Accuracy and Precision

A student titrates the same vinegar sample four times. The results are 5.20%, 5.19%, 5.21%, and 5.20% acetic acid. These are tight, reproducible numbers. The trouble is the accepted value differs. It was determined by an independent reference method: 5.02%. Every measurement was precise. Every measurement was wrong. This single scenario is why analytical chemists refuse to use "accurate" and "precise" as synonyms. Everyday speech blurs them together constantly. Accuracy asks how close a measurement sits to the true or accepted value. Precision asks how close repeated measurements sit to each other. These are logically independent properties. A data set can be precise but inaccurate (as above). It can be accurate but imprecise (scattered values that happen to average near the truth). It can be both, or neither. Recognizing which failure mode is present tells a chemist exactly where to intervene. A precision problem points to random error in technique or instrumentation. An accuracy problem points to a systematic flaw — a miscalibrated balance, a contaminated reagent, an uncorrected blank. The big picture is that every reported number in chemistry is an estimate, never the truth itself. The entire discipline of Systematic vs Random Error exists to quantify how far off that estimate might be, and why. Before any error can be classified as systematic or random, though, the chemist needs vocabulary. This means describing "how far off" in the first place — error, deviation, absolute and relative forms of each. The chemist also needs to see numerically why tight clustering says nothing about correctness. That groundwork depends on knowing what a measurement actually is, per The Nature of Chemical Measurement. It feeds directly into how we summarize data with The Mean, Median and Mode. This article builds that vocabulary from scratch. It offers precise definitions of error and deviation, plus a worked comparison of four data sets. That comparison separates accuracy from precision cleanly. It also includes a fully worked analytical example with real titration numbers. Finally, an exam-level problem forces careful distinction between the true value and the sample mean. By the end, "precise but wrong" and "accurate on average but scattered" should feel like distinct, nameable diagnoses. They should no longer feel like vague complaints.

Key takeaways

Accuracy = closeness of a measured value to the true or accepted value; quantified by error.
Precision = closeness of repeated measurements to each other; quantified by deviation and spread (standard deviation, range).
Error requires knowing (or assuming) a true value $x_t$ ; deviation requires only the data set itself, via the mean $\bar{x}$ .
A data set can be precise without being accurate — tight clustering around the wrong number is the classic trap.
Absolute error and deviation carry units; relative error and relative deviation are unitless (or in %, ppt, ppm) and let you compare across scales.
Key idea: accuracy is measured against an external truth, precision is measured against the data's own center — and confusing the two is the single most common statistical error in introductory analytical work.

Where this lives today

Every regulated analytical method — a pharmaceutical assay submitted to a drug regulator, a clinical blood-glucose meter, an environmental lab reporting lead in drinking water — must be validated separately for accuracy (often called trueness in the formal ISO 5725 standard) and precision, exactly along the lines drawn here. Trueness is established by measuring certified reference materials with known, traceable values and computing the bias of the method's mean; precision is established by running replicate measurements and computing repeatability (same analyst, same day, same instrument) and reproducibility (different analysts, days, or laboratories) standard deviations. A method can be rejected for either failure independently: a mass spectrometer that reproducibly under-quantifies a drug metabolite by 8% is precise but fails trueness; a rapid diagnostic test whose readings scatter wildly around the right average value fails precision even if unbiased overall. In clinical chemistry, point-of-care glucose meters are required by regulatory standards (such as ISO 15197) to meet both an accuracy criterion (percentage of readings within a defined window of a laboratory reference method) and a precision criterion (coefficient of variation across repeated measurements on the same sample), evaluated as two separate, non-substitutable numbers. In environmental monitoring, a mercury analyzer validated only for precision — tight replicate agreement — but never checked against a certified reference sediment or water standard could report a systematically biased "safe" reading for years, since replicate agreement alone gives no warning of a calibration drift or matrix interference. In forensic and pharmaceutical quality control, spike-and-recovery experiments (adding a known amount of analyte to a blank matrix and checking what fraction is recovered) exist specifically to probe accuracy, because internal replicate agreement cannot detect matrix effects that bias every trial the same direction — precisely the systematic-error signature worked out numerically in the vinegar and aspirin examples above, and precisely the distinction that motivates classifying errors formally into Systematic vs Random Error.

