{
  "id": "meta-analysis",
  "title": "Meta-analysis",
  "lang": "en",
  "source_lang": "en",
  "summary": "Meta-analysis statistically combines results from multiple studies to estimate effects, assess variation, and investigate the strength and limitations of a body of evidence.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Statistical meta-analysis",
    "Quantitative research synthesis"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/meta-analysis/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/meta-analysis/",
      "json": "https://aiwiki.page/zh/meta-analysis.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Meta-analysis is a family of methods in statistics for combining numerical results from multiple studies addressing related questions. Its central tasks are to estimate effects more precisely, characterize differences between studies, and determine how conclusions depend on the available evidence and the assumptions used to synthesize it. The units of analysis are usually study-level estimates, although some approaches analyze participant-level observations from several studies jointly.\n\nMeta-analysis is closely associated with systematic reviews, but the two are not synonymous. A systematic review uses explicit methods to identify, select, appraise, and synthesize research; it may or may not include a statistical combination of results. Conversely, a meta-analysis can be calculated from a collection of studies that was not assembled systematically, although its conclusions may then be vulnerable to selection. Applications span medicine, psychology, education, ecology, genetics, economics, and other fields in which individual studies provide incomplete or variable evidence.\n\nA pooled result is not simply an average of published conclusions. It depends on the definition of the outcome, the comparability of studies, their precision and risk of bias, and the statistical model connecting them. Modern meta-analysis therefore combines mathematical estimation with judgments about study design, evidence availability, and the populations to which results may apply."
    },
    {
      "heading": "Historical development and intellectual foundations",
      "text": "The problem of combining observations predates meta-analysis as a distinct research practice. Astronomers and geodesists developed methods for reconciling measurements of common quantities, contributing to the theory of least squares and weighted estimation. Combining scientific studies presents an additional difficulty: studies may differ not only through measurement error but also through populations, procedures, interventions, and definitions. A synthesis must therefore distinguish uncertainty about a shared quantity from genuine differences among quantities being estimated.\n\nIn 1904, Karl Pearson combined evidence from studies of inoculation against enteric fever, an early example of quantitative synthesis in medicine. In the 1930s, Ronald Fisher described a method for combining significance probabilities. For independent tests with probabilities \\(p_1,\\ldots,p_k\\), Fisher’s statistic is\n\n\\[X=-2\\sum_{i=1}^{k}\\log p_i.\\]\n\nUnder independent, continuous null distributions, \\(X\\) follows a chi-squared distribution with \\(2k\\) degrees of freedom. This combines evidence against joint null hypotheses; it does not estimate an average effect, and it does not automatically establish that effects have the same direction. That distinction remains important because combining statistical significance is not equivalent to combining substantive results.\n\nWilliam Cochran’s work in the mid-twentieth century helped establish methods for combining estimates and assessing their inconsistency. The Mantel–Haenszel method, introduced in 1959, provided an influential approach to combining information from stratified contingency tables and later became widely used for binary outcomes across studies. These developments supplied much of the statistical infrastructure subsequently associated with meta-analysis.\n\nGene V. Glass introduced the term meta-analysis in 1976, describing the statistical analysis of results from individual studies for the purpose of integrating findings. The approach gained prominence in psychology and education, where debates about psychotherapy and educational interventions exposed the limitations of narrative reviews that counted statistically significant findings. Standardized effects made it possible to compare results measured using different instruments, although the meaning and comparability of those effects required careful examination.\n\nDuring the 1980s and 1990s, quantitative research synthesis became increasingly institutionalized. DerSimonian and Laird’s 1986 random-effects procedure made between-study variation relatively straightforward to incorporate into applied analyses. The Cochrane Collaboration, established in 1993, promoted systematic reviews of healthcare interventions. Later developments extended synthesis to networks of treatments, correlated outcomes, participant-level datasets, and continuously updated evidence. The field consequently evolved from procedures for combining estimates into a broader methodology for examining how research findings accumulate."
