{
  "id": "sample-size",
  "title": "Sample Size",
  "lang": "en",
  "source_lang": "en",
  "summary": "Sample size is the number of observational units in a study, influencing statistical precision, power, and the reliability of inference.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Sample count",
    "Number of observations"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/sample-size/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/sample-size/",
      "json": "https://aiwiki.page/zh/sample-size.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Sample size is the number of observational or experimental units included in a statistical investigation, usually denoted by \\(n\\). Depending on the study, these units may be people, households, schools, manufactured components, transactions, or measurements taken over time. In statistics, sample size connects the design of data collection with the uncertainty of the conclusions drawn from it. It affects how precisely a quantity can be estimated, how readily a specified difference can be detected, and how complex a model can be supported by the available information.\n\nA sample size is not, by itself, a measure of study quality. A large dataset can be systematically unrepresentative, while many measurements from the same individual may contain much less information than the same number of independent observations. The relevant count therefore depends on the target of inference, the sampling or experimental design, the outcome being measured, and the proposed analysis.\n\nSample-size determination translates these features into a quantitative design. Some calculations target a specified interval width; others target the probability of detecting an effect, the accuracy of a prediction, or the expected value of collecting additional data. The familiar rule that uncertainty decreases approximately as \\(1/\\sqrt{n}\\) is central to many calculations, but it applies under assumptions that can fail under dependence, heavy-tailed distributions, selection processes, or complex models."
    },
    {
      "heading": "What is counted: observations, units, and populations",
      "text": "The simplest sample consists of \\(n\\) independently selected units from a defined population. In random sampling, the population is the collection about which an inference is intended, whereas the sample is the subset actually observed. Population size is commonly written \\(N\\), distinguishing it from sample size \\(n\\). A survey of 1,000 residents of a city with one million residents thus has \\(n=1{,}000\\), not one million. A population census attempts to enumerate the entire defined population, although nonresponse and measurement error can remain even when sampling error is absent.\n\nDifferent counts can coexist within one investigation. A study may approach 2,000 people, enroll 1,200, observe a primary outcome for 1,050, and analyze 1,030 after applying prespecified eligibility rules. These figures describe recruitment, enrollment, outcome availability, and analysis, respectively. None can automatically substitute for the others. When several outcomes have different missingness patterns, each analysis may have its own sample size. A transparent account distinguishes the planned count from the realized count and identifies which observations contribute to each estimate.\n\nThe observational unit also need not be the unit assigned to an intervention. In a school-based experiment, schools may be assigned to teaching methods while outcomes are measured for individual students. The number of students describes the volume of measurements, but the number of independently assigned schools is crucial to uncertainty about the intervention effect. Likewise, measuring ten tissue sections from each of six specimens does not create sixty independent specimens. Treating dependent subsamples as independent replicates is commonly called pseudoreplication.\n\nRepeated measurements produce another distinction. A dataset containing 100 people measured on 12 occasions has 1,200 recorded observations but only 100 people. The additional occasions may improve estimation of individual trajectories or within-person changes, yet they do not generally provide the same information as 1,200 separately sampled people. The relationship depends on the covariance among measurements and on whether the target concerns individuals, occasions, or changes over time.\n\nSample size is consequently a property of a particular inferential problem, not merely a row count in a file. In a paired experiment, the relevant count may be the number of pairs. In a study of rare events, the number of observed events may be more informative than the number enrolled. In an analysis of regional policies, thousands of individual records cannot necessarily compensate for having only a few independently varying regions. Defining the counting unit is therefore logically prior to calculating how many units are needed."
