{
  "id": "microdata",
  "title": "Microdata",
  "lang": "en",
  "source_lang": "en",
  "summary": "Microdata are records describing individual statistical units, such as people, households, businesses, or events, used to analyze variation and relationships within populations.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Unit-level data",
    "Individual-level data",
    "Record-level data"
  ],
  "generated_at": "2026-11-06",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/microdata/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/microdata/",
      "json": "https://aiwiki.page/zh/microdata.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "In statistics, microdata are data recorded at the level of individual statistical units, such as people, households, businesses, establishments, transactions, or events. Each record describes the characteristics of a unit rather than summarizing a group of units. Microdata make it possible to study distributions, relationships, and differences within populations, but their interpretation depends on how units were selected, measured, and processed. The term refers to the level of observation, not the physical size of a dataset, and is unrelated to Microdata, the HTML mechanism for embedding structured information in web pages."
    },
    {
      "heading": "Units, records, and variables",
      "text": "A microdata file commonly has one row per observed unit and one column per variable. A household survey, for example, might record household size, income, region, and housing tenure. A business file might contain employment, sales, industry, and location.\n\nThe statistical unit need not be a person. It can be an organization, a building, a geographical area, or an occurrence such as a purchase. Data describing municipalities can therefore be microdata for a study whose units are municipalities, while being aggregate data relative to the residents within them.\n\nNot all microdata fit a single rectangular table. A dataset may contain linked household and person files, repeated observations of the same unit, or separate records for transactions associated with an account. Identifiers establish relationships among records; they do not necessarily reveal the real-world identities of the units.\n\nFor a rectangular file with \\(n\\) units and \\(p\\) variables, the observations can be represented schematically as a matrix:\n\n\\[X=(x_{ij}), \\qquad i=1,\\ldots,n,\\quad j=1,\\ldots,p,\\]\n\nwhere \\(x_{ij}\\) is the recorded value of variable \\(j\\) for unit \\(i\\). In practice, columns may contain numerical values, categories, dates, text, or missing-value codes, so the file is not necessarily a numerical matrix suitable for direct algebraic operations."
    },
    {
      "heading": "Sources and forms",
      "text": "Microdata arise from several kinds of collection:\n\n- Surveys: measurements for a sample of units selected from a target population.\n- Censuses: records intended to cover all units within a defined population, such as a population census.\n- Administrative systems: records generated through activities such as taxation, benefit administration, business registration, or school enrollment.\n- Experiments: unit-level observations collected under an experimental design.\n- Operational and digital systems: transactions, service interactions, sensor readings, and other recorded events.\n\nThese sources have different strengths and limitations. Survey data are often designed around a statistical question, whereas administrative data primarily serve operational purposes. Extensive coverage does not by itself make a file representative of the population relevant to a particular analysis.\n\nCross-sectional microdata describe units at a particular time or over a defined reference period. Longitudinal microdata follow units through time. Panel data contain repeated observations of the same units, whereas repeated cross-sectional surveys may observe different samples at successive dates. A time series of national averages is not person-level microdata merely because it contains many dates.\n\nMicrodata are also distinct from raw data. A released microdata file may already have undergone validation, coding, editing, weighting, and missing-data imputation."
    },
    {
      "heading": "Relationship to aggregate data",
      "text": "Aggregate data summarize collections of units through counts, totals, percentages, averages, or other statistics. Microdata can generally be aggregated in many ways, but aggregate tables usually cannot be used to reconstruct the underlying records uniquely.\n\nFor example, an average income does not reveal how income is distributed, which households have low incomes, or how income relates to household size. Even separate tables of income and education may not identify their joint distribution.\n\nThis distinction matters when interpreting relationships. A relationship between regional averages need not hold among individuals within those regions. Inferring individual-level patterns directly from group-level patterns can produce the ecological fallacy. Microdata permit direct investigation of within-group relationships, although measurement error, selection, and omitted variables can still affect the findings.\n\nAggregation is not inherently undesirable. It can simplify analysis, match the level of a research question, and reduce disclosure risks. Whether microdata are necessary depends on the intended inference."
