Primary and Secondary Data Sources
Where Does Data Come From?
Every statistical investigation needs data — but that data has to come from somewhere. The two fundamental categories are primary data and secondary data. Knowing the difference — and the advantages and limitations of each — is essential when you plan your own enquiry and when you evaluate someone else’s.
Primary Data
Primary data is collected first-hand, specifically for your investigation. The main sources include:
- Experimental: the researcher controls conditions to test cause and effect. A laboratory experiment removes outside influences (high control). A field experiment happens in a natural setting, improving realism but reducing control. A natural experiment observes what happens when conditions change for non-research reasons.
- Simulation: a model, often computer-based, generates data when real-world collection is impossible or unethical — for example, simulating the spread of an infectious disease.
- Questionnaire: a set of written questions completed by participants. Cheap and scalable, but dependent on honest responses.
- Observation: watching and recording behaviour directly. Captures what people actually do rather than what they say they do, but the observer’s presence can change behaviour.
Primary data’s big advantage: it is collected specifically for your purpose, so it should measure exactly what you need. Its disadvantage is that collection takes time, money and effort.
Secondary Data
Secondary data already exists. Sources include published statistical databases, government censuses, academic studies, media reports and reference publications. Bodies such as the Office for National Statistics (ONS), NHS Digital and the Met Office all publish large, high-quality datasets that a researcher can use without collecting anything themselves — although access to some datasets can be restricted by cost, permission requirements or confidentiality.
The advantages of secondary data are significant: it is fast to access, often covers large populations, and can span decades of records no individual researcher could replicate. The census in England and Wales, conducted every ten years, records information about every household — a dataset of that scale is simply not achievable through primary collection for most researchers.
However, secondary data has limits. It was collected for a different purpose, so it may not measure exactly what you need, and variable definitions may have changed over time. Published figures are also often rounded before release, which reduces their accuracy for detailed calculations — always check how precisely values were recorded before relying on them. You must also acknowledge the source of any secondary data you use.
Secondary data: advantages — fast to access, large sample sizes, historical depth. Limitations — may not measure exactly what you need, figures may be rounded, and the source must be acknowledged and its quality assessed.
Census vs Sample
A census collects data from every member of the population. It gives complete data with no sampling error, but it is expensive, time-consuming, and sometimes impossible to complete fully.
In contrast, sampling collects data from a carefully chosen subset. Most practical investigations use samples — the challenge is making the sample representative enough for conclusions about the population to be valid.
Choosing Sources and Designing Data Collection Sheets
When planning where data will come from, a full list of sources is worth knowing: experimental, simulation, questionnaires, observation, reference publications, census, and sampling from a defined population. Whichever source you use, you are expected to be able to design a data collection sheet. A good sheet has categories that do not overlap, covers every possible response, and is piloted on a small group first to catch confusing questions before full collection begins.
Reliability and Validity
These words are used precisely in statistics and are often confused. Reliability is about consistency: does the measurement give the same result every time you use it? Validity is about accuracy of purpose: does the measurement capture the thing you actually want to know about?
A useful way to see the difference: imagine a bathroom scale that always reads 3 kg too heavy. It is reliable — it gives a consistent reading every time — but it is not valid, because it does not measure your true weight. A questionnaire asking “How healthy do you feel?” on a scale of 1–10 might give different answers from the same student on different days (poor reliability) and may not accurately capture their actual clinical health (poor validity).
It is easy to write “the data is reliable and valid” as if the two words mean the same thing, or to swap one for the other. The two terms have distinct meanings and should be used precisely.
If a question asks about reliability, focus on whether repeated measurements would give the same result. If it asks about validity, focus on whether the measurement actually captures what you want to measure. A questionnaire asking about income as a proxy for social class is valid if income correlates well with social class — but only reliable if the same person answers the same question consistently on different occasions.
Bias and Sensitivity
Bias is a systematic error — it pushes results in one direction rather than producing random noise. It can arise from how a sample is selected, how questions are worded, or the sensitivity of the topic.
Sensitive topics — income, health, political views, personal relationships — are especially prone to distortion. People may give socially desirable answers rather than honest ones, or refuse to answer at all. When the people who refuse to answer share a characteristic relevant to the investigation, this creates a systematic distortion in the results.
Level of Control and Minimising Distortion (Higher)
In a controlled experiment, the researcher manipulates one variable (the explanatory variable) and holds others constant, so any change in the response variable can be attributed to the manipulation. The level of control describes how completely the investigator can hold other variables constant.
High control (laboratory experiments) makes causation easier to claim but may reduce how well results transfer to real-world settings. Low control (natural experiments, observation) reflects real-world conditions better but makes cause and effect harder to isolate, because other variables are free to vary.
Extraneous variables influence the response variable without being the focus of the study, and must be managed. Techniques include holding them constant across all participants, including a control group that receives no treatment for comparison, using matched pairs so paired participants are similar on key characteristics before one receives the treatment, and randomisation, which spreads extraneous variables evenly across groups by chance.
Reliability = consistency across repeated measurements. Validity = the measurement captures what it is supposed to measure. A source can have one without the other.
Secondary data sources should always be acknowledged: record where the data came from, when it was collected or published, and who collected it.
This matters for two reasons. First, it lets the reader check the original source and judge its quality — a statistic from a government body is generally more trustworthy than one from a company’s own press release. Second, using someone else’s data without credit is a form of academic misconduct. In your investigation write-up, always note the source.