    },
    {
      "heading": "Precision-based determination",
      "text": "A major approach to sample-size determination starts with a desired precision for an estimator. Suppose \\(X_1,\\ldots,X_n\\) are independent observations with common mean \\(\\mu\\) and finite variance \\(\\sigma^2\\). The sample mean has variance\n\n\\[\\operatorname{Var}(\\bar X)=\\frac{\\sigma^2}{n},\\]\n\nso its standard error is \\(\\sigma/\\sqrt n\\). Under a normal model, or an adequate central limit theorem approximation, a two-sided confidence interval with confidence level \\(1-\\alpha\\) has approximate half-width\n\n\\[E=z_{1-\\alpha/2}\\frac{\\sigma}{\\sqrt n},\\]\n\nwhere \\(z_q\\) is the \\(q\\)-quantile of the standard normal distribution. Solving for the sample size gives\n\n\\[n\\geq \\left(\\frac{z_{1-\\alpha/2}\\sigma}{E}\\right)^2.\\]\n\nThis equation makes the square-root relationship concrete. Halving the intended margin of error requires approximately four times as many independent observations. If the population standard deviation is 10 units and the target half-width is 2 units at 95% confidence, the normal approximation gives \\(n\\geq96.04\\), rounded upward to 97. When the standard deviation is estimated rather than known, a calculation based on Student’s t-distribution may require iteration because the critical value itself depends on sample size.\n\nFor a population proportion \\(p\\), the corresponding large-sample approximation is\n\n\\[n\\geq \\frac{z_{1-\\alpha/2}^{\\,2}p(1-p)}{E^2}.\\]\n\nAt 95% confidence, an absolute margin of error of 0.05 and the conservative choice \\(p=0.5\\) produce \\(n\\geq384.16\\), or 385 observations. The choice \\(0.5\\) maximizes \\(p(1-p)\\), so it is conservative within this formula when no credible planning value is available. This calculation does not imply that every survey with 385 respondents has a five-percentage-point margin of error: it assumes an appropriate probability sample, negligible dependence, and the stated interval approximation.\n\nSampling without replacement from a small finite population reduces uncertainty. For a simple random sample, the standard error includes a finite-population correction proportional to \\(\\sqrt{(N-n)/(N-1)}\\). If \\(n_0\\) is the sample size obtained by ignoring that correction, a common adjusted expression is\n\n\\[n=\\frac{n_0}{1+(n_0-1)/N}.\\]\n\nFor \\(N=1{,}000\\) and \\(n_0=384.16\\), this gives about 278 observations after rounding upward. The correction concerns sampling from a fixed finite population; it does not necessarily apply when the inferential target is an ongoing process or a broader population.\n\nPrecision calculations also depend on the scale of the target. Estimating a prevalence of 1% within one percentage point is quite different from estimating it within 10% of its value, which means an absolute error of 0.1 percentage points. Near the boundaries of a binomial distribution, symmetric normal intervals can be inaccurate. Exact or better-calibrated interval methods, and the desired probability of attaining a specified interval width, may lead to different sample sizes from the elementary formula."
    },
    {
      "heading": "Power, effect size, and hypothesis testing",
      "text": "In statistical hypothesis testing, sample-size planning usually specifies a null hypothesis, an alternative of scientific interest, a significance level \\(\\alpha\\), and a desired statistical power \\(1-\\beta\\). Here \\(\\alpha\\) controls the probability of a false rejection under the null, while \\(\\beta\\) is the probability of failing to reject at the specified alternative. These are Type I and Type II errors. Power is therefore a property of a test under a particular data-generating condition, not a general probability that a study will reach a correct conclusion.\n\nFor two independent groups with equal sizes, a normally distributed outcome, common standard deviation \\(\\sigma\\), and a target difference \\(\\Delta\\), a familiar approximation gives the required number per group:\n\n\\[n_{\\text{group}} \\approx \\frac{2\\sigma^2 \\left(z_{1-\\alpha/2}+z_{1-\\beta}\\right)^2}{\\Delta^2}.\\]\n\nWriting the standardized effect size as \\(d=\\Delta/\\sigma\\) yields\n\n\\[n_{\\text{group}} \\approx \\frac{2\\left(z_{1-\\alpha/2}+z_{1-\\beta}\\right)^2}{d^2}.\\]\n\nFor a two-sided 5% test, 80% power, and \\(d=0.5\\), the approximation gives about 63 observations per group; an exact two-sample \\(t\\)-test calculation typically gives 64 per group.