Measurement error is the difference between a measured quantity value and a reference quantity value. It is a central concept in metrology, experimental science, engineering, and statistics because observations do not reproduce physical quantities without limitations. Instruments have finite resolution, measurement procedures depend on assumptions, environmental conditions fluctuate, and the objects being measured may change during observation. Understanding these effects is necessary for comparing results, testing theories, manufacturing interchangeable products, and interpreting relationships in observational data.
The word error does not necessarily imply a mistake. An instrument operated correctly can still produce a result that differs from its reference value. Conversely, a transcription mistake or an incorrectly selected unit is a failure in the measurement process, not simply another instance of ordinary random variation. Measurement error must also be distinguished from measurement uncertainty: error is a difference, whereas uncertainty describes the dispersion of quantity values reasonably attributable to the quantity being measured.
The treatment of measurement error combines physical models of instruments, statistical descriptions of repeated observations, comparisons with measurement standards, and analysis of how imperfect observations affect subsequent calculations. Its interpretation depends on a clear definition of the quantity sought, the reference used, and the conditions under which measurements are made.
Definition, reference values, and the measurand
A measurement begins with the specification of a measurand: the quantity intended to be measured. “The temperature of a water sample,” for example, is incomplete unless relevant conditions are identified. The result may depend on the position of the sensor, the time of observation, whether the sample is stirred, and whether the sensor has reached thermal equilibrium. Similarly, a metal component’s length depends on temperature, mechanical loading, and the precise surfaces used to define its endpoints. Some apparent disagreements between measurements therefore arise because the procedures do not realize the same measurand.
For a measured value and reference value , the measurement error is
The error has the same unit as the measured quantity and can be positive, negative, or zero. When the reference value is nonzero, relative error is defined as
Multiplication by 100 expresses it as a percentage. Relative error is useful for comparing measurements at different scales, but becomes unstable near a zero reference value. The magnitude is sometimes called the absolute error, although terminology varies and must not be confused with an uncertainty bound.
A reference value need not be an exactly known “true value.” It can be supplied by a measurement standard, a certified reference material, a higher-accuracy procedure, or an agreed conventional value. If a balance indicates grams for an object whose reference mass is grams, its estimated indication error is grams. The reference mass itself nevertheless has an uncertainty. The comparison therefore estimates the balance’s error rather than revealing it with unlimited precision.
This limitation is fundamental: for most empirical measurements, the exact true quantity value is unknown. A practical error estimate relies on a reference whose suitability and uncertainty are established. Metrological traceability connects a result to a reference through a documented, unbroken chain of calibrations, each contributing uncertainty. Traceability does not make an instrument error-free; it establishes how its results relate to a recognized reference.
Measurement error is also distinct from sampling error. Measuring the masses of selected packages imperfectly creates measurement error. Selecting a finite set of packages to estimate the production line’s average creates sampling variability. Both can affect the same reported estimate, but their causes and mathematical treatment differ. A precisely measured sample can still be unrepresentative, just as a representative sample can be measured inaccurately.
Random error, systematic error, and measurement quality
The conventional division of measurement error distinguishes random and systematic components. A random component varies unpredictably across replicate measurements under specified conditions. A systematic component remains constant or varies in a predictable manner. In a simple model,
is the quantity value sought, is a fixed systematic offset, and is a random component with zero expected value. This decomposition is useful, but it depends on which conditions are held fixed and which are allowed to vary.
For example, an instrument’s zero offset may act as a systematic effect during one afternoon’s measurements. Across many laboratories, independently varying zero offsets may instead contribute to between-laboratory variation. An environmental influence can likewise appear random when room temperature fluctuates unpredictably, but systematic when temperature rises steadily during an experiment. “Random” and “systematic” describe behavior within a measurement design, not immutable categories of physical causes.
