Science Education Systems · Article 49. Maya, Jia Jun, Hana and Ethan remain fictional Punggol learners. This article follows the data-quality layer: how Science decides whether a dataset is trustworthy enough for the claim we want to make.
The 50-second parent route
Data is not automatically evidence.
A table can be full of numbers and still be scientifically weak.
The route is:
question → measurement → identity → completeness → accuracy → consistency → timeliness → representativeness → validation → fitness for use → correction
The key question is:
Is this data good enough for this scientific purpose?
This article extends How Scientific Provenance Works, How Scientific Calibration Works and How Scientific Validation Works.
1. Data quality begins with purpose
A measurement can be perfectly adequate for one question and inadequate for another.
A kitchen scale may be fine for flour.
It is not suitable for microgram chemistry.
Quality is always judged relative to the intended use.
2. Accuracy asks whether values are close to reality
If a thermometer is systematically 3°C high, the data may be precise but inaccurate.
Calibration and method quality affect accuracy.
3. Precision asks whether repeated values cluster closely
Tight clustering can be useful.
But precision without accuracy can still produce consistently wrong data.
4. Completeness asks what is missing
Missing values.
missing time points.
missing locations.
missing subgroups.
A dataset can look large while crucial parts are absent.
5. Consistency asks whether the same thing means the same thing
One table records metres.
another centimetres.
one lab uses “growth” to mean height.
another means mass.
Without consistent definitions, merging data can create false conclusions.
6. Timeliness matters when systems change
Old data may still be valuable historically.
But it may not represent the current state of a fast-changing system.
Science should know when the data was collected.
7. Representativeness asks who or what the data describes
A perfectly measured biased sample remains weak for broad generalisation.
Measurement quality and sampling quality are separate layers.
8. Relevance asks whether the data answers the question
A precise measurement of leaf colour may not help answer a question about growth rate.
Good data can still be irrelevant data.
9. Maya’s data-quality error is number trust
If a value is written neatly in a table, she assumes it is trustworthy.
Her repair:
ask where it came from and how it was measured.
10. Jia Jun’s data-quality error is quantity worship
He prefers 10,000 records to 100 records automatically.
His repair:
ask whether the larger dataset is representative, consistent and correctly measured.
11. Hana’s data-quality error is perfectionism
One missing value appears.
She declares the whole dataset unusable.
Her repair:
judge whether the missingness materially affects the intended conclusion.
12. Ethan’s data-quality error is cleaning until the pattern looks good
He removes awkward points without a principled rule.
His repair:
define cleaning rules transparently and preserve the raw data.
13. Primary Science begins data quality with careful recording
Correct label.
correct unit.
correct trial.
correct time.
correct setup.
These simple habits protect the evidence chain.
14. Primary 3 can learn completeness
Did we record every trial?
Did we forget one object?
Is one row blank?
Children can learn that missing data changes what can be concluded.
15. Primary 4 can learn consistency
Measure every object using the same unit and procedure.
If the method changes halfway, comparisons become weaker.
16. Primary 5 can learn anomaly checking
Most readings are near 20.
One is 200.
Do not erase it automatically.
Check whether it is a real event, unit error or recording mistake.
17. Primary 6 can learn fitness for conclusion
Does the available data support “always”?
Or only “in these trials”?
Data quality and claim scope belong together.
18. Secondary Science increases data-quality complexity
Calibration.
sampling.
replicates.
data logging.
spreadsheet transformations.
statistical models.
Each stage creates possible quality loss.
19. Unit errors are data-quality failures
Milligrams recorded as grams.
minutes treated as seconds.
°C confused with K in a formula.
Unit mistakes can produce values that look numerical but are scientifically meaningless.
20. Transcription errors matter
37 becomes 73.
0.06 becomes 0.6.
Manual copying introduces error.
Double-checking or automated transfer can reduce it.
21. Duplicate records can distort totals
If one observation appears twice, averages and counts can shift.
Stable identifiers help detect duplication.
22. Missingness has different causes
Random sensor failure.
participant dropout.
unrecorded negative result.
deliberate omission.
The cause of missing data affects interpretation.
23. Missing data is not always random
If a sensor fails mainly at high temperatures, the missing values systematically hide extreme conditions.
Ignoring them can bias the conclusion.
24. Data cleaning should be documented
Which records were removed?
why?
which values were corrected?
which units converted?
Cleaning changes evidence and therefore needs provenance.
25. Outliers should be investigated, not automatically deleted
An outlier may be:
a genuine rare event;
a measurement error;
a unit error;
a different population;
a new phenomenon.
Diagnosis comes before removal.
26. Range checks can detect impossible values
Negative mass in an ordinary classroom measurement.