    },
    {
      "heading": "Defining the question and assembling the evidence",
      "text": "A meta-analysis begins with a question that determines which studies could contribute relevant evidence. For intervention research, this often specifies the population, intervention, comparator, outcomes, and eligible study designs. Other fields use different structures: an ecological synthesis might specify organisms, environmental exposures, spatial scales, and response variables. The question must be sufficiently precise to define a meaningful comparison without excluding relevant variation merely because results are inconvenient or difficult to combine.\n\nA protocol records eligibility criteria and planned analytical decisions before study results influence them. Important specifications include the primary outcome, follow-up period, preferred effect measure, handling of multiple reports, and proposed subgroup analyses. Prospective registration provides a dated record against which later changes can be assessed. Changes are sometimes necessary—for example, because outcome reporting differs from expectations—but distinguishing planned analyses from later adaptations helps expose opportunities for selective interpretation.\n\nEvidence identification generally combines searches of bibliographic databases with other sources, such as trial registries, reference lists, conference records, and regulatory documents. Database coverage, language restrictions, indexing practices, and publication delays can affect which studies are found. The number of retrieved reports is not the number of independent studies: a single experiment may generate several articles, while one article may describe several experiments. Linking related reports prevents participants or outcomes from being counted more than once.\n\nData extraction establishes the dataset on which the synthesis rests. Typical fields include sample sizes, group means, event counts, effect estimates, uncertainty measures, study characteristics, and methodological features. Independent or checked extraction reduces transcription and interpretation errors. Missing numerical information may sometimes be reconstructed from a confidence interval, test statistic, or other reported summary, but such reconstructions rely on assumptions that must remain visible. An adjusted observational estimate, for instance, cannot always be reproduced from crude event counts.\n\nStudy eligibility and statistical compatibility are separate judgments. Two studies can both address a review question yet estimate effects that should not be averaged: one may measure immediate symptom change and another long-term functional status. Conversely, different instruments may measure sufficiently similar constructs to support a standardized analysis. The decision depends on scientific meaning, not solely on whether software accepts the inputs.\n\nRisk-of-bias assessment examines how a study’s design, conduct, or reporting could systematically distort its result. In a randomized controlled trial, relevant issues include the randomization process, departures from intended interventions, missing outcomes, measurement, and selective reporting. Observational studies introduce additional concerns, especially confounding and participant selection. These assessments are usually domain-specific rather than reduced to a single numerical “quality score,” because different flaws affect different estimates in different ways."
    },
    {
      "heading": "Effect measures and the construction of comparable estimates",
      "text": "The principal input to most meta-analyses is an effect size together with an estimate of its sampling uncertainty. An effect size may be an absolute difference, a ratio, an association, or a parameter from a statistical model. Its choice determines the substantive question answered by the synthesis. For example, an absolute difference in event probabilities and a ratio of those probabilities describe different aspects of the same comparison and need not behave similarly across populations.\n\nFor continuous outcomes measured on a common scale, the mean difference preserves the original units. If studies use different scales intended to measure the same construct, a standardized mean difference divides a difference in group means by a within-study standard deviation. Hedges’ \\(g\\) applies a correction for small-sample bias to this standardized estimate. Standardization does not itself guarantee comparability: variability can differ because of eligibility criteria, measurement reliability, or population diversity, even when the underlying difference in original units is similar.\n\nFor binary outcomes, common measures include the risk difference, risk ratio, and odds ratio. If event probabilities in two groups are \\(p_1\\) and \\(p_0\\), the risk ratio is \\(p_1/p_0\\), whereas the odds ratio is\n\n\\[\\mathrm{OR} =\\frac{p_1/(1-p_1)}{p_0/(1-p_0)}.\\]\n\nRatio measures are commonly analyzed on a logarithmic scale and transformed back for presentation. An odds ratio can differ substantially from a risk ratio when events are common. Absolute effects additionally depend on baseline risk, so the same relative effect can correspond to very different absolute differences in different settings.\n\nOther applications require other transformations. Correlations are often synthesized after Fisher’s transformation,\n\n\\[z=\\frac{1}{2}\\log\\left(\\frac{1+r}{1-r}\\right),\\]\n\nwhose approximate sampling variance is \\(1/(n-3)\\) under standard assumptions. Time-to-event studies often contribute log hazard ratios, which concern event rates among those still at risk rather than probabilities over a fixed period. Diagnostic-test studies commonly require joint analysis of sensitivity and specificity because their values can trade off as the diagnostic threshold changes.\n\nFor study \\(i\\), let the selected estimate be \\(y_i\\), with estimated sampling variance \\(v_i\\). The square root of \\(v_i\\) is its standard error. These quantities establish the numerical basis for weighting, but their validity depends on the original design. Cluster-randomized trials require allowance for within-cluster similarity; crossover studies require appropriate paired analyses; adjusted estimates reflect particular covariate sets. Rare events and zero-event groups may make simple normal approximations unreliable. Corrections that add arbitrary constants to event counts can materially alter results, especially when treatment-group sizes are unequal."
