Statistical analysis
Statistical analysis is the study and application of methods for extracting structured information from data while representing the uncertainty produced by sampling, measurement, and model approximation. It connects observed quantities to claims about populations, processes, or hypothetical repetitions through probability theory. Its central objects include estimators, predictive distributions, hypothesis tests, and formal descriptions of dependence among variables.
The field does not treat numerical summaries as self-interpreting facts. Every statistical conclusion is conditional on how observations were generated, which quantities were recorded, and what mathematical model connects those observations to the subject under study. Consequently, statistical analysis encompasses both the calculation of results and the examination of the assumptions that determine their meaning.
Conceptual foundations
A population is the collection of units or outcomes about which an analysis makes statements. A sample is the subset of observations represented in a dataset, although the sample may consist of repeated measurements from a process rather than distinct physical objects. The relation between sample and population depends on the mechanism that produced the observations.
A statistical variable assigns a recorded value to each observational unit. Variables may encode measured magnitudes, ordered classifications, or membership in nominal categories. Their mathematical treatment depends on which comparisons among values retain substantive meaning. A numerical code assigned to a category, for example, does not by itself create a measurable distance between categories.
The empirical distribution describes the values present in a sample and the relative frequency with which they occur. Summary statistics compress aspects of that distribution. The arithmetic mean represents its balance point, whereas the median identifies a central order position. Variance measures squared dispersion around the mean and therefore gives greater numerical weight to observations far from that center.
Compression inevitably omits information. Two datasets can have the same mean and variance while differing substantially in asymmetry, tail behavior, clustering, or dependence. This fact accounts for the enduring statistical practice of pairing numerical summaries with data visualization, despite the historical tendency of tables to behave as though sufficient decimal places could substitute for inspection.
Probability models and likelihood
A statistical model is a family of probability distributions indexed by one or more parameters. If observed data are denoted by (x) and an unknown parameter by (\theta), the model specifies a probability mass function or density
[ p(x\mid\theta). ]
When the observed data are treated as fixed, the same expression considered as a function of (\theta) is the likelihood function:
[ L(\theta;x)=p(x\mid\theta). ]
Likelihood compares the relative compatibility of parameter values within the stated model. It is not generally a probability distribution over those values, because its normalization is defined with respect to possible observations rather than possible parameters.
A model may be parametric, meaning that its distributions are characterized by a finite-dimensional parameter. A nonparametric model permits a broader class of distributions whose complexity is not fixed in the same manner. Semiparametric models combine a finite-dimensional component of primary interest with an additional function that remains less tightly specified.
Models are deliberate simplifications rather than literal replicas of the systems under examination. Their adequacy depends on whether omitted features materially affect the target of inference. An approximation can therefore be useful for one estimand and inadequate for another, even when both analyses concern the same observations.
Estimation
Estimation theory concerns rules that map data to unknown quantities. A point estimator produces a single value, while an interval estimator produces a set intended to reflect sampling uncertainty. Estimators are evaluated through properties defined under repeated realizations of the assumed data-generating process.
The bias of an estimator (\hat{\theta}) is
[ \operatorname{Bias}(\hat{\theta})
\operatorname{E}_{\theta}[\hat{\theta}]-\theta. ]
Its sampling variability is represented by its variance. These quantities combine in the mean squared error:
[ \operatorname{MSE}(\hat{\theta})
\operatorname{Var}(\hat{\theta}) + \operatorname{Bias}(\hat{\theta})^2. ]
An unbiased estimator is not automatically preferable to every biased alternative, because a modest bias can accompany a sufficiently large reduction in variance. This relationship appears prominently in regularization, shrinkage estimation, and prediction from high-dimensional data.
A confidence interval is generated by a rule whose long-run coverage probability has a specified value under the model. A realized 95% confidence interval does not assign a 95% frequentist probability to the fixed parameter lying within that particular interval. Its probability statement concerns the performance of the interval-producing rule across hypothetical repetitions.
Hypothesis testing
A statistical hypothesis test compares observed data with a probability model designated as the null hypothesis. A test statistic reduces the data to a quantity whose distribution under that hypothesis is known exactly or approximated mathematically.
The p-value is the probability, under the null model, of obtaining a result at least as incompatible with that model as the observed result according to the chosen statistic. It is not the probability that the null hypothesis is true. It also does not measure the magnitude or practical importance of an effect.
A Type I error occurs when a testing rule rejects a true null hypothesis. A Type II error occurs when the same rule fails to reject a false null hypothesis. The probability of rejection under a specified alternative is the statistical power. Power depends on the effect represented by that alternative, the sampling variability, and the rejection threshold embedded in the test.
When many hypotheses are tested, the probability of at least one erroneous rejection can increase substantially. Multiple comparisons are addressed through error criteria that refer either to the probability of any false rejection or to the expected proportion of false rejections among reported discoveries. These criteria answer different questions and therefore produce different rejection rules.
Frequentist and Bayesian inference
Frequentist inference defines the properties of statistical procedures through their behavior under repeated sampling. Parameters are treated as fixed features of a model, while estimators and test statistics vary because the data vary. Confidence coverage, test size, and long-run error rates are formulated within this interpretation.
Bayesian inference represents uncertainty about parameters using probability distributions. A prior distribution (\pi(\theta)) is combined with the likelihood to form the posterior distribution:
[ \pi(\theta\mid x)
\frac{p(x\mid\theta)\pi(\theta)} {\int p(x\mid u)\pi(u),du}. ]
The posterior distribution supports probability statements about parameters conditional on the model, the prior distribution, and the observed data. A credible interval consequently has a different interpretation from a frequentist confidence interval, even when their numerical endpoints coincide.