Measurement bias is an estimate of systematic measurement error. A correction compensates for an estimated systematic effect: if an instrument reads units too high, the additive correction is units. Applying that correction does not eliminate all uncertainty, because the estimated bias has uncertainty and the instrument’s behavior may change. Systematic effects can also depend on the quantity being measured. A scale-factor error grows with input, while nonlinearity produces a more complicated relationship between indication and reference value.
Precision concerns agreement among replicate indications or measured values under specified conditions. Repeatability describes precision when measurements use the same procedure, operators, measuring system, location, and operating conditions over a short period. Reproducibility concerns changed conditions, such as different laboratories or measuring systems, with the changes explicitly specified. Intermediate precision describes arrangements between these cases, such as repeated measurements within one laboratory on different days.
Trueness concerns agreement between the average of an indefinitely large set of replicate measured values and a reference value; it is associated with systematic error rather than random scatter. Accuracy is the broader qualitative idea of closeness to a true quantity value and is not itself a numerical quantity in international metrological terminology. Thus, a sensor can be highly precise but substantially biased. Repeatedly obtaining with very little scatter does not establish accuracy if a suitable reference indicates . Conversely, a noisy instrument may have little average bias while individual readings remain widely dispersed.
Physical and procedural sources of error
Instrumental effects include imperfect scale factors, zero offsets, nonlinearity, drift, hysteresis, and limited resolution. Hysteresis occurs when an indication depends on the direction or history of the input: a pressure gauge may give different readings while pressure is increasing and decreasing. Drift is a gradual change in an instrument’s response over time, often associated with aging, mechanical relaxation, contamination, or changes in electronic components. Neither effect is necessarily captured by a short series of repeated readings.
Finite digital resolution creates quantization. If a display rounds an underlying indication to the nearest increment , the rounding error lies between approximately and . Under the additional assumption that values are uniformly distributed within that interval, its standard uncertainty is
The uniform model is not automatic: repeated identical displayed values do not establish that the underlying measurement is exact, nor does displaying additional decimal places necessarily provide additional information. Analog instruments have related limitations from scale spacing, pointer width, parallax, and the observer’s interpolation.
Environmental influences affect both instruments and measured objects. Thermal expansion changes dimensions, humidity changes the mass of hygroscopic materials, and ambient electromagnetic fields can disturb sensitive electronics. Precision weighing also involves air buoyancy: a body displaces air, so the balance response depends partly on the body’s volume and air density. For high-accuracy work, an apparent mass difference may require a buoyancy correction rather than being interpreted solely as a change in the object’s mass.
The act of measurement can alter the quantity being measured. A voltmeter with finite input resistance draws current from a circuit and can change the voltage at its connection points. A temperature probe exchanges heat with its surroundings and may perturb a small sample. Dynamic instruments also have response times. For a first-order sensor responding to a step from to ,
where is the time constant. Reading the sensor before equilibration produces a predictable lag relative to the final temperature.
Procedural effects include sample contamination, evaporation, inconsistent positioning, incomplete chemical extraction, and ambiguous endpoint detection. In analytical chemistry, a calibration solution may accurately characterize the detector while failing to reproduce the sample’s chemical matrix. Interfering substances can suppress or enhance the response, causing a matrix-dependent error. Transcription errors, incorrect software settings, and unit conversions require separate controls because they may generate grossly incorrect results that an ordinary noise model does not describe.
Statistical behavior and repeated observations
Repeated observations make some components of measurement error accessible to statistical analysis. If observations are independent, have the same expected value, and have finite common variance , their arithmetic mean has variance . Its standard deviation is therefore
When is unknown, the sample standard deviation is commonly used to estimate the standard uncertainty associated with the mean as . This decrease concerns the varying component represented by the observations, not every source of measurement error.
A fixed offset does not decrease through averaging. In the model , the mean approaches , not , as independent random fluctuations average out. Increasing the number of observations can consequently make a biased result appear more convincing by producing a narrow interval around the wrong center. Repetition is informative about the effects that change during repetition; it cannot by itself expose a shared calibration error, an incorrect physical model, or a constant contamination level.
