---
url: "https://questionstar.com/blog/analyzing-online-survey-results/"
title: "Analyzing online surveys: from research goal to result — QUESTIONSTAR"
description: "Analysis starts before you collect data, not after. Seven steps from research goal to a result you can rely on — from the decision question to hypothesis testing."
lastmod: "2026-07-08"
type: "blog"
---

# Analyzing online surveys — it starts before you collect data

*Which decision should your survey inform — and what steps get you to a methodologically sound result?*

The most common analysis question is: "The data's here — now what?" The honest answer: by then it's already late. The most important decisions about your analysis aren't made after data collection, but before it — when you think through which practical question your survey is actually meant to answer. Seven steps from research goal to a result you can rely on.

The most common question about analysis is: "The data's here — what do I do with it now?" The honest answer: by then it's already late. You made the most important decision about your analysis long ago — consciously or not — when you thought about which questions to ask in the first place.

Go into a survey without a clear decision question, and you'll collect answers you can't do anything with afterward. No statistical method rescues a dataset that was never collected for a specific question. Scale level, measures of central tendency, visualizations — that's all craft. But you determine the material earlier, when you set up the survey.

This article walks you through seven steps: from the research goal (step 1) through the hands-on descriptive analysis (steps 2 to 6) to hypothesis testing (step 7). The first step is the most important one — and if you've done it before you collect data, the others fall into place almost on their own.

## Step 1: Know which decision the data should inform

Before you word the first question, you need to answer three questions for yourself. Which practical decision do you want the data to support? Which answers would be actionable — would lead to a concrete measure? Which would merely be "nice to know" — interesting, but with no consequence?

Actionable means: when you see the result, you can say "so we'll do A" or "so we'll do B." Nice-to-know means: when you see the result, you say "huh, interesting" — and nothing changes. Both have their place, but you should know which kind you're collecting right now. Fill the whole questionnaire with nice-to-know items, and you're collecting expensive data that fizzles out in the report.

> A concrete example: an employee survey on satisfaction.**Approach A ("this won't work"):** "We want to know how satisfied our employees are." From that follow 30 items on general satisfaction, mood, collegiality, leadership. In the end you have a mean of 3.7 on a five-point scale. And now? Nobody knows what to do.**Approach B ("this works"):** "We want to decide whether to expand the onboarding program for new hires or leave it unchanged."From that follow targeted items on satisfaction during the first six months, on specific onboarding aspects (buddy system, technical training, feedback conversations), plus a demographic variable "start date" as a filter. The analysis goal is clear from the outset: median per onboarding aspect within the subgroup of employees with less than six months of tenure.In the end you have values from which a decision can be derived directly: which onboarding aspect gets expanded, which stays.

Both surveys cost roughly the same time and money. The second produces data that supports a decision. The first produces data that supports a report — and then sits in a folder.

A practical consequence for your preparation: write down the decision you want to support in a single sentence. If you can't, the decision isn't yet clearly formulated — and the survey will produce unclear data. Only once that sentence is in place do you derive the variables: which do you need in terms of content, which as a filter, which as a demographic classification, at which scale level, with which intended analysis operation. This groundwork takes half an hour. It saves you days later.

If you've done this step cleanly, the next six are no more than the hands-on implementation of what you've already decided. If you've skipped it, neither the right scale level nor the finest visualization will help. Then you're collecting numbers, but no answers.

## Step 2: Determine the scale level

Before you compute anything, identify the scale level for each variable. It determines which measures of central tendency, measures of dispersion and visualizations are permissible. Three levels are relevant in surveys.

**Nominal scale.** Data with no natural order. Examples: gender, state, industry, favorite brand. Here only the mode is a meaningful measure of central tendency — mean and median make no sense.

**Ordinal scale.** Data with a natural order, but without quantifiable distances. Classic example: Likert scales ("does not apply at all" to "fully applies"). Mode and median are meaningful; the mean is methodologically contested, because the distances between the steps aren't guaranteed to be equally large.

**Metric scale.** Data with a natural order and quantifiable, equally large distances. Examples: age in years, income in euros, number of employees. All three measures of central tendency are applicable.

