Repetition and replication both involve doing something again, but in science they refer to different things with different purposes. Repetition (often called repeatability) means performing the same measurement or procedure again under the same conditions, typically within the same lab, on the same equipment, by the same person. Replication means testing the same question independently, with new samples, different researchers, or a separate laboratory. The distinction matters because repetition tells you whether a measurement is consistent, while replication tells you whether a finding is real enough to generalize beyond the original setup.
How the Two Concepts Work in Practice
Think of it this way. If you step on a bathroom scale three times in a row and get three readings, that is repetition. You are checking whether the scale gives a stable number under the same conditions. If three different people each weigh themselves on three different scales in three different rooms, and the scales all agree that a known weight is accurate, that is closer to replication. The finding (the weight reading) has been confirmed under varied, independent conditions.
In a laboratory experiment, repetition often looks like running the same assay on the same sample multiple times. These repeated measurements help you estimate measurement error and give you confidence that a single reading is not a fluke of the instrument. Replication, on the other hand, means gathering new samples, running the experiment from scratch, and checking whether you reach the same conclusion. A result that replicates across independent experiments carries far more weight than one that merely repeats well on a single sample.
Biological Versus Technical Replicates
Biomedical research has its own version of this confusion, and it shows up constantly in published papers. A “technical replicate” is essentially a repetition: you take a single blood sample and measure it three times, or you split one tissue sample into three aliquots and run each through the same assay. A “biological replicate” is closer to true replication: you take samples from three different patients, or from three different mice, and run the experiment on each one independently. The distinction has become so important that journals and scientific societies have published numerous guidelines on how researchers should plan studies and report data based on which type of replicate they used.1PubMed Central. Distinguishing Between Biological and Technical Replicates in Hypertension Research on Isolated Arteries
The reason this matters is statistical. If you measure the same sample three times and treat those three numbers as independent data points, you are inflating your sample size. Your statistics will look more impressive than they should, because the variation between technical replicates is almost always smaller than the variation between genuinely independent biological samples. A study reporting “n = 9” when it actually measured three samples three times each has an effective sample size of three, not nine.
Direct Replication Versus Conceptual Replication
Even within the umbrella of replication, researchers draw a further distinction. A direct replication follows the original study’s methods as closely as possible, using different samples of the same or similar materials but otherwise trying to recreate the original procedure step by step. The goal is to test whether the original result is reproducible when you simply run the experiment again. A conceptual replication takes a different approach entirely: it tests the same underlying hypothesis using a different method, a different experimental technique, or a different population. The goal here is to see whether the explanation for a finding holds up even when you change how you test it.2PubMed Central. Making sense of replications
Both types serve science, but they answer different questions. Direct replication asks: can this result be reproduced? Conceptual replication asks: is the theoretical explanation correct, regardless of the specific method used? A finding that passes direct replication might still be wrong if the original method had a hidden confound that the replication faithfully preserved. A finding supported by conceptual replications across multiple methods gives stronger confidence in the underlying explanation, but it does not guarantee that any single original experiment was done correctly.2PubMed Central. Making sense of replications In practice, a well-supported finding ideally has both: it reproduces when you follow the original recipe, and it holds up when you test the same idea with different tools.
The Pseudoreplication Trap
One of the most common errors in experimental design is pseudoreplication, which happens when researchers treat non-independent observations as if they were independent replicates. The term was defined in a landmark ecology paper as the use of statistical tests for treatment effects with data where either treatments are not truly replicated or the replicates are not statistically independent of one another.3Ecological Monographs. Pseudoreplication and the Design of Ecological Field Experiments
A classic example: suppose you want to test whether a fertilizer increases plant growth. You apply the fertilizer to one plot and leave another untreated, then measure 30 plants in each plot. You have 60 measurements, but you do not have 60 independent replicates. The plants within each plot share the same soil, the same microclimate, and the same local conditions. They are not independent of one another. Your real sample size is two plots, not 60 plants. Treating those 60 plants as independent data points produces artificially small p-values and overconfident conclusions. This is not a rare mistake confined to student projects; pseudoreplication shows up in published research across fields, and it can quietly invalidate results if the non-independence among observations is not handled properly.4Experimental Design for the Life Sciences. Pseudoreplication
Reproducibility in Computational Work
When the research in question involves computation rather than wet-lab experiments, the terminology shifts slightly. Reproducibility, in the computational sense, means that someone can take your original data and your original code and get the same results you reported. Replicability still means what it means elsewhere: collecting new data and seeing whether the same conclusions emerge. This distinction was formalized in a widely cited framework from the U.S. National Academies of Sciences, Engineering, and Medicine, and it has become standard vocabulary in data science and computational fields.