Independence is also essential. For correlated observations,
Positive correlations generally reduce the benefit of averaging. Temperature drift, common power-supply fluctuations, and slowly changing backgrounds can make successive readings highly correlated. Taking a thousand readings in rapid succession may then contain much less information than a thousand independent measurements. Time-series analysis and measurements at different averaging durations help distinguish rapidly varying noise from slow instability.
A normal distribution often approximates the combined effect of many small contributions, but measurement errors need not be normal. Photon or radioactive-decay counts may follow a Poisson model under suitable conditions; rounding may be modeled by a uniform distribution; rare disturbances may produce heavy tails or mixtures. For Poisson counts with expected count , the standard deviation is , so relative counting fluctuations decrease approximately as . Background subtraction and detector imperfections add further components.
Statistical treatment must also distinguish an unusual observation from an invalid one. A large residual can indicate an instrument fault, but can also be an expected extreme or evidence that the assumed distribution is inadequate. Removing observations solely because they disagree with the desired result introduces selection bias. Robust statistical methods reduce sensitivity to extreme values, but do not identify their physical cause. Residual patterns, control measurements, instrument logs, and independent repetitions provide complementary evidence about whether the statistical model represents the measurement process.
Uncertainty evaluation and propagation
Measurement uncertainty expresses incomplete knowledge about a quantity value after relevant information has been considered. Unlike signed measurement error, uncertainty parameters are nonnegative. A result may have an estimated error close to zero and still have appreciable uncertainty. Conversely, a large known systematic effect may be corrected accurately, leaving a relatively small uncertainty. Treating uncertainty as merely “the possible error” obscures these distinctions and can imply a guaranteed bound where none has been established.
The Guide to the Expression of Uncertainty in Measurement, commonly called the GUM, distinguishes Type A and Type B evaluations. Type A evaluation uses statistical analysis of measured values obtained under defined conditions. Type B evaluation uses other information, such as calibration certificates, instrument specifications, reference data, and knowledge of instrument behavior. These are categories of evaluation method, not synonyms for random and systematic error. A systematic correction may have a Type A uncertainty evaluation, while a random resolution effect may be evaluated by Type B methods.
For a measurement model , first-order propagation gives the combined standard uncertainty
where
is a sensitivity coefficient and is the estimated covariance between inputs. The derivatives quantify how strongly each input affects the output. Covariances matter when inputs share a reference, a calibration factor, or environmental influences; omitting them can either overestimate or underestimate the output uncertainty.
For example, electrical resistance calculated as , with uncorrelated voltage and current inputs and sufficiently small relative uncertainties, has
With relative standard uncertainties of and , the combined relative standard uncertainty is approximately . This calculation assumes the measurement model is adequate and that the linear approximation is appropriate. Strong nonlinearities, poorly determined denominators, and asymmetric input distributions can invalidate a simple symmetric approximation. Monte Carlo methods instead propagate sampled input distributions through the measurement model, preserving modeled dependencies.
An expanded uncertainty is commonly written , where is a coverage factor. A factor near 2 often corresponds approximately to 95% coverage under suitable distributional and information conditions, but it is not a universal rule. A complete report identifies the measured value, unit, uncertainty, and relevant coverage statement. An uncertainty budget records the input quantities, corrections, distributions, sensitivity coefficients, and dependencies supporting that statement.
Calibration, traceability, and quality control
Calibration establishes a relationship between quantity values supplied by standards and corresponding instrument indications, including their uncertainties, and uses that relationship to obtain measurement results from indications. Calibration is not the same as adjustment. A calibration may reveal that an instrument reads high without changing it; adjustment alters the instrument to modify its response. Verification is different again: it establishes whether specified requirements are fulfilled. An instrument can be calibrated yet fail a particular performance requirement.
A simple calibration model is
where is the reference input, the instrument indication, an offset, and a sensitivity or scale factor. Estimating from an observed requires inversion of this relationship and consideration of uncertainty in the fitted parameters. If both reference values and indications have appreciable uncertainty, ordinary regression assumptions may be inadequate. Calibration points also need to characterize the range of use: extrapolation outside that range can be unreliable when instrument response is nonlinear.
