EDUKATE SINGAPORE · GOLD STANDARD PRACTICAL CASE STUDY 209
Standard deviation is not a 95% confidence interval for the mean.
A science poster reports sixteen learners’ mean mark of 70 and a standard deviation of 12. It then incorrectly labels ’70 ± 12′ as a 95% confidence interval. This original case explains both statistics, calculates the standard error and works through the distinct t-based confidence interval under stated assumptions.
All learners and marks are fictional. Statistical inference requires suitable sampling and model conditions; the numerical example is a teaching exercise, not an estimate of any real student population.
The corrected result at a glance
- Sample size: n = 16.
- Sample mean: 70 marks.
- Sample SD: 12 marks.
- Standard error: 12/√16 = 3 marks.
- Illustrative t multiplier: approximately 2.131 with 15 degrees of freedom.
- Illustrative 95% confidence interval: about 63.6 to 76.4 marks, conditional on assumptions.
Chapter navigation
- 1. A polished poster labels the wrong kind of uncertainty
- 2. Standard deviation describes the spread of observed scores
- 3. Standard error connects the sample size to mean uncertainty
- 4. A 95% t confidence interval gives a different worked result
- 5. NIST explains exactly what a confidence interval estimates
- 6. The two intervals answer different everyday questions
- 7. A carefully calculated interval can still be irrelevant to the wider school
- 8. The first correction should preserve the original numbers
- 9. A statistics checklist must distinguish units, purpose and assumptions
- 10. Practice: SD 12 is not a 95% margin
1. A polished poster labels the wrong kind of uncertainty
A fictional school science enrichment programme tests a short exercise on understanding climate data. Sixteen learners complete an assessment scored out of one hundred. The programme’s analyst reports a sample mean of 70 marks and a sample standard deviation of 12 marks. The figures are correctly calculated from the intended synthetic example.
The design team presents a large heading: ‘Mean = 70, 95% confidence interval = 70 ± 12.’ The bar spans 58 to 82 marks. The poster then says the average performance is known within twelve marks with 95% confidence. Yet twelve is the standard deviation of individual learner marks, not a 95% confidence interval margin for the mean.
The Gold Standard issue is a statistical label that changed meaning between analysis and communication. The observed variation among learners and uncertainty about the population mean are related but different concepts. A correct standard deviation cannot be repurposed as a confidence interval without a justified calculation.
2. Standard deviation describes the spread of observed scores
The sample standard deviation summarises how dispersed individual observations are around their sample mean. It is expressed in the same units as the scores, here marks out of one hundred. A larger standard deviation indicates greater variation in recorded results under the relevant calculation; it does not by itself describe how precisely the population mean has been estimated.
The fictional mean is 70 and the sample standard deviation is 12. The expression 70 ± 12 therefore outlines the interval from 58 to 82, which is simply one sample standard deviation below and above the mean. It should not be named a 95% confidence interval on that basis.
Even if individual scores were approximately normally distributed, roughly 68% rather than 95% of that underlying distribution would be expected within one population standard deviation under an ideal normal model. That rule of thumb itself requires assumptions and does not convert one standard deviation into a confidence limit for the population mean.
GS-R02 demands clear definitions. Reporting spread and inferential uncertainty under one word called accuracy invites readers to draw conclusions the records cannot support.
3. Standard error connects the sample size to mean uncertainty
Under an appropriate model of independent observations, the standard error of the sample mean is estimated as s divided by the square root of the sample size n. Here s = 12 marks and n = 16 learners, so the illustrative standard error is 12/√16 = 12/4 = 3 marks.
The difference between 12 and 3 matters. Standard deviation characterises variation among individual measurements in the sample, while standard error describes the uncertainty in the sample mean under the assumed sampling process. If one increases the independent sample size while the underlying variability is otherwise comparable, the uncertainty in the mean can become smaller.
The formula is not permission to count repeated or dependent observations as separate independent learners. A class selected for convenience may also fail to represent the wider intended population. The statistical design needs attention before treating the standard error as a reliable basis for a general conclusion.
GS-R01 identifies the population question and GS-R08 links the uncertainty statement to the actual study design.
4. A 95% t confidence interval gives a different worked result
The NIST formula for a conventional two-sided confidence interval for a mean under relevant conditions is sample mean ± t critical × s/√n. With n = 16, the degrees of freedom for a one-sample t interval are 15. The approximate 97.5th percentile t critical value is 2.131.
Using the fictional data gives a margin of about 2.131 × 3 = 6.393 marks. The illustrative 95% confidence interval is therefore 70 ± 6.393, or roughly 63.6 to 76.4. That differs materially from the poster’s interval of 58 to 82. A conventional t interval also depends on a suitable sampling and distribution model; it is not automatically appropriate for every school dataset.
Neither interval guarantees that the population mean lies within the calculated endpoints with a 95% probability in a fixed frequentist interpretation. The confidence level refers to the repeated-sampling coverage of the interval-producing method under its assumptions, not certainty about this one unknown parameter.
The Gold Standard asks staff to retain the calculation method, degrees of freedom, relevant assumptions and the quantity estimated alongside the final chart.