    },
    {
      "heading": "Common-effect estimation and precision weighting",
      "text": "The common-effect model assumes that every included study estimates one underlying effect \\(\\theta\\), with observed differences arising from sampling error:\n\n\\[y_i=\\theta+\\varepsilon_i, \\qquad E(\\varepsilon_i)=0, \\qquad \\operatorname{Var}(\\varepsilon_i)=v_i.\\]\n\nWhen study estimates are independent and approximately normally distributed, inverse-variance weighting gives\n\n\\[\\hat{\\theta} =\\frac{\\sum_{i=1}^{k}w_i y_i}{\\sum_{i=1}^{k}w_i}, \\qquad w_i=\\frac{1}{v_i}.\\]\n\nThe estimated variance of the pooled result is \\(1/\\sum_i w_i\\). More precise studies therefore receive greater weight, rather than each study contributing equally or solely in proportion to its participant count.\n\nThis weighting has a direct mathematical rationale. Among linear unbiased combinations of independent estimates with known variances and a common expectation, inverse-variance weights minimize variance. The calculation also corresponds to maximum likelihood estimation under a normal model with known sampling variances. In practice, those variances are usually estimated, so the optimality statement is approximate. Specialized methods, including Mantel–Haenszel procedures, can be preferable for particular data structures.\n\nA simple numerical example shows what weighting does. Suppose two independent studies estimate mean differences of \\(2\\) and \\(5\\), with standard errors \\(1\\) and \\(2\\). Their inverse-variance weights are \\(1\\) and \\(0.25\\), giving\n\n\\[\\hat{\\theta} =\\frac{1(2)+0.25(5)}{1.25} =2.6.\\]\n\nThe pooled standard error is \\(\\sqrt{1/1.25}\\), approximately \\(0.89\\). Under the normal approximation, a 95% interval is approximately \\(0.85\\) to \\(4.35\\). The result lies closer to the more precise study, not midway between the two estimates.\n\nThe common-effect model is often called a fixed-effect model, but terminology varies. In some statistical treatments, fixed-effects synthesis allows different study-specific effects while targeting a particular weighted average of them. Consequently, the label alone does not fully identify the estimand. A defensible analysis states whether it assumes a common underlying effect, targets an average restricted to the included studies, or seeks inference about a broader distribution of effects.\n\nPrecision weighting does not reward methodological validity directly. A large, systematically biased study may receive substantial weight, while a small, well-conducted study receives little. Nor does a narrower pooled interval establish that all relevant sources of uncertainty have been incorporated. Standard formulas quantify uncertainty under the specified sampling model; they do not automatically account for missing studies, measurement incompatibility, unmeasured confounding, or inappropriate inclusion decisions. The calculation can be mathematically correct while the substantive synthesis remains misleading."
