Wait, what? If one slightly different measurement would make you choose the opposite answer, how strong was your original conclusion?

A conclusion can be correct and fragile. It may depend on one uncertain data point, one hidden assumption or one interpretation of a diagram. Another conclusion may remain the same even if one weak clue changes. That second conclusion is more robust.

This preserved Hougang Primary 6 Science URL now owns one precise job: sensitivity and robustness of scientific conclusions. The old duplicated 2019 tuition advertisement, stale locality claims, grade promises and unrelated image stack have been removed.

This page is deliberately distinct from the other Hougang P6 owners. Evidence hierarchy asks which clue deserves the most weight. Evidence sufficiency asks whether there is enough evidence to conclude at all. This page asks a different question:

How much would the evidence have to change before my conclusion changes?

Sensitive conclusions change easily

A conclusion is sensitive when a small change in one input can change the final decision.

Suppose two setups differ by only 0.1°C and the thermometer itself reads only to the nearest 1°C. Declaring one setup definitely hotter may be very sensitive to measurement uncertainty.

Or suppose two explanations are equally strong except for one ambiguous diagram feature. The final choice depends heavily on how that feature is interpreted.

In both cases, the conclusion is fragile because one weak element controls too much.

Robust conclusions survive reasonable variation

A robust conclusion remains supported even when one non-critical detail changes within a reasonable range.

Example:

• three independent observations support the same explanation;
• one measurement shifts slightly;
• the other evidence still points to the same model;
• the conclusion remains unchanged.

Robustness does not mean “cannot be wrong”. It means the conclusion is not balancing on one tiny unstable detail.

The one-clue-change test

Take one piece of evidence and change it slightly.

• If a measurement were one unit higher, would the answer change?
• If one repeated trial were removed, would the trend disappear?
• If one diagram label were ambiguous, would the whole explanation collapse?
• If one assumption were false, would the mechanism still work?

If the conclusion changes immediately, identify why that evidence has such high leverage.

High leverage is not automatically bad. Sometimes one clue is genuinely decisive. The next question is whether that decisive clue is reliable.

Sensitivity and evidence hierarchy

Suppose one direct, reliable observation controls the conclusion. That may be appropriate.

But if one weak visual impression controls the conclusion despite stronger measurements pointing elsewhere, the reasoning is poorly weighted.

Ask:

• Which clue has the most influence on my answer?
• Does that clue deserve that influence?
• How direct is it?
• How reliable is it?
• Does it discriminate between the alternatives?

Sensitivity helps reveal whether evidence weighting is sensible.

Sensitivity and evidence sufficiency

If one plausible additional observation could easily reverse the conclusion, the evidence may not yet be sufficient for a strong claim.

Example:

• one trial supports A over B;
• the difference is small;
• natural variation is expected;
• another trial could plausibly reverse the ranking.

The correct response may be to collect more evidence before committing.

A fragile conclusion often signals a sufficiency problem.

Sensitivity to measurement uncertainty

Every measurement has limits.

If two values are far apart relative to the measurement resolution, the conclusion may be robust.

If two values are extremely close, small reading differences may matter.

Ask:

• How large is the observed difference?
• How precise is the instrument?
• Do repeated readings overlap?
• Would rounding change the ranking?

The closer the evidence lies to the measurement boundary, the more cautiously the conclusion should be stated.

Sensitivity to one anomalous result

Suppose four trials show 10, 11, 10 and 20.

If the conclusion depends heavily on the 20, investigate the anomaly.

• Was the method followed?
• Was the value recorded correctly?
• Can the unusual result be reproduced?
• Does excluding it change the conclusion?
• If it repeats, does the original model need revision?

This is a sensitivity analysis at Primary 6 level: test how much one unusual observation controls the result.

Sensitivity to assumptions

Some conclusions depend on assumptions that are not measured directly.