\n\nThe target effect is a substantive input, not an output of the formula. It may represent a smallest scientifically important difference, an economically consequential improvement, or a difference that would change a decision. Using an unusually large estimate from a small preliminary study can produce an unrealistically small planned sample. Planning can instead examine several plausible values of the effect and variance, making clear which conclusions depend on optimistic assumptions. Raising desired power from 80% to 90%, lowering the significance level, or targeting a smaller effect all increase the required sample.\n\nThe chosen test also matters. One-sided and two-sided tests use different rejection regions; paired and unpaired analyses have different variances; and tests for proportions can depend on both baseline risk and the alternative risk. Multiple testing introduces further requirements when several claims must be protected simultaneously. A Bonferroni correction, for example, replaces a family-level threshold with smaller thresholds for individual tests, generally increasing the sample needed for a specified per-comparison effect and power.\n\nTests of equivalence and noninferiority answer different questions from tests of a zero difference. They require a prespecified margin defining which differences count as acceptably small, and their sample-size calculations depend strongly on that margin and the assumed true effect. Failure to reject a zero-difference null does not establish equivalence.\n\nFinally, power calculated after a study by substituting its observed effect into the original power formula is not an independent assessment of evidential strength. For many standard tests, this “observed power” is largely a transformation of the p-value. The observed estimate and its interval convey the effect magnitude and remaining uncertainty more directly."
    },
    {
      "heading": "Dependence, clustering, and effective information",
      "text": "The square-root rule assumes that observations contribute information in an appropriate way. When observations are positively correlated, adding another observation often contributes less information than adding an independent one. A useful description is the effective sample size: the number of independent observations that would provide comparable precision for a particular estimator. This is not necessarily an integer and is not a universal property of the dataset. Different estimands and methods can have different effective sample sizes.\n\nFor equal-sized clusters containing \\(m\\) observations each, a common approximation assumes a shared within-cluster correlation \\(\\rho\\). The design effect is then\n\n\\[D=1+(m-1)\\rho,\\]\n\nand an approximate effective sample size is\n\n\\[n_{\\mathrm{eff}}\\approx\\frac{n}{D}.\\]\n\nIf 1,000 observations occur in clusters of 30 and \\(\\rho=0.05\\), then \\(D=2.45\\), giving an effective sample size of about 408. This expression assumes a simple exchangeable correlation structure and equal cluster sizes. Unequal cluster sizes, covariate adjustment, and small numbers of clusters require more specialized calculations.\n\nThe number of clusters can be more important than the number of observations within them. If a treatment is assigned to ten institutions, collecting substantially more measurements within those institutions does not create additional independent treatment assignments. Inferential procedures may also have few degrees of freedom, affecting critical values and interval coverage beyond what a simple design-effect multiplier captures. More clusters and more observations per cluster therefore need not be interchangeable design choices.\n\nFor a stationary time series, a related approximation for estimation of a mean is\n\n\\[n_{\\mathrm{eff}} \\approx \\frac{n} {1+2\\sum_{k=1}^{\\infty}\\rho_k},\\]\n\nwhere \\(\\rho_k\\) is the lag-\\(k\\) autocorrelation and the series must satisfy conditions that make the approximation meaningful. Positive serial correlation generally reduces effective information about the mean; some negative correlations can increase it. Estimating the autocorrelation sum can itself be difficult when the series is short or dependence persists over long periods.\n\nUnequal weights produce another notion of effective sample size. For positive weights \\(w_i\\), the expression\n\n\\[n_{\\mathrm{eff}} = \\frac{\\left(\\sum_i w_i\\right)^2}{\\sum_i w_i^2}\\]\n\nmeasures the concentration of weights under simplifying assumptions. It equals \\(n\\) for equal weights and decreases when a few observations receive disproportionately large weights. It does not by itself account for clustering or all relationships between weights and outcomes.\n\nEffective sample size is also used in Markov chain Monte Carlo to describe the information in dependent simulation draws. Ten thousand stored draws are not equivalent to ten thousand independent draws if the chain mixes slowly. This computational use concerns Monte Carlo error in approximating a distribution, rather than the number of real-world units on which the underlying analysis is based."
