The hardest question at a proposal defence is rarely about statistics. It is: “Why this design?” When the answer is “because I do quantitative work” or “because my supervisor does surveys”, the panel knows the proposal was built backwards. The method came first and the question was written to fit it.
A research design is not a personal preference. It is the logical consequence of what you want to know. A “what proportion” question cannot be answered with ten interviews, and a “why do people drop out” question is hard to answer with a questionnaire made of Likert items. What follows is a way of reading your question so that the design almost chooses itself, with a matching table and a real proposal repaired step by step.
The verb is your first clue
Underline the main verb of your research question. It almost always falls into one of five families:
- Describe: “prevalence”, “level”, “distribution”. You need a representative sample and descriptive statistics with confidence intervals.
- Compare: “how do X and Y differ”. You need clearly defined groups and enough people in each group, not just in total.
- Relate: “is associated with”, “predicts”. You measure both variables on the same units and accept that association is not causation.
- Cause: “effect of”, “impact”, “does X improve Y”. You need a controlled intervention or a credible identification strategy for observational data.
- Explore or explain: “how”, “why”, “experience of”. You need qualitative data deep enough to show process, not just outcome.
The most common mismatch in theses worldwide is a title containing “the impact of” sitting on top of a single cross-sectional survey. That data supports statements about association only. A lenient committee may let it pass; a journal reviewer usually will not.
A question–design–analysis matching table
| Question type | Fitting design | Typical analysis | What you cannot conclude |
|---|---|---|---|
| Describe a situation | Cross-sectional survey, probability sample | Proportions, means, 95% confidence intervals | Why the situation exists |
| Compare groups | Cross-sectional, stratified by group | t-test, ANOVA, chi-square | That group membership causes the gap |
| Relate or predict | Correlational, longitudinal where possible | Regression, SEM | Direction of causation from one wave |
| Intervention effect | Randomised or quasi-experiment with a control group | Post-test comparison adjusted for baseline | Generalisation far beyond the trial setting |
| Process or experience | Case study, in-depth interviews, ethnography | Thematic analysis, coding | Population percentages |
The last column matters more than the first three. Every design has a zone of forbidden conclusions. Write it down before you collect data and you have pre-empted most reviewer objections.
Unit of analysis: the quiet decision
Before choosing a design, decide what you are studying: individuals, classrooms, schools, firms or regions. A project on “organisational culture and employee engagement” that collects 300 questionnaires from 12 companies has 12 units for the culture variable, not 300. Treat them as 300 independent people and your standard errors shrink, making “significant” results suspiciously easy. With nested data like this you need multilevel models or, at minimum, cluster-robust standard errors.
Time: one snapshot or repeated measurement
Cross-sectional work is cheap and fast and suits description. But when your question contains “change”, “develop” or “after”, you need at least two measurements of the same people. A realistic compromise for a master’s project is two waves six to eight weeks apart, planning for 15–25% attrition and reporting how leavers differ from stayers. Asking respondents “how were things before?” in a single survey is not a substitute: retrospective memory bends towards the present.
A worked repair of a real-looking proposal
Original title: “The impact of social media on undergraduate academic performance”. Planned design: a questionnaire to 250 students asking about daily social media hours and self-reported grade average.
- Verb: “impact” signals a causal question, but the design only supports association.
- Measurement: self-reported grades drift upwards; self-estimated screen time is notoriously noisy.
- Confounding: students with part-time jobs may both study less and scroll more. Unmeasured, this cannot be separated out.
Two ways forward. Option one keeps the design and changes the question to “the association between the amount and type of social media use and academic performance”, uses registry grades with consent, and adds work hours and credit load as covariates. Option two keeps “impact” but changes the design: a four-week notification-limiting intervention in several course sections, with comparison sections. Option one is more feasible for a thesis; option two is more interesting but needs cooperation from instructors.
Four stress tests before you commit
- If the results come out opposite to my prediction, will this design still tell me something useful?
- What alternative explanation will the committee think of first, and have I measured it?
- Can I actually reach the right participants, in the numbers required, within my timeline?
- What is the strongest single sentence this data could justify? Write it out.
The fourth test is the most revealing. If your strongest defensible sentence sounds weaker than your title, one of them has to change.
What you do not need
You do not need to “combine qualitative and quantitative” just to look thorough; mixed methods has its own rationale and roughly doubles the workload. You do not need structural equation modelling to compare two groups. You do not need 1,000 respondents for an exploratory question. Methodological complexity never compensates for a vague question.
A one-page checklist for your methods chapter
- The research question, main verb underlined.
- Question type: describe, compare, relate, cause or explore.
- Unit of analysis and any nesting.
- Number of measurement points and why.
- Key confounders you will measure or control.
- The strongest conclusion you may write, and one you may not.
Do this today: copy your research question onto a blank page, underline the verb and fill in those six lines. If line two takes you more than half an hour, the problem is the question, not the method — take that page to your supervisor before you draft a single survey item.
Câu hỏi thường gặp
How do I know if my research should be qualitative or quantitative?
Look at what the question asks. “How much, how many, is there a relationship” points to quantitative work; “how and why” points to qualitative. Let the question decide rather than the habits of your department.
Can a cross-sectional survey show cause and effect?
On its own, no. A single survey supports claims about association. Causal claims need a controlled intervention, longitudinal data or a clear identification strategy, and even then careful wording.
What is a unit of analysis in research?
It is the entity your variables describe: a person, a class, a school or a company. Analysing company-level variables as if each employee were independent inflates significance.
Should a master’s thesis use mixed methods?
Only when the question genuinely needs two kinds of data that answer different parts of it. Adding a few interviews to pad a chapter costs time without adding value.
What do examiners ask about research design?
Typically why you chose the design, which confounders remain, whether the sample represents the population and whether conclusions go beyond the data. Prepare answers to those four and you cover most defences.