    },
    {
      "heading": "Statistical analysis and weighting",
      "text": "Microdata support descriptions of distributions and subgroups, cross-tabulations, linear regression, logistic regression, and other statistical models. They are central to much empirical work in econometrics, demographic research, and social statistics.\n\nFor sample-based microdata, the unweighted records may not represent the target population in their observed proportions. Units can have unequal selection probabilities, and some sampled units may not respond. Survey weights commonly incorporate inverse selection probabilities and adjustments for nonresponse or known population totals.\n\nIf \\(y_i\\) is an observed value and \\(w_i\\) is its survey weight, a weighted estimator of a population mean is\n\n\\[\\widehat{\\mu}_w = \\frac{\\sum_{i\\in s}w_i y_i} {\\sum_{i\\in s}w_i},\\]\n\nwhere \\(s\\) denotes the observed sample. With equal weights, this reduces to the ordinary sample mean. A weight is not simply a measure of the reliability of an individual response; its interpretation depends on the sampling and adjustment procedures.\n\nUncertainty calculations must also reflect the data design. Observations within the same sampled household, school, or geographical cluster may be dependent. Treating all records as independent observations from a simple random sample can therefore give incorrect standard errors and confidence intervals. Survey files may provide strata, cluster identifiers, replicate weights, or other information for design-appropriate estimation.\n\nA large number of records reduces some forms of sampling uncertainty but does not automatically eliminate selection bias, coverage error, or systematic measurement problems."
    },
    {
      "heading": "Research uses and inferential limits",
      "text": "The main analytical advantage of microdata is flexibility. Researchers can construct variables, define subpopulations, study interactions, and estimate quantities not included in published tables. Linked and longitudinal records can reveal transitions such as movement between employment states or changes in household composition.\n\nMicrodata also support microsimulation, in which rules or models are applied to individual units. For example, a tax-benefit model can calculate how a specified policy change would affect households with different characteristics. Such calculations depend on assumptions about coverage, behavior, and the accuracy of recorded inputs.\n\nDetailed records do not by themselves establish causation. Causal inference requires an appropriate design and identifying assumptions. Confounding, endogeneity, and selective participation remain possible even when many individual characteristics are observed."
    },
    {
      "heading": "Confidentiality and access",
      "text": "Because microdata describe individual units, their release raises issues of statistical confidentiality and data privacy. Removing names and addresses does not necessarily prevent identification. Combinations of age, occupation, location, household characteristics, or unusual events may distinguish a unit and allow linkage to outside information.\n\nAccess arrangements commonly include:\n\n- Public-use files, with restricted detail and disclosure-control measures.\n- Licensed or restricted-use files, available under specified conditions.\n- Secure research environments, where approved analyses take place under controlled access and outputs may be checked.\n- Remote execution systems, in which submitted code runs against data that users cannot directly download.\n\nDisclosure controls include suppressing values, grouping categories, limiting geographical detail, top-coding large values, and perturbing records. These measures can alter distributions and relationships, creating a trade-off between analytical usefulness and confidentiality protection.\n\nSynthetic microdata replace some or all observed values with values generated from a model. They are not automatically free from disclosure risk, and their validity for a particular analysis depends on what the synthesis process preserves. Differential privacy provides a formal framework for bounding how much an output can reveal about an individual contribution, but applying it still requires choices about privacy parameters and statistical utility."
    },
    {
      "heading": "Documentation and comparability",
      "text": "Microdata require documentation to be interpretable. Essential information includes the target population, unit of observation, sampling design, reference periods, variable definitions, coding schemes, weights, and processing history.\n\nMissing-value codes require particular attention. “Not applicable,” “not asked,” “refused,” and “unknown” can represent different processes and should not automatically be treated as equivalent. Likewise, zero may be a valid observation rather than a missing value.\n\nCombining files across places or years requires more than matching variable names. Definitions, classifications, collection modes, population coverage, and disclosure controls may change. A comparable series must account for these differences, while reproducible analysis must identify the data version and document transformations from the released records."
    }
  ],
  "links": [
    "statistics",
    "matrix",
    "population-census",
    "experimental-design",
    "panel-data",
    "time-series",
    "missing-data-imputation",
    "linear-regression",
    "logistic-regression",
    "econometrics",
    "sample-mean",
    "standard-error",
    "confidence-interval",
    "selection-bias",
    "causal-inference",
    "confounding",
    "endogeneity",
    "data-privacy",
    "synthetic-data",
    "differential-privacy"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}