that a section is skipped, pass validity and consistency checks.
No single flag proves a problem, so measure integrity per enumerator, against colleagues in the same area: interview length against the median, interviews late at night or overlapping the previous one, GPS points that need impossible travel or cluster at one spot, near-copies among their own interviews, numbers that vary less than the team's, and hardly any “don't know” answers. An enumerator who stands apart on several of these at once is the one to back-check first. See signs of fabricated interviews, interview duration, GPS checks and enumerator monitoring.
From dimensions to a data quality assessment
A data quality framework for a survey is a short written list: for each dimension, the measures, the threshold that triggers action, how often you look and who acts. A data quality assessment applies that list to the data and records what was found and done.
- Choose one or two measures per dimension from the table above, as counts or rates per 100 interviews, so days, areas and enumerators compare fairly.
- Set thresholds before fieldwork, agreed with the client, in the data quality assurance plan. Test them on the pilot data: a check that flags most pilot interviews has the wrong threshold.
- Run the checks daily on the whole dataset, while the enumerators can still go back.
- Read results by enumerator and area, comparing colleagues who work in the same area.
- Close every flag as confirmed, corrected or set aside, each with a reason, and leave the raw data as it arrived.
- Report weekly by dimension: interviews against the target, flags still open by check and enumerator, and back-check agreement.
To start from written data quality standards instead, the Field Data Quality Standard is an openly licensed checklist (CC BY 4.0) with default thresholds, in sections that follow the dimensions closely, from IDs and duplicates to back-checks. Cite its item numbers in your plan. Personal data isn't one of the six dimensions, but an assessment should cover it; see de-identifying survey data. SurveyLoopr's survey data quality checklist (SurveyLoopr also comes from ARCED International, which makes DataFlow) covers the same ground for ODK fieldwork, from form constraints to reconciling counts.
How DataFlow measures each dimension
DataFlow connects to SurveyLoopr, SurveyCTO, ODK Central and KoboToolbox, or reads CSV and Excel exports. On import it adds checks from the form and the data, and it can sync and run them on a timetable you set.
- Uniqueness: a duplicate check on each question named like an ID, which you can add to others, such as a phone number. A submission it already holds is updated in place on the next sync.
- Completeness: required questions that are always asked, skip logic, missing GPS points, interviews where over 20% of answers are “don't know” or refused, and progress against the target or a list of who to interview.
- Validity: ranges from the form's constraints, ranges suggested from the answers where the form sets none (off until someone checks them), other constraints run as rules, and “other, specify” answers to recode.
- Consistency: the form's skip logic, your own XLSForm-style rules, and impossible combinations picked from a cross-tab.
- Accuracy: outliers beyond 3 interquartile ranges, or a set number of standard deviations; GPS points less accurate than the form asks for, or 10 m; and back-checks, flagging an enumerator when 10% or more of type 1 answers differ, once at least 10 are compared.
- Timeliness: interviews sent over 7 days after they ended, or ending over an hour after they were sent, and old form versions used more than a day after the newest one's first interview.
- Integrity: interviews under 40% of the median length or 5 minutes, whichever is longer, or over 3 times the median; starts between 21:00 and 06:00; overlapping interviews by one enumerator; travel faster than 60 km/h; three or more of one enumerator's interviews within 15 m; and six statistical signs per enumerator against others in the same area. It picks interviews to back-check: a random share of each enumerator's, plus those the signs point to.
A suggested check that would flag more than 5% of interviews starts switched off, unless most of its flags come from one enumerator. Every correction needs a reason and can be undone, the raw data is never modified, and each entry in the activity log carries a hash of the one before. The Field Report shares the results as one HTML file, at a public, management or team level.
Questions
What are the six dimensions of data quality?
A common list, set out by DAMA UK and used in the UK government's Data Quality Framework, names accuracy, completeness, consistency, timeliness, validity and uniqueness. Other frameworks differ; ISO/IEC 25012 has fifteen characteristics. For survey data, add integrity: whether the interview really took place.
What is a data quality framework for a survey?
A short written list that says, for each dimension, which measures the team will compute, what threshold triggers action, how often the checks run and who acts on them. It belongs in the survey's data quality assurance plan, agreed before fieldwork.
How do you carry out a data quality assessment of survey data?
Run the agreed checks on every new batch of interviews, read the results by enumerator and area, close each flag as confirmed, corrected or set aside with a reason, and back-check a random sample of respondents. When fieldwork ends, run every check once more on the full dataset and report by dimension.
What is the difference between validity and accuracy?
Validity asks whether an answer is allowed: the right type, inside the range, on the choice list. Accuracy asks whether it is true. An age of 34 recorded for someone who is 43 is valid and wrong, so accuracy needs back-checks.
Why add integrity to the usual six dimensions?
The usual dimensions describe the values in a record. In fieldwork, an interview can be invented or cut short and still be complete, valid and consistent, so one record on its own won't show it. Integrity is measured per enumerator, from interview length, timing, GPS and answer patterns, and confirmed by back-checks.