Science Education Systems · Article 70. Maya, Jia Jun, Hana and Ethan remain fictional Punggol learners. This article follows the effect-size layer: how Science asks not merely whether two conditions differ, but how much they differ and whether that magnitude matters.
The 50-second parent route
A result can be statistically detectable and scientifically trivial.
Or practically important but uncertain because the sample is small.
Effect size answers the magnitude question.
The route is:
comparison → measured difference → unit or standardisation → effect size → uncertainty → practical meaning → replication → decision
The key question is:
How large is the effect, not merely whether one was detected?
This article extends How Scientific Internal Validity Works, How Scientific Comparison Works and How Science Decision-Making Works.
1. Effect size is about magnitude
Difference in means.
percentage change.
risk difference.
risk ratio.
odds ratio.
correlation.
standardised difference.
Different scientific questions need different effect-size measures.
2. Statistical significance is not effect size
A tiny effect measured in a huge sample can produce a very small p-value.
That does not make the effect large.
Significance and magnitude answer different questions.
3. Maya’s effect-size error is p-value worship
She sees p < 0.05 and concludes the result is important.
Her repair:
ask how large the difference is and what that magnitude means.
4. Jia Jun’s effect-size error is raw-number comparison
A 5-point improvement may be huge on a 10-point scale and small on a 1000-point scale.
His repair:
interpret magnitude relative to the measurement scale and context.
5. Hana’s effect-size error is demanding one universal threshold
She wants one number that defines “large” in every field.
Her repair:
effect importance depends on outcome, baseline risk, cost, mechanism and decision context.
6. Ethan’s effect-size error is exaggerating relative change
Risk doubles.
He assumes the absolute increase must be dramatic.
His repair:
inspect both relative and absolute effect measures.
7. Absolute difference is often the most intuitive effect
Group A average = 80.
Group B average = 75.
Absolute difference = 5 units.
The unit remains tied to the original outcome.
8. Relative difference changes the reference frame
A 5-unit increase from 10 is 50%.
A 5-unit increase from 100 is 5%.
Relative change depends on the baseline.
9. Risk difference is an absolute effect
Outcome risk falls from 10% to 7%.
Risk difference = 3 percentage points.
This tells us the absolute change in event probability.
10. Risk ratio is a relative effect
7% divided by 10% = 0.7.
The event risk is 30% lower relative to the comparison group.
The same study can therefore have both an absolute and relative description.
11. Both absolute and relative effects can be useful
Relative measures compare proportional change.
Absolute measures often communicate practical consequences more directly.
Strong reporting avoids choosing whichever sounds most dramatic.
12. Standardised effect sizes remove the original unit
A difference can be divided by a measure of variability.
This allows outcomes measured on different scales to be compared more easily.
But standardisation also makes the result less directly tangible.
13. Cohen-style standardised mean differences are one example
A difference of 0.5 standard deviations gives a scale-free description.
Rule-of-thumb labels such as small, medium and large should not replace domain-specific interpretation.
14. Correlation is an effect-size measure for association
It describes the strength and direction of a linear relationship.
A statistically significant correlation can still be weak.
15. R-squared describes explained variation in some models
It can help summarise how much variation in an outcome is accounted for by the model.
High R-squared does not prove causation or model correctness.
16. Primary Science can begin effect-size thinking qualitatively
Not just:
Did the plant grow more?
But:
How much more?
Was the difference large relative to measurement variation?
17. Primary 3 can compare visible magnitude
One ice cube melts 2 minutes faster.
Another melts 20 minutes faster.
The direction is the same.
The magnitude is not.
18. Primary 4 can use absolute change
Temperature rises from 20°C to 30°C.
Change = 10°C.
Students learn that the size of change belongs in the explanation.
19. Primary 5 can compare proportional change
A small plant grows 5 cm.
A tall plant grows 5 cm.
The absolute change is equal while the proportional change differs.
20. Primary 6 can connect magnitude to conclusion strength
A tiny difference within measurement uncertainty should be interpreted more cautiously than a large repeated difference.
21. Secondary Science makes effect-size language more quantitative
percentage change.
gradient differences.
rate ratios.
risk measures.
standardised differences.
correlations.
Magnitude becomes a formal analytical layer.
22. Effect size depends on the outcome scale
A 2 mm change may be trivial for one engineering component and catastrophic for another.
The physical meaning determines importance.
23. Practical significance depends on consequences
A tiny improvement can matter when applied to millions of people.
A large improvement may matter little if the outcome itself is unimportant.
Magnitude and context interact.
24. Clinical significance is one domain-specific form
An effect can be statistically detectable yet too small to change patient experience meaningfully.
Thresholds for meaningful change depend on the outcome and decision context.
25. Educational significance also needs context
A modest average score gain may be meaningful if it closes a critical conceptual gap or persists across years.
A larger short-term gain may matter less if it disappears quickly.
26. Duration is part of effect interpretation
Immediate effect.
one-week effect.
one-year effect.
The same initial magnitude can have different long-term value.
27. Effect heterogeneity matters
Average improvement = 5.
But one subgroup improves 15 and another changes 0.
The average can hide scientifically important variation.
28. Average effects should not erase individual response distributions
Some people improve.
some remain unchanged.
some worsen.
Population averages describe one layer of the system.
29. Effect size and external validity are connected
A direction may replicate across contexts while magnitude changes.