Computational reproducibility might sound trivial, but it often is not. Code breaks, software versions change, random-number seeds get lost, data-cleaning steps go undocumented, and the gap between “I ran this analysis” and “someone else can run this analysis and get the same output” can be surprisingly large. The push for reproducibility in computational work is less about questioning whether the finding is true and more about making sure the work is transparent enough to be checked.
What Large-Scale Replication Projects Have Found
The urgency behind these distinctions became vivid in the 2010s, when several major replication efforts put well-known findings to the test. The most famous of these, conducted across 100 psychology studies, found that replication effect sizes were about half the magnitude of the original effects. While 97% of the original studies had reported statistically significant results, only about 36% of the replications reached significance.5PubMed. Estimating the reproducibility of psychological science That gap is striking. It does not mean most psychology is wrong, but it does mean that many published effect sizes were inflated, and some findings that seemed solid were probably noise.
Cancer biology saw similar results. A large project attempting to replicate preclinical cancer studies found that for positive effects, the replication effect sizes were roughly 85% smaller than the originals, and about 92% of replication effect sizes came in below the original. Taking positive and null effects together, about 46% of the replications were judged successful.6PubMed Central. Investigating the replicability of preclinical cancer biology These are not fringe findings being retested. These were high-profile, highly cited papers from top journals.
One factor that predicted replication success in the psychology project was the strength of the original evidence: studies with larger original effects and stronger statistical support replicated better than those scraping by just below the significance threshold.5PubMed. Estimating the reproducibility of psychological science This is not surprising in hindsight, but it underscores why the difference between repetition and replication matters so much. A finding repeated in one lab may look robust locally because the same measurement error, the same quirks of the sample, and the same subtle biases all carry over. Only when independent replication strips those shared conditions away does the true signal emerge.
When Replication Fails, Is It Always a Problem?
Not every failed replication means the original finding was wrong. Some research topics are inherently sensitive to context. A psychology experiment that depends on cultural norms, local language nuances, or the specific population tested may not replicate in a different country or demographic for legitimate reasons. Research has found that the contextual sensitivity of a topic is associated with replication success even after adjusting for methodological differences like statistical power and effect size.7PubMed Central. Contextual sensitivity in scientific reproducibility In other words, some findings are genuine but not universal. They reflect something real about one population or setting that does not transfer to another.
This is where the direct-versus-conceptual replication distinction becomes practically useful. A direct replication that fails in a different country might indicate boundary conditions, not fraud or error. A conceptual replication that tests the same hypothesis with a completely different method and finds support for it in that different population would then suggest the underlying idea is sound even if the specific experimental result does not travel well. The two types of replication, used together, paint a much richer picture than either does alone.
Statistical Power and the False Positive Problem
Behind many replication failures lies a more fundamental issue: the original studies were underpowered. When a study has low statistical power, it needs a lucky break to produce a significant result. That means the studies that do squeak past the significance threshold are disproportionately likely to have overestimated the real effect or to have caught a statistical mirage. Large surveys of published research reveal that statistical power is often inadequate, and when the research record also includes selective reporting or publication bias, even conventional meta-analyses can be unreliable. Research has found that high median retrospective power (above roughly 50% to 60%) is associated with credible, replicable findings, while meta-analyses built on low-power studies should be interpreted with great caution or discounted altogether.8PubMed. Retrospective median power, false positive meta-analysis and large-scale replication
This connects back to the repetition-versus-replication distinction in a practical way. Repeating a measurement within a single low-powered study does not fix the power problem. It just gives you a more precise estimate of whatever that one study happened to find, noise and all. Replication with new, adequately powered studies is the only real corrective. This is why the scientific community’s growing emphasis on preregistered, well-powered replication studies is not just methodological housekeeping but a genuine shift in how we decide what counts as established knowledge.