    },
    {
      "heading": "Random-effects models and between-study heterogeneity",
      "text": "A random-effects model allows underlying effects to differ across studies. A conventional hierarchical formulation is\n\n\\[y_i\\mid\\theta_i\\sim N(\\theta_i,v_i), \\qquad \\theta_i\\sim N(\\mu,\\tau^2),\\]\n\nwhere \\(\\theta_i\\) is a study-specific effect, \\(\\mu\\) is the mean of the distribution of effects, and \\(\\tau^2\\) is the between-study variance. Marginally, \\(y_i\\) has variance \\(v_i+\\tau^2\\), producing weights\n\n\\[w_i^*=\\frac{1}{v_i+\\hat{\\tau}^2}.\\]\n\nCompared with common-effect weighting, this generally reduces the relative dominance of the most precise studies. It does not mean that small studies are intrinsically more informative; it reflects the assumption that even very precise studies reveal only particular realizations from a distribution of effects.\n\nBetween-study variance can be estimated by several methods. The DerSimonian–Laird estimator uses a moment-based calculation, while restricted maximum likelihood is widely used in modern applications. Estimation is difficult when few studies are available, and different methods can produce materially different uncertainty assessments. A zero estimate of \\(\\tau^2\\) does not establish that effects are identical: the data may simply contain little information about their dispersion.\n\nCochran’s \\(Q\\) statistic measures weighted disagreement with a common-effect estimate:\n\n\\[Q=\\sum_{i=1}^{k}w_i(y_i-\\hat{\\theta})^2.\\]\n\nUnder suitable common-effect assumptions, it is approximately chi-squared with \\(k-1\\) degrees of freedom. It has limited ability to detect heterogeneity with few imprecise studies and can detect small departures with many precise studies. A related descriptive statistic is\n\n\\[I^2=\\max\\left(0,\\frac{Q-(k-1)}{Q}\\right)\\times100\\%,\\]\n\nwith a conventional value of zero when \\(Q=0\\). \\(I^2\\) expresses inconsistency relative to sampling variability; it is not the percentage of studies that disagree or a direct measure of substantive importance. Its uncertainty can be considerable.\n\nThe distinction between a mean effect and an effect in a new setting is crucial. A confidence interval around \\(\\mu\\) describes uncertainty about the average, whereas a prediction interval describes the range anticipated for an underlying effect in an exchangeable new study. Approximate prediction intervals incorporate both uncertainty in the estimated mean and estimated between-study variation. They can span effects in opposite directions even when the mean is estimated relatively precisely. With few studies, their calibration and sensitivity to distributional assumptions require particular caution.\n\nRandom-effects analysis is not a universal remedy for incompatibility. Its interpretation assumes that a meaningful distribution connects the studies; arbitrary mixtures of different questions do not become coherent through an added variance parameter. Moreover, applying random-effects weights can increase the influence of small studies affected by selective publication. Model choice therefore follows the scientific estimand and expected variation, rather than a mechanical rule that switches models according to whether a heterogeneity test crosses a significance threshold."
    },
    {
      "heading": "Explaining variation and handling dependent results",
      "text": "Researchers often seek to explain heterogeneity using study characteristics. Meta-regression extends the synthesis model by relating effects to moderators:\n\n\\[y_i=\\beta_0+\\beta_1x_{i1}+\\cdots+\\beta_px_{ip} +u_i+\\varepsilon_i,\\]\n\nwhere \\(x_{ij}\\) are characteristics such as intervention intensity, average age, follow-up duration, or study design, and \\(u_i\\) represents residual between-study variation. Categorical moderators yield subgroup comparisons; continuous moderators retain information that can be lost through arbitrary categorization. Both approaches require enough independent studies to support the number and complexity of parameters.\n\nThe interpretation of a moderator differs from that of a randomized comparison. Studies with one characteristic may differ systematically in many others. For example, newer studies may use different interventions, recruit different populations, and have lower risks of bias simultaneously. A meta-regression coefficient for publication year cannot separate those mechanisms without additional assumptions. Multiple moderator searches also create opportunities for chance findings, especially when analyses are selected after the results are known.\n\nStudy-level associations need not describe participant-level relationships. A synthesis might find larger effects in studies whose participants have a higher average age, yet the treatment-by-age interaction within each study could be absent or reversed. This is an instance of the ecological fallacy. Conversely, variation in study averages may be too small to reveal an important individual-level modifier. Participant-level data can help separate within-study interactions from between-study associations, although they do not eliminate problems of missingness, measurement, or confounding.\n\nDependence is another central complication. A study may report several outcomes, repeated follow-ups, or multiple intervention groups sharing one comparator. If all estimates are treated as independent, the same underlying information is counted repeatedly and uncertainty is usually understated. Investigators may select one prespecified estimate, combine related outcomes, or use a model that represents their covariance. For instance, two treatment comparisons sharing a control group have correlated sampling errors even when the treatment groups themselves are distinct.\n\nMultivariate meta-analysis models several related outcomes jointly, while multilevel models represent estimates nested within studies or other clusters. Robust variance estimation provides an alternative when within-study covariance is incompletely known, using independent study clusters as the basis for uncertainty calculations. These procedures still require adequate numbers of clusters and suitable small-sample corrections. A multilevel random effect alone does not automatically repair all sampling correlations; the covariance structure and the intended estimand must be specified consistently.\n\nDependence can also occur across publications through overlapping cohorts, reused administrative databases, or shared participant samples. Identifying such overlap is partly a documentary task rather than a purely statistical one. The relevant sample size is the amount of independent information, not the number of effect estimates entered into a program."