5. NIST explains exactly what a confidence interval estimates
The US National Institute of Standards and Technology’s e-Handbook section on Confidence Limits for the Mean gives the form sample mean ± a t multiplier times s/√n. Its explanation distinguishes a confidence-interval method’s long-run coverage from an assertion that the fixed population mean has a 95% probability of being inside this particular observed interval.
NIST’s statistics glossary separately defines standard deviation as a dispersion measure and standard error as a standard deviation of a statistic’s sampling distribution. Those definitions are the precise conceptual distinction the fictional poster loses.
The official sources do not certify the school’s data or say that a simple one-sample t interval is valid for every possible sample. The method presumes suitable sampling, independence and statistical conditions. A qualified reviewer must determine whether those assumptions hold for an actual assessment dataset.
The Gold Standard connects the quantity on the chart to what the calculation and evidence genuinely represent, rather than allowing a familiar percentage label to make the result seem stronger.
6. The two intervals answer different everyday questions
If a teacher asks how varied the sixteen learners’ scores were, a standard deviation of twelve marks is relevant descriptive information alongside the distribution and mean. If the teacher asks how precisely the mean estimates a defined population mean, a confidence interval may be relevant if the sample and analytical assumptions support that inference.
The poster’s ’70 ± 12′ could be retained as mean and sample standard deviation when presented with the correct explanation, but its label should not suggest 95% mean uncertainty. The illustrative ‘63.6 to 76.4’ t interval belongs to a different question. Neither interval measures an individual learner’s ability with 95% certainty.
Readers may also mistake a confidence interval for a prediction interval for the next student. These intervals differ because a new student’s score varies around the population mean as well as the estimate of that mean. The institution should specify which variable or statistic each range describes.
GS-R07 requires that statistical meaning survive editing and publication, not merely that a number be copied without typographical error.
7. A carefully calculated interval can still be irrelevant to the wider school
The sixteen learners were chosen because they attended one enrichment session. If the intended claim concerns every student in the district, that convenience group may not justify representative population inference even when the t interval is computed correctly. Statistical precision from a formula does not erase problems of selection, clustering or missing responses.
A review should identify how learners were included, whether scores are independent under the model, what population is relevant and whether the distributional assumptions are plausible for the size and shape of the data. The fact that scores are bounded between zero and one hundred can also matter in some datasets.
The Gold Standard’s outcome is not simply substituting 6.4 for 12 on the poster. It is ensuring that the study design and inferential purpose are appropriate before a numerical interval becomes a public statement about the wider population.
A responsible school may choose to report the sixteen sample results descriptively without a population confidence interval when its collection design does not support that broader claim.
8. The first correction should preserve the original numbers
The school keeps the source scores and calculation sheet. It corrects the statement that ±12 is a 95% confidence interval, while preserving the standard deviation as a valid descriptive result. Qualified staff decide whether the intended sampling design supports the separate t-interval example and how its assumptions should be communicated.
A bounded explanation might read: ‘The sixteen students had a sample mean of 70 marks and sample standard deviation of 12 marks. A conventional t confidence interval for the mean, if its conditions apply, would use a standard error of 3 and a margin around 6.4, not 12. No claim about all district learners is established without reviewing the sampling design.’
The centre should not blame the designer for misconduct merely because the supplied spreadsheet used ambiguous labels. Responsibility for scientific interpretation, design presentation and approval needs to be connected. GS-R09 calls for a correction, while GS-R10 addresses prevention in the reporting template.
9. A statistics checklist must distinguish units, purpose and assumptions
The revised template asks whether a number is a sample mean, sample standard deviation, standard error, confidence interval, population prediction interval or something else. It records sample size, intended population, actual sampling method, t multiplier where appropriate and a plain-language interpretation of each result.
An effective chart can show the average score, distribution of individual results and an appropriate interval separately. The visual should not force a single whisker to perform several incompatible roles. Reporting the method in small text can help advanced readers without making beginners guess what the shaded region means.
Teachers can use a simple analogy: how different are pupils’ heights is not the same question as how precisely a sample estimates the average height of a wider population. The algebra of the two questions is related, but the inference differs.
GS-R08 guards the label; GS-R12 requires renewed review if the sample design, size or calculation method changes.
10. Practice: SD 12 is not a 95% margin
Exercise. Sixteen independent observations under a suitable illustrative one-sample model have mean 70 and sample standard deviation 12. A poster labels 70 ± 12 as a 95% confidence interval for the mean. Compute the standard error and a conventional t-based 95% interval with t critical approximately 2.131 for fifteen degrees of freedom.
The standard error is 12/√16 = 3. The approximate margin is 2.131 × 3 = 6.393. The corresponding illustrative interval is 70 ± 6.393, or about 63.6 to 76.4. The original ±12 is the sample standard deviation, not the correct interval margin under the specified model.
The reviewer must still consider the underlying sampling and distribution assumptions before making an inference about a real population. The Gold Standard lesson is that a mathematically correct number becomes misleading when it is labelled as a different statistic.
Frequently asked questions
Does SD = 12 mean the mean is accurate within twelve marks?