Ask the tutor: If I only have one measurement and no repeated trials, can I say anything at all about its precision?

The physical picture

Picture a dartboard. The bullseye is the true value $x_t$ — the concentration, mass, or absorbance that a perfect, error-free measurement would yield. Each dart throw is one trial in a replicate measurement. Where the darts land, individually and as a cluster, tells the whole story. If all four darts land in a tight cluster right on the bullseye, the archer is both accurate and precise — small error, small spread. If the four darts land in a tight cluster off to one side, the archer is precise (the throws agree with each other) but inaccurate (they agree on the wrong spot) — this is the vinegar titration above, and it is the pattern produced by systematic error: a bias that pushes every trial the same way by roughly the same amount. If the darts are scattered widely but their average position happens to be near the bullseye, the archer is accurate on average but imprecise — the pattern produced by random error, which pushes trials unpredictably in both directions and tends to cancel upon averaging. If the darts are scattered and off-center, both accuracy and precision have failed. The key formal move is to attach numbers to "how far off." For a single measurement $x_i$ compared against a known true value $x_t$ , define the error:

E_i = x_i - x_t

For a single measurement compared against the data set's own mean $\bar{x}$ (no true value required, or none known), define the deviation:

d_i = x_i - \bar{x}

Error is a statement about accuracy because it references an external, independently-established truth. Deviation is a statement about precision because it references only the internal consistency of the replicates. This distinction — external truth versus internal consistency — is the entire conceptual content of "accuracy versus precision," and everything else in this article is built from it.

Figure

Four dartboards: the four accuracy/precision combinations, bullseye = true value

Figure

Number line: error (vs true value) versus deviation (vs mean) for one data point

Notice how the two arrows in the second figure differ in length by more than an order of magnitude. That gap — a small deviation next to a large error — is precisely the mathematical signature of "precise but inaccurate," and it is the reason a chemist can never certify accuracy just by looking at how tightly replicates cluster.

The core ideas

Absolute error. For a single measurement $x_i$ against the true value $x_t$ :

E_i = x_i - x_t

where: $E_i$ carries the same units as $x_i$ and can be positive (measurement too high) or negative (too low). Reading: error is a signed quantity that tells you both the size and the direction of the mistake.

Relative error, which allows comparison across different magnitudes and is usually reported as a percentage or in parts-per-thousand:

E_{i,\text{rel}} = \frac{x_i - x_t}{x_t} \times 100\%

where: the denominator is the true value, never the measured value or the mean. Reading: a 0.02 mL error is trivial in a 250 mL volumetric delivery but enormous in a 0.05 mL microdroplet — relative error normalizes for this.

Absolute deviation of a single measurement from the sample mean of $N$ replicates:

d_i = x_i - \bar{x}, \qquad \bar{x} = \frac{1}{N}\sum_{i=1}^{N} x_i

where: $\bar{x}$ is computed purely from the data in hand, with no reference to any external truth. Reading: deviation measures internal scatter, and by construction the deviations of a data set sum to zero, $\sum_i d_i = 0$ .

Relative deviation (often as a percentage of the mean), and its close cousin the mean deviation used as a rough single-number precision estimate:

d_{i,\text{rel}} = \frac{x_i - \bar{x}}{\bar{x}} \times 100\%, \qquad \overline{|d|} = \frac{1}{N}\sum_{i=1}^{N} |x_i - \bar{x}|

where: absolute value bars are needed for the mean deviation because the signed deviations would otherwise cancel to exactly zero — this is why real precision statistics (standard deviation, discussed alongside The Mean, Median and Mode) use squared or absolute deviations rather than signed ones.

