Definition
Experimental design is the discipline of planning a study so that its results can support a causal conclusion. The designer specifies which variable will be manipulated (the treatment or independent variable), which outcome will be measured (the dependent variable), how participants or units will be assigned to conditions, what other variables will be held constant or balanced across groups, and how many observations will be collected. Each of these choices is a defence against a specific threat to the inference.
The classical building blocks come from R. A. Fisher's work on agricultural trials: randomisation, replication, blocking, and controlled comparison. The same logic now underwrites randomised controlled trials in medicine, A/B testing in software, controlled field experiments in development economics, and laboratory studies across every empirical discipline. Where a question can be settled by deliberate intervention, a well-designed experiment is the most efficient way to settle it.
Why it matters
How it works
A standard randomised experiment runs in five stages. The researcher first defines the question and chooses an outcome that can be measured cleanly. They specify a treatment whose effect on that outcome is being tested, and an appropriate control condition. They calculate the sample size needed to detect a plausible effect with adequate statistical power. They then randomly assign units to the treatment and control groups, ideally with blinding in both directions. Finally, they collect data, analyse the difference between groups using a method specified in advance, and report effect sizes alongside any test of significance.
Randomisation is the engine that makes the design work. By assigning units to conditions independently of any of their other characteristics, it ensures that the treatment and control groups differ systematically only in the treatment itself. Any subsequent difference in outcomes can therefore be attributed to the treatment with quantified uncertainty. Replication — running the same experiment multiple times, ideally by independent teams — guards against the chance that a single significant result was a fluke. Blocking, stratification, and factorial designs let a single study answer more than one question without losing the protections that randomisation provides.