PSLE Science Reality Lab Vol No.304 | “RSD = 2%” — Does That Mean the Method Is 98% Accurate?

Series ID: PSLE-SCI-REALITY-0304

Wait, What? Five Results Nearly Match — and They Can Still All Be Wrong Together

A laboratory report shows five repeated measurements and one tidy summary: RSD = 2%. A student reads the number and says, “Great. The method is 98% accurate.” It sounds mathematical. It is also the wrong mathematical story.

Relative standard deviation, commonly shortened to RSD, describes how spread out repeated results are compared with their mean. It is a relative measure of precision. It does not, by itself, tell us how close the results are to the quantity we intended to measure. A method can produce a tight cluster of values that all sit too high or too low because of bias, calibration error, contamination, an incorrect blank, a wrong conversion factor or another systematic effect.

This Reality Lab owns one exact learner job: how to evaluate a laboratory chart or quality statement that turns a small RSD into a percentage-accuracy claim.

Quick Answer

RSD = 2% does not mean 98% accurate.
RSD compares the standard deviation of repeated results with their mean, usually expressed as a percentage.
A smaller RSD generally means the repeated results are more tightly clustered relative to the mean under the stated conditions.
Tight clustering is evidence about precision, not automatically about closeness to a reference or accepted value.
Repeated results can be precise but biased.
RSD is only meaningful when the repetitions, calculation basis and measurement conditions are clear.
When the mean is very close to zero, dividing by that small mean can make RSD unstable or misleading.
There is no universal rule that one RSD percentage is acceptable for every scientific method and every purpose.

Owned Learner Job — and the Non-Ownership Boundary

This page does not become the canonical owner of standard deviation, uncertainty, calibration, accuracy, repeatability or general statistics. Those ideas remain with their existing owners. Reality Lab applies them to a specific real-world communication object: a report that advertises or highlights a low RSD as evidence of measurement quality.

The current 2026 PSLE Science objectives require learners to interpret and analyse information, evaluate observations, information and methods, and communicate explanations and reasoning. A Primary 5/6 learner does not need advanced statistical theory to build the transferable habit here. The child needs one disciplined question: Does this number describe how closely repeated results agree with one another, or how closely they agree with a reference?

Rebuild the Evidence Object

Imagine a fictional laboratory checking the mass of a reference object whose accepted value is 100.0 g. Two instruments each measure the object five times.

Instrument	Five results (g)	Pattern
P	99.8, 100.1, 100.0, 99.9, 100.2	Tightly clustered and close to the reference
Q	104.8, 105.1, 105.0, 104.9, 105.2	Tightly clustered but shifted high

Both sets can have similarly small relative spread. Yet Q is not giving the same measurement story as P. Q repeats itself well, but its centre is about 5 g above the reference. A low RSD can therefore coexist with a substantial bias.

Reality Lab habit: agreement with yourself is not the same as agreement with the reference.

What RSD Actually Compares

A common form of the calculation is:

RSD (%) = standard deviation ÷ mean × 100

The standard deviation summarises how far the repeated values tend to spread around their mean. Dividing by the mean makes the spread relative to the size of the measured quantity. Multiplying by 100 expresses that ratio as a percentage.

Notice what is missing from the formula: there is no accepted reference value. That is why RSD cannot, by itself, answer the question “How close are we to the correct or assigned value?”

Observed, Calculated, Inferred and Overclaimed

Layer	Example
Observed	Repeated instrument results: 49.7, 50.1, 50.0, 50.2, 50.0.
Calculated	The mean, standard deviation and RSD.
Supported inference	The repeated results are tightly clustered relative to their mean under these conditions.
Overclaim	“RSD = 2%, therefore the method is 98% accurate.”

Case 1 — Precise but Biased

A temperature sensor is checked in a stable reference bath at 20.0°C. It reads 22.1, 22.0, 22.1, 22.0 and 22.1°C. The readings have tiny spread. Their RSD is small. Yet they are centred about two degrees above the reference.

The correct diagnosis is not “bad precision”. It is “good repeatability under this test, but evidence of positive bias relative to the reference”. Different quality questions require different evidence.