The accuracy–precision independence statement. For a data set of $N$ replicates with mean $\bar{x}$ and true value $x_t$ , define the bias as the error of the mean:

\text{Bias} = \bar{x} - x_t

where: bias is a population-level accuracy statistic, while individual deviations $d_i$ are precision statistics — and the equation shows explicitly that bias can be large even while every $d_i$ is tiny (all measurements clustered tightly, but the cluster itself sits far from $x_t$ ), or bias can be exactly zero while every $d_i$ is large (scattered measurements whose errors happen to cancel in the average).

Derivation AWhy deviations must sum to zero, and what that means for measuring precision

Goal. Show algebraically that $\sum_i d_i = 0$ for any data set, and explain why this forces precision to be quantified with $|d_i|$ or $d_i^2$ rather than $d_i$ itself.

Step 1 — Write the sum of deviations. By definition, for $N$ replicate measurements $x_1, x_2, \dots, x_N$ with mean $\bar{x} = \frac{1}{N}\sum_i x_i$ , the sum of the signed deviations is

\sum_{i=1}^{N} d_i = \sum_{i=1}^{N} (x_i - \bar{x}) = \sum_{i=1}^{N} x_i - N\bar{x}

Step 2 — Substitute the definition of the mean. Since $\bar{x} = \frac{1}{N}\sum_i x_i$ , it follows that $N\bar{x} = \sum_i x_i$ . Substituting,

\sum_{i=1}^{N} d_i = \sum_{i=1}^{N} x_i - \sum_{i=1}^{N} x_i = 0

Step 3 — Interpret the result. The mean is defined as exactly the point that balances the positive and negative deviations, the way a fulcrum balances a see-saw. This is not a physical law about measurements — it is a pure algebraic consequence of how $\bar{x}$ is defined. It holds for any data set, however scattered or however biased, precise or imprecise.

Step 4 — Consequence for quantifying precision. Because $\sum_i d_i = 0$ always, a "mean deviation" computed with signed values would report zero spread for a wildly scattered data set, which is obviously useless. Precision statistics must therefore strip the sign, either by taking absolute values (giving the mean absolute deviation $\overline{|d|} = \frac{1}{N}\sum_i |d_i|$ ) or by squaring (giving the variance $s^2 = \frac{1}{N-1}\sum_i d_i^2$ , which underlies the standard deviation). This is exactly why the standard deviation formula — developed alongside The Mean, Median and Mode — is built from squared deviations rather than raw ones: squaring is the simplest operation that is both always non-negative and analytically tractable (differentiable, additive in variance for independent errors).

Go deeper · Derivation B · 1 minWhy error, not deviation, can reveal a systematic bias — and why replication alone cannot fix it

Derivation BWhy error, not deviation, can reveal a systematic bias — and why replication alone cannot fix it

Goal. Show that increasing the number of replicates $N$ shrinks the precision uncertainty of the mean but leaves a constant systematic error completely unchanged, proving that precision and accuracy are governed by different mechanisms.

Step 1 — Model a measurement with both error types. Write each measurement as the true value, plus a constant systematic bias $\beta$ (the same for every trial, e.g. from an uncalibrated pipette), plus a random fluctuation $\varepsilon_i$ that varies unpredictably from trial to trial and averages to zero over many trials:

x_i = x_t + \beta + \varepsilon_i

Step 2 — Average over $N$ replicates. Take the mean of both sides:

\bar{x} = \frac{1}{N}\sum_{i=1}^{N} x_i = x_t + \beta + \frac{1}{N}\sum_{i=1}^{N}\varepsilon_i

Step 3 — Let $N$ grow large. As $N \to \infty$ , the random fluctuations $\varepsilon_i$ , having no directional preference, average toward zero by the law of large numbers:

\frac{1}{N}\sum_{i=1}^{N}\varepsilon_i \;\longrightarrow\; 0 \quad \text{as } N \to \infty

so that

\bar{x} \;\longrightarrow\; x_t + \beta

Step 4 — Compute the bias of the mean in this limit. Using the bias definition from the core equations,

\text{Bias} = \bar{x} - x_t \;\longrightarrow\; \beta

The random component has vanished from the bias entirely — replication drives it out — but the systematic term $\beta$ survives untouched, no matter how large $N$ becomes. This is the mathematical proof of a fact every analyst learns the hard way: you cannot average away a systematic error. Only precision (the spread of $\varepsilon_i$ around zero, quantified by the standard deviation of the mean, which shrinks as $1/\sqrt{N}$ ) improves with more replicates. Accuracy, once a constant bias $\beta$ is present, requires finding and removing that bias — recalibrating the instrument, correcting for a reagent impurity, running a blank — not simply repeating the measurement. This asymmetry is the seed of the full classification into Systematic vs Random Error.

