Why a correct figure, an accurate measurement, or a convincing statistic may still fail to support the conclusion drawn from it.
A school reports that its students achieved an average examination score of 82%.
A company announces that customer satisfaction increased by 15%.
A medical device displays a precise numerical reading.
A researcher presents a chart showing a strong relationship between two variables.
Each statement contains a number.
Each may be accurate.
But what, exactly, has been demonstrated?
Does an average examination score establish that students understand the subject?
Does a higher satisfaction score prove that the company’s service improved?
Does a precise measurement establish that the correct phenomenon was measured?
Does a statistical relationship demonstrate that one factor caused another?
Not necessarily.
Numbers can provide powerful evidence. They can also create an impression of certainty that exceeds what the underlying information justifies.
The difference lies not in whether a number looks scientific, but in the relationship between what was observed, how it was measured, what the number represents, and what conclusion someone attempts to draw from it.
A number becomes useful evidence only within an argument that connects it to a specific claim.
And even then, the strength of that evidence depends on the quality of the connection.
1. A number is not yet an explanation
Imagine a teacher announces:
“The average score on this mathematics test was 82%.”
This is a quantitative statement.
It tells us something about the results of a particular assessment.
But it leaves many questions unanswered.
What did the test measure?
Were the questions easy or difficult?
Did the assessment examine procedural skills, conceptual understanding, or both?
Were students allowed to use calculators?
How were partial answers graded?
Did all students take the same assessment under comparable conditions?
The number 82 does not contain these answers.
It is a result produced by a particular process.
Without understanding that process, we cannot reliably determine what the result means.
This distinction is fundamental:
A numerical result describes an outcome under specified conditions. An explanation identifies the mechanisms or relationships that produced it.
The two may support one another, but they are not interchangeable.
2. Measurement begins before the instrument produces a value
Consider a thermometer.
It displays:
[
23.4^\circ\text{C}
]
The reading looks precise.
But before treating it as evidence, we need to know what was measured.
Was the instrument measuring indoor air temperature?
The surface temperature of a wall?
The temperature inside a refrigerator?
Was the sensor positioned correctly?
Had it reached thermal equilibrium with its surroundings?
Was it calibrated appropriately for the task?
A numerical display cannot answer these questions.
Measurement involves more than obtaining a value.
It requires defining the quantity of interest and using a suitable procedure to estimate it.
In metrology, the measurand is the quantity intended to be measured.
For example, “the temperature of this room” may be too vague for a demanding technical application.
Air temperature varies with position, height, ventilation, nearby surfaces, and time.
A more carefully specified measurand might be the air temperature at a particular location and height during a defined interval.
The instrument can then provide a reading relevant to that quantity.
The number acquires meaning through the measurement definition.
3. Precision is not the same as accuracy
A digital scale displays:
[
75.000\ \text{kg}
]
Another scale displays:
[
75.0\ \text{kg}
]
The first display contains more decimal places.
Does that make its result more accurate?
No.
The number of displayed digits is not sufficient evidence of measurement accuracy.
The first scale could have a systematic error.
It might consistently display a value that is too high.
Repeated measurements could be highly consistent while remaining systematically wrong.
This is the distinction between two important concepts.
Precision concerns how closely repeated measurements agree under specified conditions.
Accuracy concerns how closely a measurement result agrees with the true value of the quantity being measured.
A measurement process may be precise without being accurate.
It may also produce a result that happens to be close to the true value without being sufficiently repeatable for the intended application.
In technical work, measurement uncertainty provides another essential perspective.
It characterizes the dispersion of values that could reasonably be attributed to the measurand, given the available information.
A reported value without an appropriate understanding of its uncertainty can communicate more confidence than the measurement warrants.
The important lesson is not that numerical measurements are unreliable.
It is that their reliability must be established rather than assumed from their appearance.
4. A correct measurement can answer the wrong question
Imagine a school wants to know whether students have developed a deep understanding of mathematical reasoning.
The administration chooses a test consisting entirely of routine calculations.
Students achieve high scores.
