What is quantitative data analysis?
Quantitative data analysis uses numerical summaries, comparisons and models to answer questions about measurements or counts. It can describe what happened, estimate a quantity, examine an association, test a hypothesis or support a prediction. The interpretation depends on how the data was collected and what assumptions the analysis requires.
Good analysis begins before choosing a statistical test. Define the question, unit, population, measures and comparison. Then inspect the data, choose an appropriate approach and report the size and uncertainty of the result.
Numbers can support useful decisions on their own. Qualitative material may add context when the question calls for it, but a comment does not automatically prove the cause of a numerical pattern. Likewise, statistical significance does not establish that a result is true, important or causal.
Descriptive, inferential and predictive analysis
Scroll horizontally to see all columns →
| Purpose | Example question | Possible output | Limit to explain |
|---|---|---|---|
| Descriptive | How many requests were completed, and how long did they take? | Counts, proportions, distributions, averages and spread | The summary describes the observed records and their coverage |
| Inferential | What can this sample tell us about a wider population or model? | Estimates, confidence intervals and hypothesis tests | Inference depends on design, assumptions and how the sample was obtained |
| Predictive | How well can available information predict a later event? | A model assessed on suitable held-out or future data | Prediction accuracy does not establish the cause of the event |
Not every operational report needs inference or prediction. If you need to describe all recorded requests in a particular month, a carefully defined descriptive summary may answer the question. Adding a test does not repair incomplete records or an unclear denominator.
1. Define the analysis before opening the spreadsheet
Write an analysis question that identifies the measure, population and period. For example: “Among customers who started service in April, what proportion completed onboarding within 14 days?” is more useful than “Analyze onboarding.”
Specify what counts as starting service, completing onboarding and being eligible for the calculation. Decide how to handle someone who has not yet had a full 14-day observation period. These definitions determine the calculation.
For comparisons, state what differs between the groups and what else might explain the result. Two regions may have different account types, support needs or data coverage. A difference in averages is a finding to interpret in that context.
2. Prepare and inspect the data
- Identify the unit. A row might represent a person, account, visit, application or reporting period. Repeated visits are not independent people.
- Check types and units. Confirm dates, numeric fields, category codes and units such as minutes versus hours.
- Review duplicates. Distinguish accidental copies from legitimate repeated observations.
- Inspect missingness. Separate not asked, not applicable, not answered and not yet observed where the source allows it.
- Investigate unusual values. Correct errors using evidence; do not delete a valid extreme value because it changes the average.
- Retain a record of changes. Preserve the source and explain transformations, exclusions and revisions.
Missing is not zero. If a follow-up outcome is unavailable, entering zero can change both the apparent performance and the meaning of the measure. Plan the handling of missing data around the analysis rather than choosing whichever treatment gives the preferred result.
3. Describe the distribution before comparing
Start with the number of observations and the distribution of each important variable. For a categorical result, show counts and proportions with a defined denominator. For a numerical measure, inspect the center, spread and shape.
Scroll horizontally to see all columns →
| Summary | What it helps show | What to consider |
|---|---|---|
| Mean | The arithmetic average | It can be influenced by large or small values |
| Median | The middle of an ordered distribution | It can help describe a skewed distribution but does not summarize every feature |
| Range or interquartile range | Different descriptions of spread | The range depends on extremes; the interquartile range describes the middle half |
| Proportion | The share meeting a defined condition | The numerator and denominator must refer to compatible units and eligibility |
| Distribution chart | Shape, gaps, clustering or unusual observations | Use a chart appropriate to the variable and sample |
For an ordered rating item, category frequencies often communicate more than an average alone. If you summarize a multi-item scale or model ratings numerically, follow the instrument's scoring guidance and explain the assumptions. Numeric category labels do not automatically become measurements with equal intervals.
Worked example: completion rates across two sites
The following figures are fictional. Both sites use the same eligibility and completion definition for the same period, and the records are complete for this illustration.
