Choosing the right research design is probably the most important decision you will make in any study, and also the one with the least room for subsequent correction. You can change the statistical analysis after collecting the data, you can reformulate the hypotheses, you can even completely rewrite the theoretical framework. But you cannot change the design. If you chose a cross-sectional design when you needed a longitudinal one, or an observational design when you needed an experimental one, no statistical analysis can compensate for that limitation.
In psychology, research designs can be organized along several dimensions. The most fundamental is the distinction between experimental designs (where the researcher manipulates at least one variable and randomly assigns participants to conditions) and observational designs (where the researcher measures variables as they occur naturally, without manipulation or randomization). This distinction has direct implications for the conclusions you can draw: only experimental designs allow legitimate causal inferences, because only they control the threats to internal validity that prevent attributing observed effects to the manipulated variable.
Experimental and quasi-experimental designs
The classic experimental design with random assignment and a control group is the gold standard for establishing causal relationships. If you randomly assign 100 participants to a treatment group or a control group, and the treatment group shows a significantly greater improvement, you can conclude with considerable confidence that the treatment caused the improvement. Randomization ensures that, in expectation, the groups are equivalent on all variables before the intervention, including variables you have not measured. No observational design can offer this guarantee.
However, experimental designs are not always possible or ethical. You cannot randomly assign people to experience trauma in order to study post-traumatic stress disorder. You cannot manipulate a child's attachment style to see how it affects their social development. In these situations, quasi-experimental designs offer a reasonable alternative. They compare groups that naturally differ on the variable of interest (for example, people who have experienced a traumatic event versus people who have not), but without the guarantee of equivalence that randomization provides. This means that there is always the possibility that the observed differences are due to confounding variables, and this limitation must be explicitly acknowledged.
Cross-sectional vs. longitudinal designs
Another crucial dimension is temporal. Cross-sectional designs collect data at a single point in time, while longitudinal designs follow the same participants over time. Cross-sectional designs are faster and cheaper, but they have severe limitations for studying processes of change, development, or causality. If you find a correlation between work stress and depression in a cross-sectional study, you cannot know whether stress causes depression, depression causes the perception of stress, or both are consequences of a third factor.
Longitudinal designs allow you to address some of these limitations, although they are not a perfect solution either. A longitudinal panel design, where you measure the same variables in the same participants on multiple occasions, allows you to assess whether changes in one variable predict subsequent changes in another (known as cross-lagged effects). This comes closer to causality than a simple cross-sectional correlation, although it does not establish it definitively because it does not control for all possible confounding variables that vary over time.
Longitudinal studies have their own practical challenges: participants drop out of the study (attrition), the passage of time can produce changes unrelated to your variables of interest (maturation), and repeated exposure to measurement instruments can alter responses (practice or sensitization effects). These problems do not invalidate longitudinal designs, but they require attention in both the design and analysis of the data.
How to choose the right design
The choice of design should derive directly from your research question. If your question is causal ("Does this treatment reduce anxiety?"), you need an experimental or at least quasi-experimental design. If your question is about associations ("Are self-esteem and academic performance related?"), a cross-sectional correlational design may be sufficient as a first step, although a longitudinal design would be more informative. If your question is about processes of change ("How does adaptation evolve after a divorce?"), you need a longitudinal design with multiple measurement points.
It is also important to consider practical limitations. A 5-year longitudinal design with monthly measurements would be ideal for many questions, but it may be unfeasible due to time, budget, or attrition constraints. In these cases, the key is to be honest about the limitations of the design you can implement and not draw conclusions that your data cannot support. A well-executed and honestly interpreted cross-sectional design is better science than a poorly planned longitudinal design or an experimental design with implementation problems. What matters is not that your design is perfect (none is), but that it is coherent with your question and that you interpret the results within the limits that the design allows.
That coherence between question and design is exactly what the Methodologist checks: you paste the project before collecting a single data point and it flags the design problems worth fixing, whether the sample size supports what you want to estimate, and whether the analysis plan really answers the question you wrote. It is the moment when changing the design is still free.
And if your project has to go through a research ethics committee, the ethics committee simulator anticipates the objections they will raise, which paragraphs are missing from the application, and whether you are working with a vulnerable population without having said so. It does not replace real approval, but it saves the round of corrections that delays fieldwork by a month.
If you are choosing the design for your dissertation and have doubts about which one best fits your research question, my statistical support for doctoral theses reviews this decision in the design phase, before a methodological mistake becomes irreversible.