Measurement traceability commonly connects working instruments to laboratory standards, national standards, and realizations of the International System of Units. Since 20 May 2019, the SI has been defined through fixed numerical values of seven defining constants. This makes the definitions exact, but does not make their laboratory realizations exact. Realizing a unit, transferring it between standards, and applying it to an actual object each introduce uncertainty. A traceable measurement therefore remains conditional on its documented calibration chain and the associated measurement procedure.
Routine quality control addresses changes between calibrations. Check standards, blank samples, duplicate measurements, and control charts can reveal drift, contamination, or changes in variability. A blank assesses signals that arise without the analyte of interest, while a reference material provides information about performance for a material with assigned properties. An internal check may reveal instability without establishing the absolute correctness of the measurement scale. Independent references are especially valuable for detecting errors shared by the routine method and its usual controls.
Interlaboratory comparisons test agreement under broader conditions. Persistent differences can expose method-dependent bias, inconsistent corrections, or underestimated uncertainties. Agreement alone, however, does not prove absence of error, because laboratories may share the same reference or physical model. In conformity assessment, uncertainty also affects decisions near tolerance limits. Guard bands shift acceptance boundaries to control specified decision risks; their meaning depends on the decision rule, the uncertainty model, and assumptions about the measured population or item.
Measurement error in regression and observational research
Measurement error affects not only individual reported values but also estimates of relationships between variables. In an errors-in-variables model, the measured predictor differs from the underlying predictor of interest. Suppose
where is observed instead of . Under the classical assumptions that has zero mean and is independent of and the outcome disturbance, ordinary least-squares regression of on estimates a slope attenuated toward zero.
In this simple setting, the large-sample slope is
The variance ratio is often called a reliability ratio. If measurement-error variance equals the variance of the true predictor, the estimated slope is half the underlying slope under these assumptions. This is known as regression dilution or attenuation. It explains why a weak observed association need not imply a weak underlying relationship, but the formula does not establish how large the correction should be without additional information.
Error in an outcome has different consequences. Independent, zero-mean additive error in the outcome of an ordinary linear regression generally increases residual variance without producing the same slope attenuation. Errors correlated with predictors or disturbances can nevertheless bias estimates. With multiple predictors, nonlinear models, or differential measurement error, bias need not be toward zero; its direction depends on the model and the relationships among the errors. Measurement error in a variable used to control confounding may also leave residual confounding in studies of causation.
A different structure is Berkson error, often represented as , where a recorded or assigned value describes a group or target and actual individual values vary around it. Assigning a common exposure estimate to everyone in an area can sometimes approximate this structure. In a simple linear model with appropriate independence assumptions, Berkson error does not cause classical slope attenuation, although it increases variability. Nonlinear models can behave differently, so the direction of the error relationship is substantively important.
Correction methods include replicate measurements, validation subsamples, regression calibration, simulation–extrapolation, and likelihood-based or Bayesian models. Their validity depends on information about the measurement mechanism. Replicates can estimate within-person or within-instrument variability but may not identify a shared bias. A validation procedure must itself be characterized rather than automatically treated as exact. Without external information or restrictive assumptions, the true predictor variance, error variance, and relationship of interest may not be separately identifiable from the observed data.
Domain-specific forms and limits of measurement
In spectroscopy and mass spectrometry, measured signals must be translated into wavelengths, masses, or concentrations through instrument models and calibration. Errors can arise from background subtraction, overlapping peaks, imperfect detector response, and variation in sample preparation. A high signal-to-noise ratio does not exclude interference from another substance. Likewise, a narrow fitted peak can have a precisely determined center while the entire wavelength scale remains offset. Detection, quantification, resolution, and accuracy describe different aspects of performance.
