A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Correlational designs measure two or more variables without manipulation to describe relationships. Longitudinal designs follow individuals over time to infer developmental trajectories. Observational designs systematically record behavior in natural or structured settings. None establishes causality directly, but each provides rich descriptive and predictive information, often with higher ecological validity than experiments.
Identify correlational, longitudinal, and observational elements in published studies. Discuss how longitudinal data can suggest (but not prove) causality. Design an observational study with clear coding protocols.
From your study of research design, you know that the randomized experiment is the gold standard for establishing causation — random assignment eliminates confounding, and the only systematic difference between groups is the manipulation. But the real world constantly presents questions where random assignment is impossible, unethical, or deeply artificial. You cannot randomly assign children to be raised in poverty or affluence. You cannot randomly assign people to smoke for 20 years. You cannot randomly assign a nation to experience civil war. For these questions — often the most important questions in psychology — correlational and observational designs are not a fallback. They are the appropriate tool.
A correlational design measures two or more variables and assesses the statistical relationship between them. A correlation coefficient tells you the direction and strength of association, but it cannot, by itself, distinguish among three causal stories: X causes Y, Y causes X, or a third variable Z causes both. This is the third-variable problem, and it is why the mantra "correlation does not imply causation" exists. But correlational data are not useless for causal thinking — they are the first filter. If childhood poverty is not correlated with adult health outcomes at all, we have strong evidence against a causal relationship. A significant correlation is necessary but not sufficient for causality. Statistical techniques like partial correlation (controlling for Z) and structural equation modeling can test whether a causal model is *consistent with* the data, even if they cannot definitively confirm it.
Longitudinal designs add the dimension of time, which provides partial leverage on the causation question. The minimum requirement for X to cause Y is that X must precede Y. A cross-sectional study (measuring X and Y at the same moment) cannot establish temporal precedence. A longitudinal design, by measuring X at time 1 and Y at time 2, can. Cross-lagged panel models extend this by measuring both X and Y at multiple time points, allowing researchers to test whether X at time 1 predicts Y at time 2 *controlling for* Y at time 1 (and vice versa). When prior-period self-prediction is controlled, a remaining cross-lagged effect is evidence for a directional relationship. It is still not proof of causation — unmeasured confounders remain possible — but the argument is substantially stronger than a cross-sectional correlation.
Observational designs shift the focus from variables to behaviors, recording what people actually do in natural or structured settings rather than asking them to report it. Their strength is ecological validity: you are measuring the phenomenon in its real context, not in a laboratory approximation. Systematic observation requires clear operational definitions of the target behavior, reliable coding schemes, and attention to the observer's potential effect on what is being observed (reactivity). Used well, observational data can capture phenomena — the precise timing of parent-infant interaction, the structure of peer conflict on a playground — that no questionnaire or experiment can replicate. The limitation is that observation describes what happens without explaining why, making it most powerful in combination with correlational or experimental follow-up.
The critical insight is that "non-experimental" does not mean "weak." These designs are weaker for answering the question "does X cause Y?" than a well-run experiment. But they are often *stronger* for answering questions like "how does X relate to Y in the real world?", "what patterns of behavior characterize this population?", and "does this relationship persist over years?" Choosing a design means matching the design to the question — and the questions that most of psychology cares about are ones where experimental control is either impossible or would destroy the very phenomenon under investigation.
No topics depend on this one yet.