Accurate data are close to the true value; reproducible data give the same result when the measurement is repeated. These are independent properties, and data can be one without the other. A measurement can be reproducible yet wrong — a miscalibrated instrument that returns the same biased value every time — or accurate on average yet irreproducible, scattering widely around the truth with no consistency. Accuracy is judged against an external reference standard; reproducibility is judged only against the measurement's own repeat behavior. Both are required before data support a scientific claim, and neither guarantees the other.
The Mechanism: Bias and Variance in Two Different Directions
Every measurement result can be decomposed into the true value plus bias plus random error. Accuracy is governed mainly by bias, the systematic offset that pushes every reading the same direction. Reproducibility is governed mainly by random error, the scatter that changes from run to run.
This decomposition is why the two properties diverge. A systematic offset of +2.0 mg on a balance shifts the mean of 100 readings away from the certificate value even though each reading sits 0.05 mg from its neighbor. Precision within the run stays high; the mean stays wrong. Conversely, an unbiased method with a standard deviation of 1.5 units produces a mean close to truth over many replicates, while any single pair of measurements may differ by 4 units or more.
In plain terms: accuracy asks whether the arrow hits the target center, reproducibility asks whether the arrows land in the same place. A tight cluster off-center is reproducible and inaccurate.
The formal vocabulary splits further. Repeatability is agreement under identical conditions — same operator, same instrument, same day. Reproducibility is agreement across changed conditions — different operators, laboratories, or reagent lots. ISO 5725 defines both, and requires a reproducibility standard deviation that is almost always larger than the repeatability standard deviation. In interlaboratory studies the ratio commonly runs between 1.5 and 3, meaning scatter across labs is two to three times the within-lab scatter.
Quantified Comparison: Where the Two Properties Diverge by Orders of Magnitude
The clearest real-world numbers come from calibration and proficiency testing. In clinical chemistry, a common acceptance limit for total allowable error is set as a percentage of the target value, often 10 to 20 percent depending on the analyte and the source guideline. A laboratory can meet a precision target far tighter — a coefficient of variation of 2 percent — while a persistent calibration bias of 8 percent still passes the error budget. The method is reproducible to 2 percent and systematically 8 percent inaccurate, and both facts coexist without contradiction.
Bias detection requires an external check. A 2021-style proficiency testing round with 20 participating laboratories using the same method can show a mean result 6 percent below the assigned value while individual laboratories cluster within 1.5 percent of each other. All 20 are reproducible; all 20 share a bias traceable to one reference material batch. No amount of repeat measurement reveals it.
| Property | What it measures | Reference needed | Typical metric | Detects calibration drift? |
|---|---|---|---|---|
| Accuracy | Closeness to true value | Yes, external standard | Bias as percent of target | Yes |
| Repeatability | Same conditions | No | Within-run standard deviation | No |
| Reproducibility | Changed conditions | No | Inter-lab standard deviation | Only partly |
| Trueness | Mean vs. reference | Yes, certified material | Mean deviation | Yes |
| Precision | Scatter only | No | Coefficient of variation | No |
The practical consequence is that a laboratory can report a coefficient of variation of 1.8 percent and a bias of 9 percent simultaneously. The first number reassures; the second drives the patient result. Accuracy errors of this size propagate directly into downstream decisions, while precision errors partly average out.
A Correction to a Common Misreading: Reproducibility Is Not Accuracy's Proof
A frequent claim in method validation documents is that good reproducibility demonstrates accuracy. It does not. Reproducibility constrains random error only. The bias term is invisible to any analysis that compares a method against itself.
The 2015 reproducibility project in psychology made this visible at scale. Repeating 100 published studies with high statistical power found that only 36 percent replicated their original significant result, and the replicated effect sizes averaged roughly half the original estimates. A separate 2018 replication of 21 social science experiments in Nature and Science produced significant effects in 13 of 21 cases, with mean effect sizes about 75 percent of the originals. In neither case was the failure a precision problem. The original measurements were internally consistent; they simply did not correspond to a stable external truth.
