eduKate Singapore

VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Next useful routeCanonical hubHELP

Mathematics in Bukit Timah | Simpson’s Paradox, Weighted Averages and Hidden Groups

—

by

eduKate Asia
Original mathematical editorial illustration of two grids with the same numbers of light and dark cells arranged in checkerboard and grouped-block patterns.

Did you know that one fictional method can have a higher success rate in every difficulty group and still have a lower overall success rate? The arithmetic is not broken. The groups are represented in different proportions, so the total rates are weighted differently. This is Simpson’s paradox, one of the most useful examples of why readers should look beyond a single headline percentage.

This Mathematics in Bukit Timah guide develops weighted averages, subgroup rates, aggregation, confounding, standardisation, comparisons with common weights, denominators, selection effects and the difference between description and causation. We use invented methods A and B and imaginary Easy and Hard puzzles. The counts are not school results, tuition outcomes, real learner performance, medical studies or estimates from Bukit Timah residents. The examples demonstrate exact arithmetic without creating evidence about real people.

Quick answer: Method A succeeds on 9 of 10 Easy and 24 of 40 Hard puzzles: 90% and 60%, overall 33/50=66%. Method B succeeds on 72 of 90 Easy and 5 of 10 Hard puzzles: 80% and 50%, overall 77/100=77%. A has a ten-percentage-point advantage in each difficulty group but B has an eleven-point advantage in the pooled headline. Different difficulty mixes reverse the apparent ranking.

Article contents

Original conceptual comparison of two differently grouped data patterns, accompanying fictional Simpson's paradox and weighted-rate examples.

Start by separating the groups

A headline rate is a fraction: successes divided by attempts.

If the attempts contain categories of unequal difficulty, the overall fraction depends both on how each method performs in a category and on how many attempts of each category were assigned to it.

That second ingredient—the group mix—can dominate the overall comparison. To understand the headline, inspect both numerator and denominator for every subgroup.

Why the Bukit Timah setting is fictional

Families around Sixth Avenue, Beauty World and King Albert Park may see statistical comparisons in articles about education and technology.

But no school or tutor is being assessed here. “A” and “B” are abstract puzzle methods, with toy counts chosen to make the mechanism transparent.

A mathematical reversal in invented numbers is not evidence that any real institution’s results are misleading or that one method truly causes better performance.

Define success and difficulty consistently

We imagine a puzzle trial produces either success or failure.

There are two predefined difficulty categories, Easy and Hard. Every attempted puzzle belongs to exactly one category.

For the arithmetic to be interpretable, “success” must mean the same thing across both methods, and difficulty categories must have the same definitions. Otherwise the rates cannot be compared cleanly.

The complete fictional table

DifficultyMethod A successes / attemptsA rateMethod B successes / attemptsB rate
Easy9 / 1090%72 / 9080%
Hard24 / 4060%5 / 1050%
All33 / 5066%77 / 10077%

Every cell is an invented puzzle count. The table is neither an eduKateSG score report nor a claim about learning systems.

Worked example 1: compare Easy puzzles

Method A’s Easy success rate is 9/10=90%.

Method B’s Easy rate is 72/90=80%.

Thus A’s Easy rate exceeds B’s by 10 percentage points. The difference uses percentages as numerical proportions; it is not a relative ten-percent increase.

Worked example 2: compare Hard puzzles

Method A completes 24 of 40 Hard puzzles, giving 24/40=60%.

Method B completes 5 of 10 Hard puzzles, giving 5/10=50%.

A again has a 10-percentage-point advantage. It is higher in both specified difficulty categories.

Worked example 3: calculate A’s overall result

Method A succeeds on nine Easy plus 24 Hard puzzles, totalling 33 successes.

It attempts ten Easy plus forty Hard puzzles, totalling 50 attempts.

Overall success rate is 33/50=66%. The result is lower than both methods’ Easy rates because A spends most of its attempts on Hard puzzles.

Worked example 4: calculate B’s overall result

Method B succeeds on 72 Easy plus five Hard puzzles, totalling 77 successes.

It attempts ninety Easy plus ten Hard puzzles, totalling 100 attempts.

Overall success rate is 77/100=77%. It is higher than A’s 66%, even though B’s within-difficulty success rates are lower.

There is no arithmetic contradiction

Method A’s two conditional rates are 90% and 60%. Its overall rate is 66%, between those two values.

Method B’s two conditional rates are 80% and 50%. Its overall rate is 77%, also between those two values.

Neither method’s total escapes the range of its own subgroup rates. The reversal occurs because their weights are different.

A weighted average explains every total

If groups have success rates r₁,r₂ and proportions w₁,w₂ of attempts, where w₁+w₂=1, the overall success rate is

R=w₁r₁+w₂r₂.

The weights are attempt shares, not an arbitrary equal average of the two percentages. For actual success counts, this identity follows by dividing the sum of successes by the sum of attempts.

Worked example 5: derive A’s weights

Method A has 10 Easy attempts out of 50, so Easy weight is 10/50=0.2.

Hard weight is 40/50=0.8.

Therefore A’s overall rate is 0.2×0.9+0.8×0.6=0.18+0.48=0.66.

