Survey Data Processing and Analytics: From Raw Responses to Reliable Answers

Why data processing deserves attention

Once fieldwork ends, raw survey data looks deceptively complete: thousands of rows, hundreds of columns, ready for charts. In reality, it contains duplicates, inattentive respondents, inconsistent answers, uncoded text, imbalanced samples and variables that need recoding before they mean anything.

Data processing is the disciplined work of turning that raw data into a clean, documented, analysis-ready dataset, then producing tables and analyses that answer the research questions. Mistakes here can silently change conclusions, so it deserves the same care as questionnaire design and fieldwork.

This guide explains each step of our data processing & analytics work, and what to ask of any supplier.

Step 1: Data checks and validation

Before analysis, every dataset is checked against the questionnaire and routing. Do the right respondents see the right questions? Are values within valid ranges? Do skip patterns hold? Are quotas met? These checks catch programming errors and data export problems early.

Validation continues throughout fieldwork, not only at the end. Checking interim data within the first day of fieldwork allows issues to be fixed while most of the sample is still to come.

Step 2: Cleaning and quality removal

Cleaning removes or corrects responses that would distort results. Cleaning rules should be agreed before fieldwork and applied consistently, with every removal recorded.

• Speeders: completes far faster than a realistic reading time
• Straight-liners: identical answers across grids
• Failed attention or trap questions
• Inconsistent answers between related questions or with profile data
• Nonsense, copied or irrelevant open-ended answers
• Duplicates detected by identity, device or pattern checks
• Implausible values, such as impossible quantities or ages

Removed completes are replaced during fieldwork where possible. Transparent suppliers report how many were removed and why. Our online data collection guide explains the quality checks that run during fieldwork.

Step 3: Coding open-ended answers

Open-ended questions capture what people say in their own words, which is often the richest part of a survey. To analyse them at scale, answers are coded into themes. A codeframe is built from an initial sample of responses, refined as coding progresses, and applied consistently, with multiple codes allowed where answers mention several ideas.

Quality checks include a second coder reviewing a sample for consistency, and nets that group related codes into higher-level themes. Text analytics tools can speed up coding of very large volumes, but human review remains important for accuracy and nuance, especially across languages.

Step 4: Recoding and derived variables

Many analysis variables do not exist directly in the questionnaire. Age is grouped into bands, scale questions are summarised into top-two box scores, NPS is calculated from 0–10 ratings, brand funnels are built from awareness and usage questions, and segments are assigned from combinations of answers.

Each derived variable should be defined clearly and checked, because errors here propagate through every table that uses it.

Step 5: Weighting

Weighting adjusts the influence of each respondent so the sample matches known population characteristics, or restores natural proportions after boosted samples. Common methods include cell weighting and rim weighting (raking) across several variables at once.

Weights should be modest and trimmed if extreme. The effective sample size after weighting should be reported, because heavy weighting reduces precision. Weighting can correct imbalances among people who were reached; it cannot correct for groups that were missed entirely. Our sampling guide explains how design and weighting work together.

Step 6: Tabulation and significance testing

Banner tables show every question broken down by key subgroups, such as region, age, gender, segment or customer type, side by side. They are the workhorse of survey analysis, letting researchers scan for patterns quickly.

Significance testing marks differences between columns that are unlikely to be due to chance. Tables should show unweighted and weighted bases, flag small bases and apply appropriate tests. Remember that statistical significance indicates a difference is probably real, not that it is large or important.

Deliverable formats

Datasets should always come with documentation: a codebook, variable definitions, cleaning rules, removals and weighting specifications, so results can be reproduced and reused.

Advanced analytics

These techniques answer questions simple tables cannot, but they must be planned at the design stage, because they depend on how questions are asked. Our research design cycle guide explains why the analysis plan comes first.

• Driver analysis: which factors most influence satisfaction, preference or loyalty
• Segmentation: grouping respondents by needs, attitudes or behaviour using cluster or latent class methods
• Conjoint and MaxDiff: preferences, trade-offs and willingness to pay
• Price sensitivity analysis
• TURF: the combination of products or messages that reaches the most people
• Regression and correlation analysis for relationships between variables
• Text analytics on large volumes of open-ended answers

Driver analysis in practice

Suppose a satisfaction survey asks customers to rate ten aspects of service and their overall satisfaction. Simply asking which aspects are “important” tends to rate everything as important. Driver analysis instead estimates which aspects are most strongly related to overall satisfaction, using techniques such as relative importance analysis.

Combining derived importance with current performance produces a priority map: aspects that matter a lot and perform poorly are the first to fix. This turns a long list of scores into a focused action plan.

Segmentation in practice

Segmentation divides a market into groups that share needs or behaviours, so products, messages and channels can be tailored. Good segmentations are distinct, sizeable, stable and actionable: you can describe each segment, estimate its size and reach it.

The analysis involves choosing input variables, testing several solutions, and selecting the one that is both statistically sound and makes business sense. A short typing tool then lets future surveys or sales teams assign people to segments using a few questions.

