What is data analysis? A practical introduction with examples
Data analysis turns observations into answers. Learn the steps, common types of analysis and how to turn a question into a defensible conclusion.
Data analysis is the process of examining observations to answer a question. It includes deciding what to measure, checking the data, looking for patterns, applying suitable methods and explaining what the evidence supports. A chart or regression can be part of an analysis, but neither is a conclusion by itself.
The question determines what counts as a useful answer. A transport planner asking how long commuters wait needs a different analysis from one asking whether a new timetable reduced waiting times.
Start with a question you can measure
Consider a hypothetical university studying student travel. It has survey responses with journey duration, travel mode, distance and departure time. Before calculating an average, the analyst needs to settle several choices:
- Does a journey include the walk to the bus stop?
- Is the unit one student, one journey or one student-day?
- Does the survey cover all students or only people who answered an email?
- Is the question about a typical journey, the longest delays or the effect of a timetable change?
Those decisions change the analysis. Students who travel every day contribute more observations to a journey-level dataset than students who attend once a week. An average over journeys and an average over students need not match.
Common types of data analysis
These labels describe different goals. A project can use several of them.
| Goal | Question | Possible output | Main caution |
|---|---|---|---|
| Descriptive | What happened in the observed data? | Median journey time by travel mode | A summary may hide differences within groups |
| Exploratory | What patterns or problems should we investigate? | Distributions, missingness checks and unusual observations | A discovered pattern needs further scrutiny |
| Inferential | What can the sample tell us about a wider population? | A population mean estimate with a confidence interval | Sampling and dependence assumptions matter |
| Predictive | What outcome should we expect for a new case? | A journey-time prediction evaluated on held-out journeys | Strong performance on the training data is insufficient |
| Causal | What would change if an intervention occurred? | An estimated timetable effect under a stated design | Association alone does not identify an effect |
NIST describes exploratory data analysis as an approach to investigating structure, unusual observations and assumptions, often through graphics. That is a useful starting point before choosing a model. The distinction between descriptive and inferential statistics explains why a sample summary and a population claim are separate outputs.
A practical analysis workflow
- Define the target. Write the population, period, outcome and intended comparison in one sentence. For example: weekday door-to-door journey duration among enrolled students this term.
- Document the source. Keep the survey questions, collection dates, units and response rules alongside the data. Preserve the original file.
- Check quality. Look for duplicate submissions, impossible durations, missing values and inconsistent categories. Record decisions rather than silently deleting inconvenient rows.
- Describe before modelling. Examine the distribution, group sizes and coverage. Compare medians as well as means when journeys have a long upper tail.
- Choose a method for the question. A comparison of observed groups may need only a table. A prediction task needs evaluation on new observations. A policy question needs a credible comparison design.
- Check the answer. Ask whether different reasonable cleaning rules or definitions change the conclusion. Report uncertainty where the target requires inference.
- Explain and preserve. State what was measured, what was excluded, what the result means and how someone else could reproduce it.
The steps often loop back. Finding repeated submissions may require revisiting the unit of observation; a large amount of missing travel-mode data may make a planned comparison uninformative.
A small example: the average is not the whole answer
Suppose five invented journey times are 10, 15, 20, 25 and 80 minutes. Their mean is 30 minutes; their median is 20. Both calculations are correct. The mean reflects total time across journeys, while the median locates the middle journey. Reporting only one number could hide the long delay.
Do not discard the 80-minute journey just because it changes the average. It could be a recording error, a genuine disruption or a different kind of journey. Check the source and explain any exclusion. NIST's discussion of location measures explains why different summaries react differently to extreme values.
Where econometrics fits
Econometrics applies statistical reasoning to economic questions. If the travel project asks whether a subsidy changed commuting choices, economic behaviour and the way eligibility was assigned become central. More observations or a more complicated regression cannot by themselves settle that causal question.
Start practising with the Statistics measurement lab, then follow the econometric analysis workflow when the question calls for modelling economic relationships.
References
- NIST/SEMATECH. What is exploratory data analysis?
- NIST/SEMATECH. Measures of location