250°C room temperature.
A human height of 18 metres.
Simple domain constraints catch errors early.
27. Logic checks can detect inconsistent combinations
End time before start time.
age decreasing over time.
part greater than total.
Scientific data can violate logical relationships even when each number looks plausible alone.
28. Cross-source comparison can reveal problems
One sensor reports a sudden jump.
Nearby sensors do not.
Is the event local, or is one instrument failing?
Independent measurements help diagnose quality.
29. Data quality and provenance are inseparable
If we cannot trace a suspicious value back to its source, repair becomes difficult.
See How Scientific Provenance Works.
30. Data quality and calibration are inseparable
A perfectly recorded measurement from a drifting instrument can still be wrong.
See How Scientific Calibration Works.
31. Data quality and sampling are inseparable
Accurate measurements from an unrepresentative sample can answer the local question well and the population question badly.
See How Scientific Sampling Works.
32. Data quality and reproducibility are inseparable
If transformations are undocumented, another researcher cannot reconstruct which data produced the result.
See How Scientific Reproducibility Works.
33. Data quality can decay over time
Sensors drift.
definitions change.
software updates.
records are copied.
metadata disappears.
Quality must be maintained, not assumed permanently.
34. Data lineage helps locate quality loss
Raw reading → cleaned table → derived variable → graph → model.
If the graph is wrong, trace backward to find where the error entered.
35. Quality checks should occur before analysis
Do not build an elaborate model first and inspect the data later.
Basic validation checks belong upstream.
36. Quality checks should also occur after analysis
A strange result may expose hidden data problems missed earlier.
Scientific workflows loop backward when necessary.
37. Data quality is multidimensional
A dataset can be:
accurate but incomplete;
complete but outdated;
current but biased;
representative but poorly calibrated.
No single score captures everything.
38. Fitness for use combines the dimensions
For this decision, which quality dimensions matter most?
A real-time safety system may prioritise timeliness and low false negatives.
A historical study may tolerate age but demand provenance.
39. Quality thresholds should match consequences
A classroom demonstration and a medical device should not require identical data standards.
The cost of error changes the required evidence quality.
40. AI systems are built on data quality
Training data.
evaluation data.
live sensor data.
retrieval sources.
Garbage in can become fluent garbage out.
41. AI can magnify hidden quality problems
A biased dataset can be processed at huge scale.
An incorrect source can be repeated thousands of times.
Automation multiplies both good and bad inputs.
42. AI can also help inspect data quality
Useful prompts:
“Find missing values and inconsistent units.”
“List impossible or suspicious ranges.”
“Identify duplicate records.”
“Which variables changed definition across files?”
Human scientific judgement remains necessary.
43. Parents can teach data quality through household records
Track plant height for a week.
Use the same ruler.
measure at the same time.
label the pot.
record every day.
Children see how consistency improves interpretation.
44. Small-group tuition can run data-quality audits
Give students a deliberately messy dataset.
Missing unit.
duplicate row.
impossible value.
unclear label.
outlier.
Ask what must be checked before analysis.
45. A compact data-quality checklist
- What scientific purpose will this data serve?
- Are values measured accurately enough?
- Are important records missing?
- Are units and definitions consistent?
- Is the data current enough?
- Is the sample representative enough?
- Are duplicates present?
- Are any values physically or logically impossible?
- Were cleaning rules documented?
- Can suspicious values be traced to source?
- Does uncertainty match the intended claim?
- Is the dataset fit for the decision or model?
46. Frequently asked questions
What is scientific data quality?
It is the degree to which data is accurate, complete, consistent, relevant, timely, representative and fit for the scientific purpose in which it will be used.
Is more data always better?
No. More biased, duplicated or poorly measured data can increase confidence in the wrong conclusion.
Should outliers always be removed?
No. They should be investigated first because they may represent errors, rare events or scientifically important phenomena.
How does data quality help PSLE Science?
It supports careful tables, correct units, repeated trials, fair comparisons and evaluating whether conclusions match the observations.
How does data quality change in Secondary Science?
It expands to sampling, calibration, missing data, digital processing, statistical analysis and data lineage.
47. Continue the Science Education Systems series
- How Scientific Anomaly Detection Works
- How Scientific Parameter Estimation Works
- How Scientific Standards Work
Conclusion: Data earns scientific value through quality, not quantity alone
Maya sees the number.
Jia Jun sees the size of the dataset.
Hana sees what is missing.
Ethan traces how the values were transformed.
Science needs all four.
Measure carefully.
label clearly.
check completeness.
inspect consistency.
trace the source.
Then decide whether the data is fit to carry the claim.
