An experiment is a systematic investigation in which researchers deliberately establish or vary conditions and observe the resulting outcomes. Experiments can test a hypothesis, measure a quantity, compare alternatives, or discover relationships not previously recognized. They are an important component of the scientific method, especially when questions concern how changing one condition affects another. Their evidential value depends on the arrangement of comparisons, the quality of measurements, and the assumptions connecting observations to conclusions. (itl.nist.gov)
Purpose and scope
Experimental investigation differs from purely observational research because the investigator intervenes in the system being studied rather than only recording existing conditions. This intervention can help distinguish causation from correlation: two quantities may vary together without one producing changes in the other. Nevertheless, an intervention alone does not establish a causal effect; the experiment must provide a credible comparison and address alternative explanations. (nobelprize.org)
Experiments may be exploratory, identifying promising factors or unexpected behavior, or confirmatory, evaluating predictions specified beforehand. Other experiments optimize a manufacturing process or determine which operating conditions produce a desired response. These purposes influence the choice of factors, measurements, and analysis. Experimental design is therefore not simply a procedure for collecting data, but a plan connecting the research objective to interpretable evidence. (itl.nist.gov)
Variables, treatments, and experimental units
A factor is a condition whose influence is investigated, and its levels are the values or categories included in the experiment. A treatment may consist of one factor level or a combination of levels. The response, sometimes called the dependent variable, is the measured outcome. For example, an investigation might vary processing temperature and record product yield. Factors of interest must be distinguished from nuisance factors, which affect the response but are not the main subject of investigation. (itl.nist.gov)
The experimental unit is the entity to which a treatment is independently assigned, such as a specimen, production batch, person, or agricultural plot. Multiple measurements from one unit are not necessarily independent treatment replications. Recognizing the correct unit matters because analyses based on an exaggerated number of independent observations can understate uncertainty. Measurements, treatment allocation, and the structure of variation must therefore be considered together. (itl.nist.gov)
Controls, randomization, and replication
A control condition supplies a reference against which an intervention is evaluated. Depending on the question, it may involve no intervention, an established treatment, or a different operating condition. Comparisons become difficult to interpret when treatment groups also differ systematically in other relevant respects. Such confounding can make a response attributable to several explanations rather than to the intended factor alone. (nobelprize.org)
Randomization uses a chance mechanism to allocate treatments or determine the order of experimental runs. It helps prevent systematic associations between treatments and uncontrolled influences, although it does not guarantee identical groups in a particular experiment. Blocking addresses known sources of variation by grouping similar units and comparing treatments within those groups. For example, treatments may be compared within production batches rather than across batches made under different conditions. (itl.nist.gov)
Replication repeats treatment applications across experimental units, providing information about variability and improving the precision of estimated responses. Repeated instrument readings can assess measurement variation, whereas independent experimental runs also capture differences between units or occasions. Replication, randomization, and blocking serve distinct purposes and are often combined rather than treated as interchangeable safeguards. (itl.nist.gov)
Major experimental designs
A completely randomized design assigns treatments across the available units without forming blocks. A randomized block design incorporates groups defined by a nuisance factor. A factorial design investigates combinations of multiple factors, allowing researchers to examine both individual effects and interactions. An interaction occurs when the effect of one factor depends on the level of another. Fractional factorial designs investigate selected combinations, reducing the required number of runs at the cost of restrictions on which effects can be separately estimated. (itl.nist.gov)
A randomized controlled trial applies random assignment to comparisons of interventions, including studies in medicine, education, and economics. A natural experiment, by contrast, exploits circumstances outside the investigator’s control that create potentially informative differences in exposure. Despite its name, it is not an investigator-assigned experiment; its use in causal inference depends on assumptions about how those differences arose. (nobelprize.org)
Analysis and interpretation
Statistics connects experimental observations to estimates and uncertainty. Depending on the design, researchers may use analysis of variance, regression, or other models to compare responses. Analysis must reflect treatment allocation, blocking, replication, and dependence among observations. Otherwise, apparently precise results can rest on an inappropriate representation of the experiment. (itl.nist.gov)
Hypothesis tests and confidence intervals address different aspects of inference. A p-value expresses the incompatibility of observations with a specified statistical model; it is not the probability that the tested hypothesis is true. Nor does statistical significance measure effect size or practical importance. Interpretation consequently involves the estimated magnitude, uncertainty, experimental context, and adequacy of the assumptions, rather than a significance threshold alone. (doi.org)
Verification and ethical constraints
Detailed records of procedures, materials, data processing, and analysis enable others to scrutinize experimental findings. Terminology varies, but the US National Academies distinguishes reproducibility—obtaining consistent computational results using the same data and methods—from replicability, which concerns consistent findings from studies collecting new data. Agreement need not mean numerically identical outcomes, because experiments include uncertainty and variation. (nationalacademies.org)
Research ethics places constraints on what experiments may be conducted and how participants are treated. The 1979 Belmont Report identifies respect for persons, beneficence, and justice as principles for human-subject research, connecting them to informed consent, assessment of risks and benefits, and fair participant selection. Animal research has a distinct framework commonly expressed as replacement, reduction, and refinement: using alternatives where possible, obtaining robust information with appropriately limited animal numbers, and minimizing suffering while improving welfare. (hhs.gov)