Choosing a method that fits your question.

Most first projects choose a method the way people choose a route home: whichever one they already know. Surveys, usually, because a form is easy to build. Then the question turns out to be about why people did something, and a survey cannot answer why without the respondent doing the analysis for you.

Method follows question. Here is what each of the four common options actually buys.

Surveys

Good for: how common something is, and how two things vary together across a lot of people.

Cannot tell you: why, or what caused what. A correlation in a survey is a starting point for an explanation, not the explanation.

Realistic scope: for a first project, 80–150 responses is a normal haul, and you will need to distribute for two to three weeks. Every extra question costs you completions, so cut the "nice to know" items.

Where they go wrong: asking people to report on their own behavior over long periods, and double-barrelled items / "the feedback was clear and timely" measures two things and gives you one number.

Interviews

Good for: how people understand something, the sequence of how a decision happened, and anything where you do not yet know the answer categories.

Cannot tell you: how widespread it is. Twelve interviews cannot support "most students".

Realistic scope: 8–15 interviews of 30–45 minutes. Then double your time estimate, because an hour of interview is roughly four hours of transcription and coding. This is the number people get wrong.

Where they go wrong: leading questions, and a schedule so tight it becomes a spoken survey. Ask about specific past events / "walk me through the last time" / rather than general opinions.

Secondary and existing data

Good for: trends over time, comparisons across places, and anything where the collection would be beyond your reach. Policy documents, published statistics, institutional records, existing datasets.

Cannot tell you: anything the original collectors did not measure, and they measured it for their purposes, not yours.

Realistic scope: often the best choice for a short project, because your time goes into analysis rather than collection. Budget the first week for understanding how the data was defined before touching a single number.

Where they go wrong: definition drift. A category counted one way in 2016 and another way in 2021 produces a dramatic trend that is entirely bureaucratic.

Experiments and quasi-experiments

Good for: causal claims. This is the only one of the four that earns the word "because" cleanly.

Cannot tell you: whether it holds outside the setup you built.

Realistic scope: hard but not impossible at small scale, especially with two intact groups you did not assign / two class sections, two cohorts. Say quasi-experiment in that case and name the resulting threat.

Where they go wrong: a difference between the groups other than the one you introduced. If the morning section got the intervention and the afternoon section did not, time of day is now inside your finding.

Matching them up

• "How many / how often / does X relate to Y" → survey or existing data
• "How do people experience / decide / describe" → interviews
• "Has this changed over time" → existing data or documents
• "Does X cause Y" → experiment, or an honest correlational claim

On mixing methods

Combining two methods is genuinely stronger when each answers a different part of the question / a survey to establish the pattern, six interviews to explain it. It is a mistake when it means doing two thin studies instead of one solid one. For a first project, one method done properly beats two done at 60%, every time.

---

Before you commit, write the sentence you hope to publish, then ask whether your chosen method could support it. "Students revise more when feedback is fast" needs a comparison. "Students describe fast feedback as easier to act on" needs interviews. The sentence tells you the method, and it takes five minutes to find out you were about to build the wrong instrument.