Data cleansing in survey research is the process of finding and fixing faulty, duplicate, or fake responses so your results reflect what real respondents think. Skip it, and a handful of bots or careless answers can quietly shift your averages, your segments, and your decisions.
Picture a satisfaction survey where 200 identical bot answers drag your score toward whatever they chose. Nobody sees the problem until a decision goes wrong.
In this guide, we’ll cover what to look for, when to check, a step-by-step process with real thresholds, and how to measure the result.
What is data cleansing in survey research?
Data cleansing in survey research is the process of detecting, correcting, or removing faulty responses so a dataset is accurate, consistent, and ready for analysis. It covers formatting errors, duplicates, logic conflicts, low-effort answers, and fake submissions.
Cleansing goes beyond basic data cleaning, which fixes obvious errors such as typos. It also asks whether answers are plausible and consistent across the whole dataset. For a side-by-side comparison, see our guide to data cleansing vs data cleaning.
Three nearby terms are easy to confuse with it:
- Validation checks entries against rules at the moment of entry.
- Scrubbing removes records that should not be in the dataset at all.
- Wrangling is the wider job of reshaping data for analysis, and cleansing is one step inside it.
Why clean survey data matters
Clean survey data matters because bad responses change results, not just add noise. Three effects show up again and again:
- Cost: Poor data quality costs organizations at least $12.9 million a year on average.
- Bias: Bogus respondents made up about 4% to 7% of opt-in samples. Removing them moved one national-mood question by two percentage points.
- Trust: Charts and dashboards inherit every error. One visibly wrong number makes stakeholders doubt the rest.
Small subgroups suffer most. When a segment has only 80 respondents, a few fake answers can flip its result.
Common data issues in survey responses
Most survey data problems fall into nine patterns. Each has a recognizable signal you can check for.
| Issue | What it looks like | How to catch it |
|---|---|---|
| Incomplete responses | Respondent quits partway, key questions blank | Completion-rate rule, required questions |
| Speeding | A 20-second finish on a 10-minute survey | Compare to median completion time |
| Straight-lining | The same answer down an entire grid | Pattern check on matrix questions |
| Duplicate submissions | Same IP address, identical text, repeat entries | Duplicate IP and duplicate text checks |
| Bots and fake respondents | Copy-pasted or machine-written text | Plagiarism checks, red herring questions |
| Contradictory answers | Age 17 with 10 years of work experience | Logic checks across related questions |
| Out-of-range values | Age entered as 150 | Numeric validation |
| Inconsistent labels or scales | “NY” and “New York,” or 1-5 mixed with 1-10 | Recoding rules |
| Low-quality open ends | One-word answers, random characters | Length and gibberish checks |
When to cleanse survey data: Before, during, and after fieldwork
Cleanse at three points: before fielding to prevent errors, during fieldwork to catch them early, and after collection to audit what remains. Prevention is always cheaper than repair.
Before fielding: Prevent bad data
Good survey design removes many problems before they exist:
- Set validation rules, such as a numeric range for age.
- Use skip logic so respondents only see questions that apply to them.
- Mark the questions your analysis depends on as required, and keep that list short.
- Add one or two attention or red herring questions.
- Test the survey with dummy data to confirm logic and quotas work.
During fieldwork: Catch problems early
Check early data after a soft launch of roughly 10% of the sample or 100 completes, whichever is smaller. Then check again at about 60% and 90% of data collection.
Catching 10 bad responses mid-field beats finding 200 at the end. Wait until after the soft launch to set speeder cutoffs, because thresholds set too early can screen out valid respondents.
After collection: Audit before analysis
Once fieldwork closes, run a full audit before anyone builds a chart:
- Review all flagged responses and their reasons.
- Re-check logic conflicts and duplicates across the full dataset.
- Read a sample of open-ended answers.
- Compare totals and key percentages with what you expected.
- Lock a cleaned version and keep the raw file untouched.