    },
    {
      "heading": "Allocation, recruitment, and incomplete observations",
      "text": "A fixed total sample can be distributed in different ways. In a two-group comparison with independent outcomes, the variance of the estimated difference in means is\n\n\\[\\operatorname{Var}(\\bar X_1-\\bar X_0) = \\frac{\\sigma_1^2}{n_1} + \\frac{\\sigma_0^2}{n_0}.\\]\n\nWhen the groups have equal outcome variances and equal per-unit costs, equal allocation minimizes this variance for a fixed total \\(n_0+n_1\\). If variances differ, the variance-minimizing ratio is \\(n_1/n_0=\\sigma_1/\\sigma_0\\). If per-unit costs are \\(c_1\\) and \\(c_0\\), allocation under a fixed sampling budget instead satisfies\n\n\\[\\frac{n_1}{n_0} = \\frac{\\sigma_1}{\\sigma_0} \\sqrt{\\frac{c_0}{c_1}},\\]\n\nunder the same simple variance model.\n\nUnequal allocation may nevertheless be part of a design for operational or scientific reasons. It can increase information about a less common exposure, support comparisons involving a shared control group, or accommodate limited availability of particular units. Its consequences depend on the full design. With equal variances and a fixed total sample, a 2:1 allocation gives less precision for a single treatment difference than a 1:1 allocation, although the loss is smaller than the allocation ratio alone might suggest.\n\nStratified sampling divides a population into groups before drawing samples. Under standard assumptions, allocation proportional to stratum population size provides a self-weighting sample, while allocation proportional to population size times within-stratum standard deviation minimizes the variance of an estimated total for a fixed sample size. Disproportionate sampling may be used to ensure that small subpopulations have enough observations for separate analysis. Appropriate weights are then needed when estimating population-wide quantities.\n\nRecruitment targets commonly exceed the number of complete observations required by an analysis. If \\(n_{\\text{complete}}\\) usable observations are needed and each recruited unit is assumed to supply a usable outcome with probability \\(r\\), a basic inflation is\n\n\\[n_{\\text{recruit}} = \\left\\lceil\\frac{n_{\\text{complete}}}{r}\\right\\rceil.\\]\n\nThus 400 complete outcomes and an expected completion rate of 80% imply a target of 500 recruited units. This calculation provides the desired count in expectation; it does not guarantee that at least 400 outcomes will be observed. A design requiring a specified probability of reaching that count can model the number completing as a binomial random variable.\n\nInflation addresses lost quantity, not necessarily lost validity. If missingness depends on unobserved outcomes, enrolling more units does not automatically remove the resulting bias. Missing-data imputation can use information from incomplete records under explicit assumptions, but its precision depends on the missingness pattern and on how informative the observed variables are. Similarly, recruiting many additional respondents cannot ensure representativeness when participation systematically excludes relevant parts of the target population."