Scroll horizontally to see all columns →
| Site | Eligible accounts | Completed | Completion rate |
|---|---|---|---|
| A | 80 | 60 | 60 ÷ 80 = 75% |
| B | 20 | 18 | 18 ÷ 20 = 90% |
| Combined | 100 | 78 | 78 ÷ 100 = 78% |
The combined rate is 78%, not the simple average of the two site rates, which would be 82.5%. The eligible groups are different sizes, so combine the compatible counts or use the equivalent denominator-weighted average.
Site B's rate is 15 percentage points higher than Site A's. That is a descriptive difference in these records. It does not establish that B's approach caused better completion. Account mix, support needs and other conditions may differ.
A responsible next step is to examine those conditions and the practical importance of the gap. If a wider inferential question is needed, choose an appropriate design and analysis rather than attaching a test to the percentages after the fact.
4. Choose a comparison or model that fits the question
Scroll horizontally to see all columns →
| Question | Methods an analyst might consider | Design detail that matters |
|---|---|---|
| How do numerical outcomes differ between groups? | Difference estimates, suitable mean-comparison or rank-based methods, regression | Independent versus paired observations, distribution and variance assumptions, relevant covariates |
| How do proportions or categories differ? | Proportion estimates, contingency-table methods or suitable categorical models | Counts, sample design, small cells and repeated observations |
| How are two numerical measures associated? | A scatterplot, correlation or regression | Relationship shape, influential observations and confounding |
| How did the same units change over time? | Paired summaries or a suitable repeated-measures model | Correct matching, missing follow-up, time and within-unit dependence |
| Can we predict a future outcome? | A suitable predictive model and validation plan | Data leakage, future performance, calibration and consequences of errors |
This table helps identify the questions to discuss with an analyst; it is not an automatic test selector. Sample size alone does not determine the correct method. A large dataset can still have biased selection, incompatible measures or dependence between observations.
For repeated data, see longitudinal data analysis. For definitions and measurement types, see quantitative data.
5. Interpret uncertainty without treating a p-value as a verdict
A p-value describes how incompatible the observed result is with a specified statistical model, using the test's definition of more extreme results. It is not the probability that a hypothesis is true, nor does it measure effect size or practical importance. The American Statistical Association's statement explains these limits and the need for fuller reasoning and reporting.
Report the estimated difference or association in understandable units. Include uncertainty where the analysis supports it and explain the assumptions. A narrow interval does not correct biased collection or a poorly defined measure.
A non-significant result is not automatically evidence of no meaningful difference. A small study may leave substantial uncertainty. Conversely, a large dataset can produce a small p-value for a difference that is too small to matter for the decision.
If many outcomes or subgroup comparisons were examined, report that fact and address the implications. Selecting only the most favorable result can give a misleading picture of the evidence.
6. Keep association, prediction and causation separate
An association shows that measures vary together in the analyzed data. A prediction estimates an outcome from available information. A causal claim asks what would change because of an intervention or exposure, compared with a suitable alternative.
Quantitative designs can support causal inference when appropriately designed and analyzed. Qualitative work can contribute to understanding processes and experiences. Neither a regression coefficient nor an explanatory quotation automatically proves causation.
For example, customers with more support contacts may also be more likely to cancel. That does not establish that contacting support causes cancellation. Underlying service difficulty may influence both. Read the operational context and use a design suited to the claim.
7. Bring in other evidence when it answers a real question
Comments, interviews and documents can help investigate a surprising result or understand how a process was experienced. Explain how those sources relate to the numerical data. An interview subset may be selected for depth rather than representativeness.
Keep disagreement visible. A numerical improvement alongside critical comments may reflect different experiences, coverage or questions. Do not discard one source to make the story simpler. See mixed-methods data analysis for planning that integration.
Build comparable data across recurring collections
Local teams do not always need one identical survey. Agree on the small shared core needed for aggregation: a measure definition, unit, eligible population, period and relevant context. Use a data dictionary to document those decisions and retain local questions where they serve local needs.
Collect stable reference information once where appropriate, update changing details with dates and preserve historical context. If a definition changes, document whether earlier and later results can still be compared. A matching field name is not sufficient evidence that two measures mean the same thing.
Record calculations and review decisions so another person can reproduce the reported figure from the permitted source data. Reproducibility is a valuable check, but it does not by itself prove that the measure or interpretation is valid.