Understanding why effect size varies helps map transportability.
30. Effect size and internal validity are inseparable
A precisely measured large difference is meaningless causally if the groups were incomparable or measurement was biased.
Magnitude should be interpreted only after the design is credible.
31. Effect size and confidence intervals are inseparable
An estimate of 5 units is incomplete without uncertainty.
Is the plausible range 4.8–5.2?
Or −2 to 12?
The next article on confidence intervals follows this logic.
32. Effect size and statistical power are connected
Small effects require more information to detect reliably.
Large effects can often be detected with fewer observations.
Power calculations therefore need an expected effect size.
33. Tiny effects require careful measurement
If the expected difference is smaller than instrument noise or natural variation, the study may be incapable of resolving it meaningfully.
Measurement resolution sets a floor.
34. Large effects can still be biased
An enormous difference does not excuse poor design.
Systematic error can be enormous too.
35. Relative effects can look stable while absolute effects change
A risk ratio may remain similar across populations with different baseline risks.
The absolute number of events prevented can therefore differ substantially.
36. Baseline risk is crucial in decision-making
A 50% relative reduction from 2% to 1% is different in absolute impact from 50% reduction from 40% to 20%.
Same relative effect.
very different practical consequence.
37. Number needed to treat translates absolute risk reduction
In suitable medical contexts, inverse absolute risk reduction can estimate how many people need treatment for one additional event to be prevented over a defined period.
Interpretation requires the time horizon and baseline population.
38. Effect sizes can be transformed
Log scales.
odds ratios.
standardised units.
The mathematical representation may change while the underlying comparison remains.
39. Meta-analysis combines effect sizes across studies
Different studies estimate the same underlying effect with different uncertainty.
Weighted synthesis can estimate a pooled effect while preserving heterogeneity.
40. Weight should reflect information, not fame
Larger, more precise studies often contribute more statistical weight than smaller uncertain ones.
Scientific synthesis should not weight studies by prestige alone.
41. Publication bias can exaggerate pooled effect size
Small studies with weak or null effects may remain unpublished.
The visible literature can then overstate the average magnitude.
42. Regression to the mean can exaggerate apparent improvement
If participants are selected because they scored extremely low, some improvement may occur naturally on retest.
Effect size should be compared against a credible control.
43. Ceiling effects can compress effect size
If participants already score near the maximum, improvement cannot be expressed fully on the scale.
The measurement instrument limits the observed magnitude.
44. Floor effects create the opposite problem
If the scale cannot represent values below a minimum, deterioration can be hidden.
Measurement range shapes observed effects.
45. AI benchmarks need effect sizes, not only rankings
Model A scores 92.
Model B scores 91.8.
The ranking differs.
Is the difference meaningful relative to benchmark variability?
Magnitude and uncertainty matter.
46. AI can exaggerate tiny benchmark gains
A headline may say “new model beats previous state of the art.”
The actual gain may be very small or within evaluation noise.
Readers should inspect the effect size.
47. AI can help students practise effect-size reasoning
Useful prompts:
“Give me the same relative risk reduction at two different baseline risks.”
“Create a statistically significant but practically tiny result.”
“Give me two studies with the same effect size but different uncertainty.”
“Ask whether an average hides subgroup heterogeneity.”
48. Parents can teach effect size through everyday comparisons
A new study routine saves 3 minutes per day.
Is that meaningful?
Maybe not.
A routine improves sleep by 45 minutes.
That may matter more.
Ask about magnitude before celebrating direction.
49. Small-group tuition can separate score change from skill change
A total mark improves by 4 points.
Was that one lucky question?
Or did explanation quality improve across many items?
Effect size should connect to the learning mechanism.
50. A compact effect-size checklist
- What outcome is being compared?
- What is the absolute difference?
- What is the relative difference?
- What unit does the effect use?
- Would a standardised measure help comparison?
- How uncertain is the effect estimate?
- Is the magnitude practically meaningful?
- What is the baseline risk or baseline level?
- Does the effect vary across subgroups?
- Does it persist across time?
- Could measurement ceilings or floors distort it?
- Does the effect replicate elsewhere?
51. Frequently asked questions
What is an effect size?
An effect size is a quantitative measure of the magnitude of a difference, relationship or treatment effect.
Is effect size the same as statistical significance?
No. Statistical significance addresses compatibility with a null model under stated assumptions; effect size addresses magnitude.
Why report both absolute and relative effects?
Because relative effects show proportional change while absolute effects often communicate real-world consequence more directly.
Can a small effect matter?
Yes. Small effects can matter when outcomes are important, exposure is widespread or effects accumulate over time.
How does effect-size thinking help PSLE Science?
It encourages learners to describe how much change occurred rather than saying only that one condition was higher or lower.
How does it change in Secondary Science?
Students can interpret percentage change, gradients, rate differences, standardised effects, risk measures and quantitative uncertainty more formally.
52. Continue the Science Education Systems series
- How Scientific Internal Validity Works
- How Scientific Statistical Power Works
- How Scientific Confidence Intervals Work
Conclusion: Science asks how much before deciding how much it matters
Maya sees significance.
Jia Jun measures magnitude.
Hana checks uncertainty.
Ethan asks whether the effect changes the real system enough to matter.
Science needs all four.
Measure the difference.
choose the right scale.
show uncertainty.
compare with consequences.
Then let importance be earned by magnitude and context, not by a threshold alone.