    },
    {
      "heading": "Outcome distributions and specialized designs",
      "text": "Different outcomes carry different amounts and forms of information. A continuous outcome with moderate variability may support precise estimation from a sample that is inadequate for estimating an uncommon binary event. For independent binary observations with event probability \\(p\\), the probability of seeing no events is\n\n\\[P(X=0)=(1-p)^n.\\]\n\nTo have at least a 95% chance of observing one or more events when \\(p=0.001\\), the required count satisfies\n\n\\[n\\geq\\frac{\\log(0.05)}{\\log(0.999)},\\]\n\nwhich gives 2,995 after rounding upward. Observing at least one event, however, is a much weaker objective than estimating its probability precisely.\n\nIf no events occur, the absence of observations does not establish that the population event probability is zero. Solving \\((1-p)^n=0.05\\) gives a one-sided 95% upper confidence bound\n\n\\[p_{\\mathrm{upper}}=1-0.05^{1/n}.\\]\n\nFor sufficiently large \\(n\\), this is approximately \\(3/n\\), often called the rule of three. Thus zero events among 100 independent observations remains compatible with an event probability near 3% at this confidence level. The approximation assumes the same event probability across observations and complete ascertainment of the event.\n\nFor event counts modeled by a Poisson distribution, exposure time matters alongside the number of units. If a homogeneous process has rate \\(\\lambda\\) over total exposure \\(T\\), the expected event count is \\(\\lambda T\\). Precision for the estimated rate is therefore closely tied to total exposure and the number of events. If counts are more variable than the Poisson model allows, calculations based on equality of mean and variance can overstate precision.\n\nTime-to-event studies introduce censoring: an individual's event time may not be observed before follow-up ends. Under proportional-hazards assumptions, an approximate event requirement for comparing two groups is\n\n\\[D\\approx \\frac{\\left(z_{1-\\alpha/2}+z_{1-\\beta}\\right)^2} {q(1-q)(\\log H)^2},\\]\n\nwhere \\(D\\) is the total number of events, \\(q\\) is the allocation proportion, and \\(H\\) is the assumed hazard ratio. Translating an event target into a participant target requires assumptions about recruitment timing, follow-up duration, event incidence, and loss to follow-up. Equal enrollment can therefore produce very different information under different observation periods.\n\nDesign structure also changes the calculation. In paired observations, the variance of a difference is \\(\\sigma_1^2+\\sigma_0^2-2\\rho\\sigma_1\\sigma_0\\), so positive within-pair correlation can substantially improve precision. In a factorial design, information for an interaction is not the same as information for either main effect. A study adequate for an overall average effect may consequently be inadequate for interactions, subgroup-specific effects, or rare outcomes recorded within the same sample."
    },
    {
      "heading": "Regression, prediction, and high-dimensional data",
      "text": "For linear regression, sample size interacts with the number of parameters and with the geometry of the predictors. In the classical model \\(Y=X\\beta+\\varepsilon\\), with independent errors having common variance \\(\\sigma^2\\), the conditional covariance of the ordinary least squares estimate is\n\n\\[\\operatorname{Var}(\\hat\\beta\\mid X) = \\sigma^2(X^{\\mathsf T}X)^{-1},\\]\n\nprovided the design matrix has full column rank. Increasing the number of observations often increases information, but highly correlated predictors can leave particular coefficients imprecise even in a large sample. Sample size alone does not determine the eigenvalues of the information matrix or resolve near-collinearity.\n\nIf there are \\(p\\) fitted coefficients, the usual residual variance estimate has \\(n-p\\) degrees of freedom. Having \\(n>p\\) is a mathematical requirement for ordinary full-rank fitting with positive residual degrees of freedom, not a guarantee of useful estimation. The necessary sample can be much larger when the signal is weak, the predictor distribution is narrow, or the intended target is a precisely estimated individual coefficient. Polynomial terms, categorical levels, and interactions each contribute parameters; counting named variables can therefore understate model complexity.\n\nIn logistic regression, total sample size and the number of events both matter. A historical rule of thumb suggested roughly ten outcome events per fitted parameter, but it is not a universal adequacy criterion. Performance also depends on event prevalence, predictor distributions, effect magnitudes, separation, and the extent of model selection. With complete or quasi-complete separation, ordinary maximum likelihood estimation can produce unbounded coefficient estimates even when the overall dataset appears substantial.\n\nPrediction introduces targets beyond coefficient precision. A model may be evaluated for calibration, discrimination, prediction error, or performance in a distinct population. Data used to select features or tune model settings are not an independent test set. Cross-validation helps estimate predictive performance, but repeatedly using the same validation results to make design choices can make the reported performance optimistic. The size of a held-out evaluation sample determines the uncertainty of the performance estimate, separately from the sample used to fit the model.\n\nIn machine learning, the number of recorded features can greatly exceed the number of observations. Regularization and structural assumptions such as sparsity can make estimation possible in such settings, but do not make sample requirements disappear. They change the relationship between complexity, assumptions, and predictive error. There is no single observations-per-parameter rule that covers penalized regression, decision trees, and large neural networks.\n\nLearning curves describe how performance changes with training-sample size and can reveal diminishing returns or continuing data limitations. Such curves are specific to the task, data distribution, algorithm, and evaluation criterion. Adding highly redundant records may yield little improvement, while adding observations from previously underrepresented conditions can change performance substantially. The relevant quantity is therefore not only the size of training data, but also its coverage of the conditions under which the model will be evaluated."