No. Standard deviation describes variation among individual observations. It is not automatically the margin of error for the mean.
Why does the standard error become three?
Because, under the independent-sample formula, s/√n = 12/4 = 3 for sixteen observations.
Is a 95% confidence interval the range containing 95% of pupils’ marks?
No. A confidence interval for the population mean is about the estimate of that mean, not a prediction of individual pupil scores.
Does 95% mean this one fixed interval probably contains the mean?
In conventional frequentist terms, it describes the method’s long-run coverage under assumptions. The fixed population mean is not being treated as random.
Can a school use a t interval for any sixteen volunteers?
A calculation is possible, but a defensible population inference needs suitable sampling and model conditions. A convenience group may not represent the intended population.
Why not simply replace twelve with 6.4 and publish?
Because the study’s sampling, independence and distribution assumptions, as well as the intended claim, must also be checked.
The statistics-publication quality gate
- Name the intended population, sampled units and inclusion method.
- Distinguish observed-score spread from uncertainty in the estimated mean.
- Record n, sample standard deviation and standard-error calculation.
- Use a suitable confidence method and disclose assumptions and multiplier.
- Label chart whiskers according to what they actually represent.
- Avoid population or causality claims that the selected sample cannot justify.
A helpful classroom analogy compares two questions about a class’s heights: ‘How spread out are the students?’ and ‘How precisely does this class estimate the average height of a larger population?’ The first concerns variation among people, and the second requires a sampling model. A learner may be capable of calculating both while still misunderstanding the difference. The Gold Standard aims to make the question behind each number visible before the graph is coloured.
For an independent review exercise, give students three statements: ‘The sample mean was 70’, ‘The sample SD was 12’, and ‘A t-based 95% interval for the population mean is approximately 63.6–76.4 under the example assumptions.’ Ask them to explain what each says and what it does not. A strong answer does not claim that 95% of pupils scored in that last range or that the class automatically represents every learner in the district.
The accepted finding for the report
The source statistics can remain in the record: mean 70 and SD 12. The published label ‘95% confidence interval 70 ± 12’ is unsupported. A correctly identified t-interval for the mean under the selected model has a margin of about 6.4, but a real inferential claim also requires professional review of independence, sampling and other assumptions.
The Gold Standard correction preserves the useful descriptive result without replacing one kind of uncertainty with another. Better statistics communication teaches what a number can say, not only how to calculate it.
Deeper explanation: distinguish individual spread and the precision of a mean
A simple classroom comparison can make the statistical error visible. Ask how much the sixteen pupils’ marks vary around their average: the standard deviation is relevant. Next ask how precisely those sixteen observations estimate a defined wider population mean under an appropriate sampling model: the standard error and an inferential interval may be relevant. Both questions use the same underlying scores, but they are not the same question. A graph that labels one standard deviation as a 95% confidence interval erases the distinction.
The standard error in the fictional example is 12 divided by the square root of 16, or 3. With fifteen degrees of freedom the illustrative two-sided 95% t multiplier is about 2.131. The interval margin is therefore about 6.393, and the calculated interval is approximately 63.6–76.4. Reporting 58–82 as the confidence interval because it equals 70 ± 12 uses the descriptive spread as though it were inferential uncertainty. The error is conceptual, even when every arithmetic step was copied exactly from the spreadsheet.
A confidence interval also does not promise that 95% of individual students will score between its endpoints. Nor, in the conventional frequentist interpretation, does it assign a probability to the already fixed population mean. The 95% level refers to the long-run performance of an interval-producing procedure under appropriate assumptions. Students can learn this by imagining many hypothetical samples and the intervals each would produce, rather than a single magic range that certifies the class.
The source data and sampling plan must support any population claim. Sixteen volunteers from one workshop might provide useful descriptive information without being a representative sample of every student in a district. Computing a correctly formatted t interval does not repair selection bias, clustering or unrelated missing observations. An intellectually honest school can keep the sample mean and standard deviation while declining to generalise the result beyond the people actually measured.
A practical reviewer should ask the author to label each field: sample mean, standard deviation, standard error, confidence level, method, sample size and intended population. The final chart must state what its whiskers actually mean. This is a better quality control than checking whether a percentage sign appears next to the word confidence. The Gold Standard respects valid descriptive work and protects readers from a more precise-looking claim than the study design allows.
NIST sources and connected Gold Standard casework
- NIST · Confidence Limits for the Mean
- NIST · Statistics Glossary, Standard Deviation and Standard Error
- The Gold Standard of Statistics
- The Gold Standard of Teaching Statistics
- Gold Standard · Evidence, Claims and Verification
- Gold Standard · Core Requirements
- Case 163 · Sixty Seedling Measurements and Pseudoreplication
- Case 188 · Average Salary Rose but Nine of Ten Workers Earned Less
- Case 193 · Screening Sensitivity and False Alarms
- Gold Standard Casebook
NIST’s e-Handbook supplies the confidence-interval method. Its validity for a particular dataset must be assessed under the relevant conditions. GS-R01–GS-R12 is eduKate’s educational management framework.