Worked examples

Example 1 — Foundational: computing error and deviation for a single replicate set. A student determines the density of an unknown metal block by mass and volume displacement, running the measurement four times. The accepted literature density of the metal (aluminum) is $x_t = 2.70\ \text{g/mL}$ . The four trials give:

x_1 = 2.74,\ x_2 = 2.71,\ x_3 = 2.73,\ x_4 = 2.72\ \text{(all in g/mL)}

Step 1 — mean.

\bar{x} = \frac{2.74 + 2.71 + 2.73 + 2.72}{4} = \frac{10.90}{4} = 2.725\ \text{g/mL}

Step 2 — error of each trial against the true value $x_t = 2.70$ g/mL.

E_1 = 2.74 - 2.70 = +0.04, \quad E_2 = 2.71-2.70=+0.01, \quad E_3=2.73-2.70=+0.03, \quad E_4=2.72-2.70=+0.02

All four errors are positive — every trial reads high. This pattern (all errors sharing one sign, of similar magnitude) is the signature of a systematic error, perhaps from residual air trapped under the block during volume displacement, which would make the measured volume too small and the density too large. Step 3 — bias of the mean.

\text{Bias} = \bar{x} - x_t = 2.725 - 2.700 = +0.025\ \text{g/mL}, \qquad \text{relative bias} = \frac{0.025}{2.70}\times100\% = 0.93\%

Step 4 — deviation of each trial against the sample mean $\bar{x}=2.725$ g/mL (this is the precision calculation, and it does not need $x_t$ at all).

d_1 = 2.74-2.725=+0.015, \quad d_2=2.71-2.725=-0.015, \quad d_3=2.73-2.725=+0.005, \quad d_4=2.72-2.725=-0.005

Check: $d_1+d_2+d_3+d_4 = 0.015-0.015+0.005-0.005 = 0$ , confirming the algebraic identity from the derivation above. Step 5 — mean absolute deviation as a precision summary.

\overline{|d|} = \frac{0.015+0.015+0.005+0.005}{4} = \frac{0.040}{4} = 0.010\ \text{g/mL}

Interpretation. The precision is good: replicates agree to within about $0.010$ g/mL of their own mean, or $0.010/2.725 \times 100\% \approx 0.37\%$ relative — tight clustering. But the accuracy is worse: the mean itself sits $0.025$ g/mL (0.93%) above the true density. This is a mild version of "precise but not fully accurate," almost certainly driven by a systematic technique flaw rather than random noise, since all four errors share the same sign.

Example 2 — Applied: the vinegar titration from the overview, quantified. Four titrations of a household vinegar sample for acetic acid content, against a certified reference value of $x_t = 5.02\%$ (w/v), give:

x_1 = 5.20\%,\ x_2 = 5.19\%,\ x_3 = 5.21\%,\ x_4 = 5.20\%

Step 1 — mean.

\bar{x} = \frac{5.20+5.19+5.21+5.20}{4} = \frac{20.80}{4} = 5.200\%

Step 2 — deviations (precision).

d_1 = 5.20-5.200=0.00,\quad d_2=5.19-5.200=-0.01,\quad d_3=5.21-5.200=+0.01,\quad d_4=5.20-5.200=0.00

\overline{|d|} = \frac{0+0.01+0.01+0}{4} = 0.005\%

Relative precision: $0.005/5.200 \times 100\% \approx 0.10\%$ — excellent reproducibility by any standard. Step 3 — errors (accuracy), against $x_t = 5.02\%$ .