Why Replication Studies Are So Rare
If replication is so important, you might wonder why it does not happen more often. The short answer is that the incentive structure of science punishes it. Under the current academic reward system, a researcher choosing between writing an original paper and conducting a replication study faces lopsided returns. The expected citation payoff for replication work is lower than for original research, and citations drive hiring, tenure, and grant funding.9Research Policy. Replication studies in economics—How many and which papers are chosen for replication, and why? Journals, too, have historically favored novel findings over confirmatory work. A researcher who spends a year carefully replicating someone else’s study may end up with a result that is both less publishable and less career-enhancing than a quick-and-dirty original finding, even though the replication contributes more to the reliability of the overall knowledge base.
Some fields have begun to push back. Registered Reports, a publishing format where journals commit to publishing a study based on its design before the results are known, have grown in popularity across psychology, neuroscience, and other disciplines. Several journals now explicitly invite replication studies. Funding agencies in some countries have started earmarking grants for replication work. These are still relatively small-scale changes compared to the overall volume of research, but they represent a genuine recognition that the traditional reward system was failing to produce the verification that science depends on.
How Measurement Error Fits In
One practical reason researchers repeat measurements within a study is to reduce measurement error. If a biological variable like blood pressure or cholesterol fluctuates from moment to moment, a single reading gives you the true value plus some random noise. Taking multiple readings and averaging them shrinks that noise. Research on epidemiological studies has confirmed that using the average of two error-prone measurements slightly decreases measurement error compared to relying on a single reading.10PubMed Central. Reflection on modern methods: calculating a sample size for a repeatability sub-study to correct for measurement error in a single continuous exposure This is a perfectly legitimate use of repetition. The key is understanding what it does and does not accomplish: it makes your individual measurements more precise, but it does not tell you whether the relationship you find between variables would hold up in a different group of people.
This is maybe the cleanest summary of the difference. Repetition improves precision. Replication tests generalizability. Both are necessary, and neither substitutes for the other. A study with beautifully precise measurements that never replicates in another lab is not much use. A study that replicates across labs but uses wildly imprecise instruments each time is messy and hard to interpret. The strongest evidence comes when both precision and generalizability have been demonstrated.
Terminology Varies by Field, and That Causes Confusion
One genuinely frustrating aspect of this topic is that different scientific fields use these terms differently. In metrology and engineering, “repeatability” refers to variation under identical conditions (same operator, same instrument, short time span) while “reproducibility” refers to variation under changed conditions (different operator, different instrument, different lab). In computational science, “reproducibility” means getting the same results from the same data and code, while “replicability” means getting consistent results from new data. In the life sciences, the biological-versus-technical replicate distinction maps roughly onto replication-versus-repetition but with its own vocabulary. And in everyday scientific conversation, researchers sometimes use “replication” and “reproducibility” interchangeably, even though formal definitions distinguish them.
If you are reading across disciplines, the safest approach is to look at what is actually being described rather than relying on the label. Is the researcher talking about redoing a measurement on the same sample? That is repetition, regardless of what they call it. Are they talking about testing the same hypothesis with new, independent data? That is replication. Are they talking about re-running someone else’s analysis code on the original dataset? That is computational reproducibility. The concepts are stable even when the labels drift.
When Industry Relies on the Distinction
Outside of academic research, the repetition-replication distinction shows up in applied settings where measurement quality has direct financial or safety consequences. In manufacturing, gauge repeatability and reproducibility studies (often called gauge R&R) are a standard quality-control tool. Repeatability asks whether the same operator measuring the same part with the same gauge gets consistent readings. Reproducibility asks whether different operators measuring the same part get consistent readings. If repeatability is poor, the instrument needs calibration or replacement. If reproducibility is poor, the operators need better training or the measurement protocol needs standardization.
In pharmaceutical development, the distinction is built into regulatory expectations. A drug company must show that an assay gives repeatable results within a single run (intra-assay precision) and that results replicate across different runs, different days, and different analysts (inter-assay precision). Regulatory bodies treat these as separate validation criteria, each with its own acceptance thresholds. A test that passes one but fails the other does not get approved. The logic is the same as in basic science, just formalized with regulatory teeth: consistency within a run is not enough if the result falls apart when conditions change even slightly.