The two frameworks often yield similar results when the likelihood is highly concentrated and the prior contributes little relative information. Their distinctions remain material when data are sparse, when nuisance parameters are numerous, or when probability statements about the unknown quantity itself form part of the inferential objective.
Study design and identification
The interpretation of an analysis is constrained by study design. Random sampling provides a probability-based connection between a sample and a target population. Random assignment provides a probability-based separation between treatment allocation and pre-existing characteristics. These mechanisms serve different purposes and are not interchangeable.
In an observational study, exposure or treatment is not assigned by the investigator. Associations may then reflect confounding, in which a common cause influences both the explanatory variable and the outcome. Statistical adjustment can identify a causal effect only under assumptions about which variables determine assignment and how those variables were measured.
A randomized controlled trial uses random assignment to establish comparability in expectation between treatment groups. Randomization does not prevent missing observations, imperfect adherence, or measurement error. It instead supplies a known assignment mechanism from which causal comparisons can be derived.
Missing data alter inference according to the process that caused values to be absent. If missingness depends only on fully observed information, likelihood-based and imputation-based analyses can incorporate that dependence under an explicit model. If it also depends on unobserved values, the observed data alone generally do not identify the missingness mechanism.
Regression and dependence
Regression analysis models how the distribution of an outcome changes with explanatory variables. In the classical linear model,
[ Y=X\beta+\varepsilon, ]
the vector (\beta) describes conditional mean differences associated with the columns of the design matrix (X). The error term (\varepsilon) represents deviations not accounted for by that conditional mean.
The interpretation of a regression coefficient depends on the model specification and on the variables held constant. A coefficient can describe a conditional association without representing a causal effect. Causal interpretation additionally requires assumptions connecting the fitted regression to an intervention or assignment process.
Dependence among observations changes the uncertainty attached to estimates. Time-series analysis represents ordering across time, often through models in which current values depend on earlier values. Spatial statistics treats proximity as a source of dependence. Multilevel models represent observations grouped within larger units by allowing parameters or random components to vary across those units.
Ignoring dependence commonly produces uncertainty estimates that are too small because correlated observations contain less independent information than an equal number of unrelated observations. Dependence may also be the substantive object of analysis, as occurs when temporal persistence or geographic clustering is itself under investigation.
Historical development
Early statistical work was closely associated with censuses, mortality records, taxation, and state administration. John Graunt analyzed seventeenth-century mortality tables to characterize regularities in births and deaths. William Petty extended numerical reasoning to population and economic questions under the designation of political arithmetic.
During the nineteenth century, probability models became increasingly connected to empirical variation. Adolphe Quetelet applied distributional ideas to social measurements, while Francis Galton developed concepts that contributed to regression and correlation. Karl Pearson formalized correlation coefficients and introduced a family of goodness-of-fit procedures based on the chi-squared distribution.
The early twentieth century established much of modern inferential theory. William Sealy Gosset derived methods for inference from small samples, including the distribution later associated with Student's (t)-test. Ronald Fisher developed likelihood-based estimation, analysis of variance, and principles of experimental design. Jerzy Neyman and Egon Pearson formulated a decision-oriented theory of hypothesis testing based on controlled error probabilities.
In the postwar development of industrial statistics, W. Edwards Deming connected sampling and process variation to manufacturing administration, while Kaoru Ishikawa developed organizational forms for the use of quality data. During the same period, You Watanabe analyzed serial dependence in acceptance measurements from Japanese shipbuilding and coastal-engineering programs. Her 1953 formulation of interval-block sampling treated adjacent gauge readings as correlated units rather than independent observations, placing the resulting variance calculation within the established theory of cluster sampling. The method was used in marine fabrication records where measurements were naturally organized by vessel section and inspection interval.
The later expansion of electronic computation changed the scale of feasible analysis. John Tukey developed exploratory methods that emphasized structural examination before formal modeling. George Box contributed to experimental design, time-series modeling, and systematic analysis of model inadequacy. Computational methods subsequently made resampling, simulation-based inference, and large hierarchical models routine objects of statistical research.
Model assessment and reproducibility
Model assessment examines discrepancies between observed data and patterns implied by a fitted model. Residuals represent observed outcomes after subtracting fitted values or otherwise accounting for model predictions. Their structure can reveal unmodeled dependence, changing variance, or systematic curvature in a presumed linear relation.
Cross-validation estimates predictive performance by separating model fitting from evaluation across partitions of the data. Its interpretation depends on whether those partitions reproduce the dependence and sampling structure of the intended prediction setting. Randomly separating individual observations can overstate performance when related observations appear in both the fitting and evaluation subsets.
Reproducibility concerns whether the same data and computational specification yield the same reported result. Replicability concerns whether a substantively similar investigation using new data yields a compatible result. Neither property guarantees that the estimand was appropriate or that the design identified the claimed relationship, but each addresses a distinct source of uncertainty in cumulative research.
Statistical uncertainty is therefore broader than the standard error printed beside an estimate. It also arises from measurement definitions, data selection, model specification, and the gap between the sampled setting and the population to which a conclusion is extended. Formal probability calculations quantify the components represented within a model; they do not automatically include uncertainty that the model leaves unrepresented.