The scores may accurately represent their performance on those calculations.
But do they establish that students can construct mathematical models, interpret unfamiliar problems, or justify their reasoning?
Not by themselves.
The test may have measured procedural fluency effectively while providing limited evidence about conceptual understanding.
This problem is closely related to construct validity.
A construct is an attribute or capability that cannot always be observed directly, such as mathematical understanding, language proficiency, or critical thinking.
An assessment uses observable performance to draw conclusions about that attribute.
The central question is whether the interpretation of the scores is supported by evidence and an appropriate theory of what is being assessed.
A perfectly administered test can still provide weak evidence for a claim it was not designed to evaluate.
Correct measurement does not guarantee that the intended question has been answered.
5. What is being counted?
Suppose a language school reports:
“Student engagement increased by 40%.”
Before interpreting the claim, we need to ask what engagement means.
Does it refer to attendance?
Homework submissions?
Messages sent to teachers?
Time spent in an online classroom?
Participation in discussions?
These indicators are not equivalent.
A student may attend every lesson without actively engaging.
Another student may participate deeply while sending very few messages.
A third may spend a long time on an assignment because the instructions were unclear.
The difficulty lies in operationalization: defining how an abstract concept will be represented through observable indicators.
Operationalization is necessary for many kinds of research.
But choosing an indicator does not make that indicator identical to the underlying concept.
A measure can be useful without capturing every aspect of what we want to understand.
The danger begins when the distinction disappears.
For example:
“Students submitted more assignments.”
This is an observable claim.
“Students became more motivated.”
This is a broader psychological interpretation.
The first may contribute evidence for the second, but it does not automatically establish it.
6. The denominator can change the story
A company announces:
“Complaints decreased from 100 to 80.”
That is a reduction of 20 complaints.
But what happened to the number of customers?
Suppose the company previously served 1,000 customers and now serves 400.
Initially, the complaint rate was:
[
\frac{100}{1000}=10%
]
Later, it became:
[
\frac{80}{400}=20%
]
The total number of complaints decreased.
The complaint rate doubled.
Both statements are numerically correct.
They describe different aspects of the situation.
A count tells us how many events occurred.
A rate relates those events to an appropriate reference population, exposure, or opportunity.
Neither is automatically the better measure.
The relevant choice depends on the question.
If the company needs to estimate the total workload of its complaint-handling team, the count may matter.
If it wants to compare the frequency of complaints relative to customers served, the rate may be more informative.
The key is to identify the denominator and determine whether it represents a meaningful basis for comparison.
7. A percentage increase can conceal a small absolute change
A school announces:
“The number of students achieving a perfect score increased by 100%.”
That sounds dramatic.
But imagine the number increased from two students to four.
The relative increase is indeed 100%.
The absolute increase is two students.
Neither figure is false.
Yet the interpretation changes when both are known.
Now consider a different school where the number increases from 200 to 400.
The relative increase is again 100%.
But the absolute increase is 200 students.
The same percentage can describe very different situations.
This is why responsible numerical communication often requires both relative and absolute information.
Percentages describe relationships.
They do not necessarily reveal the scale of the underlying quantities.
8. An average can hide the distribution
Consider two groups of five students.
Group A receives these scores:
[
80,\ 80,\ 80,\ 80,\ 80
]
Group B receives:
[
40,\ 60,\ 80,\ 100,\ 120
]
The mean score is 80 in both groups.
But the distributions are different.
Group A has no variation.
Group B has substantial variation.
If the assessment permits scores above 100, as some point-based assessments do, both sets are mathematically possible.
The example illustrates that an average alone does not describe how values are distributed.
In practical educational settings, the score scale would need to be specified.
For an examination scored from 0 to 100, we could instead compare:
Group A:
[
60,\ 60,\ 60,\ 60,\ 60
]
Group B:
[
20,\ 40,\ 60,\ 80,\ 100
]
Both groups have a mean of 60.
Yet their educational situations differ considerably.
In Group A, all students may need similar support.
In Group B, students may have very different needs.
The mean is correct in both cases.