Case 2 — Accurate on Average but Noisy

Another sensor reads 18.2, 20.9, 21.4, 19.1 and 20.4°C in the same 20.0°C reference condition. The mean may sit near 20°C, but the individual readings are spread widely. The average can look reassuring while the precision is poor.

This flips the previous case. One method can be tightly grouped but shifted. Another can average near the reference while individual results jump around. “Measurement quality” is not one single number.

Case 3 — The Replicates Are Not Independent

A report takes one prepared sample, leaves it in the instrument and records the display ten times within thirty seconds. RSD is 0.4%. The headline says, “Excellent reproducibility.” That language may overreach.

The ten readings show short-term repeatability of one setup. They do not test what happens when a new sample is collected, prepared by a different person, measured on another day or run in another laboratory. Precision itself has conditions. If the claim widens, the repetitions must widen with it.

Case 4 — The Mean Is Almost Zero

Suppose blank-corrected measurements are 0.01, −0.01, 0.02 and −0.02 units. Their mean is near zero. A relative measure that divides spread by the mean can become huge or undefined even though the absolute spread is small.

This is an important model limit. Relative statistics depend on the denominator. When the mean approaches zero, RSD may stop being a useful way to summarise variability. A responsible report can use an absolute standard deviation or another fit-for-purpose measure instead.

Case 5 — The “Pass” Label Without a Job

A method summary says “RSD below 5% = PASS” but gives no source for the acceptance rule. Is 5% always good? No. The answer depends on the measurand, concentration, instrument, application, regulatory or method requirements and the consequence of error.

Scientific reasoning does not replace one magic threshold with another. It asks what requirement applies to this job and why.

Representation Check: What Does the Table Hide?

How many repeated results produced the RSD?
Are the individual values shown?
What is the mean?
What is the measurement unit?
Were repetitions made from one sample or independently prepared samples?
Were they measured by one operator, one instrument and one day, or across changed conditions?
Was a reference material measured too?
Does the report show bias or recovery evidence separately from precision?
Is the acceptance limit justified for this method and purpose?

Comparison and Baseline Check

Imagine Method A has RSD 1.5% and Method B has RSD 2.0%. It is tempting to declare A “better”. That conclusion is too broad. If A is biased by 8% and B agrees closely with a reference, B may be more useful for the intended measurement. If A was tested only over five minutes and B across several days, their precision claims are not even operating at the same scope.

Compare methods on matched evidence: same quantity, similar concentration range, comparable replication, relevant reference values and the quality dimensions required by the decision.

Alternative Explanations for a Low RSD

The method genuinely has low random variation.
The sample was measured repeatedly without independent preparation, so only one source of variation was tested.
The instrument rounds results coarsely, making repeats appear identical.
The measurement range is large compared with the small variation.
A stable systematic error shifts every result together without increasing spread.
Data were selectively removed or filtered before the RSD was reported.

A low RSD is useful evidence. Healthy scepticism asks what kind of evidence it is, not whether it should be dismissed.

Evidence That Strengthens a Measurement-Quality Claim

Individual repeated values are reported, not only the final percentage.
The replication design matches the claimed scope.
A reference or assigned value is measured so bias can be assessed separately.
The method is tested across the relevant range, not at one convenient level.
Different days, operators or instruments are included when reproducibility is claimed.
Outlier handling is stated rather than hidden.
The acceptance criterion is tied to a named method, specification or fit-for-purpose requirement.

Evidence That Weakens It

RSD is subtracted from 100 and renamed “accuracy”.
No raw or replicate values are shown.
One sample is reread many times but the claim says the entire method is reproducible.
The mean is near zero and the RSD is still treated as a stable percentage.
A universal pass threshold is asserted without a method or purpose.
A reference value is available but never compared with the measured mean.

How Far Can the Conclusion Travel?

A scientifically bounded sentence could be:

“The repeated results had an RSD of 2% under the stated conditions, indicating small relative spread. This supports a precision claim for those conditions but does not by itself establish accuracy or absence of bias.”

The sentence is not weaker because it is careful. It is stronger because each clause is tied to evidence.