> 💡 **Why scale level comes first**
> 
> A common source of error is computing a mean over Likert data without checking the distribution. The result looks like a meaningful metric — "identification with the company: 3.7 out of 5" — but can be misleading if the distribution is bimodal, that is, many very satisfied and many very dissatisfied respondents, but hardly anyone in the middle.
> 
> The mean is then a methodological artifact, not a finding. Only the frequency distribution and the median show the true picture. Rule of thumb: never report a mean without looking at the accompanying distribution.

## Step 3: Compute frequencies

Frequencies answer the basic question: how do the answers distribute? Three variants need to be distinguished.

**Absolute frequency.** The raw count: how often each value occurs. In a survey on preferred communication channels with 200 participants, 80 might choose "email," 60 "phone," 50 "chat," 10 "letter." The absolute frequency is 80, 60, 50, 10.

**Relative frequency.** The absolute frequency divided by the total number of answers — usually expressed as a percentage. In the example: 40%, 30%, 25%, 5%. Relative frequencies are the standard form of presentation in reports, because they're comparable across samples of different sizes.

**Cumulative frequency.** The running sum of the relative frequencies. Useful for ordinal- or metric-scaled data. Example: in an age survey, 30% are under 30 years, another 35% between 30 and 45, another 25% between 45 and 60, the remaining 10% over 60. The cumulative values: 30%, 65%, 90%, 100%. From them you read off that two thirds of the sample are under 45 years of age.

For variables with very many categories — such as income or age in single years — grouping into classes is advisable. Instead of "38 years, 39 years, 40 years …" you look at age groups "25 — 34," "35 — 44," "45 — 54." The grouping should be defined before data collection, not adjusted afterward to produce a desired result.

## Step 4: Compute measures of central tendency

Measures of central tendency condense a distribution into a single number that describes the center. Three are relevant in practice.

### Arithmetic mean (average)

The sum of all values divided by their count. Applicable only to metric data. Highly susceptible to outliers: a single extreme value can distort the result considerably. Compute the mean over all participants in an income survey, and you get a value pulled upward by a few very high incomes — one that overstates where the typical participant sits.

### Median

The value that splits the size-sorted distribution into two equal halves: 50% of the data lie below it, 50% above. Applicable to ordinal and metric data. Robust against outliers — a single extreme value does not change the median. With a skewed distribution (income, waiting times, response durations) the median is almost always more informative than the mean.

### Mode

The most frequently occurring value. Applicable to all scale levels, the only meaningful measure of central tendency for nominal-scaled data. With multimodal distributions there can be several modes — in which case reporting a single "central value" is misleading anyway, and the mode warns against it.

> Rule of thumb for the choice: for nominal data, only the mode. For ordinal data, mode and median. For metric data, compute all three — and check whether mean and median differ markedly. If the difference is large, the distribution is skewed, and the median is the more honest metric.

## Step 5: Take dispersion into account

A measure of central tendency alone is rarely enough. Two samples can have the same mean and still look completely different — in one all values lie close together, in the other the dispersion is enormous. Measures of dispersion quantify these differences.

**Range.** Maximum minus minimum. The simplest measure, but easily distorted by outliers. Quick to compute, quick to communicate — methodologically more of an orientation value.

**Quartiles and interquartile range.** The quartiles split the sorted data into four equal groups. The lower quartile (Q1) marks the boundary to the lowest 25%, the upper quartile (Q3) to the highest 25%. The interquartile range (Q3 minus Q1) captures the middle 50% of the distribution and is robust against outliers.

**Standard deviation and variance.** The mean squared deviation of the values from the mean. Applicable only to metric data. The standard deviation is the square root of the variance and in the same unit of measurement as the original data — which makes it more understandable in reports than the variance itself.

> 💡 **What dispersion means in practice**
> 
> Picture two teams that both report employee satisfaction of 4.0 on a five-point scale. In the first team, everyone answers between 3.5 and 4.5. In the second team, half answer with 5, the other half with 3. The mean is identical, but the reality behind it completely different.
> 
> In the first team the mood is homogeneously positive. In the second team there's a very satisfied and a distinctly less satisfied half — a classic case in which the standard deviation supplies the actual signal. Report only the mean, and you miss it.

