Precision, Accuracy and Bias

Introduction

A laboratory can make nearly identical measurements that are all wrong, or scattered measurements whose average lies near the true value. These situations require different remedies. Precision concerns repeat variation; accuracy concerns agreement with a suitable reference; bias is a persistent displacement. Chemical measurements should report the conditions under which precision was assessed and the evidence supporting an accuracy claim.

Core explanation

Precision describes how close repeated results are to one another. Repeatability refers to short-term measurements under nearly unchanged conditions, such as the same analyst, instrument and sample preparation. Intermediate precision may include different days or analysts within a laboratory; reproducibility can involve different laboratories. A small standard deviation from repeated instrument readings does not guarantee precision across independent sample preparation. The level of replication must match the intended use.

Accuracy involves a reference value with its own uncertainty. A certified reference material, validated orthogonal method or well-prepared spike can help test accuracy. In ordinary speech “accurate” may mean both close to truth and reliable; in measurement science, it is safer to state the actual observed agreement and uncertainty. Bias is the difference between the expected or average measurement and the reference under defined conditions. It may come from a wrong calibration stock, incomplete extraction, matrix suppression, contaminated blanks or a model that omits an interference.

Random variation and bias can coexist. If a balance drifts between runs, repeated masses may scatter; if it is consistently miscalibrated, values may cluster tightly around the wrong mass. More repeats reduce uncertainty of the mean for independent random errors, but do not remove a fixed calibration bias. A laboratory that repeats the same biased procedure one hundred times can report a very precise wrong average. Cross-checks must challenge the common source of error, for instance with an independent reference material.

Simple numerical summaries are useful but need context. The sample mean is the sum of readings divided by their count. The sample standard deviation describes spread among the observed independent results, using n−1 in the denominator for an ordinary estimate from n values. Relative standard deviation is standard deviation divided by mean, often expressed as a percentage; it is not meaningful when the mean is near zero. Bias can be reported as an absolute difference or percentage of the reference, provided the reference and denominator are stated.

In chemistry, sample heterogeneity complicates interpretation. Three aliquots from one well-mixed solution may show excellent repeatability, while three field samples from different locations differ greatly for real environmental reasons. That larger spread is not necessarily instrument imprecision. Separate sampling variation, preparation variation and instrument variation with a nested design. The NIST measurement-process characterization handbook covers repeatability, reproducibility, stability, calibration and uncertainty as related but distinct properties.

When comparing methods, ask whether the difference matters for the decision. A 1% bias may be negligible for a rough synthesis check but unacceptable for a reference assay. The method's validated target, concentration range and matrix determine acceptable accuracy and precision. A claimed improvement should be compared with both random variation and possible shared bias.

Step-by-step reasoning

Identify the measurand and reference. Make independent measurements at the variation level of interest: repeated injections for instrument repeatability, separate preparations for method precision, or separate laboratories for broader reproducibility. Calculate mean and spread. Compare the mean with a reference while considering reference uncertainty and any known recovery limits. Use blanks and independent standards to seek bias. Report results and limitations with explicit units and test conditions.

Visual explanation

Draw four target diagrams. Tight points centered on the target are precise and accurate; tight points off center are precise but biased; broad points centered on average are less precise but have little apparent bias; broad off-center points are neither. Label the target as a reference with uncertainty, not an infallible exact truth. Alongside, draw a time series to show a balance drifting gradually even if small groups of nearby readings appear tight.

Real-world analogy

A clock that is always seven minutes slow gives consistent readings but wrong time. A clock that jumps around by several minutes is imprecise; averaging many readings may help only if its errors are random and centered. Calibrating against a trustworthy time reference tests bias. Chemical instruments behave similarly, though matrix effects can make the bias depend on the particular sample.

Real-world example

A lab measures a 10.00 mg/L certified metal reference three times and obtains 9.49, 9.50 and 9.51 mg/L. The spread is tiny, but the mean of 9.50 mg/L is 0.50 mg/L low, a 5% negative bias relative to the certified value. The team checks standard preparation and discovers the extraction step loses some metal. It changes the preparation and validates recovery rather than collecting more identical low results.

Why?

Why can an independent standard reveal an error that repeatability misses? All repeats of one sample may share the same wrong stock solution or sample-preparation loss. An independently prepared, known material challenges that common pathway. Agreement across independent references and methods is stronger evidence of accuracy than a narrow cluster from one setup.

Common misconception

“Precise means accurate.” Tight results can be biased. “The average of enough repeats always approaches the correct value” holds only for suitable independent random errors without persistent bias. “A certified value has no uncertainty” ignores uncertainty in reference assignment and stability. State how the reference was established and whether it matches the sample matrix.

Worked example

Five independent assays of a 20.0 mg/L reference yield 19.6, 19.7, 19.5, 19.6 and 19.6 mg/L. The mean is 19.6 mg/L. The deviations are 0, +0.1, −0.1, 0 and 0, so the sample standard deviation is sqrt[(0.01+0.01)/(5−1)] ≈ 0.071 mg/L . Relative standard deviation is about 0.071/19.6 × 100% = 0.36% . Bias relative to 20.0 mg/L is −0.4 mg/L , or −2.0%. The method is repeatable under these conditions yet shows an apparent negative bias that warrants investigation against the reference uncertainty and acceptance target.

Quick check

1. Why will taking more instrument readings not by itself correct a misprepared calibration stock? Answer: Every reading uses the same systematic concentration error. Repetition can narrow random scatter but preserves the bias; an independent standard or stock preparation is needed.

Exam focus

Name the replication level and calculate mean, standard deviation and bias with units. Distinguish a narrow spread from agreement with a reference. Explain how a certified or independent check can reveal systematic error. If the sample is heterogeneous, separate real sample variation from instrument noise. Avoid claiming “accuracy” when no suitable reference has been tested.

Advanced insight

Bias may vary with concentration, matrix or time, so one reference point may not validate an entire working range. A method-comparison study can examine differences across concentrations rather than only overall correlation; two biased methods can correlate strongly. Uncertainty budgets combine random components and corrected or bounded systematic components, but correction does not make the uncertainty of the correction disappear.

Summary

Precision measures agreement among repeats under stated conditions; accuracy addresses closeness to a reference; bias is systematic displacement. Replication reveals random variation at the level repeated, while independent references and controls reveal common errors. More repeats cannot fix a shared bias. Report both spread and reference agreement, with conditions and uncertainty appropriate to the intended decision.

Practice questions

1. A method returns 4.98, 4.99 and 5.00 mg/L for a true 6.00 mg/L reference. Describe its precision and apparent bias. Answer: The readings are tightly grouped, so short-term precision is high. Their mean is 4.99 mg/L, about 1.01 mg/L low relative to the reference, showing substantial negative bias if the reference is reliable.

2. Why are three injections of one solution insufficient to estimate day-to-day method precision? Answer: They omit new sample preparation, different days and possible changes in analyst or instrument state. Run independent preparations across days to measure that level of variation.

3. A reference material has uncertainty ±0.2 mg/L and a method mean differs by 0.1 mg/L. Can the method be declared biased from that difference alone? Answer: No. The difference is smaller than the stated reference uncertainty and measurement uncertainty must also be considered. More evidence is needed before claiming a systematic displacement.