Tempting Reasoning That Fails

Tempting statement	What went wrong	Repair
RSD 2% means 98% accurate.	RSD describes relative spread, not closeness to a reference.	Keep precision and reference agreement separate.
Lowest RSD means best method.	Bias, range and intended use may differ.	Compare multiple quality dimensions.
Ten readings prove reproducibility.	Ten immediate readings may test only repeatability.	Change days, operators, preparations or laboratories when the wider claim requires it.
Small spread proves no systematic error.	Systematic error can move all results together.	Use suitable reference evidence.
5% is always acceptable.	Acceptance is purpose- and method-dependent.	Use a justified requirement for the job.

Model and Measurement Limits

RSD compresses a set of repeated measurements into one relative-spread number. Compression is useful because a reader can compare variability at a glance. Compression also hides things: the shape of the data, unusual values, time trends and the exact source of variation. Two datasets can share the same RSD while having different scientific histories.

IUPAC defines precision in terms of closeness of agreement among independent test results under stipulated conditions and warns against confusing precision with accuracy. NIST measurement guidance likewise separates these ideas. The practical lesson for a Primary Science learner is simple: first ask what the statistic was built to describe.

PSLE-Style Transfer Case — Three Thermometers

Three fictional digital thermometers are tested five times in a stable 25.0°C reference condition.

Thermometer	Readings (°C)	Evidence pattern
A	25.0, 25.1, 25.0, 24.9, 25.0	Close together and close to reference
B	27.0, 27.1, 27.0, 26.9, 27.0	Close together but shifted high
C	23.8, 25.9, 24.7, 25.6, 25.0	Wider spread

Which appears most precise? A and B both have tight clustering; calculation would distinguish them more exactly.
Which gives strongest evidence of agreement with the reference? A.
Could B have a small RSD? Yes. Small spread can coexist with bias.
Would “100 − RSD” give accuracy? No. That subtraction has no scientific basis here.
What extra evidence would help diagnose B? Calibration or checks against trusted reference values across the working range.

Explained Practice

Practice 1 — Tight Cluster, Wrong Centre

Five mass results are 10.50, 10.51, 10.49, 10.50 and 10.50 g. The reference is 10.00 g. What does the pattern show?

Answer: Very tight repeatability but a substantial positive difference from the reference. Precision evidence is strong; accuracy cannot be claimed from the tight cluster.

Practice 2 — Same Mean, Different Spread

Set P is 99, 100, 101. Set Q is 90, 100, 110. Both average 100. Which has lower relative spread?

Answer: P. A matching mean does not mean matching precision.

Practice 3 — Near Zero

Blank-corrected values average almost zero. Why may RSD be a poor summary?

Answer: RSD divides by the mean. When the mean is close to zero, the relative percentage can become unstable or extremely large even for small absolute variation.

Practice 4 — Better Caption

Repair “RSD 1.8%, therefore 98.2% accurate.”

Answer: “Repeated results had RSD 1.8% under the stated conditions, indicating low relative spread. Reference evidence is needed to assess bias or agreement with an assigned value.”

Delayed Independent Return

What does RSD compare?
Why can a biased method still have a small RSD?
Why do replication conditions matter?
Why should you not subtract RSD from 100 to obtain accuracy?

If you can answer those without looking back, the central distinction has transferred: spread around the mean is not distance from a reference.

Route to Existing Canonical PSLE Science Owners

For repeated measurements, use How to Read Repeated PSLE Science Results When the Measurements Do Not Match Exactly. For checking an instrument against a reference value, use How to Use a Reference Value to Check a PSLE Science Measuring Instrument Before Trusting Its Readings. Those pages own the general skills; this Reality Lab applies them to an RSD quality claim.

Parent and Tutor Teaching Guide — Darts Without Darts

Instead of the familiar dartboard analogy, use number cards. Put a target card marked 100 on the table. Place five result cards tightly around 105. Ask: “Do these agree with one another?” Yes. “Do they agree with the target?” No. Then place a second set around 100 but spread widely. The child can now see two different questions before any formula appears.

Only after that visual distinction should you show RSD as a way to summarise relative spread. Ask the learner to say aloud what RSD does not contain: it does not contain the reference value. That one observation prevents the 100 − RSD mistake.

Authoritative Sources

The Quiet Return

A small RSD is not a fake number. It can be excellent evidence that repeated results stay close together relative to their mean. It simply answers a narrower question than the word “accuracy” suggests.

Ask whether the results agree with one another before asking whether they agree with the reference.

Small Group Tutorials