Ask:

• What am I assuming about the diagram?
• Am I assuming other conditions stayed the same?
• Am I assuming there is only one pathway?
• Am I assuming the sample is typical?
• Am I assuming the process responds immediately?

Then flip one assumption:

If this assumption were false, would my conclusion survive?

If not, that assumption deserves explicit checking.

Sensitivity to interpretation

The same graph or diagram can sometimes support more than one interpretation unless the task is read carefully.

For example, a graph may show:

• higher final value;
• greater amount of change;
• earlier onset;
• steeper change during one interval.

If the conclusion changes depending on which interpretation is selected, return to the task and define the requested relationship precisely.

Interpretive sensitivity is often a task-definition problem.

Sensitivity to time window

A conclusion may hold over one interval and not another.

Suppose Setup A changes faster early but reaches a plateau, while Setup B changes more slowly and later overtakes it.

“A changes faster” may be true during one interval and false across the entire observation period.

Ask:

• Which time interval is the conclusion about?
• Would a longer observation change the ranking?
• Is the final state different from the early trend?

Robust conclusions specify their time boundary.

Sensitivity to sample choice

If the conclusion changes depending on which one plant, animal, material sample or trial is chosen, the sample may be too narrow.

Ask:

• Is this one sample representative?
• Would another sample reasonably behave differently?
• Is natural variation expected?
• Do several samples support the same pattern?

A broad conclusion should not rest on a highly sample-sensitive observation.

Sensitivity to one missing control

A causal conclusion may collapse if one important uncontrolled factor differed between setups.

Suppose the intended comparison changes light level, but water amount also differs.

Ask:

If water, rather than light, caused the result, would the same evidence still appear?

If yes, the causal conclusion is not robust to the missing control.

This is why method quality matters so much to conclusion strength.

Robustness through independent evidence

A conclusion becomes stronger when several independent lines of evidence point to the same account.

For example:

• a direct measurement supports the claim;
• a changed representation shows the same relationship;
• a counterfactual test produces the predicted result;
• a repeated trial reproduces the pattern.

If one weak clue is removed, the rest still support the conclusion.

Independent convergence creates robustness.

Repeated copies of one clue do not create the same robustness

Three statements derived from the same graph are not three independent evidence sources.

If the graph is wrong or misread, all three collapse together.

Ask:

• Do my evidence pieces come from independent observations?
• Or are they different descriptions of the same underlying source?

Robustness depends on independence, not just count.

The tipping-point idea

Sometimes the conclusion changes only after a threshold is crossed.

Example:

• At 20°C, A is higher than B.
• At 21°C, A is still higher.
• At 22°C, they are equal.
• At 23°C, B becomes higher.

The threshold around which the ranking changes is a tipping point for the conclusion.

Primary 6 students can use this idea qualitatively:

How far can this condition change before my conclusion stops being true?

Robustness is claim-specific

A data set may robustly support a narrow claim and weakly support a broad one.

Example:

• Robust narrow claim: Setup A had a higher measured value in all five trials.
• Less robust broad claim: Setup A will always outperform Setup B under every possible condition.

The second claim extends far beyond the tested range.

Robustness should be evaluated inside the claim’s boundary.

The leave-one-out test

Take several evidence pieces and remove one at a time.

• Remove Trial 1. Same conclusion?
• Remove one graph point. Same trend?
• Ignore one indirect clue. Same explanation?
• Remove one assumption. Does the mechanism still stand?

If every single removal reverses the conclusion, the reasoning is highly fragile.

If most reasonable removals leave the conclusion unchanged, robustness is higher.

The perturbation table

Change tested	Reasonable change?	Does conclusion change?	What does that reveal?
Measurement +1 unit	?	?	?
Remove one trial	?	?	?
Flip one assumption	?	?	?
Use another sample	?	?	?

This is a teaching scaffold for analysing fragile conclusions. It is not an exam format requirement.

Sensitivity in MCQ

A difficult MCQ often contains one controlling condition.