Geophysical and environmental measurements often confront spatial and temporal representativeness. A temperature sensor measures conditions near its sensing element, whereas the desired quantity may be an area average or a long-term trend. Changes in station location, shielding, observation time, or instruments can introduce discontinuities. Separating measurement effects from genuine environmental change requires metadata, overlapping observations, and comparisons with related measurements. Variation within the region is not automatically instrument error, but it becomes relevant when a point observation is used to estimate a regional quantity.
In medicine and epidemiology, laboratory measurements, exposure estimates, and questionnaire responses have different error structures. Within-person biological variation must be distinguished from analytical variation: repeated values can differ because the person’s physiological state genuinely changes. For categorical observations, errors are often described through misclassification probabilities rather than additive differences. Sensitivity and specificity characterize particular conditional probabilities, not the overall probability that every reported classification is correct. Misclassification can affect estimated associations in ways that depend on prevalence, study design, and whether errors differ between groups.
Psychological and educational measurement adds a distinction between random score variation and whether a procedure measures the intended construct. A test can yield highly consistent scores while systematically missing important aspects of the attribute it is intended to assess. Such problems concern validity and the operational definition of the measurand, not merely instrument noise. A detailed numerical uncertainty analysis cannot by itself establish that the chosen observations adequately represent an abstract construct.
At microscopic scales, quantum mechanics introduces further distinctions. Intrinsic variation in outcomes for identically prepared systems is not simply a classical calibration defect. The uncertainty principle constrains statistical spreads associated with quantum observables, while measurement back-action concerns changes caused by measurement interactions. These concepts are related to, but not interchangeable with, instrument error or uncertainty in estimating a parameter. In particular, the familiar position–momentum uncertainty relation is not a universal formula for an apparatus’s measurement error; operational error–disturbance relations require separately defined quantities and assumptions.
Historical development and modern terminology
The systematic study of measurement error developed alongside astronomy, surveying, and geodesy. Repeated observations of celestial positions rarely agreed exactly, making it necessary to combine imperfect readings and assess their reliability. Roger Cotes discussed weighted combinations of observations in work published posthumously in 1722. Thomas Simpson’s work in the 1750s examined the benefits of averaging under explicit assumptions about observational errors. These investigations helped make observational variation a mathematical subject rather than merely an inconvenience of instruments.
Adrien-Marie Legendre published the method of least squares in 1805 in connection with determining cometary orbits. Carl Friedrich Gauss presented his treatment in 1809 and stated that he had used the method since 1795. The method chooses parameters to minimize a sum of squared residuals. Under a normal error model, this criterion has a maximum-likelihood interpretation, although least squares can also be justified through other statistical properties. A residual is not the same as a true measurement error: it is a difference from a fitted model whose parameters have themselves been estimated from observations.
Pierre-Simon Laplace developed probabilistic arguments supporting the combination of many observations and the appearance of approximately normal distributions under suitable conditions. Friedrich Wilhelm Bessel and other nineteenth-century astronomers investigated observer-dependent effects, including the “personal equation” in the timing of astronomical events. These differences showed that an observation could depend systematically on the observer even when established procedures were followed. Studies of instruments, atmospheric effects, and observer behavior gradually broadened error analysis beyond repeated numerical scatter.
Industrial standardization and national measurement laboratories expanded the practical importance of calibration and comparability. Walter A. Shewhart’s statistical quality-control work in the 1920s distinguished stable process variation from evidence of special disturbances. In this setting, measurement-system variation had to be separated from changes in the production process itself. Comparisons between laboratories further exposed the limitations of equating a small within-laboratory standard deviation with complete knowledge of measurement quality.
International terminology increasingly distinguished error from uncertainty. A 1980 recommendation by a working group convened by the International Bureau of Weights and Measures helped establish the approach later developed in the GUM, first published in 1993. The third edition of the International Vocabulary of Metrology, published in 2008, formalized distinctions among measurement error, uncertainty, precision, trueness, and related concepts. These frameworks connect statistical evaluation with physical modeling: repeated observations characterize some influences, while reference standards, corrections, environmental information, and model assumptions account for others.