A research design is the logical structure of the enquiry: what you will compare, to what, measured when, on whom. It sits between your research question and your methods, and it is what determines whether your conclusion can be supported at all.
This post covers where design sits relative to methodology and method, the four decisions that fix a design, the main designs with what each can and cannot support, the validity threats that come with each, and how to write the paragraph.
From design to a written methodology chapter
ThesisAI drafts full academic documents with chapter structure and inline citations verified against the papers they came from, so the design you state and the methods you describe stay consistent throughout.
Zacznij pisaćDesign, Methodology and Method
Three words used interchangeably in student writing and distinguished carefully in marking criteria.
| Term | What it is | Example |
|---|---|---|
| Methodology | The reasoning that justifies your approach, including its philosophical grounding | A critical realist approach, seeking mechanisms rather than regularities |
| Design | The structure of the enquiry: comparisons, timing, allocation, units | A quasi-experimental pre-test/post-test design with a non-equivalent comparison group |
| Method | The specific procedures for collecting and analysing data | A 24-item questionnaire administered twice, analysed with a mixed ANOVA |
Your methodology chapter moves through them in that order. The wider chapter is covered in our research methodology guide, and the philosophical layer in research philosophy.
The Four Decisions That Fix a Design
Every named design is a combination of answers to these four questions. Work through them and the name falls out; start from the name and you will usually pick the one you have heard of.
- Do you manipulate anything? If you actively change a condition, the design is experimental or quasi-experimental. If you observe what is already there, it is non-experimental.
- Is there a comparison, and how was it formed? Random allocation, matched groups, a naturally occurring comparison, or none at all. This single answer does most of the work in deciding whether you can claim causation.
- How many time points? One measurement is cross-sectional. Two or more on the same units is longitudinal, and only that supports claims about change.
- How many units, studied how deeply? Many units shallowly supports generalisation; few units deeply supports mechanism. You cannot have both in a thesis.
The Main Designs
| Design | Supports | Cannot support | Feasible in a thesis? |
|---|---|---|---|
| True experimental | Causal claims, with random allocation to conditions | Behaviour outside the controlled setting | Only in lab-based fields |
| Quasi-experimental | Plausible causal claims where allocation was not random | Ruling out selection differences entirely | Often, and the realistic causal option |
| Correlational | Strength and direction of association | Which variable causes which | Yes, the most common student design |
| Descriptive / survey | Prevalence, distribution, profile of a population | Explanation of any of it | Yes |
| Case study | Mechanism and context in depth, single or comparative | Statistical generalisation to a population | Yes, and frequently the strongest choice |
| Longitudinal / panel | Change over time, and temporal ordering | Anything, if you cannot fund the follow-up | Only using existing panel data |
| Ethnographic | Practice as it is actually performed, over time, in place | Comparison across sites, usually | Rarely, given fieldwork time |
| Mixed methods | Both distribution and explanation, integrated by design | Either half done well on a short timeline | Yes, at roughly double the work |
Experimental and Quasi-Experimental
An experiment manipulates an independent variable, allocates units to conditions at random, and measures the effect on a dependent variable. Random allocation is the whole point: it makes the groups equivalent on everything, measured and unmeasured, so a difference afterwards can be attributed to the manipulation.
A quasi-experiment does everything except the random allocation, because it usually cannot. You compare a department that received the new training with one that did not, or the same unit before and after a policy change. This is realistic for student research and legitimate, provided you name what random allocation would have controlled and address it - typically by matching on observed characteristics, by baseline measurement, or by demonstrating equivalence on plausible confounders.
Common quasi-experimental structures: pre-test/post-test with a non-equivalent comparison group; interrupted time series; difference-in-differences where a change hit one group and not another.
Non-Experimental Designs
Correlational designs measure two or more variables as they occur and quantify the relationship. They are efficient, ethical where manipulation would not be, and permanently limited: association is compatible with X causing Y, Y causing X, and Z causing both. Write "is associated with", never "leads to", and the chapter holds up.
Descriptive designs establish what is there - prevalence, distribution, a profile. Undervalued by students who assume description is not enough for a thesis. It is, when nobody has measured the thing in your context, and it is far better done than a weak explanatory study.
Comparative designs set two or more naturally occurring groups against each other. The design question is always the same: what else differs between these groups besides the thing you are interested in?
Case Study and Longitudinal Designs
A case study is an in-depth examination of one or a few bounded cases in their real context, usually drawing on several data sources. The two decisions that make or break it are the boundary - what is inside the case and what is context - and the selection logic: a typical case, an extreme case, a critical case or a deviant one, each supporting a different kind of claim. Case studies generalise to theory rather than to populations, and saying so explicitly protects you in a viva.
Longitudinal designs measure the same units repeatedly. They establish temporal order, which is a precondition for causal argument, and they cost time a thesis does not have. The realistic route for a student is an established panel study or administrative data with a time dimension. A cohort design follows a group sharing a starting characteristic; a panel design follows the same individuals; a trend design repeats on fresh samples from the same population and cannot track individual change.