Ask:

• Which single fact eliminates the distractor?
• If that fact were absent, would two options remain plausible?
• Is my answer depending on a strong clue or on an assumption?

This helps distinguish a decisive condition from a fragile guess.

Sensitivity in structured answers

Before finalising an open-ended conclusion, ask:

• Which evidence clause carries the answer?
• Is that evidence direct and reliable?
• Would a reasonable alternative interpretation change the mechanism?
• Does the answer acknowledge the relevant boundary?

If the conclusion is sensitive, the wording should be more carefully bounded.

Sensitivity during checking

Checking should not mean changing answers because they feel uncertain.

Use sensitivity questions:

• What evidence would have to be different before I should change this answer?
• Did I discover such evidence?
• Or am I only feeling less confident?

A robust answer should not be changed without a real evidential reason.

Five Primary 6 robustness failure modes

1. One-clue dependency

The conclusion rests entirely on one weak detail. Repair with leverage and reliability checks.

2. Tiny-difference certainty

A very small measured difference is treated as decisive despite instrument limits. Repair with measurement-sensitivity checks.

3. Assumption-insensitive thinker

Hidden assumptions are never tested. Repair by flipping one assumption at a time.

4. Evidence-count robustness

Several dependent clues are mistaken for independent confirmation. Repair by tracing each clue back to its source.

5. Unbounded conclusion

A result robust inside the tested conditions is extended to every condition. Repair by keeping the claim inside its operating boundary.

A Phase 4 Primary 6 sensitivity lesson

• Claim: state the conclusion precisely.
• Leverage: identify which evidence most controls it.
• Perturb: change one measurement or clue slightly.
• Assumption: flip one hidden assumption.
• Sample: test whether another sample could reverse the result.
• Time: change the observation interval where relevant.
• Leave out: remove one evidence source at a time.
• Independent: check whether remaining evidence is genuinely independent.
• Boundary: define where the conclusion is valid.
• Calibrate: strengthen or weaken confidence according to how much the conclusion survives.

Why small groups help sensitivity reasoning

Three students may reach the same answer for different reasons.

• One depends on a direct measurement.
• One depends on a diagram assumption.
• One depends on a remembered keyword.

Change one clue and see whose answer survives. The class learns that identical final answers can have very different robustness underneath.

What parents can practise at home

• Ask “which clue is doing most of the work?”
• Ask “what if that measurement were slightly different?”
• Ask “what assumption are you making?”
• Ask whether one new trial could reverse the conclusion.
• Ask whether several clues are independent.
• Ask where the conclusion stops being valid.
• During checking, ask what evidence would justify changing the answer.

How to tell whether robustness reasoning is improving

• High-leverage evidence is identified.
• Small differences are checked against measurement resolution.
• Assumptions are tested rather than hidden.
• Independent evidence is distinguished from repeated copies of one source.
• Conclusions are tested across reasonable variations.
• Tipping points and boundaries are recognised.
• Fragile conclusions are stated more cautiously.
• Robust answers are not changed merely because confidence fluctuates.

How this page fits the Hougang Science network

This eduKateSingapore page owns sensitivity and robustness of conclusions. It complements evidence sufficiency and conclusion calibration, evidence hierarchy and weighting, multi-source evidence integration, and scientific checking and uncertainty.

For the complete P3-to-PSLE map, use Hougang Primary Science Learning Library.

Official curriculum reference

The Ministry of Education’s Science Teaching & Learning Syllabus: Primary Three to Six develops analysis, evaluation, inference and communication alongside scientific knowledge. Sensitivity testing is used here as an age-appropriate scaffold for checking whether conclusions remain supported when reasonable uncertainties or assumptions are changed.

---

A strong Primary 6 conclusion should not collapse because one weak clue moved slightly. Identify what the answer depends on, perturb the uncertain parts, test the assumptions, remove one evidence source at a time and see whether the scientific account still stands.