{
  "id": "preregistration",
  "title": "Preregistration",
  "lang": "en",
  "source_lang": "en",
  "summary": "Preregistration is the time-stamped recording of research questions, methods, and analysis plans before the relevant data or results are examined.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Research preregistration",
    "study preregistration",
    "pre-registration"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/preregistration/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/preregistration/",
      "json": "https://aiwiki.page/zh/preregistration.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Preregistration is the practice of recording a research plan in a time-stamped, preserved record before collecting data or, in some settings, before accessing or analyzing the data relevant to the planned tests. The record typically specifies research questions, hypotheses, study design, outcome measures, and analysis procedures. Its central purpose is to distinguish decisions made independently of the results from decisions informed by them. Used in statistics and empirical research, preregistration supports transparent interpretation of evidence but does not by itself establish that a study is well designed or its conclusions correct."
    },
    {
      "heading": "Purpose and statistical rationale",
      "text": "Research commonly involves choices about measurements, exclusions, transformations, statistical models, and which results to report. When these choices depend on observed outcomes, the final analysis may appear more decisive than its statistical justification permits. This can occur without deliberate misconduct: investigators may reasonably explore several approaches but report the selected analysis as though it had been the only one considered.\n\nPreregistration makes the intended analysis independently inspectable. It helps separate confirmatory research, which evaluates specified predictions using planned procedures, from exploratory research, which uses observations to develop questions, hypotheses, or models. Both are legitimate, but they support different interpretations of evidence.\n\nOne concern is multiple testing. Suppose a researcher conducts \\(m\\) independent tests of true null hypotheses, each at significance level \\(\\alpha\\). The probability of obtaining at least one false rejection is\n\n\\[1-(1-\\alpha)^m.\\]\n\nFor \\(m=20\\) and \\(\\alpha=0.05\\), this probability is approximately \\(0.64\\). The calculation assumes independence and does not describe every form of analytical flexibility, but it illustrates why selecting a favorable result from many opportunities can be misleading.\n\nA sufficiently detailed preregistration can identify the intended family of tests, distinguish primary from secondary outcomes, and specify adjustments. It does not itself perform those adjustments. Nor does simply recording a plan guarantee that a reported p-value has its claimed interpretation: that also depends on the design, assumptions, analysis, and adherence to relevant decision rules.\n\nPreregistration can also make hypothesizing after the results are known more detectable when a result-generated hypothesis is presented as an advance prediction. It preserves the distinction between explaining an observed pattern and predicting it independently."
    },
    {
      "heading": "What a preregistration contains",
      "text": "The appropriate level of detail varies across fields and research designs. A substantive preregistration commonly includes:\n\n- Research questions and hypotheses: the relationships or effects under investigation, including directional predictions where relevant.\n- Study design: the population, recruitment procedures, experimental conditions, measurements, and any randomization or masking procedures.\n- Sample size and stopping rules: the planned sample size, its justification, and the conditions under which recruitment or data collection will end.\n- Outcomes and variables: primary and secondary outcomes, their measurement times, scoring procedures, and any composite measures.\n- Data-processing rules: eligibility criteria, observation exclusions, treatment of outliers, transformations, and handling of missing data.\n- Statistical analyses: model specifications, contrasts, covariates, interactions, inferential thresholds, and multiplicity procedures.\n- Contingency plans: procedures for anticipated problems, such as model nonconvergence, assumption violations, or unexpectedly sparse observations.\n\nA sample-size justification may use statistical power, a target precision for a confidence interval, resource constraints, or another explicitly stated criterion. A Bayesian plan may specify the likelihood, prior distributions, model comparisons, and decision criteria.\n\nOperational precision matters. “A regression will test the hypothesis” leaves open the outcome definition, predictors, transformations, interactions, and model family. A more complete plan identifies the intended analysis and explains when alternatives will be used. Analysis code tested on simulated or otherwise independent data can reduce ambiguity."
    },
    {
      "heading": "Timing, preservation, and accessibility",
      "text": "The defining feature is an identifiable boundary between planning and exposure to relevant evidence. Preregistration before data collection usually provides a clear boundary, but the practice also applies to existing datasets.\n\nFor secondary-data research, investigators may register a plan before receiving the data, before accessing restricted variables, or before examining the relationship of interest. Prior familiarity with a dataset matters: a plan written after extensive exploration does not acquire the evidential status of an independent prediction merely because the final statistical test has not yet been run.\n\nPreregistration records therefore often describe:\n\n- whether data already exist;\n- what the investigators have previously accessed or examined;\n- when the plan was registered;\n- which research decisions remained independent of the relevant results.\n\nA registry or repository normally preserves a dated version that cannot be silently overwritten. Some records are immediately public; others are embargoed temporarily. Public availability supports scrutiny, whereas a time stamp establishes timing. These are related but distinct functions.\n\nA private, editable document can support project planning, but it offers weaker independent evidence of what was specified at a particular time."