How Sopact reduces coding and reporting work
An assistant becomes more useful when it can filter the actual records, calculate from the selected data and bring back the relevant text. A plausible narrative without that path is not enough.
A workflow with repeated manual work
- Define from an initial sampleRead material and agree on the codebook.
- Apply it across the datasetCode responses and check the result.
- Revise a definitionReturn to affected material and recode it.
- Reconnect the numbersReconcile coded results with ratings and context, then rebuild the view.
The Sopact workflow
- Your team owns the definitionsDecide what each code means and improve it as you learn.
- Apply coding across the eligible dataAutomate application; people review quality and exceptions.
- Reprocess after a definition changesReapply the revised definition across the configured scope instead of recoding each response by hand.
- Ask across coded text and numbersKeep the response, rating and relevant record context connected; inspect the evidence behind the result.
This compares workflow patterns, not a claim that every research tool requires manual coding or separate files. Some already automate parts of this work; compare the complete cycle.
For this codebook-based workflow, the main saving is repeated application and reconnection—not the removal of human judgment. A changed definition can be reapplied across the configured data while reviewers concentrate on quality, exceptions and interpretation. Coded text stays connected to the relevant ratings and context.
Count the recurring work in ownership cost. Include setup, coding, recoding after revisions, source reconciliation, review and reporting, plus your actual platform and processing expenses. A worked scenario of four cycles of 4,000 responses illustrates 272 fewer annual staff hours; it is an assumption-based example, not a customer benchmark. Existing automation, review needs and implementation effort can substantially change the result.
Adjust the workload assumptions and compare total effort →
A reliable assistant should calculate from the selected records and let a reviewer open the supporting evidence. Check the data scope, definition, denominator and access permissions. Reproducible arithmetic does not make every AI interpretation correct.
What the software workflow should support
Spreadsheets, statistical packages, databases and reporting tools can all contribute. Choose around the analysis and the team's recurring work: preparation, definitions, linked context, calculations, review and sharing.
Evaluate Sopact where that collection-to-review workflow needs to stay manageable for an operating team. Ask to see how your records, measures and review process work in practice. Verify the calculations and controls you need; do not assume a platform supports every statistical model.
AI-generated calculations and explanations should be checked against the data and method. A polished answer is not a substitute for correct eligibility rules, appropriate analysis and an accountable reviewer.
Report the result so someone can use it
Include the question, population, period, source, coverage, method, main estimate, uncertainty where appropriate, limitations and next action. Show the denominator beside a percentage. Distinguish a finding from a hypothesis the team will investigate.
Use How to Write an Impact Report to structure the wider narrative and browse report examples for presentation ideas.
Watch: bringing qualitative context into a review
These related videos discuss qualitative evidence within a connected workflow. They are companion context, not tutorials on statistical testing.
Connected Data Intelligence: Why Qualitative Data Gets Ignored
Why Qualitative Analysis Stays Small — And How to Scale It
Frequently asked questions
What are the main steps in quantitative data analysis?
Define the question and measures, inspect and prepare the data, describe the distribution, choose an appropriate analysis, assess uncertainty and report the interpretation and limits.
Do I always need inferential statistics?
No. A descriptive analysis may answer an operational question. Use inference when the question and design call for conclusions beyond the observed records or about a statistical model.
Does a small p-value prove the result is real?
No. A p-value is interpreted within a specified model and assumptions. It does not establish truth, importance or causation.
Can I average percentages from different groups?
Only if the averaging method matches the question. To calculate a combined rate from compatible groups, combine the numerators and denominators rather than giving differently sized groups equal weight by default.
Do comments prove why a numerical result occurred?
No. They can provide accounts and context that help investigation, but they do not automatically establish the cause of the result.
How do I compare local data without imposing one survey?
Agree on the limited shared measures and context needed for comparison, document definitions and differences, and retain local questions. Combine only results with compatible meanings and coverage.
Choose your next step
ASME
Explore how a membership organization brings quantitative and qualitative evidence into review. The numerical example above is fictional.
Read the customer story →