    },
    {
      "heading": "Missing evidence, reporting bias, and sensitivity analysis",
      "text": "The studies available for synthesis may be an unrepresentative subset of all studies conducted. Publication bias arises when the probability or timing of publication depends on findings. Selective reporting can operate within published studies through choices among outcomes, time points, subgroups, or statistical models. These mechanisms differ from ordinary sampling variation because the observed evidence has been filtered according to its results or associated characteristics.\n\nA funnel plot displays study effects against a measure of precision or sample size. Under simplified conditions, estimates from smaller studies scatter more widely around a central effect, producing an approximately symmetric funnel. Asymmetry may indicate missing unfavorable results, but it can also arise from genuine effect differences, methodological weaknesses in small studies, chance, or mathematical relationships between an effect measure and its standard error. A visually symmetric plot does not establish the absence of reporting bias.\n\nStatistical procedures for detecting small-study effects include Egger-type regression tests and other measure-specific methods. Their validity depends on the outcome and effect metric, and they generally have little power when few studies are available. Some standardized effect sizes and binary-outcome measures can generate structural associations with their estimated precision. Accordingly, a generic asymmetry test applied indiscriminately across data types can produce misleading evidence about publication processes.\n\nSeveral methods attempt to evaluate or adjust for missing evidence. Trim-and-fill estimates hypothetical missing studies from funnel asymmetry, but its assumptions can fail in heterogeneous datasets. Selection models explicitly represent how observation probabilities depend on results, often through significance or effect direction. Other approaches examine the distribution of reported significance probabilities. None can identify all missing evidence from the observed studies alone: corrections necessarily depend on assumptions about an unobserved selection mechanism.\n\nSensitivity analysis assesses how conclusions change under alternative defensible choices. Examples include changing the heterogeneity estimator, excluding studies with particular risks of bias, using alternative within-study correlations, and comparing adjusted with unadjusted estimates. Leave-one-out analysis repeats the synthesis while omitting each study in turn. Influence diagnostics can identify studies that strongly affect the mean, heterogeneity, or model fit, but influence does not by itself justify exclusion. A highly influential study may be unusually informative rather than erroneous.\n\nBias in the primary studies also survives pooling. Combining many observational estimates affected by similar confounding can yield a highly precise estimate of a biased quantity. Quantitative bias analyses may explore plausible departures from assumptions, but they cannot convert unsupported assumptions into observed information. Statistical robustness, methodological validity, and completeness of the evidence are distinct properties, each requiring separate examination."
    },
    {
      "heading": "Individual-participant, network, and Bayesian extensions",
      "text": "Individual-participant-data meta-analysis uses original participant-level records rather than relying exclusively on published summaries. This permits harmonized outcome definitions, common adjustment strategies, direct examination of participant-level effect modifiers, and consistent time-to-event analyses. A two-stage approach first analyzes each study and then combines its estimates; a one-stage approach fits a joint model while representing study structure. With compatible specifications, their results may be similar, but sparse data, nonlinear models, and different weighting or adjustment choices can produce differences.\n\nAccess to participant data does not guarantee complete or unbiased evidence. Investigators may be unable to obtain some datasets, and available datasets may not represent all eligible studies. Harmonization can also be limited by variables that were measured differently or not measured at all. Pooling records without accounting for study membership can destroy the protection of within-study comparisons, especially when treatment availability or participant characteristics differ across studies. Participant-level synthesis therefore remains a study-structured analysis rather than an ordinary analysis of one large dataset.\n\nNetwork meta-analysis combines direct and indirect evidence across multiple interventions. If studies compare A with B and others compare B with C, an indirect A–C comparison can be constructed on an appropriate additive effect scale. When \\(d_{AB}\\) denotes A minus B and \\(d_{BC}\\) denotes B minus C,\n\n\\[d_{AC}=d_{AB}+d_{BC}.\\]\n\nA connected network permits estimation of a coherent set of relative effects, while multi-arm studies require explicit allowance for correlated comparisons.\n\nThe validity of indirect comparisons depends on transitivity: the different sets of comparisons must be sufficiently comparable in important effect modifiers. Consistency is the related statistical agreement between direct and indirect evidence where both exist. For example, if A–B trials recruit substantially more severe cases than B–C trials and severity modifies relative effects, an indirect A–C estimate may not answer the intended question. Treatment rankings summarize model output but can be unstable when evidence is sparse and do not themselves express effect magnitude or certainty.\n\nBayesian inference provides another framework for common-effect, random-effects, network, and participant-level models. A prior distribution for the mean effect, heterogeneity, and other parameters is updated through the study likelihood to obtain a posterior distribution. Computation often uses Markov chain Monte Carlo. Bayesian models can integrate external information and express uncertainty in heterogeneity directly, but sparse evidence may make conclusions sensitive to prior choices.\n\nOther extensions address specialized data structures. Diagnostic-test meta-analysis often jointly models sensitivity and specificity or uses hierarchical summary receiver-operating-characteristic models. Dose–response synthesis represents effects across exposure levels while accounting for shared reference groups. Multivariate models can synthesize correlated outcomes, and prevalence synthesis requires models suited to bounded proportions and varying populations. Each extension changes the likelihood or dependence structure to match the evidence rather than merely applying a different label to the same weighted average."