But a decision based only on the mean may ignore important differences.
An average summarizes a distribution. It does not replace the distribution.
9. The same average can arise from different mechanisms
Suppose two schools report identical average examination results.
One school has a relatively uniform student population.
The other has a wide range of prior preparation, language backgrounds, and educational needs.
Even if the average scores are identical, the processes that produced them may differ.
This matters because numerical similarity does not necessarily imply causal similarity.
Two systems can produce the same observed result through different combinations of factors.
A comparison becomes more informative when we examine relevant context, variation, and possible mechanisms.
However, context must not become an excuse for dismissing inconvenient data.
The purpose of contextual analysis is to improve interpretation, not to make every numerical result infinitely negotiable.
10. Correlation is evidence of association, not automatic proof of causation
Imagine a study finds that students who spend more time reading tend to achieve higher examination scores.
There is an association between reading time and examination performance.
Does reading more necessarily cause the higher scores?
It may contribute.
But other explanations are possible.
Students with stronger prior skills may read more.
Family resources may influence both reading opportunities and academic performance.
Motivation may affect both variables.
The direction of influence may also be more complex than a simple one-way relationship.
Correlation can provide evidence that two variables are related.
It does not, by itself, establish the causal mechanism.
Causal inference requires additional reasoning, assumptions, and often research designs capable of addressing alternative explanations.
Randomized experiments can strengthen causal inference under appropriate conditions, but they are not always feasible or ethical.
Observational studies can also provide meaningful causal evidence when their design and assumptions are sufficiently strong.
The important principle is to match the strength of the conclusion to the strength of the evidence.
11. Why a graph can be accurate and still mislead
A graph may display real data without containing any fabricated values.
Yet its presentation can strongly influence interpretation.
Consider a bar chart showing sales increasing from 98 units to 102 units.
If the vertical axis begins at zero, the change may appear modest.
If it begins at 97, the visual difference may appear dramatic.
Both graphs can represent the same numbers.
But they emphasize different aspects of the change.
For bar charts, where bar length encodes magnitude, a nonzero baseline can exaggerate visual differences.
Other graph types may legitimately use restricted axes when the purpose is to reveal small variations, provided the scale is clear.
Graphical interpretation therefore requires attention to scale, labeling, omitted categories, time intervals, and the choice of visual representation.
A chart is not merely a neutral container for numbers.
It is a designed representation of information.
12. Sample size matters—but it is not everything
A company surveys 12 customers.
Ten report satisfaction.
The company announces:
“83% of our customers are satisfied.”
The calculation is approximately correct:
[
\frac{10}{12}\times100%\approx83.3%
]
But can the company generalize the result to all customers?
That depends on how the 12 respondents were selected and whether they adequately represent the population of interest.
A small sample can produce substantial statistical uncertainty.
A larger sample can reduce random sampling uncertainty under suitable conditions.
However, sample size alone cannot eliminate selection bias.
A survey of 10,000 people drawn from a systematically unrepresentative group may still provide a misleading estimate of the broader population.
This distinction matters:
More observations can improve precision without necessarily correcting a biased sampling process.
A large dataset is not automatically a representative dataset.
13. Statistical significance is not the same as practical importance
Suppose a large study detects a small difference between two teaching methods.
The difference is statistically significant under the study’s chosen analysis.
Does that mean the difference is educationally important?
Not necessarily.
Statistical significance concerns the relationship between the observed data and a statistical model, including its assumptions and hypothesis-testing framework.
It does not directly measure the practical value of an effect.
A very small difference can be statistically detectable in a sufficiently large dataset.
A potentially important effect may remain statistically uncertain in a small study.
Responsible interpretation should therefore consider effect sizes, uncertainty intervals, research design, and practical consequences.
A numerical threshold does not replace substantive judgment.
14. A forecast is not a fact about the future
A company predicts that demand will increase by 12% next year.
The forecast may be based on historical data, market assumptions, and a statistical model.
But it is still a prediction.
It describes an expected outcome under particular assumptions.
