Design, method and methodology
A methodology chapter justifies the way a study is carried out; methods are the specific procedures, such as sampling, questionnaires or statistical tests. The research design sits between them: it is the overall plan that links the research question to the data, and in quantitative research it determines what kind of conclusion the data can support (Creswell & Creswell, 2023). In a proposal the design is what you plan to do; in the thesis it is what you did, including any changes and why they were made.
Textbooks group quantitative designs in slightly different ways. The five categories below follow a common classification in education and the social sciences (Creswell & Guetterman, 2024) and are equally useful in health, business and engineering. The key question in placing a study is: does the researcher control who receives the possible cause, and if so, is it allocated at random?
- 1Is a variable deliberately manipulated or an intervention delivered?
- 2No: are you comparing groups that already differ?
- 3No groups, only variables: descriptive or correlational
- 4Existing groups: causal-comparative
- 5Yes, intervention delivered: are units randomly assigned?
- 6No: quasi-experimental; Yes: experimental
The five designs
Descriptive
Descriptive designs measure and summarise characteristics of a population or phenomenon at one time or over time: prevalence of a condition, levels of an attitude, distribution of an outcome. They answer "what is" questions and usually test no causal hypothesis. Data are often collected by survey, structured observation or existing records. The main design decisions concern the population, the sampling frame and the precision of the estimates. If patterns found in the data suggest hypotheses, those hypotheses should be reported as exploratory and tested in new data. Descriptive work is not a lesser design: a well-sampled prevalence estimate, with its confidence interval and a clear account of non-response, is often exactly what a policy question needs, and it provides the baseline that later correlational or experimental studies build on.
Correlational
Correlational designs measure two or more variables in the same group and examine how they vary together, using correlation or regression. Nothing is manipulated. They can describe the strength and direction of a relationship and support prediction, but not, on their own, a causal claim: the association may be produced by a third variable or by reverse causation. Cross-sectional data cause particular problems when the hypothesis concerns a process over time, such as mediation (Maxwell & Cole, 2007). See Reporting correlation analyses for how to present the results.
Causal-comparative (ex post facto)
Causal-comparative designs compare groups that already differ on the independent variable, such as students who did or did not attend a summer programme, or patients with and without a condition, and examine differences in an outcome. The researcher cannot assign the independent variable because it has already happened or cannot ethically be manipulated. The central threat is selection: the groups may differ in other ways that explain the outcome. Matching, statistical adjustment and careful measurement of likely confounders reduce, but do not remove, this threat (Antonakis et al., 2010).
Quasi-experimental
In a quasi-experiment an intervention is delivered, by the researcher or by a programme or policy, but units are not randomly assigned. Common forms include non-equivalent group designs with pre- and post-tests, interrupted time series and regression discontinuity. Because the timing and delivery of the intervention are known, quasi-experiments can support stronger causal inference than causal-comparative studies, especially when there are measurements before the intervention. Bias in non-randomised studies of interventions can be appraised with ROBINS-I (Sterne et al., 2016).
Experimental
In a true experiment the researcher manipulates the independent variable and assigns units to conditions at random. Random assignment makes the groups comparable in expectation on all characteristics, measured or not, so differences in outcomes can be attributed to the manipulation within the limits of chance. Randomised controlled trials in health research are experiments, and their reporting now follows CONSORT 2025 (Hopewell et al., 2025). See How to design experimental research in psychology and Single-case experimental designs for detailed guidance.
Comparing the designs
| Design | Independent variable | Assignment | Typical question | Strength of causal claim |
|---|---|---|---|---|
| Descriptive | None (variables described) | None | What is the level or prevalence of X? | None intended |
| Correlational | Measured | None | Is X associated with Y? | Weak; association only |
| Causal-comparative | Pre-existing group difference | Already determined | Do groups that differ on X differ on Y? | Weak to moderate, depending on adjustment |
| Quasi-experimental | Delivered intervention | Not random | Did the intervention change Y? | Moderate, stronger with pre-tests and time series |
| Experimental | Manipulated | Random | Does X cause a change in Y? | Strongest, within the sample and setting |
A survey can appear in every row. A questionnaire might measure prevalence (descriptive), relate attitudes to behaviour (correlational), compare graduates of two routes (causal-comparative) or serve as the outcome measure in a randomised trial (experimental). Naming the survey as the "design" hides the decision that matters, which is how the comparison is set up (Bell et al., 2022).