    },
    {
      "heading": "Interpretation, reporting, and continuing methodological questions",
      "text": "The first interpretive question is what quantity the synthesis estimates. A pooled relative effect among eligible randomized trials is not automatically an absolute effect in a current population. An average association from observational studies is not automatically causal. Causal inference depends on design and assumptions about treatment assignment, confounding, measurement, and selection; meta-analysis can organize evidence about those issues but cannot bypass them. Similarly, external validity depends on whether included participants, interventions, and settings support application elsewhere.\n\nStatistical significance is only one aspect of interpretation. An interval that includes no effect may also include effects large enough to matter, while a very precise nonzero effect may be substantively small. Relative and absolute scales can illuminate different aspects of a result. Converting a relative effect to an absolute difference requires a relevant baseline risk and assumptions about how the relative effect transfers across settings. Heterogeneity further separates the precision of the average from the variability of effects expected in practice.\n\nA forest plot presents study estimates, their uncertainty intervals, and usually a pooled estimate. The size of a study marker often reflects its statistical weight, while a diamond represents the pooled interval. Such plots help reveal incompatible directions, imprecise studies, and the influence of large studies, but they do not display every source of bias. Tables describing study designs, populations, outcomes, and risk-of-bias judgments provide information that the numerical plot cannot encode.\n\nReporting standards make analytical decisions inspectable. The PRISMA statement, first published in 2009 and updated as PRISMA 2020, specifies reporting items for systematic reviews, including reviews containing meta-analyses. Its flow diagram distinguishes records identified, reports retrieved, and studies included. Transparent reporting also identifies effect measures, model specifications, variance estimators, handling of dependence, excluded data, protocol deviations, and the distinction between planned and exploratory analyses. PRISMA is a reporting guideline, not a certificate that the analysis is methodologically sound.\n\nEvidence-certainty frameworks, including GRADE in healthcare, assess more than the width of the pooled interval. Relevant considerations include risk of bias, inconsistency, indirectness, imprecision, and missing evidence. Sharing extracted datasets and analysis code allows numerical checking and facilitates updates, subject to data-governance restrictions. Software implementations in R, Stata, SAS, and dedicated review programs support many established procedures, but their defaults differ, particularly for heterogeneity estimation and small-sample inference.\n\nContinuing methodological work addresses reliable inference with few studies, sparse events, non-normal effect distributions, selective availability, and complex dependence. Living systematic reviews incorporate newly available evidence, creating further questions about repeated analyses, version control, and changing eligibility criteria. Automated screening and data-extraction systems can accelerate parts of the workflow, but errors in study identification, outcome interpretation, or duplicate detection can propagate directly into the pooled result. Reproducible decision records and explicit model assumptions remain central to evaluating those evolving methods."
    }
  ],
  "links": [
    "statistics",
    "ronald-fisher",
    "confidence-interval",
    "randomized-controlled-trial",
    "confounding",
    "effect-size",
    "standard-deviation",
    "odds-ratio",
    "variance",
    "standard-error",
    "maximum-likelihood-estimation",
    "degrees-of-freedom",
    "prediction-interval",
    "covariance",
    "publication-bias",
    "sensitivity-analysis",
    "bayesian-inference",
    "prior-distribution",
    "posterior-distribution",
    "markov-chain-monte-carlo",
    "causal-inference",
    "external-validity"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}