    },
    {
      "heading": "Bayesian, simulation-based, and sequential determination",
      "text": "In Bayesian inference, sample-size planning can target properties of the future posterior distribution, such as the width of a credible interval, the probability that an interval excludes a specified value, or the expected loss associated with a decision. These calculations combine a sampling model with a prior distribution. Prior information can affect the amount of new data required, although its influence depends on the model and on whether the prior is compatible with the eventual observations.\n\nA normal example makes the information accounting explicit. Suppose \\(X_i\\mid\\mu\\) are independent normal observations with known variance \\(\\sigma^2\\), and \\(\\mu\\) has a normal prior with variance \\(\\tau^2\\). The posterior variance is\n\n\\[\\operatorname{Var}(\\mu\\mid X) = \\left(\\frac{1}{\\tau^2}+\\frac{n}{\\sigma^2}\\right)^{-1}.\\]\n\nThe prior contributes precision \\(1/\\tau^2\\), while the data contribute \\(n/\\sigma^2\\). This permits a direct calculation of \\(n\\) for a desired posterior variance. Interpreting prior precision as an equivalent sample size is straightforward in this conjugate example, but no single equivalence applies to every prior, likelihood, or inferential target.\n\nBayesian assurance averages a study's probability of satisfying a success criterion over a distribution of uncertain design parameters. Unlike conventional power evaluated at one fixed effect, it can incorporate uncertainty about the effect itself and about nuisance quantities such as the variance. The distribution used to represent design uncertainty may differ from the prior used in the final analysis. A high target assurance can be unattainable if the design distribution assigns substantial probability to effects that would not satisfy the success criterion even with very large samples.\n\nWhen analytical formulas are unavailable, the Monte Carlo method can evaluate candidate sample sizes. A simulation generates datasets under specified scenarios, applies the intended analysis, and records outcomes such as interval coverage, power, bias, or predictive error. Repeating this process estimates the operating characteristics for each design. Simulation error must also be considered: if a true success probability is about 0.8, 10,000 independent simulated studies estimate it with a standard error of approximately 0.004. Plausible alternative scenarios are important because an accurately simulated but unrealistic model can still produce an unsuitable design.\n\nSequential designs do not always fix the final sample in advance. A group-sequential design examines accumulating data at prespecified stages and may stop early for efficacy or futility. It therefore has both a maximum sample size and an expected sample size that varies with the true effect. Repeatedly applying an ordinary fixed-sample significance threshold without adjustment generally changes the overall false-positive probability.\n\nAdaptive designs can modify allocation or sample size using accumulating information under defined rules. For example, a design may re-estimate an initially uncertain outcome variance. Such procedures are evaluated as whole algorithms, including their adaptation and stopping rules. Their uncertainty cannot generally be understood by inserting the eventual final count into a formula derived for a fixed design."