Cross-sectional and longitudinal timing
Time is a second dimension that cuts across all five designs. A cross-sectional study measures each participant once; a longitudinal study measures the same participants repeatedly, as in a panel or cohort. If the question concerns change, sequence or development, such as whether stress precedes burnout, repeated measurement is usually needed, because cross-sectional data cannot show which came first and can give biased estimates of processes that unfold over time (Maxwell & Cole, 2007). Longitudinal designs cost more and lose participants between waves, so the attrition plan belongs in the proposal. Describe a study by both dimensions, for example "a longitudinal correlational design" or "a cross-sectional causal-comparative design".
How strong a causal claim can you make?
Researchers using non-experimental designs are often advised to avoid causal language altogether. Hernán (2018) argues that this is unhelpful: many observational studies are asked causal questions, and replacing "effect" with "association" disguises the question rather than improving the answer. The better course is to state the causal question honestly, set out the assumptions needed to answer it, address confounding in the design and analysis, and describe how far the assumptions are likely to hold. In management and the social sciences, Antonakis et al. (2010) reviewed common threats such as omitted variables, simultaneity and measurement error, and the design and statistical remedies for each.
Observational studies, including correlational and causal-comparative designs, should be reported with STROBE (von Elm et al., 2007) or the relevant extension, and sample size should be justified by a power analysis or a precision target for the main estimate (Faul et al., 2007).
Worked example (illustrative)
An illustrative case: a doctoral researcher in education asks whether a peer-tutoring programme improves first-year mathematics results. Four designs are possible, each answering a slightly different question.
- Descriptive: survey first-year students to estimate how many use peer tutoring and how often.
- Correlational: relate hours of tutoring to examination marks in one cohort, adjusting for prior attainment. A positive association could reflect motivation rather than tutoring.
- Causal-comparative: compare marks of students who chose to attend with those who did not. Self-selection is the main threat; matching on prior attainment reduces it.
- Quasi-experimental: the programme is introduced in two departments one year before the others; compare the change in marks before and after, against departments not yet offering it.
- Experimental: randomly offer places when demand exceeds supply, and compare outcomes of those offered and not offered.
The researcher chooses the quasi-experimental design, because the phased introduction is real and randomisation is not permitted by the university. Her methodology chapter names the design, explains why stronger designs were not feasible, describes the comparison departments and states the assumption that, without the programme, marks would have changed in parallel. For how to write this up, see How to write a research design.
What to report
| Item | What to state |
|---|---|
| Design | The design named precisely, with the reason it fits the question |
| Independent variable | Whether it was measured, pre-existing, delivered or manipulated |
| Assignment | How units came to be in each group; randomisation method if used |
| Confounding | Likely confounders, how they were measured and how they were handled |
| Sample size | Power analysis or precision target and the assumptions behind it |
| Guideline | CONSORT 2025 for randomised trials; STROBE for observational designs |
| Causal language | The causal question, its assumptions and the limits of the conclusion |
Common mistakes
- Calling a study "survey research" without stating whether it is descriptive, correlational or comparative.
- Labelling a non-randomised comparison as an experiment.
- Treating causal-comparative and quasi-experimental designs as interchangeable.
- Drawing causal conclusions from a cross-sectional correlation without discussing confounding.
- Avoiding the word "effect" while still implying causation in the discussion.
- Omitting the sample size justification.
Key takeaways
- Quantitative designs are best classified by whether the independent variable is measured, pre-existing, delivered or manipulated, and whether assignment is random (Creswell & Guetterman, 2024).
- A survey is a data collection method that can serve several designs, not a design in itself (Bell et al., 2022).
- Correlational designs need two or more variables, not two groups; causal-comparative designs compare groups that already differ.
- Quasi-experiments deliver an intervention without randomisation and can support stronger inference than causal-comparative studies (Sterne et al., 2016).
- State causal questions honestly and address confounding rather than hiding behind the word "association" (Hernán, 2018; Antonakis et al., 2010).