    },
    {
      "heading": "Deviations, amendments, and exploratory analyses",
      "text": "Preregistration does not prohibit changes. Equipment failures, recruitment difficulties, unexpected data structures, or errors in the original plan may justify revisions. The important distinction is between transparent revision and undisclosed replacement of the original plan.\n\nAn amendment preserves the earlier record, states what changed, and explains the timing and reason for the change. A revision made before relevant results are examined differs from one made after an unexpected pattern becomes visible.\n\nReports can distinguish among:\n\n- analyses conducted as preregistered;\n- preregistered analyses that were modified or omitted, with explanations;\n- additional exploratory analyses.\n\nReporting the planned analysis alongside a justified alternative can clarify whether conclusions depend on analytical choices. Sensitivity analysis can examine such dependence without treating every departure from the plan as inherently invalid.\n\nExploration remains valuable. A new pattern may motivate a hypothesis, but testing that hypothesis on the same observations generally does not provide the same evidential separation as testing it on independent data. A new study or a genuinely untouched holdout dataset can provide a subsequent test. This separation is related to, although not identical with, the separation of training and evaluation data in machine learning."
    },
    {
      "heading": "Related research practices",
      "text": "Trial registration and analysis plans\n\nProspective registration became particularly prominent in clinical research, where registries document studies and make their existence and intended outcomes discoverable. It helps identify completed studies whose results have not been published and can expose discrepancies between registered and reported outcomes.\n\nRegistration of a randomized controlled trial may include its design, eligibility criteria, interventions, and outcomes without specifying every statistical decision. A detailed statistical analysis plan serves a complementary function. Study registration, protocol publication, and analysis preregistration overlap, but they are not interchangeable.\n\nRegistered reports\n\nRegistered reports are a publication format in which journals evaluate the research question and methods before the main results are known. Following initial review, a journal may offer in-principle acceptance, conditional on requirements such as adequate execution and transparent reporting rather than a particular result.\n\nThis differs from ordinary preregistration, which need not involve peer review or any publication commitment. Registered reports address incentives associated with publication bias more directly, although they do not guarantee correct design or execution.\n\nReproducibility and replication\n\nPreregistration is associated with open science, but it is distinct from sharing data, materials, or code. Those practices help others inspect or reproduce an analysis; preregistration documents when analytical commitments were made.\n\nA preregistered study can be difficult to reproduce if its materials are unavailable. Conversely, a fully reproducible analysis may have been selected after extensive exploration. Preregistration also differs from replication: an original study can be preregistered, and a replication can be conducted without preregistration."
    },
    {
      "heading": "Development and applications",
      "text": "Prospective protocols and advance analysis plans predate the contemporary term. Public clinical-trial registration became increasingly institutionalized in the late twentieth and early twenty-first centuries. During the 2010s, preregistration expanded across psychology, economics, political science, and other empirical disciplines as concerns about selective reporting and analytical flexibility received greater attention.\n\nApplications extend beyond conventional experiments. Researchers preregister observational studies, surveys, analyses of administrative records, and meta-analyses. Review protocols can specify eligibility criteria, search methods, outcomes, and synthesis procedures before the evidence is assembled.\n\nAdaptive studies can also be preregistered. Their plans specify rules governing adaptations rather than fixing every eventual action. For example, a design can permit sample-size changes according to a defined procedure while accounting for the statistical consequences."
    },
    {
      "heading": "Limitations and methodological debate",
      "text": "Preregistration is a transparency mechanism, not a certificate of research quality. A registered study may still have poor measurement, inadequate power, inappropriate models, or confounding. In particular, advance specification does not by itself justify causal inference.\n\nIts effectiveness depends on specificity, accessibility, and comparison with the final report. Vague plans can leave substantial discretion, while detailed plans can still be ignored. Registering many possible analyses and selectively emphasizing favorable ones can preserve some of the problems that registration is intended to reveal.\n\nThere is also a tension between advance commitment and methodological learning. An unforeseen data feature may make the planned model inappropriate. Strict adherence can then produce a weaker analysis than a transparently justified revision. The relevant distinction is not simply “planned” versus “unplanned,” but whether decisions and their relationship to observed results are accurately disclosed.\n\nPreregistration is not universally required for valid inference. Some procedures explicitly accommodate sequential decisions, model selection, or exploration under stated assumptions. Conversely, advance planning does not make an invalid procedure valid. Its contribution is to make the research decision process more visible, allowing readers to assess the relationship between the questions asked, the analyses performed, and the evidence reported."
    }
  ],
  "links": [
    "statistics",
    "multiple-testing",
    "significance-level",
    "p-value",
    "randomization",
    "sample-size",
    "statistical-power",
    "confidence-interval",
    "bayesian-inference",
    "prior-distribution",
    "sensitivity-analysis",
    "machine-learning",
    "randomized-controlled-trial",
    "publication-bias",
    "meta-analysis",
    "confounding",
    "causal-inference"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}