Research Designs for Activity and Health
A practical guide to defining activity-and-health questions, choosing and evaluating study designs, and interpreting results without overstating causal conclusions.
From Research Question to
Start with a question that can be measured
A useful activity-and-health question identifies who is being studied, what activity is examined, which health outcome is measured, and over what period. For example: “Among adults without type 2 diabetes, is higher weekly moderate-to-vigorous physical activity associated with a lower incidence of diabetes over five years?” This specifies the group, activity, outcome, and follow-up period. It also distinguishes a new diagnosis during follow-up from diabetes that was already present.
Next, define the variables so they can actually be assessed. Activity might be described by frequency, duration, intensity, and domain—work, transportation, or leisure. A questionnaire can capture context and is practical for large surveys; accelerometers and pedometers measure movement using devices, but require procedures for wearing and processing them. These methods capture different aspects of activity, so the choice should fit the question. Health outcomes need definitions too: for instance, researchers should state how and when they will measure blood pressure, fitness, depressive symptoms, falls, or a new diagnosis. Inaccurate measurement—or measurement that differs between groups—can distort the observed relationship.
Match the design to the comparison
The design determines what can be compared and what conclusions are reasonable:
Cross-sectional study: Measures activity and health at about the same time. A survey comparing current activity with current blood pressure can estimate how common an outcome is and identify an association. Because the measurements are concurrent, it often cannot establish which came first.
Cohort study: Classifies people by activity and follows them to observe later outcomes. A prospective cohort can establish that measured activity preceded a new diagnosis. However, differences between activity groups may explain some or all of the association, so the timing alone does not prove that activity caused the outcome.
Case-control study: Starts with people who have an outcome (cases) and comparable people who do not (controls), then compares their earlier activity. This can be efficient for rare outcomes, but recalling past activity may be difficult or inaccurate.
Randomized controlled trial: Assigns participants at random to an activity intervention or comparison condition, then measures outcomes. Random assignment tends to balance other characteristics between groups and strengthens causal inference. Trials can be expensive, may not represent everyday activity, and cannot ethically or practically assign every exposure of interest.
These designs answer different questions: cross-sectional studies describe a snapshot, cohorts track later outcomes, case-control studies look back from an outcome, and trials test assigned interventions. None guarantees a causal result by itself. Measurement quality, participant selection, missing follow-up, and analysis also affect how strongly the findings support a conclusion.
Who Is Compared—and What Can Distort the Comparison?
Who do the findings describe?
The is the group researchers want their findings to describe; the sample is the subset actually studied. How that sample is recruited affects whether the results can speak to the . With probability sampling, eligible people have a known chance of selection. This can support inference to the when recruitment and analysis are appropriate. By contrast, volunteers recruited from one gym form a convenience sample: it may be useful for preliminary research, but may not represent the wider population.
Recruitment is not the only issue. People who do not participate, or who leave before follow-up ends, can make the observed sample differ from the intended population. Thus, a study’s findings may be informative about its participants while still having limited reach beyond them.
Make a useful comparison
A study needs groups that allow the research question to be addressed. In a cohort study, researchers might compare people with different activity levels; in a trial, they might compare an intervention group with a control group. Groups should be comparable in relevant ways. The sample size should be sufficient to detect a meaningful difference with reasonable precision. A larger sample can reduce random , but it does not by itself remove bias or confounding: adding more participants cannot correct a systematically unrepresentative sample or inaccurate measurement.
Separate confounding from other problems
A is related to both activity and the health outcome and can create or distort their observed association. Age is one example: it may be related both to activity levels and to cardiovascular risk. If age is not accounted for, an apparent activity–health difference may partly reflect differences in age rather than activity.
Researchers can address plausible confounders during through randomization, restriction, or matching, and during analysis through stratification or statistical adjustment. The choice of variables to adjust for should rest on subject knowledge and a clear causal rationale—not merely on whether a variable is statistically associated with the outcome. A mediator is different: it is a step on the pathway from activity to the outcome. Adjusting for a mediator can remove part of the relationship being studied, so the appropriate adjustment depends on the research question.
Other problems require distinct attention:
concerns who enters or remains in the study.
Information bias concerns inaccurate measurement or recall. For example, an activity measure that is inaccurate or differs between groups can distort a comparison.
Reverse causation occurs when poor health reduces activity, rather than low activity preceding poor health.
Statistical adjustment for selected variables does not necessarily resolve these issues. A careful interpretation considers sampling, participation, measurement, and the direction of events as well as confounding.
Analyzing Results and Judging Causal Evidence
Compare outcomes in a way that fits the question
Analysis begins by comparing health outcomes across activity levels while accounting for the and relevant confounders. The appropriate comparison depends on the outcome: researchers might compare proportions or risks, average measurements, or rates over time. Regression models can estimate an association while adjusting for specified variables, but what is adjusted for should be chosen with the research question and a sound rationale in mind.
Results should report an —for example, a risk ratio, odds ratio, or mean difference—along with its , commonly expressed as a confidence interval. The estimate summarizes the observed relationship; the helps communicate how imprecisely it has been estimated. Neither number, by itself, establishes that activity caused the outcome.
Researchers may also ask whether outcomes vary across increasing activity levels, a , or whether results differ across subgroups. These analyses can add useful detail, but their interpretation still depends on the study and its limitations.
Judge causal evidence cautiously
An association is not proof of cause and effect. Interpret an estimate alongside the design, activity and health measurements, sampling, potential confounding, and alternative explanations. For example, a cohort can show that measured activity preceded a new diagnosis, but differences between activity groups may still explain some or all of the association. Random assignment in a trial tends to balance other characteristics and strengthens causal inference, yet a trial can still have limitations, including those related to its measurements, participants, or follow-up.
A pattern that remains consistent across sensible analyses can strengthen the evidence, but it does not erase concerns about measurement, selection, or other explanations. The sound conclusion is therefore calibrated: describe what the comparison found, report its , and explain how the study’s design and limitations affect what can be inferred.
: A research question that identifies the population, activity, health outcome, and time period.