Research Design

Research design is the overall strategy, framework, and structure that a researcher uses to address a research question, encompassing the type of study (experimental, observational, or qualitative), the data collection methods, the sampling approach, and the analytical plan.^[1] It provides the logical architecture that connects the research question to the evidence needed to answer it.

The choice of research design determines the types of conclusions that can be drawn from a study. Experimental designs (particularly randomized controlled trials) support causal inference. Observational designs (cohort, case control, cross sectional) identify associations but are limited in their ability to establish causation. Qualitative designs provide depth and context but do not produce generalizable numerical estimates. Selecting an inappropriate design for the research question wastes resources, produces inconclusive evidence, and can mislead readers. Conversely, a well chosen and well executed design maximizes the study's contribution to knowledge while efficiently using available resources.

Research design decisions proceed in a logical sequence. The research question comes first, determining whether the study should be descriptive (what is the prevalence?), associational (what factors are related to the outcome?), or causal (does the intervention cause the outcome?). The type of question guides the choice of design. Questions about experiences, meanings or social processes (how do parents experience a new school program?) call for qualitative designs, and mixed methods designs combine qualitative and quantitative components when a question needs both.^[1]

For causal questions, the randomized controlled trial is the gold standard. When randomization is not feasible, quasi experimental designs (such as regression discontinuity, interrupted time series, or difference in differences) provide alternatives with varying levels of causal credibility. For associational questions, cohort and case control studies are the primary designs; cross sectional studies can also estimate associations, but because exposure and outcome are measured at the same time they cannot show which came first.^[2] For descriptive questions, cross sectional surveys and descriptive epidemiological studies (such as analyses of surveillance or registry data) are appropriate.

Within each broad category, specific design features are selected: prospective versus retrospective data collection, individual versus cluster level allocation, parallel versus crossover treatment assignment, and single versus multicenter implementation. Each choice involves tradeoffs between internal validity, external validity, feasibility, and cost.

The data collection methods (surveys, clinical measurements, observations, medical records review) and the sampling strategy (probability or nonprobability) are integral parts of the design. The statistical analysis plan, specified in advance as part of the protocol, ensures that the analytical methods match the design and the type of data collected.

Consider research design at the earliest stage of study planning, before any data are collected. The design should be determined by the research question, the level of evidence needed, the feasibility of different approaches, and ethical considerations. For intervention studies, use the strongest design that is ethically and practically feasible. For observational studies, select the design that best controls for confounding and bias. For exploratory or hypothesis generating research, qualitative designs, mixed methods or exploratory quantitative work (such as pilot surveys or analyses of existing datasets) may be most appropriate.

A team wants to determine whether a new school lunch program reduces childhood obesity. After considering the options, they select a cluster randomized trial, randomly assigning 40 schools (20 treatment, 20 control) and measuring age and sex standardized body mass index (BMI z score) at baseline and after one academic year. They choose cluster randomization because the intervention operates at the school level and individual randomization is impractical. The sample size is calculated to account for clustering (intraclass correlation = 0.03), and the analysis uses multilevel modeling to account for the nested data structure.