The mechanistic lesson is that reproducibility has a ceiling. When the true effect is small relative to measurement noise, high reproducibility still permits large bias. A method can be repeatable to three significant figures and off the reference value by two.
A Timeline View: How Accuracy and Reproducibility Decay Over a Study
Neither property is static. Instruments drift, reagents degrade, and operators change. The decay profile differs between the two, which is why both must be monitored on separate schedules.
| Stage | Typical elapsed time | Accuracy risk | Reproducibility risk | Monitoring action |
|---|---|---|---|---|
| Initial validation | Day 0 | Established against reference | Established within run | Full method validation |
| First month | 30 days | Slow calibration drift | Operator-to-operator scatter | Weekly control samples |
| Quarterly | 90 days | Reference material lot change | Instrument service effects | Proficiency testing round |
| Annual review | 365 days | Cumulative bias accumulation | Seasonal environmental shifts | Revalidation against standard |
| Method transfer | Variable | New-site calibration offset | Cross-site dispersion | Inter-laboratory comparison |
| Long-term archive | 1 to 5 years | Reference standard expiry | Reagent lot turnover | Periodic re-qualification |
Calibration drift rates vary widely by instrument. A well-maintained analytical balance drifts less than 0.1 percent of reading over 12 months, while some electrochemical sensors shift 2 to 5 percent within a single quarter. The drift sets the accuracy monitoring interval; nothing about reproducibility changes that schedule.
Edge Cases Where the Two Properties Trade Off Deliberately
Some fields knowingly accept one property at the expense of the other, and the choice is defensible. High-throughput screening often favors reproducibility because the goal is ranking, not absolute truth. A compound ranked 1 of 500 by a biased but consistent assay is still useful in relative terms, even if the absolute potency estimate is 15 percent off.
Another case is field measurement under uncontrolled conditions. Portable air-quality sensors can achieve repeatability within 5 percent on a fixed unit but show 20 to 30 percent inter-unit reproducibility, because each unit's electrochemical cell ages differently. The bias is corrected by co-location against a reference monitor for 24 to 48 hours, after which accuracy improves without improving unit-to-unit reproducibility.
A third case is biological replication. Cell-line studies that replicate within one passage show low technical variability but poor biological reproducibility across passages, because gene expression drifts. The two properties are separated by a factor the researcher controls: replication at the technical level, the biological level, or the laboratory level.
What to Actually Do: Concrete Actions for Each Property
The actions separate cleanly because the properties do.
- Establish accuracy with a certified reference material whose assigned value is traceable to a national standard, and record the bias as an absolute figure, not a pass or fail.
- Establish reproducibility by repeating the measurement under at least two changed conditions: a different operator and a different day.
- Track both on separate control charts. A shift in the mean signals bias; widening limits signal increased scatter.
- Set acceptance limits per analyte from a published error budget rather than from a fixed single number.
- Re-qualify the reference standard whenever the bias changes by more than half the allowable total error.
- Report both a bias estimate and a reproducibility standard deviation in any method transfer document.
The reconciliation point matters here. A single allowable total error limit, such as 10 percent, must be split deliberately between bias and precision; if bias alone consumes the full 10 percent, the method fails even with zero scatter. Laboratories that set one combined threshold without allocating it between the two components routinely misclassify methods as acceptable.
Limitations and What Neither Property Can Tell You
Accuracy and reproducibility share a blind spot: neither addresses validity. A method can be accurate against its reference and reproducible across laboratories while measuring the wrong construct entirely. A questionnaire that consistently overstates pain intensity will show excellent test-retest reliability and good agreement with a flawed criterion, yet capture something other than the intended variable.
There is also a statistical floor. Below a signal-to-noise ratio of roughly 2, reproducibility estimates themselves become unstable, and a laboratory may need 10 or more replicate runs to distinguish real scatter from chance. With 3 replicates, the confidence interval on a standard deviation spans a factor of about 2.5.
Finally, the properties mean nothing without a defined measurement interval and a defined set of changed conditions. Reproducibility across the same laboratory is a weaker claim than reproducibility across laboratories. The exact conditions must be stated or the number cannot be interpreted.