A survey export contains two submissions with the same identifier, a negative task duration and several blank ratings. Those records need attention, but they do not all call for deletion. One may be a duplicate, another a tracking fault and the blanks a legitimate consequence of routing.
Data cleaning identifies and handles errors, inconsistencies and other quality issues so that data can be analysed appropriately. It includes checking structure, values, duplicates and missingness, with decisions documented well enough to reproduce and assess.
Preserve the original and understand the fields
Keep an unchanged source copy and perform transformations in a separate, controlled process. Read the definitions before changing values: “not asked”, “declined” and “not applicable” may each appear as missing but represent different circumstances.
The UK Data Service’s QAMyData project illustrates systematic checks for issues such as missing-value definitions and inconsistent metadata. Automated checks can identify candidates for review, but the study context determines what constitutes an error.
Verify units and joins. A duration recorded in milliseconds but interpreted as seconds can look like extreme behaviour, while joining tables at the wrong level can duplicate legitimate participants many times.
Treat quality signals as evidence to assess
Fast completion, repeated ratings and short comments can indicate poor engagement, but none proves it on its own. A knowledgeable respondent may answer quickly, and identical ratings may honestly describe similar experiences. Review signals together using rules suited to the study and considerate of different abilities and response styles.
Set eligibility and exclusion criteria before fieldwork where possible. If a new issue requires a later decision, document the reason and whether the outcome was already visible. Avoid changing thresholds until a preferred result appears.
Investigate outliers separately from impossible values. A valid extreme experience may be the most important one for the product team to understand.
Handle missingness without inventing certainty
A blank rating is not a neutral rating, and a missing event is not necessarily non-use. Examine where and why information is absent. The appropriate handling may differ across descriptive reporting, modelling and particular survey designs.
Imputation requires assumptions and should reflect the uncertainty it introduces. Replacing every blank with an average can distort distributions and relationships, so it should not be treated as a routine repair that leaves the analysis unaffected.
Keep a log of corrections, exclusions and transformations, including affected counts. Report consequential decisions with the findings and use sensitivity checks when reasonable alternatives could change the conclusion. The cleaned dataset should make the evidence easier to trust, with a visible account of how it differs from the original.