E_1 = 5.20-5.02=+0.18,\quad E_2=5.19-5.02=+0.17,\quad E_3=5.21-5.02=+0.19,\quad E_4=5.20-5.02=+0.18

\text{Bias} = \bar{x}-x_t = 5.200-5.02=+0.18\%, \qquad \text{relative error} = \frac{0.18}{5.02}\times100\% \approx 3.6\%

Step 4 — diagnosis. The relative deviation ( $\approx 0.10\%$ ) is roughly 36 times smaller than the relative error ( $\approx 3.6\%$ ). This numeric ratio is exactly what "precise but not accurate" looks like in real data: the random-error contribution is tiny, but a systematic offset of about $+0.18\%$ (w/v) — an amount consistent, for instance, with an indicator endpoint that is consistently reached a few drops past true equivalence, or a titrant standardized against a slightly-off primary standard — pushes every trial the same direction by nearly the same amount. Because the sign and magnitude of the four errors are so consistent, an analyst should suspect calibration of the NaOH titrant or a systematic indicator error, not random pipetting noise, and should not expect additional replicates to fix it (per the second derivation).

Example 3 — Exam-level: distinguishing the mean's accuracy from its precision when $x_t$ must be estimated by an independent method. A pharmaceutical lab assays the potency of an aspirin tablet (labeled 325 mg) by UV-spectrophotometry, running six replicate tablets, and separately verifies four of the same batch by an independent HPLC reference method assumed to give the true content. Spectrophotometric results (mg aspirin per tablet):

319.8,\ 321.4,\ 318.6,\ 320.9,\ 322.1,\ 317.9

HPLC reference results (assume this method is accurate, i.e. its mean approximates $x_t$ ):

323.5,\ 324.1,\ 323.8,\ 324.4

Step 1 — establish $x_t$ from the reference method.

x_t \approx \frac{323.5+324.1+323.8+324.4}{4} = \frac{1295.8}{4} = 323.95\ \text{mg}

Step 2 — mean and deviations of the spectrophotometric method (precision).

\bar{x} = \frac{319.8+321.4+318.6+320.9+322.1+317.9}{6} = \frac{1920.7}{6} = 320.117\ \text{mg} \approx 320.12\ \text{mg}

Deviations from this mean:

d_1=-0.32,\ d_2=+1.28,\ d_3=-1.52,\ d_4=+0.78,\ d_5=+1.98,\ d_6=-2.22 \ \text{(mg, each rounded to 2 d.p.)}

Check on the sum: $-0.32+1.28-1.52+0.78+1.98-2.22 = -0.02$ , which is zero within the rounding used at each step — consistent with the exact-arithmetic identity $\sum d_i = 0$ . Mean absolute deviation:

\overline{|d|} = \frac{0.32+1.28+1.52+0.78+1.98+2.22}{6} = \frac{8.10}{6} = 1.35\ \text{mg}

Relative precision: $1.35/320.12 \times 100\% \approx 0.42\%$ — reasonably precise for a UV assay. Step 3 — bias of the spectrophotometric mean against the HPLC-derived true value.

\text{Bias} = \bar{x} - x_t = 320.12 - 323.95 = -3.83\ \text{mg}, \qquad \text{relative error} = \frac{-3.83}{323.95}\times100\% \approx -1.18\%

Step 4 — the exam trap, made explicit. A careless analyst might compute "error" using each spectrophotometric reading against the spectrophotometric mean (that is just deviation again, disguised) or might compute deviations of the HPLC readings against the spectrophotometric mean (mixing two different data sets' reference frames). Both are conceptually wrong. The correct comparison for accuracy always crosses from the method under test to an independently established truth ( $\bar{x}_{\text{UV}}$ vs. $x_t$ from HPLC); the correct comparison for precision always stays within one method's own replicate set ( $x_i$ vs. that method's own $\bar{x}$ ). Here, the spectrophotometric method is reasonably precise (0.42% relative spread) but carries a real negative bias of about $-1.18\%$ relative to the reference — meaning the UV method systematically under-reports aspirin content by roughly 3.8 mg per 325 mg tablet, a bias worth investigating (perhaps an incorrect molar absorptivity, or a decomposition product absorbing at the same wavelength) before the method is validated for release testing.

