Analyse survey data — from raw export to finished report

Every export has one row per respondent and one column per variable. A question, however, rarely matches exactly one column. A matrix question with six items becomes six variables. A multiple-choice question with eight options becomes eight 0/1 columns, one per option.

That is why analysis starts with the codebook. It records which variable belongs to which question, what the values mean and which codes stand for "no answer". Without this record, code 99 ("don't know") eventually ends up in a calculation as if it were a rating.

• One row per respondent, one column per variable
• Multiple response as 0/1 columns, not a single column
• Codebook: variables, values, labels and missing codes
• Import from Excel and CSV or SPSS including labels

Preparation comes before the first number. Define missing values properly, because a "don't know" is not a 3 on the scale. Collapse ages into groups, reverse negatively worded items and weight a sample that is not representative.

Repeatability matters. A second wave needs the same preparation. Manual spreadsheet steps are often neither documented nor reproducible. In DataLion, capture them as recodes and in the codebook script. The script can be applied to every later wave.

• Define missing values instead of computing with them
• Recode: build groups, reverse items
• Weighting for representative results
• Preparation as a script — reusable on every wave

First, inspect the frequency distribution for each question: how many answered what, in absolute numbers and percentages? Check the base. Does it include all respondents or only those who saw the question? Filters can change the result substantially. A percentage without its base size is worthless.

Next comes the crosstab. It breaks a question down by an attribute, such as satisfaction by age group, recommendation by region or purchase intent by customer segment. Defensible findings usually come from this comparison, not from an overall figure.

A significance test inside the table checks whether a difference between two groups is statistically reliable. Confidence levels range from 80 to 99 %.

• Frequencies in absolute numbers and percentages, always with the base
• Choose the base carefully: all respondents or only those filtered in
• Crosstabs by audience, region, wave
• Significance at 80–99 % directly in the table

Scale questions ("How satisfied are you …?", 1 to 5) are common and easy to analyse badly. The usual reflex is to report the mean. A Likert scale is ordinal, so the distance between steps is not guaranteed to be equal. A mean also hides the distribution. A value of 3.0 can represent a group of undecided respondents or two opposing groups.

The top-2 box shows the share of the two agreeing steps. The bottom-2 box shows the opposite. Both are easy to communicate. In DataLion, define them once as a top/bottom box code and use them in every table and chart.

For many items side by side, use a polarity chart. Agreement and disagreement run from a shared centre in opposite directions. This makes differences between items visible.

• Report top-2 box and bottom-2 box instead of a mean
• Show the distribution, not just the central tendency
• Polarity chart for many items side by side
• Define scales once, reuse them everywhere

Closed questions tell you where the problem is. Open-ended answers explain why. Manual coding takes time: with 800 verbatims, somebody reads for a day and assigns categories that can be difficult to reproduce later.

AI sentiment and topic analysis condenses verbatims into topics, assigns a sentiment and writes both back as new variables. Open-ended answers can then be broken down by segment in a crosstab or tracked across waves.

• Topics and sentiment coded automatically instead of by hand
• Results as variables — crosstabbable like any question
• Topic frequency over time as an early indicator
• Configurable AI endpoint, self-hostable in the EU

The final step is where most of the time disappears: moving results into a report. Charts get copied one by one, reformatted, and at the next wave it all starts over.

Build the analysis once as a dashboard. Stakeholders can filter it by segment themselves. For presentations, export the same analysis natively to PowerPoint as real, editable slides. Set up recurring studies as a tracker. Each new wave uses the same structure.

• Interactive dashboard instead of a static deck
• Native PowerPoint export — editable slides
• Tracker: new wave, same analysis
• Excel, PDF and SPSS as export formats too

Excel works well for small surveys. A one-off survey with 100 responses and 15 questions is faster to analyse with pivot tables than it is to set up a tool. We would not recommend anything else.

The limit is repetition. When the same analysis is produced per segment, wave or location, the workload grows each time.

• Excel: fine for one-off, small surveys
• Awkward for multiple response, weighting and significance
• Tipping point: recurring waves and many segments