Step-by-step guide to cleaning survey data
Follow these seven steps in order, and work on a copy so the raw data stays untouched. Flag first and delete second.
- Back up and set rules.
Save the raw export and write your cleaning rules before you look at results. - Remove incomplete or invalid responses.
Apply a documented completion rule, such as dropping responses under 20% complete. - Find duplicates and suspected bots.
Combine IP, text, timing, and pattern checks. - Check logic.
Flag answers that contradict each other, such as a stated age that cannot match stated experience. - Handle missing values and outliers.
Apply a stated rule for each, as described below. - Normalize text and scales.
Merge spelling variants and recode mismatched rating scales. - Document and re-check.
Record what you removed and why, then re-run the checks on the final file.
Thresholds and rules of thumb for quality checks
No universal cutoffs exist, but practitioners share common starting points. Treat these as defaults and tune them for your survey length and audience.
| Check | Common starting rule |
|---|---|
| Speeders | Under one-third of median completion time; some teams use 25% to 40% |
| Straight-lining | Same answer on 5 or more rows of a grid |
| Soft launch size | 10% of sample or 100 completes, whichever is smaller |
| Open-ended checks | About one open end per five minutes of survey time |
| Combined flags | Remove at two or more failed checks, review single failures |
| Incomplete responses | Example rule: drop under 20% complete, keep partials that answered key questions |
How to handle missing data and outliers
Missing data and outliers need a rule, not a gut call. Decide the rule before you see how it affects your results.
Missing data: Delete, estimate, or flag
Pick one of three approaches per question:
- Delete the response when the respondent skipped most of the survey, or when the share missing is small.
- Estimate the value only with a documented method. Imputation means replacing a missing value with an estimate from other answers, such as the median.
- Keep as a category when not answering is itself informative, such as “prefer not to say” on income.
Also fix the cause. If one question is blank for many people, the wording or placement is probably the problem.
Outliers: Keep, cap, or remove
An outlier is a value far from the rest, such as a household income of $100 million in a consumer survey. It is not always an error. Check it against the respondent’s other answers first.
Remove it if the data is impossible. If it is plausible but extreme, cap it. Winsorizing means replacing values above a set percentile with that percentile’s value, for example the 90th, as Sawtooth describes. Report results both ways when the choice moves your conclusion.
How to clean open-ended survey responses
Open-ended responses need cleaning because they feed theme coding and text analysis. Open-ended questions are also where bots and low-effort answers show up most clearly.
Work through the text in this order:
- Remove gibberish, such as random keystrokes, and one-word replies to questions that need a sentence.
- Fix obvious typos and merge spelling variants, such as “NYC” and “New York City.”
- Group similar answers into one category.
- Code the remaining answers into themes. Coding means sorting answers into labeled categories.
- Keep the original verbatims in a separate column so nothing is lost.
Review a sample by hand even if software does the first pass. Genuine respondents sometimes write short or unusual answers that are still valid.
Survey data cleansing example: From 8,500 responses to 7,200
This illustrative case comes from a retail product satisfaction survey. The team collected more than 8,500 responses and needed reliable Net Promoter Score (NPS) results, a loyalty metric based on how likely customers are to recommend a product.
- Collected: About 8,500 responses.
- Incomplete: Roughly 12%, or about 1,000 responses, removed.
- Bot-like: More than 200 identical entries removed.
- Gibberish or conflicting: The remaining 80 or so removed.
- Scales: Mixed rating scales recoded to one standard.
- Result: About 7,200 usable responses, a retention rate near 85%.
The identical block of bot-like entries would have pulled the score toward whatever answer the bots chose. After cleansing, the NPS reflected real customers, and product insights became clearer.
Cleansing priorities by survey type
The same checks do not matter equally everywhere. Each survey type has a different failure pattern, so rank your checks accordingly.
- Customer satisfaction and NPS surveys: Watch for duplicate submissions and label drift across waves, such as one product appearing under 12 names. A repeated dashboard needs stable categories more than aggressive fraud screening.