Limiting cases, scope & common traps

Scenario	Deviations $d_i$	Errors $E_i$	Diagnosis
All replicates cluster tightly near $x_t$	small	small	accurate and precise (ideal)
Replicates cluster tightly, off from $x_t$	small	large, same sign	precise, not accurate → suspect systematic error
Replicates scattered widely, mean near $x_t$	large	small (for the mean)	"accurate on average," not precise → suspect random error
Replicates scattered, mean far from $x_t$	large	large	neither — both error types present
$N \to \infty$ with constant bias $\beta$ present	shrinks toward pure random spread	bias of mean $\to \beta$ (unchanged)	replication cannot cure inaccuracy (see second derivation)
$\beta = 0$ (no systematic error)	unaffected by this condition	shrinks toward 0 as $N\to\infty$	replication does cure imprecision, given enough trials
True value $x_t$ unknown	computable directly from data	undefined — cannot compute error without a reference	precision can always be assessed; accuracy cannot, without a standard

The chemist behind it

The vocabulary of error and precision as used in analytical chemistry today traces most directly to the mathematical treatment of observational error built up across the 18th and 19th centuries by astronomers and geodesists — scientists forced to reconcile many imperfect measurements of a single fixed quantity, such as a star's position or a survey baseline. Carl Friedrich Gauss formalized the treatment of random measurement error and the method of least squares, giving the field its normal (Gaussian) distribution model for random scatter, while the distinction between constant, determinate error and variable, indeterminate error was sharpened by 19th- and 20th-century physical scientists translating astronomical error theory into laboratory practice. Analytical chemistry inherited this apparatus wholesale in the 20th century, and modern textbook treatments — most influentially those of Skoog, West, and Holler in the United States — codified "accuracy" and "precision" as the two independent axes along which every measurement method must be validated, alongside the language of systematic and random error that the next level of this vocabulary makes precise.

Origins · 1 minCarl Friedrich Gauss and how this was discovered

Carl Friedrich Gauss

1777–1855 · German mathematician and astronomer whose theory of errors underlies the standard-deviation and least-squares machinery used throughout modern analytical chemistry

Gauss developed the method of least squares (independently of Legendre) while reconciling repeated astronomical observations of the same celestial object, and derived what is now called the Gaussian (normal) distribution as the natural probability law for random observational error. His insight that random errors, however individually unpredictable, follow a predictable statistical distribution in aggregate is the theoretical foundation beneath every modern claim that "precision improves with more replicates" — a claim made rigorous in the second derivation above.

History

- 1795 — Gauss begins using least-squares reasoning to reconcile repeated astronomical measurements.

- 1809 — Gauss publishes Theoria Motus, formally deriving the normal distribution as the law governing random observational error.

- 1900s–1930s — physical and analytical chemists adapt astronomical error theory into laboratory error analysis, formalizing "determinate" (systematic) versus "indeterminate" (random) error.

- 1963 — first edition of Skoog & West's analytical chemistry textbook codifies "accuracy" and "precision" as the two independent validation criteria still taught today.

- Present — accuracy and precision, formalized as trueness and precision under ISO 5725, remain the two pillars of method validation in every regulated analytical laboratory (pharmaceutical, environmental, clinical, forensic).

Is "precision" the same thing as "reproducibility"?

They are closely related but not identical in careful usage. Precision is the general term for how closely repeated measurements agree with each other, usually quantified by a standard deviation, range, or mean deviation. Reproducibility is a specific, formally defined type of precision: agreement between measurements made under changed conditions — different analyst, different day, different instrument, sometimes different laboratory. Its counterpart, repeatability, is precision measured under unchanged conditions (same analyst, same session, same instrument). Both are precision statistics in the sense used in this article — neither says anything about closeness to a true value — but they probe different sources of random variability, which matters when deciding whether a method's scatter comes from the chemistry itself or from operator-to-operator variation.

