A clean survey dataset is not one with every blank filled. It is one whose structure, codes, missing states, corrections and exclusions can be explained and reproduced from the source instrument.
1. Confirm the unit, keys and file structure
Define what one row represents and verify the expected number of records. Check unique IDs, duplicate submissions, repeat groups and joins between household, person, facility or follow-up files.
In Excel, avoid merged cells and multiple header rows in the analysis table. In SPSS, verify variable types, widths, measurement levels, labels and user-missing definitions.
- One row has a defined meaning
- Identifiers are unique where required
- Repeat tables reconcile
- Dates and numbers have consistent types
- Raw source is preserved
2. Validate codes, ranges and skip logic
Compare every categorical value with the deployed choice list. Distinguish legitimate values from export labels, typos and legacy codes. Set range rules from the questionnaire and study context rather than convenient generic limits.
Use cross-field checks for age and age group, consent and interview completion, dates and study period, or parent and follow-up questions. A downstream answer after a skipped parent question needs review, not automatic acceptance.
3. Profile missingness without erasing meaning
Calculate missingness by variable and relevant operational group such as site, form version or collection period. Overall completeness can hide a systematic gap in one team or version.
Keep not asked, not applicable, refused, do not know and accidentally blank distinct when the instrument supports that distinction. Do not replace missing values with zero unless zero is genuinely the observed value.
4. Build derived variables from written definitions
Document reverse scoring, composite scores, categories, denominators and indicator logic. Preserve source variables and test boundary values so categories cannot overlap or leave unintended gaps.
Use SPSS syntax, Power Query, protected formulas or another repeatable transformation. Manual editing can be valid for reviewed corrections, but every change needs an audit trail.
5. Reconcile before analysis begins
Report the starting rows, excluded rows by reason and final analysis rows. Review outcome distributions, small groups, missing predictors and the assumptions of the planned method.
Package the approved data with a dictionary and cleaning log. If an assumption fails, revise the method or withhold the test rather than forcing output because it appeared in the proposal.
- Row counts reconcile
- Exclusions have reasons
- Outcome coding is verified
- Small cells and outliers are reviewed
- Approved version is named
See the Data Rescue and Results Pack
We can review your actual form, dataset and reporting requirement, then scope the smallest service that solves the risk.
Explore the service ↗