Choosing: A Decision Path
- What does your question ask for? "How many" needs description. "Is X related to Y" needs correlational. "Does X cause Y" needs experimental or quasi-experimental. "How and why does X happen here" needs a case study. Start from your research question, not from the design.
- Can you manipulate the variable, ethically and practically? If not, a causal design is closed to you and the honest move is to reframe the question as associational.
- Does a comparison already exist in the world? A policy that hit one region and not another is a natural experiment, and it is the cheapest causal leverage a student can get.
- Do you need change, or a snapshot? If change, check for existing panel data before designing collection.
- Breadth or depth? Decide deliberately. The failure mode is thirty interviews analysed shallowly, which delivers neither.
- Does your timeline survive it? Halve your estimate of how long recruitment will take, and check the design still fits.
Threats to Validity by Design
Each design comes with characteristic weaknesses. Naming yours before an examiner does is a straightforward way to gain credit rather than lose it.
| Design | Main threat | Standard mitigation |
|---|---|---|
| Experimental | Weak ecological validity; demand characteristics | Blinding; a realistic task; replication in a field setting |
| Quasi-experimental | Selection: the groups differed already | Baseline measurement, matching, statistical control, and a stated limitation |
| Correlational | Confounding and reverse causation | Measure plausible confounders; use associational language throughout |
| Descriptive / survey | Non-response bias; unrepresentative sampling | Report the response rate; discuss who is likely missing |
| Case study | Researcher interpretation; boundary drift | Triangulate sources; member checking; state the case boundary early |
| Longitudinal | Attrition, and attrition that is not random | Report dropout and compare leavers with stayers |
| Mixed methods | Two shallow strands instead of one integrated study | State the integration point before collecting anything |
The full framework - internal, external, construct and statistical conclusion validity, with reliability alongside - is in our guide to reliability and validity.
Writing the Research Design Paragraph
This sits early in the methodology chapter, after the research questions and the research type and before the sampling and instruments. Name the design, justify it from the question, and concede its principal limit.
Weak: "A quantitative research design was used for this study. Quantitative research uses numerical data and statistical analysis. This design was chosen because it is appropriate for the research questions and allows the collection of objective data."
"Quantitative" is not a design. The justification is circular, and "objective data" is an assertion an examiner will want defended.
Strong: "The study uses a quasi-experimental pre-test/post-test design with a non-equivalent comparison group. RQ2 asks whether the coaching programme changes self-reported autonomy, which requires a before-and-after comparison against units that did not receive it. Random allocation was not possible: the programme was rolled out by department on an existing schedule fixed before the study began. Two departments that had not yet received it serve as the comparison, matched on size, mean tenure and grade distribution. Because allocation was not random, unobserved differences between departments cannot be excluded, and baseline autonomy scores are therefore reported and controlled for in the analysis. Measurement is at two time points eight weeks apart, chosen because the programme runs for six weeks."
Every sentence does work: the design, the reason from the question, why the stronger option was unavailable, how the comparison was formed, the residual threat, and the timing with its rationale. Our guide to writing a methodology section covers what follows it.
FAQs About Research Design
What is a research design?
The logical structure of a study: what is compared with what, measured when, on which units, and whether anything is manipulated. It sits between the research question and the methods, and it determines what conclusions the data can support.
What is the difference between research design and methodology?
Methodology is the reasoning that justifies your whole approach, including its philosophical assumptions. Design is the structure of the enquiry itself. Method is the specific procedure. A methodology chapter states all three, in that order.
What are the main types of research design?
Experimental, quasi-experimental, correlational, descriptive, comparative, case study, longitudinal, ethnographic and mixed methods. Which family you are in is set by whether you manipulate anything and how your comparison group was formed - see also our guide to the types of research.
Can I claim causation from a correlational design?
No. You can report a strong association, argue for a plausible direction from theory or timing, and recommend a design that could test it. Writing "leads to" or "impacts" where you mean "is associated with" is the single most common overclaim in student theses.
Is a case study a valid research design for a thesis?
Yes, and often the strongest available one. It requires a clearly stated case boundary, an explicit selection logic, more than one data source, and an upfront statement that it generalises to theory rather than to a population.
How do I choose between a quantitative and a qualitative design?
By what the question needs, not by preference. Prevalence, distribution and relationships between variables need quantitative work; meaning, process and mechanism need qualitative work. If you genuinely need both, that is a mixed methods design and roughly twice the work.
Can I change my research design partway through?
Yes, with supervisor agreement and usually an ethics amendment. Report what you actually did and why it changed. A documented change with a stated reason is ordinary research practice; an undocumented one is a problem in the viva.
Work the four decisions before you look at any list of design names, and let the name be the output rather than the input. A design you can justify from your own research question, with its principal weakness named in the same paragraph, is one an examiner has very little to attack. Sampling comes next, in our guide to sampling methods.