If those assumptions change, the forecast may no longer be appropriate.
For example, a model based on stable customer behavior may perform poorly after a major regulatory or technological change.
This does not make forecasting useless.
Forecasts are valuable precisely because they help people reason about uncertain futures.
But their usefulness depends on understanding the conditions under which they were produced.
A forecast should not be communicated as though the future had already been measured.
15. Evidence always supports a particular claim
Consider the statement:
“The temperature in this classroom was 24°C at 10:00 a.m.”
A properly performed measurement may provide strong evidence for that specific claim.
But does it establish that the classroom was comfortable for every student throughout the day?
No.
Comfort may depend on humidity, air movement, clothing, activity, individual differences, and other conditions.
The temperature reading may be relevant evidence.
It is not sufficient evidence for every conclusion about the classroom environment.
The relationship between evidence and claim is therefore specific.
A measurement that strongly supports one statement may provide weak support for another.
This is one reason scientific and professional communication must define the question before selecting the evidence.
16. Why the same data can support different legitimate conclusions
Imagine a school evaluates a new instructional program.
The average score rises.
At the same time, the difference between high-performing and low-performing students becomes larger.
One researcher focuses on the improvement in the average.
Another focuses on the increased inequality of outcomes.
Both may be identifying real features of the same dataset.
The disagreement may concern which outcome matters for the decision being made.
That does not mean every interpretation is equally defensible.
Claims still require appropriate evidence and valid reasoning.
But data do not independently determine which goals a decision-maker should prioritize.
A numerical analysis can clarify trade-offs.
It cannot always decide what should be valued.
17. Measurement errors and reasoning errors are different
Suppose an instrument reports an incorrect value because it was improperly calibrated.
That is a measurement problem.
Now suppose the instrument reports a reliable value, but a researcher draws an unsupported conclusion from it.
That is an inference problem.
The two require different corrections.
Improving the instrument may solve the first.
It will not necessarily solve the second.
Likewise, better statistical calculations cannot repair a research question that was poorly defined from the beginning.
A rigorous investigation must examine several stages:
The question being asked.
The quantity or phenomenon being observed.
The measurement procedure.
The collected data.
The analysis.
The interpretation.
The conclusion.
Each stage can introduce limitations.
And accuracy at one stage does not guarantee validity at the next.
18. The role of language in numerical evidence
Numbers often appear to transcend language.
The expression:
[
2+2=4
]
does not depend on whether the reader speaks English, German, Ukrainian, or Spanish.
But the interpretation of numerical evidence frequently depends on language.
Compare:
“Sales increased by 15%.”
“Sales increased to 15%.”
“Sales increased by 15 percentage points.”
These statements describe different mathematical relationships.
Suppose a market share rises from 20% to 35%.
The increase is 15 percentage points.
The relative increase is:
[
\frac{35-20}{20}\times100%=75%
]
Calling this a 15% increase would communicate a different numerical relationship.
In multilingual business communication, such distinctions must be preserved.
A translator who replaces percentage points with percent may change the quantitative meaning of the statement.
An editor who removes a qualifying phrase may make a conclusion sound stronger than the evidence permits.
Numerical accuracy and linguistic accuracy are therefore closely connected.
This is particularly important in technical, academic, financial, and professional documents.
19. When a correct number becomes a misleading claim
A number can be correct in isolation and misleading in context.
This may happen when:
- the denominator is omitted;
- a relative change is presented without its absolute scale;
- an average conceals important variation;
- a correlation is described as causation;
- a measurement is used to support a claim about a different construct;
- uncertainty is ignored;
- a forecast is presented as an established outcome;
- relevant conditions or limitations are omitted.
These problems do not all have the same cause.
Some arise from poor measurement.
Others arise from statistical misunderstanding.
Others result from communication choices.
And some may involve deliberate manipulation.
The appropriate response is not to distrust every number.
It is to examine how the number is connected to the claim.
20. A practical framework for evaluating numerical evidence
When someone presents a numerical result, seven questions can help determine its evidential value.
1. What is the claim?
Identify exactly what the number is supposed to establish.
