In experimental design, replication is the application of a treatment or treatment combination to more than one experimental unit. It provides information about variation among units receiving the same treatment, allowing treatment effects to be distinguished from background variability. Genuine replication occurs at the level at which treatments are independently applied or assigned; taking several measurements from a single treated unit does not ordinarily create additional treatment replicates.
Experimental units and observations
An experimental unit is the smallest unit to which a treatment can be independently assigned. Depending on the experiment, it may be a person, an agricultural plot, a manufacturing batch, a culture vessel, a classroom, or an entire building. The unit observed or measured need not be the experimental unit.
For example, suppose a fertilizer is randomly assigned to agricultural plots, and ten plants are measured within each plot. The plots are the experimental units; the plants provide observations within those units. Five independently treated plots per fertilizer constitute five treatment replicates, not fifty.
The distinction depends on the actual assignment procedure. If individual plants in separate pots receive independently assigned treatments, the pots may be experimental units. If treatment is applied through a shared irrigation system, the relevant experimental unit may instead be the group of pots served by that system.
Replication therefore concerns the structure of an experiment, not simply the number of rows in a dataset. Its interpretation depends on independence, shared sources of variation, and the level at which treatment comparisons are made.
Statistical purposes
Estimating experimental error
Units receiving the same treatment usually produce different outcomes. These differences can arise from biological heterogeneity, material variation, environmental conditions, and measurement error. Replication supplies information for estimating the variation relevant to treatment comparisons, often called experimental error.
In a simple experiment with treatments and independent replicates per treatment, a one-way model is
where is an overall mean, is the effect of treatment , and is a random error. A conventional identifying constraint is .
If errors are independent, have mean zero, and have common variance , variation within treatments provides an estimate of . In a one-way analysis of variance, the within-treatment error has degrees of freedom. Normality is an additional assumption for the usual exact finite-sample tests.
Without replication within treatment combinations, pure within-treatment error cannot be estimated directly from those combinations. An analysis may still be possible using assumptions about the response model, negligible interactions, or an independently justified error estimate, but the uncertainty then depends on those additional assumptions.
Improving precision
For independent observations with common variance , the sample mean has variance
Its standard error is therefore . Increasing replication improves precision, but with diminishing returns: under this model, halving the standard error requires four times as many replicates.
For two independent treatment groups,
This relationship underlies the construction of confidence intervals and many sample-size calculations. Other things being equal, more genuine replicates increase statistical power to detect a specified effect size.
These gains do not imply that replication removes systematic errors. Increasing the number of units does not by itself eliminate a biased instrument, an inappropriate control, or confounding between treatment and another factor.
Replication, randomization, and blocking
Replication is commonly presented alongside randomization and blocking as a central principle of experimental design. These principles have distinct functions:
- Replication supplies information about variability and increases precision.
- Randomization determines treatment allocation through a chance mechanism, supporting treatment comparisons that are not systematically tied to pre-existing differences.
- Blocking groups similar units and compares treatments within those groups, reducing the influence of known nuisance variation.
A field experiment may place one plot of each treatment in every block, with treatment positions randomized separately within blocks. Treatment replication is then distributed across blocks. The analysis must account for the block structure rather than treat all observations as an unstructured collection.
Modern accounts of these principles owe much to Ronald A. Fisher and the development of agricultural experimentation in the early twentieth century. Their scope now extends to laboratory, industrial, ecological, and behavioral experiments.
Forms and levels of replication
Terminology differs across disciplines, so descriptions of the experimental structure are more informative than labels alone.
Biological and technical replicates
Biological replicates generally refer to distinct biological samples or experimental units intended to represent biological variation. Technical replicates are repeated measurements or processing steps applied to the same underlying sample, such as repeated instrument readings or multiple assay wells prepared from one extract.
Technical replication can characterize measurement variability and improve the precision of a sample-level result. It does not necessarily provide information about variation among independently treated organisms or preparations.
The distinction is contextual. Separate cultures independently assigned to treatments may be experimental replicates even if they share an original source. Conversely, several samples taken from one treated organism may remain subsamples rather than independent treatment replicates.
