Choosing a study design from the research question
What question does this help answer?
Given a defined aim, which family of designs can produce an answer that means what the aim needs it to mean, with the data, time and resources that are actually available?
Start with the kind of question
- Describing: how common is something, what does it look like and who has it? Case series, prevalence surveys and descriptive cohorts answer this. They do not estimate effects.
- Explaining a cause: does an exposure or intervention change an outcome? Randomized trials answer this directly; observational and quasi-experimental designs answer it only under explicit assumptions.
- Evaluating a test: how accurately does it classify people, or does using it change decisions and outcomes? Accuracy and impact are different questions with different designs.
- Predicting: what is the chance of an outcome for an individual? Prediction model development and validation answer this, and a good predictor need not be a cause.
- Understanding experience or process: why and how does something happen? Qualitative and mixed-methods designs answer this.
- Summarizing what is known: what do existing studies show? Systematic and scoping reviews answer this.
From the question to a design
For a causal question, follow how the exposure arrived
If the intervention can ethically and practically be assigned, ask first whether it can be randomized, including randomizing clusters or the order of a rollout. If it cannot, the way it was assigned suggests the options. A threshold rule points to regression discontinuity. A change on a known date with repeated measurements points to an interrupted time series. A change in some sites but not others points to a controlled before-and-after or difference-in-differences design. A treatment decision recorded in routine data points to a cohort analysed as an emulated target trial. Each replaces randomization with an assumption, and that assumption has to be defensible for this question.
Constraints that legitimately change the choice
- A rare or slowly developing outcome favours sampling on the outcome, as a case-control study does.
- A rare exposure favours sampling on the exposure, as a cohort does.
- If only existing records are available, the design must be one those records can support, including a definable time zero.
- With few sites or clusters, designs that rely on comparing clusters become fragile.
- A short timeline may rule out designs that need long follow-up. That changes the question that can be answered, not the rigour required to answer it.
When not to choose yet
If the question is still a topic, such as readmissions, define the population, outcome, comparison and time horizon first. A design chosen before the target quantity tends to answer whatever its data happen to allow.
How this connects to Mallard
Mallard's design brief proposes a design with alternatives from the study description, and the investigator chooses before a plan is drafted. The brief applies these considerations to one study; this page explains the reasoning so the choice can be checked rather than simply accepted.
Clinical example
An investigator asks whether a sepsis alert reduced time to antibiotics. The alert went live across the hospital on a fixed date, and monthly data exist for two years before and after. Patients can no longer be randomized, and no comparable hospital lacks the alert. An interrupted time series fits how the alert arrived; a single before-and-after comparison would not separate the alert from a trend already under way. This illustrates the reasoning and assesses no real alert.
What to watch out for
A design is chosen to answer a defined question, not picked from the top of a hierarchy. The highest-ranked design is the wrong one if it cannot measure the outcome that matters, and when the feasible design cannot support the intended claim, the claim should change rather than borrow the language of a stronger design.
Before using this in your workspace
State each aim's question type first. Then record why the chosen design fits how the exposure was assigned and what data exist, which assumption replaces randomization if nothing was randomized, and which designs were considered and set aside, with the reason.
- An overview of clinical research: the lay of the land (opens in a new tab) · Lancet · doi 10.1016/S0140-6736(02)07283-5
- The C-Word: Scientific Euphemisms Do Not Improve Causal Inference From Observational Data (opens in a new tab) · Am J Public Health · doi 10.2105/AJPH.2018.304337
- Clinical research study designs: The essentials (opens in a new tab) · Pediatr Investig · doi 10.1002/ped4.12166