How to Choose a Research Methodology
Choose the method from the question, not from the tools you already know: experiment, observational study, qualitative study, or a review, and say what would count as an answer.
By Dr Raktim Mondol · 29 September 2026 · 3 min read
A methodology is the reason a reader should believe your answer. The method is how you will produce that answer. People mix the two up and then pick a method because they already know the software. Start from the question in How to formulate a research question. The question decides the method. The method does not get to rewrite the question so that your favourite tool still fits.
Match the method to the kind of claim
| The claim you want to make | A method that can support it | What it cannot support |
|---|---|---|
| Changing A changes B, in this setting | An experiment or a quasi-experiment, with a comparison you control or can argue is fair. | A description of one system that “worked well”. |
| A and B occur together in existing data | An observational study. Say clearly that you are not claiming you caused the pattern. | Causal language copied from an experimental paper. |
| People in this setting describe the problem this way | Interviews, focus groups, or a careful reading of documents, with the sample and the questions written down. | A percentage, unless you also did a count and can defend the sample. |
| The published evidence, taken together, says this | A systematic or scoping review, with rules fixed before the search. | A handful of papers you already agreed with. |
Mixed methods are not a way to avoid the choice. Use two methods when each answers a different part of the question, and say which part. “We did interviews and a model” is not a methodology until you say what the interviews were allowed to decide and what the model was allowed to decide.
Write the answer condition before you collect anything
For an experiment, name the comparison, the metric, and the difference you would treat as real, including how you will handle run-to-run variation. For an observational study, name the factors you will record and the ones you know you cannot see. For interviews, name the questions and how you will decide that a theme is present. This is the sentence you must not edit after you have seen the outcome. Editing it later is how a study quietly changes into a different study.
Prefer the smallest method that can answer the question
- A benchmark and a baseline answer “does this change help on this task?” They do not answer “should a hospital deploy this?”
- A user study with a handful of people can find failures. It cannot estimate a population rate.
- A case study can explain how something happened once. It is a weak way to claim it happens in general.
- A simulation answers the question inside the assumptions of the simulation. State those assumptions next to the result, not in an appendix nobody is pointed to.
Check feasibility and ethics in the same pass
A method you cannot carry out is the wrong method, even if it is the textbook one. Data you are not allowed to use, a sample you cannot recruit, compute you cannot afford, and a skill nobody on the project has are all reasons to change the design now. If the work involves people, their records, or a community that can be harmed by the result, the ethics path is part of the method. “We will get approval if it works” is not a plan.
Want guidance, not just a guide?
Statistics & Experiment Design for ML Research
Design experiments and prove your results actually hold up.
3 Hours live online · Third Saturday of every month · $299 USD
View the course