## Step 6: Choose fitting visualizations

Which form of presentation is the right one depends on the scale level and on the message. The most important options for survey data:

**Column and bar chart.** The standard form for absolute or relative frequencies of nominal- or ordinal-scaled data. Columns vertical, bars horizontal — the choice is usually a question of labeling: long category names read better horizontally.

**Pie chart.** Suitable when the division of a whole into a few (at most five, six) categories is to be shown. With more categories the pie chart loses legibility — a bar chart is then almost always the better choice.

**Box plot.** Condenses a distribution into five metrics: minimum, lower quartile, median, upper quartile, maximum. Ideal for comparing several groups — for instance when you want to visualize the distribution of customer satisfaction across several branches. Dispersion, skewness and outliers are recognizable at a glance.

**Scatter plot.** Shows the relationship between two metric variables. Each point represents one answer; its position in both dimensions is the answer to both questions. Useful for a visual check of correlations before computing formal correlation coefficients.

**Radar chart.** Displays several dimensions of an assessment at once. Frequently used for image and competence profiles: "how is our company perceived on the dimensions innovation, service, price, quality, sustainability?" Works well for four to eight dimensions.

**Word cloud.** For open-ended answers. The frequency of the words is represented by font size. A quick overview of the dominant themes — but no substitute for a content-analytic evaluation that takes context and connotation into account.

In QUESTIONSTAR all of the visualizations named are available directly in the analysis module. You switch between display types with a click, without exporting the data. For structural detail analyses — subgroup comparisons, crosstabs, filter combinations — export to Excel, CSV or SPSS is additionally available.

## Step 7: Test hypotheses

Up to this point the analysis describes. What it doesn't yet do: test whether observed differences are statistically robust. That's exactly what inferential statistical procedures are for.

For comparing **means** — for instance satisfaction in two departments, or before and after a measure — the t-Test is the standard procedure. You'll find a detailed guide with examples in our article [The t-Test — comparing means the methodologically sound way](/blog/the-t-test).

For comparing **frequency distributions** — for instance whether the distribution of preferred communication channels differs between age groups — the Chi Square test is the fitting procedure. Details in the article [The Chi Square Test in online surveys](/blog/the-chi-square-test).

For the **aggregation** of several items into a scale value — the typical case with multi-item Likert scales for constructs such as employee identification or brand loyalty — you'll find methodological guidance in the article [Aggregating multi-item Likert scales](/blog/aggregating-multi-item-likert-scales).

## Conclusion

A clean analysis follows a clear sequence: formulate the research goal, clarify the scale level, compute frequencies, determine measures of central tendency and dispersion, choose fitting visualizations, and where appropriate apply inferential statistical procedures. Work through this sequence — and make the first decision before you collect data — and you avoid the most common methodological errors and arrive at results that hold up before a specialist audience, management and auditor.

In QUESTIONSTAR the descriptive analysis steps are automated: frequencies, means and visualizations appear directly in the analysis module. For deeper analyses you export the raw data to Excel or SPSS and work there with the full range of inferential statistical methods. If you're unsure which method fits, write to us — we answer method questions even without a contract.

## Sources

- Bortz, Jürgen, and Christof Schuster: *Statistik für Human- und Sozialwissenschaftler*. 7th edition. Springer, 2010.
- Eid, Michael, Mario Gollwitzer and Manfred Schmitt: *Statistik und Forschungsmethoden*. 5th edition. Beltz, 2017.
- Fahrmeir, Ludwig, Christian Heumann, Rita Künstler, Iris Pigeot and Gerhard Tutz: *Statistik. Der Weg zur Datenanalyse*. 8th edition. Springer Spektrum, 2016.

## Related articles

- [the-t-test](/blog/the-t-test)
- [the-chi-square-test](/blog/the-chi-square-test)
- [aggregating-multi-item-likert-scales](/blog/aggregating-multi-item-likert-scales)

*Author: Dr. Paul Marx — see [About](/about).*