- Employee surveys: Protect anonymity first. Cleaning should never expose who answered, and you should set a minimum group size before reporting any subgroup.
- Market research and panel studies: Prioritize bots, speeders, and straight-lining. Inattentive respondents distort advanced methods too, and Sawtooth notes that failing to remove them can exaggerate willingness-to-pay in conjoint studies, which test how people trade off product features.
- Academic and public-interest surveys: Document every rule, because reviewers and co-authors may ask for your cleaning log. Your response-rate reporting also depends on what you removed.
How sample source changes your cleansing needs
Where respondents come from changes how much fraud you should expect. A list of your own customers carries little bot risk. An open link shared on social media or a large paid incentive carries a lot.
Research shows the gap. After speeding and attention-check removals, opt-in samples still had 3% to 7% bogus respondents, versus 1% in address-recruited samples. Address-recruited means people were invited through a random sample of home addresses rather than by volunteering online.
| Sample source | Typical risk | Cleansing focus |
|---|---|---|
| Your own customer or employee list | Duplicates, outdated contacts | Scrub the list, check duplicate submissions |
| Online panel | Speeders, straight-lining, professional respondents | Timing, pattern, and trap checks |
| Open link or social media | Bots, repeat entries | Duplicate IP, plagiarism, and fraud checks |
| High-incentive studies | Fraud risk rises as the reward grows | Stricter screening and manual open-end review |
How to decide what to remove, flag, or keep
Remove a response only when the evidence is strong. Otherwise flag it and decide at analysis time. Single checks are weak evidence, for a clear reason.
Pew tested this in 2020. 84% of bogus respondents passed a trap question, 87% were not flagged as speeders, and 76% passed both checks. Checks help, but they miss most fake cases on their own. Speeding also flags real people who read fast.
So combine signals. Remove a response when it fails two or more independent checks, or when its data is impossible. Flag and keep it when it trips one soft signal. Keep incomplete responses if the questions you need are answered.
Finally, run your key results with and without the flagged responses. If the conclusion holds both ways, your cleaning rules did not drive the answer.
How to measure survey data quality after cleansing
Measure cleansing with a handful of numbers you set before fielding. Vague goals like “clean enough” cannot be checked, but these can:
- Retention rate: Responses kept divided by responses collected. The example above kept about 85%. Investigate any wave where the rate swings sharply.
- Flag rate: The share caught by each check. Compare it with Pew’s 4% to 7% range for bogus respondents in opt-in samples.
- Duplicate rate: Zero exact duplicates in the final file.
- Key-question completeness: 100% on the questions that drive your analysis.
- Label consistency: Exactly one label per product, region, or segment.
- Data quality score: If your platform scores responses, set a review cutoff before fielding rather than after.
Track these across waves. A sudden jump in any one is a signal to check your sample source or survey design.
How to check that your cleaned sample still represents your audience
Cleansing can quietly change who is in your data. If it removes far more young respondents than older ones, your results tilt toward the older group even though every remaining response is genuine.
After cleaning, compare your sample with a trusted benchmark. For US general-population surveys, the Census Bureau’s American Community Survey publishes annual demographic figures for age, income, and region. For customer surveys, compare against your own customer file instead.
If the gap is large, look at which rule removed the most people in the underrepresented group. Then consider weighting, which adjusts each respondent’s influence so the sample matches known population shares. Report any weighting you apply.
Tools for cleansing survey data
Match the tool to the size and frequency of the job. Small one-off projects need little. Recurring studies need repeatable steps.
- Spreadsheets: Excel or Google Sheets handle filters, duplicate removal, and conditional formatting for a few thousand rows.
- Python or R: Pandas (a Python data library) and R let you script every rule so each wave is cleaned the same way.
- Survey platform checks: Built-in quality filters catch problems at collection.