    },
    {
      "heading": "Historical development, limitations, and reporting",
      "text": "Modern sample-size determination developed alongside mathematical accounts of sampling error and statistical testing. William Sealy Gosset's 1908 paper, published under the name “Student,” established the small-sample distribution central to inference about a normal mean with unknown variance. Ronald Fisher developed influential methods of experimental design in the 1920s and 1930s, emphasizing randomization, replication, and the organization of experiments. Jerzy Neyman and Egon Pearson's 1933 work on testing hypotheses formalized the framework connecting significance levels, alternative hypotheses, and power.\n\nSurvey sampling supplied a complementary tradition focused on precision, allocation, and cost. Neyman's 1934 work developed principles of stratified sampling, including allocation based on within-stratum variability. Later computational methods made it increasingly feasible to determine sample sizes for nonstandard outcomes, multilevel designs, and adaptive procedures. Jacob Cohen's books on statistical power analysis, first published in 1969 and revised in 1988, helped make explicit power calculations widely used in behavioral research.\n\nThese developments did not establish a universal threshold separating “small” and “large” samples. Such descriptions are relative to a model and a task. Thirty observations may be adequate for one precisely measured quantity and inadequate for another. The familiar claim that a sample of 30 automatically justifies normal approximations has no general mathematical basis. Convergence depends on the underlying distribution, the statistic, and the portion of its distribution being approximated.\n\nLarge samples reduce some kinds of random error but do not automatically remove selection bias, confounding, or measurement error. For an estimator \\(\\hat\\theta\\), the decomposition\n\n\\[\\operatorname{MSE}(\\hat\\theta) = \\operatorname{Var}(\\hat\\theta) + \\left(\\operatorname{Bias}(\\hat\\theta)\\right)^2\\]\n\nshows why increasing \\(n\\) can leave a substantial mean squared error when bias persists. A very precise estimate of the wrong target remains wrong. Similarly, causal inference requires identification assumptions or an appropriate design; a larger observational sample cannot alone establish an unmeasured causal contrast.\n\nVery large samples also separate statistical detectability from practical importance. Under regular conditions, a fixed nonzero effect becomes easier to distinguish from zero as \\(n\\) increases, even if its magnitude is negligible for the scientific question. Conversely, a small study can yield an inconclusive result despite an effect that would matter substantively. Effect estimates, uncertainty intervals, and the meaning of the target difference are therefore distinct from the sample count and the significance decision.\n\nA reproducible sample-size account specifies the primary estimand or hypothesis, analysis method, significance level or interval criterion, desired power or precision, assumed effect and variability, allocation, dependence adjustments, and allowance for incomplete data. It also identifies any software or simulation procedure, rounding rules, and sensitivity analyses. Reporting the realized numbers at recruitment, follow-up, and analysis makes it possible to distinguish an original design assumption from what the investigation actually achieved."
    }
  ],
  "links": [
    "statistics",
    "random-sampling",
    "population-census",
    "covariance",
    "estimator",
    "variance",
    "sample-mean",
    "standard-error",
    "central-limit-theorem",
    "confidence-interval",
    "normal-distribution",
    "students-t-distribution",
    "binomial-distribution",
    "statistical-hypothesis-testing",
    "null-hypothesis",
    "significance-level",
    "statistical-power",
    "type-i-and-type-ii-errors",
    "effect-size",
    "multiple-testing",
    "bonferroni-correction",
    "p-value",
    "effective-sample-size",
    "degrees-of-freedom",
    "time-series",
    "markov-chain-monte-carlo",
    "missing-data-imputation",
    "poisson-distribution",
    "correlation",
    "factorial-design",
    "linear-regression",
    "ordinary-least-squares",
    "logistic-regression",
    "maximum-likelihood-estimation",
    "cross-validation",
    "machine-learning",
    "regularization",
    "training-data",
    "bayesian-inference",
    "posterior-distribution",
    "credible-interval",
    "prior-distribution",
    "monte-carlo-method",
    "ronald-fisher",
    "experimental-design",
    "selection-bias",
    "confounding",
    "mean-squared-error",
    "causal-inference"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}