2. What was actually measured or counted?
Distinguish the observed indicator from the broader concept being discussed.
3. How was the information obtained?
Consider instruments, procedures, sampling, definitions, and data quality.
4. What is the relevant reference point?
Check denominators, baselines, comparison groups, and time periods.
5. What uncertainty or variation exists?
Consider measurement uncertainty, sampling variability, and differences within the data.
6. What alternative explanations remain possible?
Especially when causal claims are made, identify other mechanisms that could produce the observed pattern.
7. Does the conclusion go beyond the evidence?
Determine whether the statement is appropriately limited to what the data can support.
These questions are not a substitute for specialized statistical or scientific methods.
They are a way to organize the reasoning that must accompany those methods.
21. Why this matters in education
Students are often taught how to calculate an answer.
They may receive less practice deciding whether the answer is relevant to the original question.
A student can compute an average correctly without understanding what the average conceals.
A student can apply a formula accurately without checking whether its assumptions hold.
A student can interpret a graph without noticing that the scale distorts the visual impression.
These are not merely arithmetic difficulties.
They concern mathematical modeling, scientific reasoning, language, and evidence evaluation.
The goal of education should therefore include more than producing correct numerical outputs.
Students should learn to explain what a result means, what it does not mean, and what additional information would be needed to support a stronger conclusion.
This approach connects mathematics with physics, biology, economics, social research, and professional communication.
It also helps students transfer knowledge beyond familiar classroom exercises.
22. Why this matters in professional language services
A professional translator may encounter statistical claims in research reports.
An editor may work with percentages in financial documentation.
A copywriter may be asked to communicate performance indicators.
A multilingual project may require technical information to be understood by audiences with different linguistic backgrounds.
In each case, the task involves more than reproducing numbers.
The relationships expressed by those numbers must remain intact.
A translator should not silently change a relative increase into a percentage-point increase.
An editor should not remove a qualification that limits a scientific conclusion.
A professional writer should not turn a measured association into an unsupported causal promise.
This does not mean language professionals should independently certify scientific findings outside their expertise.
It means they should recognize when a linguistic change could alter the meaning of a quantitative claim and seek clarification where necessary.
Levitin Language School provides Professional Language Services covering translation, proofreading, editing, copywriting, and related work.
For assignments involving specialized terminology or quantitative information, the subject matter, intended audience, and required level of review should be established before the work begins.
23. The difference between information and justified knowledge
A number can tell us what was observed.
A statistical model can help us describe relationships.
A measurement can estimate a physical quantity.
An experiment can provide evidence about a causal hypothesis.
But none of these automatically eliminates the need for interpretation.
Knowledge requires a justified connection between evidence and conclusion.
That connection may be strong or weak.
It may depend on assumptions.
It may be revised when new information becomes available.
And it may support only a narrower claim than the one someone initially wanted to make.
This is not a weakness of scientific reasoning.
It is one of its strengths.
Good reasoning makes the boundaries of a conclusion visible.
Conclusion: evidence is a relationship, not a number
A number does not become evidence simply because it is precise, official, statistically calculated, or displayed in a professional chart.
Its evidential value depends on what it represents and what someone is trying to establish with it.
A correct calculation can describe the wrong quantity.
An accurate measurement can address the wrong question.
A large dataset can be systematically unrepresentative.
A statistically significant difference can have little practical importance.
And a convincing graph can communicate a stronger impression than the underlying data justify.
None of this means that numbers are untrustworthy.
It means that numbers require disciplined interpretation.
A number becomes evidence when it is connected to a clearly defined claim through an appropriate method and a defensible chain of reasoning.
The stronger that connection, the stronger the evidence.
The weaker the connection, the less certainty the number can legitimately provide.
The real question is therefore not merely:
“Is this number correct?”
It is:
“What does this number actually allow us to conclude?”
Tymur Levitin
Founder & Director — Levitin Language School
Global Learning. Personal Approach.
Levitin Language School
Professional Language Services
© Tymur Levitin. All rights reserved.