Replication across batches, sites, and times
Treatments may be repeated across production runs, laboratory days, locations, seasons, or operators. Such replication can reveal variation that an experiment conducted in a single setting cannot estimate.
The design determines what can be learned. If each treatment occurs in every batch, treatment effects can be compared within batches. If one treatment occurs only in one batch and another treatment only in a different batch, treatment and batch are confounded.
Replication across settings can also reveal interactions: a treatment effect may differ by site or season. An average effect and the consistency of that effect across settings are separate quantities.
Replicated factorial experiments
In a factorial experiment, replication usually means applying each treatment combination to multiple experimental units. This supports estimation of error separately from the modeled main effects and interactions.
An unreplicated factorial experiment can still estimate selected effects, but testing them may require assumptions about which interactions are negligible. Unused model terms are not automatically equivalent to independently observed experimental error.
Pseudoreplication and correlated observations
Pseudoreplication occurs when observations that do not provide independent replication for a treatment comparison are analyzed as though they do. Common examples include:
- treating many organisms in one treated tank as independent treatment replicates;
- counting repeated measurements of one unit as independently treated units;
- comparing one treated site with one untreated site while treating samples within each site as independent treatment replicates.
For instance, one treated aquarium and one control aquarium do not provide replicated aquarium-level treatments, even if each contains many fish. Differences between aquariums cannot be separated from treatment using that comparison alone.
Correlation illustrates why observation counts can exaggerate information. Suppose there are independent clusters, each containing observations, with common marginal variance and common within-cluster correlation . Then
Relative to independent observations, the variance is multiplied by the design effect
For positive , collecting additional observations within existing clusters provides less information than the same number of independent observations. The associated effective sample size, , is a variance-based approximation for this particular structure, not a universal measure of replication.
Correlated data are not inherently invalid. A mixed-effects model, a cluster-level analysis, or another method matched to the design can account for dependence. Such methods cannot, however, create independent treatment assignments that the experiment never contained.
Replication versus repeated measurements
Repeated measurements record outcomes from the same unit at multiple times or under multiple conditions. They can improve precision by revealing within-unit changes, but they create a dependence structure that must be represented in the analysis.
If twenty participants each provide ten readings after assignment to a treatment group, the experiment has twenty participant-level experimental units, not two hundred independently treated units.
In crossover or within-subject experiments, the same unit may receive several treatments in a randomized sequence. Here the treatment contrast can be defined within units, and its uncertainty depends on within-unit variation, treatment order, and possible carryover effects. Thus, “one unit, one observation” is not a general requirement; the essential requirement is to match the analysis to the assignment and dependence structure.
Allocation and practical limits
Under a fixed budget, experimental design involves a trade-off between increasing the number of independent units and measuring each unit more thoroughly. Additional independent units improve estimation of between-unit variation, whereas subsampling and technical repetition can reduce uncertainty in each unit’s measured outcome.
For two independent groups with equal variances and equal per-unit costs, balanced allocation minimizes the variance of their mean difference for a fixed total number of units. Different variances or costs can favor unequal allocation.
No fixed number of replicates is adequate for every experiment. Relevant considerations include the target treatment contrast, expected variation, desired precision or power, dependence among observations, and the number of settings over which conclusions are intended to apply.
Replication also does not guarantee causal inference or broad generalizability. Many replicates within a narrow set of conditions can estimate an effect precisely for those conditions while providing little direct evidence about other populations, environments, or treatment implementations.
Distinction from replication of a study
Within-experiment replication differs from conducting a new study to examine an earlier finding. A replication study collects new evidence, potentially using a closely matched procedure or a different operationalization of the same hypothesis. It may assess the stability of an effect across samples, investigators, or settings.
The two meanings are related but not interchangeable. A single experiment can contain many treatment replicates without having been repeated by another research group. Conversely, a later study can repeat an earlier procedure while retaining inadequate replication at the treatment-assignment level.
Combining independent studies through meta-analysis addresses variation between studies as well as uncertainty within them. It does not replace the need for each study to identify its experimental units and estimate uncertainty at the appropriate level.