- ETL and orchestration: ETL (extract, transform, load) tools move and reshape data between systems. Data orchestration automates cleaning and routing of responses to CRMs and dashboards.
- Text analysis: NLP (natural language processing) tools group open-ended answers by theme and flag gibberish.
Whatever you choose, script or document the steps. A repeatable process beats a heroic cleanup.
Common data cleansing mistakes to avoid
Most cleansing errors come from process, not tools. These are the ones that do the most damage.
- Cleaning without a raw backup: A wrong rule is only fixable if the original survives.
- Deleting on a single flag: One failed check is a lead, not a verdict.
- Changing rules after seeing results: Cutting responses until the numbers look right is cherry-picking. Write the rules first.
- Over-cleaning open ends: Short or blunt answers can be valid. Removing them erases real views.
- Ignoring who gets removed: If cleansing drops one age group far more than others, your sample no longer represents your audience.
- Skipping documentation: If nobody can say which rule removed which response, nobody can defend the results.
Privacy and consent when cleaning survey data
Cleansing touches personal data, so privacy rules apply to the cleaning itself. The core rule is simple: only keep the personal details you need, for as long as your consent terms allow.
Remove or separate direct identifiers, such as names and email addresses, before sharing cleaned data with analysts who do not need them. Store the raw file securely and delete it when your retention period ends.
US privacy law also shapes the process. Under the California Consumer Privacy Act, consumers have the right to delete personal information and the right to correct inaccurate information a business holds about them. A respondent’s request to correct or delete their data should flow through your cleaning log, so you can show what changed. Check with your legal team for rules that apply to your audience.
How QuestionPro supports survey data cleansing
QuestionPro’s Data Quality module screens responses as they arrive, so cleansing starts at collection rather than after it. You switch it on under Analytics, then Manage Data, then Data Quality. It only screens responses collected after you enable it, so turn it on before launch.
The checks include duplicate IP addresses, speed traps, duplicate text, gibberish words, one-word answers, all checkboxes selected, patterned responses, plagiarism detection, and quality terminates for red herring questions. Here is how the workflow runs:
- Review: Each flagged response shows the question, the answer, the reason, and a Data Quality Score from 0 to 100. A higher score means poorer quality.
- Filter: Dashboards can show flagged or unflagged responses only, using a saved Data Quality Flag filter.
- Export: Flags export as short codes, so your cleaning log travels with the data.
- Override: You can unflag or re-flag any response manually.
That workflow supports the flag-first approach described above. You see the evidence, then decide. Learn more about the QuestionPro data quality tool and how it fits into our survey software.
Clean data starts before the first response arrives
Data cleansing works best as a habit built into the whole survey, not a rescue job at the end. Design to prevent bad data, check early while you can still adjust, and audit before you analyze.
The teams that trust their results do three things consistently. They write rules before seeing numbers, they flag before they delete, and they log every change. That discipline turns a messy export into findings you can defend in a meeting.
Frequently Asked Questions (FAQs)
Usually not. A single failed check is weak evidence, since fast readers can trip speed traps and honest respondents can miss trap questions. Flag the response, review it, and remove it when it fails two or more independent checks.
It varies by sample source and survey design. Research found bogus respondents made up roughly 4% to 7% of opt-in samples, before counting incomplete answers. Track your own removal rate per project and investigate any wave that jumps.
It is risky. Changing rules after seeing results lets the answer you want steer the cleaning. Write your rules before analysis, apply them uniformly, and if you must adjust, report results both ways and note the change.
Yes. Keep an untouched, versioned copy of the raw export, stored according to your retention policy and respondent consent terms. It lets you audit decisions, rebuild after a rule change, and answer questions about exactly what you removed.
Partly. Text analysis can flag gibberish, copied, and bot-written answers, as with QuestionPro’s plagiarism detection check. Treat flags as leads, not verdicts, because real respondents sometimes write short or unusual answers that a human reviewer should confirm.



