Measurement error

Measurement error - is the difference between the value obtained by measuring something and the value it really has. The International Vocabulary of Metrology puts it in one line: measurement error is the "measured quantity value minus a reference quantity value"^[1]. The reference may be a true value, a certified standard or an agreed conventional value.

Every measurement carries some error. A kitchen scale, a laboratory balance, a customer satisfaction questionnaire, an inventory count, a timesheet. None of them is perfect. The question is never whether error exists but how big it is, which direction it pushes the results and whether anyone has bothered to check. In statistics and research methods the concept explains why observed relationships look weaker than real ones; in quality management it decides whether a good part gets scrapped or a bad one gets shipped; in business data it sits quietly behind wrong forecasts and misleading dashboards.

Definition and formula of measurement error

The simplest way to write it down:

\[ x_{observed} = x_{true} + e \]

where e is the error. Rearranged, the absolute error is

\[ e = x_{observed} - x_{true} \]

It carries the units of the measurement (grams, euros, hours) and a sign. A positive value means the instrument over-reads, a negative one that it under-reads.

Absolute error says little on its own. Two grams wrong on a 500 g bag of coffee is a nuisance; two grams wrong on a 5 g pharmaceutical dose is a disaster. That's why people also use the relative error:

\[ e_{rel} = \frac{x_{observed} - x_{true}}{x_{true}} \times 100\% \]

There is a catch. The true value is almost never known. In practice it gets replaced by a reference value from a better instrument or a calibrated standard, and the VIM notes that measurement error is known only in that situation; when the measurand is supposed to have a unique true value nobody can observe, the error itself remains unknown^[1].

Classical test theory

Psychometrics and survey research use the same idea in slightly different clothes. Classical test theory, which grew out of Charles Spearman's early work on correlation^[2], assumes that any observed score X is the sum of a true score T and a random error E:

\[ X = T + E \]

The error is assumed to have a mean of zero and to be uncorrelated with the true score. Under those assumptions the variance splits neatly:

\[ \sigma^2_X = \sigma^2_T + \sigma^2_E \]

The share of observed variance that comes from true differences between people is the reliability of the measure. That concept has its own article, reliability of measurement, and it returns below when we look at regression.

Types of measurement error

Random versus systematic error

The most important distinction is between errors that scatter and errors that shift. The VIM defines systematic measurement error as the component that "in replicate measurements remains constant or varies in a predictable manner", and random measurement error as the component that "varies in an unpredictable manner"^[3]. An estimate of the systematic part is called bias.

Random error makes repeated readings spread around a central value. It averages out: take enough readings and their mean drifts towards the value the instrument would give without noise. It hurts precision, not the centre.
Systematic error moves all readings in the same direction. A scale that adds 2 g to every weighing will still add 2 g after a thousand repetitions. Averaging doesn't help. Only comparison with something better (a reference standard, a different method) reveals it.

The NIST/SEMATECH handbook makes the same point in quality-engineering language: bias is a quantitative term for the difference between the average of measurements made on the same object and its true value, while accuracy is a qualitative term describing agreement between a measurement and that value. Bias is typically identified by calibration against a reference laboratory's standards^[4].

Gross errors (blunders)

Some errors are neither random nor systematic in the statistical sense. They're mistakes. Someone reads 15.8 as 18.5, writes kilograms in a column meant for pounds, weighs a product with the packaging still on. Gross errors usually produce values far from the rest and are often caught by plausibility checks. Not always. A blunder that lands inside the normal range can survive every automated filter.

Sources of error

By origin, errors are commonly grouped into four families:

Instrument errors - wear, drift, poor resolution, zero offset, a sensor that has not been calibrated for months.
Observer (operator) errors - parallax when reading a dial, different operators clamping a part differently, rounding habits, fatigue at the end of a shift.
Environmental errors - temperature, humidity, vibration, electromagnetic interference. Steel gauges expand with heat.
Method errors - a measurement procedure that doesn't actually capture the quantity of interest, or an approximation built into the formula used. A survey question that measures "intention to buy" when management wants to know actual purchases is a method error, not a respondent error.

In laboratory work the same families are often discussed under the heading of experimental error.

Measurement error in surveys and business data

Surveys are measurement instruments too, and fairly noisy ones. Biemer and Lyberg place measurement error inside the total survey error framework next to sampling, coverage, nonresponse and processing errors. Among the sources they discuss are the questionnaire itself, the interviewer, the respondent and the data-collection mode^[5].

Typical forms:

Response bias - a general tendency to answer in a particular way regardless of content, such as agreeing with every statement or always picking the middle of the scale.
Social desirability bias - respondents present themselves favourably. In their review of research on sensitive questions, Tourangeau and Yan conclude that misreporting on such topics is quite common, depends on the situation and on survey design, and is a more or less motivated process in which respondents edit their answers to avoid embarrassment in front of an interviewer^[6].
Interviewer bias - tone, wording or the interviewer's own characteristics nudge answers.
Recall error - people forget, or misplace events in time. Small expenses vanish from memory altogether.
Common method bias - when the same respondent rates both the cause and the effect on the same questionnaire, shared response styles inflate the correlation between them.
Coding and data-entry errors - an open answer coded into the wrong category, a transposed digit, a decimal comma read as a thousands separator. In business systems this category keeps growing: duplicated customer records, wrong units in an ERP, timestamps in mixed time zones.