Can a measurement be accurate without being precise?

Yes, and it happens whenever random errors of roughly equal magnitude occur in both directions across replicate trials, so that they largely cancel when averaged. A student's four titration readings might be 5.02%, 4.85%, 5.31%, 4.95% — scattered widely (poor precision) — yet their mean of 5.033% sits almost exactly on an accepted value of 5.02% (excellent accuracy of the mean). This pattern typically signals random error: inconsistent technique, a poorly controlled variable, or an unstable instrument, rather than any consistent bias. It is riskier to rely on than the tight, biased case, because a single measurement drawn from this data set could be far from the truth even though the average is fine.

How do absolute and relative error differ, and when should each be reported?

Absolute error, $E_i = x_i - x_t$ , carries the same units as the measurement (grams, molarity, percent) and tells you the literal size of the discrepancy. Relative error, $E_{i,\text{rel}} = (x_i-x_t)/x_t \times 100\%$ , is unitless (or expressed as %, ppt, or ppm) and normalizes the discrepancy against the scale of the quantity being measured. Relative error is almost always the more meaningful number for comparing methods or reporting quality, because an absolute error of 0.5 mg is negligible for a 500 mg tablet assay but catastrophic for a 2 mg trace-metal determination. Regulatory and journal reporting conventions typically demand relative error (or percent recovery, a closely related quantity) precisely for this reason.

Why does the sum of deviations always equal zero, but the sum of errors does not?

The sum of deviations is zero purely by the algebraic definition of the mean: $\bar{x}$ is defined as the value that exactly balances the sum of $(x_i - \bar{x})$ to zero, so this holds for any data set regardless of how good or bad the measurements are. Errors are computed against an independent, externally fixed value $x_t$ that has no algebraic relationship to the particular data set at hand. There is no mathematical reason for $x_i - x_t$ to sum to zero across replicates — and in the presence of a systematic bias, the sum of errors is strongly and predictably nonzero, which is exactly the numerical signature used to detect systematic error in real data.

Does more replication always improve precision as much as we'd like?

Precision, quantified through the standard deviation of the mean, improves proportionally to $1/\sqrt{N}$ , not $1/N$ — so quadrupling the number of replicates only halves the spread of the mean, and improvements have strongly diminishing returns. Doubling from 4 replicates to 8 shrinks the standard error by a factor of only $\sqrt{2}\approx1.41$ , not 2. This is why analytical protocols typically specify a modest, fixed number of replicates (often 3 to 6) rather than chasing arbitrarily tight precision through brute-force repetition, and why, once a systematic bias is suspected, resources are better spent finding and removing that bias than running yet more trials.

Is a certified reference material always required to assess accuracy?

A certified reference material (CRM) with a rigorously established, traceable value is the gold standard, but it is not the only route. Accuracy can also be assessed by comparison against an independent, well-validated reference method (as in the aspirin example, where HPLC served as the accuracy benchmark for a UV assay), by spike-and-recovery experiments that add a known quantity of analyte to a blank or real matrix and check the recovered fraction, or by interlaboratory comparison exercises where consensus values across many laboratories approximate the truth. What all these approaches share is the same logical structure: some value external to the replicate set under test must be brought in, because — as the article's core distinction insists — accuracy cannot be assessed from a data set's own internal scatter alone.

If my data is both accurate and precise, am I done validating the method?

Not necessarily. Accuracy and precision as defined here describe performance under the specific conditions tested — a given concentration range, matrix, analyst, and instrument. A method might be both accurate and precise at high analyte concentrations but fail at trace levels (a linearity and limit-of-detection issue), or might perform well on a clean standard solution but suffer matrix interference in a real sample (a selectivity issue). Full method validation therefore checks accuracy and precision across the intended working range and in representative sample matrices, not just at one convenient concentration — accuracy and precision are necessary criteria, but they are evaluated as functions of concentration and matrix, not single fixed numbers.