Data security and confidentiality

Survey datasets can contain personal information and commercially sensitive results. Processing should follow strict security practice: access limited to the project team, identifiable details separated from responses, secure storage and transfer, and deletion or anonymisation according to agreed retention periods.

Our Data Protection policy describes these practices, supported by our information security certification.

An example: rescuing a messy tracker dataset

A client brings three years of brand tracker data from different suppliers. Variable names differ, codeframes have changed, weighting was inconsistent and some waves contain obvious speeders.

The team builds a harmonised data map, recodes variables to consistent definitions, re-applies agreed cleaning rules, aligns codeframes with documented bridges, and re-weights every wave to the same targets. The result is a single clean dataset with a dashboard showing reliable trends, and documentation explaining every adjustment. Some apparent “changes” in earlier reports turn out to have been method differences, not market shifts.

Common data processing mistakes

Each of these undermines trust in results, and each is avoided with disciplined, documented processes.

• Deciding cleaning rules after seeing the results
• Not documenting removals and weighting
• Extreme weights that quietly reduce precision
• Codeframes that change without notice between waves
• Reporting small bases without warnings
• Treating significance as importance
• Datasets delivered without a codebook

Handling missing data

Some respondents skip questions or answer “don’t know”. How these are treated affects results. Tables should make clear whether percentages are based on all respondents or only those who answered, and “don’t know” answers should be shown rather than silently removed.

For advanced modelling, missing values may be handled by excluding incomplete cases, using imputation methods or modelling them explicitly. Each choice has implications, so it should be documented and, where results are sensitive to it, tested both ways.

Tracker data and consistency

For tracking studies, processing must be identical wave after wave: the same cleaning rules, codeframes, derived variables and weighting targets. Any change must be documented and, ideally, bridged by processing a wave both ways so trends remain interpretable.

Automated processing scripts help here. Once validated, they apply exactly the same steps each wave, reducing manual effort and the risk of inconsistent treatment.

Visualising results clearly

Good charts make results easy to understand: the right chart type for the comparison, sorted bars, clear labels, consistent colours for brands or segments, bases shown, and significant differences highlighted. Avoid three-dimensional effects, crowded legends and charts that require explanation to read.

Our reporting & insights team turns processed data into charts and stories that decision-makers can act on.

Quality assurance in processing

Processing outputs should be checked before delivery. Common checks include comparing table totals with raw counts, verifying bases against routing, spot-checking derived variables, confirming weights hit their targets, and having a second analyst review a sample of tables. Automated checks catch mechanical errors; human review catches results that look implausible and need investigation.

These checks take time, but they protect the credibility of everything built on the data.

Reproducibility and audit trails

Every number in a report should be traceable back to raw data through documented steps. Processing scripts, version-controlled specifications and change logs make results reproducible, so anyone can see how a figure was produced and rerun the analysis if questions arise months later. This is especially important for trackers, published research and studies that inform significant decisions.

Working with a data processing partner

Ask about cleaning rules, removal reporting, coding quality checks, weighting methods, table specifications, software used and documentation provided. Request a sample table set and codebook.

We provide data processing as a standalone service for agencies and in-house teams, as well as within full studies. Read our quantitative research guide, or contact our data team.

Frequently asked questions

• What is a codebook?

  A document listing every variable in the dataset with its question text, labels, codes and any derived definitions, so others can understand and reuse the data.

• Can you harmonise data from different waves or suppliers?

  Yes. We map variables, align codeframes and definitions, apply consistent cleaning and weighting, and document every adjustment.

• Do you offer text analytics for large volumes of comments?

  Yes. We combine automated text analysis with human review to code large volumes of open-ended responses accurately, including in multiple languages.

• What does survey data processing include?

  Validation, cleaning, removal of poor-quality responses, coding of open-ended answers, recoding, weighting, tabulation with significance testing and documentation.

• What software do you use?

  Industry-standard tools for tabulation and statistics such as SPSS, Q and R or Python, with outputs in Excel, SPSS, CSV and dashboards.

• How do you code open-ended answers?

  We build a codeframe from initial responses, apply it consistently with multiple codes where needed, and check a sample with a second coder.

• What is weighting?

  Adjusting each respondent’s influence so the sample matches known population characteristics or restores natural proportions after boosted samples.

• Can you process data from another supplier’s survey?

  Yes. We regularly clean, code, weight and tabulate data collected by clients or other suppliers.

• Do you provide significance testing?

  Yes. Tables include appropriate significance tests, with weighted and unweighted bases and small-base warnings.

• Can you build dashboards?

  Yes. We build interactive dashboards for trackers and large studies, with filters and documentation.

• How quickly can tables be delivered?

  Standard tables are usually delivered within a few working days of fieldwork closing, depending on complexity and coding volume.

• Is our data kept confidential?

  Yes. Access is restricted, identifiable data is separated, and storage and transfer are secure under confidentiality agreements.

Still curious? Ask the assistant Talk to a specialist →

Explore our Market Research services