Nobody re-measures data that arrived from a transaction system, so errors of this last kind tend to stay.

Measurement error versus sampling error

The two are easy to confuse and quite different. Sampling error arises because only part of the population was observed; it shrinks as the sample grows and would vanish in a full census. Measurement error arises because each individual observation is wrong to some degree. It doesn't shrink with sample size if it is systematic. A census of 10 million households using a question everyone misunderstands produces a very precise wrong answer. Selection problems such as sample selection bias belong to a third group again: the right people weren't asked.

Effects of measurement error

Random error in a single reading is just noise. The trouble starts when noisy variables enter analysis.

Attenuation bias (regression dilution)

Suppose a manager wants to estimate how much sales Y rise per unit of some driver T (training hours, say), but only has a noisy measure X = T + U, with error U independent of T and of the regression error. Regressing Y on X by ordinary least squares gives, in large samples, not the true slope β but

\[ \beta_{observed} = \lambda \cdot \beta, \qquad \lambda = \frac{\sigma^2_T}{\sigma^2_T + \sigma^2_U} = \frac{\sigma^2_T}{\sigma^2_X} \]

Fuller calls λ the reliability ratio^[7]. Since it lies between 0 and 1, the estimated slope is pulled towards zero. Epidemiologists call the same effect regression dilution bias^[8]. Some points that often get missed:

Random error in the outcome variable does not bias a simple linear slope. It only makes the estimate less precise and lowers the coefficient of determination.
Correlations are attenuated too. Spearman's classic result says the observed correlation equals the true one multiplied by the square root of the product of the two reliabilities^[2].
With several predictors, the bias no longer goes only towards zero. Error in one variable can spill over into the coefficients of others, in either direction.
Systematic and differential error (error that depends on the outcome, as when people who had a bad experience remember it differently) can bias results in any direction.
In nonlinear models such as logistic regression, the effects are more complicated, and Carroll and co-authors devote a whole monograph to them^[9].

There is also a different error structure, the Berkson error, where the true value varies around an assigned value (for example, every employee in a department is given the department's average exposure). In a linear model this type does not bias the slope, which is exactly why it has to be told apart from the classical type before any correction is applied^[9].

Practical consequences

In quality control, measurement variation adds to process variation. A process that is actually capable can look incapable, and parts close to the specification limits get misclassified: good ones rejected, bad ones accepted. Control charts built on noisy data show wider limits and miss real shifts. In management, a KPI that wobbles because of data errors triggers reactions to noise, while a KPI that is biased lets a problem grow unnoticed.

A note on uncertainty

Error and uncertainty are not the same thing. Error is a single (usually unknown) difference between a result and the true value; uncertainty is a parameter describing the range of values that could reasonably be attributed to the measurand, given everything known about the measurement. Modern metrology reports results with their uncertainty precisely because the error itself can't be known exactly.

How to detect and reduce measurement error

No single technique fixes everything. Each one targets a particular type.

Calibration - comparing an instrument against a reference standard of known value. It is the main defence against systematic error. Calibration intervals, traceability to national standards and records of adjustments are standard requirements in quality control systems.
Repeated measurement - averaging n independent readings reduces the standard deviation of the mean by a factor of √n. It helps against random error only.
Gage R&R studies - a structured experiment in which several operators measure the same parts several times, separating repeatability (equipment variation) from reproducibility (operator variation)^[10]. The method and its acceptance criteria are covered in measurement system analysis, a standard step in Six Sigma projects before process capability is assessed.
Reliability analysis - test-retest correlation, internal consistency or inter-rater agreement for questionnaires and assessments.
Validity checks - is the instrument measuring the right thing at all? A perfectly consistent measure can still be consistently wrong. See validity and reliability.
Triangulation - comparing results from different methods or sources: survey answers against transaction records, self-reported hours against system logs. Disagreement points to error in at least one of them. The choice of data collection methods matters here.
Training and procedures - clear operating instructions, standard question wording, interviewer training, double data entry, validation rules in input forms.
Monitoring - statistical process control applied to check standards shows when an instrument starts drifting.

When error cannot be removed at the source, statistical correction is possible. Dividing an attenuated slope by an estimate of λ is the simplest approach. More advanced ones include regression calibration, simulation-extrapolation (SIMEX) and instrumental variables^[9].

Example of measurement error

A coffee roaster fills 500 g bags and checks them on a checkweigher. During a routine check the quality technician places a certified 500.0 g reference weight on the scale ten times:

Reading	Value (g)	Deviation from mean (g)
1	502.1	+0.1
2	501.8	-0.2
3	502.4	+0.4
4	501.9	-0.1
5	502.2	+0.2
6	502.0	0.0
7	501.7	-0.3
8	502.3	+0.3
9	502.1	+0.1
10	501.5	-0.5

The ten readings add up to 5020.0 g, so the mean is 502.0 g.

Absolute error of the mean (estimated bias): 502.0 - 500.0 = +2.0 g.
Relative error: 2.0 / 500.0 × 100% = 0.4%.
Sample standard deviation of the readings: about 0.28 g (random component).
Standard error of the mean: 0.28 / √10 ≈ 0.09 g.

What does this tell the roaster? The random part is small; any single reading is usually within a few tenths of a gram of the mean. The systematic part is not. Because the filling line is adjusted until bags read 500 g, they actually contain about 498 g on average. Every bag. Taking more readings would never reveal it, since all ten agree with one another nicely. Only the reference weight did. After recalibration the scale's mean reading moves back to 500.0 g and the filling target can be set honestly.

Now suppose the same company uses customer survey scores to explain repeat purchases. If the variance of true satisfaction is 16 and the variance of random response error is 4, then λ = 16 / (16 + 4) = 0.8. A true effect of 0.50 extra purchases per satisfaction point would show up in the regression as 0.8 × 0.50 = 0.40. A fifth of the effect disappears into noise. If the reliabilities of two survey measures were 0.8 and 0.7 and their true correlation 0.60, the observed correlation would be about 0.60 × √(0.8 × 0.7) ≈ 0.45.

Advantages of accounting for measurement error

Decisions rest on the actual process, not on the quirks of the instrument.
Fewer false rejections and fewer escaped defects in inspection.
Regression and correlation estimates are not systematically understated, so the value of an intervention (training, a marketing campaign) isn't underestimated.
Survey results can be compared across modes, countries and years with more confidence.
It forces the organisation to document its measurement procedures, which supports measurement of performance and audits.

Limitations of correction methods

Corrections need extra information: a reliability estimate, replicate measurements, a validation subsample or a reference standard. Without it, the size of the error is a guess.
The classical model assumes error that is random, independent of the true value and of other variables. Real data often break these assumptions; social desirability bias, for instance, depends on the true answer.
Correcting for attenuation increases the variance of the estimate. A corrected coefficient is less biased but less precise, and with low reliability it can become unstable.
Using the wrong error model (classical instead of Berkson, or the reverse) can make things worse rather than better.
Statistical correction does nothing about blunders and biased instruments. Those have to be found and fixed at the source.
Calibration, Gage R&R studies and validation surveys cost time and money.

References

Biemer P.P., Lyberg L.E. (2003), Introduction to Survey Quality, John Wiley & Sons, Hoboken
Carroll R.J., Ruppert D., Stefanski L.A., Crainiceanu C.M. (2006), Measurement Error in Nonlinear Models: A Modern Perspective, 2nd edition, Chapman & Hall/CRC, Boca Raton
Fuller W.A. (1987), Measurement Error Models, John Wiley & Sons, New York
Hutcheon J.A., Chiolero A., Hanley J.A. (2010), Random measurement error and regression dilution bias, BMJ, Vol. 340, c2289
JCGM 200:2012 (2012), International vocabulary of metrology - Basic and general concepts and associated terms (VIM), 3rd edition, Joint Committee for Guides in Metrology, Sèvres
NIST/SEMATECH (2012), e-Handbook of Statistical Methods, National Institute of Standards and Technology, Gaithersburg
Spearman C. (1904), The proof and measurement of association between two things, The American Journal of Psychology, Vol. 15, No. 1
Tourangeau R., Yan T. (2007), Sensitive questions in surveys, Psychological Bulletin, Vol. 133, No. 5, pp. 859-883

Footnotes

↑ JCGM 200:2012, International vocabulary of metrology (VIM), 3rd edition, 2012, clause 2.16 and Note 1
↑ Spearman C., The proof and measurement of association between two things, The American Journal of Psychology, 1904
↑ JCGM 200:2012, International vocabulary of metrology (VIM), 3rd edition, 2012, clauses 2.17-2.19
↑ NIST/SEMATECH, e-Handbook of Statistical Methods, section 2.1.1.3 "Bias and Accuracy"
↑ Biemer P.P., Lyberg L.E., Introduction to Survey Quality, John Wiley & Sons, 2003
↑ Tourangeau R., Yan T., Sensitive questions in surveys, Psychological Bulletin, 2007, pp. 859-883
↑ Fuller W.A., Measurement Error Models, John Wiley & Sons, 1987
↑ Hutcheon J.A., Chiolero A., Hanley J.A., Random measurement error and regression dilution bias, BMJ, 2010
↑ Carroll R.J., Ruppert D., Stefanski L.A., Crainiceanu C.M., Measurement Error in Nonlinear Models, Chapman & Hall/CRC, 2006
↑ NIST/SEMATECH, e-Handbook of Statistical Methods, section 2.4 "Gauge R & R studies"

Measurement error — recommended articles

Reliability of measurement — Measurement system analysis — Systematic error — Random error — Sampling error — Validity and reliability — Statistical process control