Worked example 6: derive B’s weights

Method B has 90 Easy attempts out of 100, so Easy weight is 0.9.

Hard weight is 10/100=0.1.

B’s overall rate is 0.9×0.8+0.1×0.5=0.72+0.05=0.77.

What exactly changed between the methods?

A handles 20% Easy and 80% Hard.

B handles 90% Easy and 10% Hard.

So A performs more of the difficult category, while B performs more of the easy one. This different composition changes what the overall averages describe.

A’s conditional advantage is real within the stipulated table

We calculated that A’s Easy rate is higher than B’s Easy rate.

We also calculated that A’s Hard rate is higher than B’s Hard rate.

Those are exact properties of the toy counts. They do not by themselves establish that assigning a puzzle to A instead of B causes a higher success chance, because the attempts were not described as randomly assigned or otherwise comparable beyond the difficulty label.

The pooled ranking answers a different question

The headline A=66%, B=77% answers:

“What fraction of all attempts actually listed under each method succeeded?”

It does not answer:

“What happens if both methods receive the same distribution of puzzle difficulty?”

Those are different comparisons. Neither fraction is computationally invalid; the mistake is presenting one as though it answered the other.

Standardisation creates a common comparison mix

One way to compare category-specific rates is to apply the same chosen weights to both methods.

For example, let both methods be evaluated on a hypothetical mix of 50% Easy and 50% Hard puzzles.

This is direct standardisation, a mathematical reweighting exercise. It does not produce new observed experiments.

Worked example 7: standardise A to a 50–50 mix

A’s adjusted rate under equal Easy and Hard weights is

0.5×90%+0.5×60%=45%+30%=75%.

This differs from its observed pooled 66% because the hypothetical common mix contains fewer Hard puzzles than A actually faced.

Worked example 8: standardise B to the same 50–50 mix

B’s adjusted rate is

0.5×80%+0.5×50%=40%+25%=65%.

Under this equal mix, A is 75% and B is 65%, so A is ahead by ten percentage points—the same category advantage seen in the original table.

Common weights prevent this reversal when all subgroup advantages agree

Suppose A has greater rate than B in every comparable group.

If both overall rates use the same nonnegative weights summing to one, then the weighted A-minus-B difference is a sum of nonnegative weighted subgroup differences.

It cannot be negative. Therefore unequal group weights are essential to this particular kind of reversal when A leads in every stratum.

Prove it with two groups

Let A’s rates be a₁,a₂ and B’s be b₁,b₂, with a₁>b₁ and a₂>b₂.

With shared weights w and 1−w, where 0≤w≤1, subtract the adjusted rates:

R_A−R_B=w(a₁−b₁)+(1−w)(a₂−b₂).

Each term is nonnegative, and with positive weight on at least one strictly positive difference, the result is positive. The ranking cannot reverse under that common composition.

A different common mix gives another adjusted number

Using exactly half Easy and half Hard is one defensible fictional comparison, but it is not the only possible reference mix.

A common 70% Easy and 30% Hard mix yields different absolute rates.

The important methodological condition is that both methods use the same stated reference weights, not that the weights must be fifty-fifty.

Worked example 9: 70% Easy, 30% Hard

For A, adjusted rate is 0.7×0.9+0.3×0.6=0.63+0.18=81%.

For B, adjusted rate is 0.7×0.8+0.3×0.5=0.56+0.15=71%.

A again leads by ten percentage points. The absolute adjusted percentages differ from the 50–50 result because the reference mix changed.

Pooling the observed attempts supplies a possible reference

Across the two methods, Easy attempts total 10+90=100.

Hard attempts total 40+10=50. There are 150 attempts overall.

Thus the combined observed difficulty mix is two-thirds Easy and one-third Hard. It can serve as a transparent common standard for this exercise, though other standards may suit different questions.

Worked example 10: adjust to the pooled difficulty mix

For A: (2/3)×0.9+(1/3)×0.6=0.6+0.2=0.80.

For B: (2/3)×0.8+(1/3)×0.5=0.533333…+0.166666…=0.70.

Under the common pooled mix, A is 80% and B 70%. The original observed headline was A 66% and B 77%.

Standardised rates depend on what comparison is meaningful

If the question concerns a future workload with an expected 70% Easy share, the 70–30 standard may be relevant.

If the question concerns balanced competence across categories, the 50–50 comparison may be informative.

If the question simply asks what happened in the observed dataset, the raw totals 66% and 77% are valid descriptions. Choose the weights to match the question and disclose them.

Why this is called a paradox

A paradox feels surprising because people expect a comparison made in every subgroup to carry over to the whole.

But an aggregate rate is a weighted average, and the two methods’ weights differ.

The seeming contradiction disappears once the denominator composition is shown. The phenomenon is a warning about aggregation, not a violation of algebra or probability.

The difference is eleven percentage points in the pooled result

Pooled A rate is 66%. Pooled B rate is 77%.

Subtract A−B=66%−77%=−11 percentage points.

Inside each group, A−B=+10 percentage points. The sign reversal is the defining feature of this example.

Percentage points and relative percentage changes differ

A move from 50% to 60% is an increase of 10 percentage points.

The relative increase is (60−50)/50=20%.

When discussing Simpson’s paradox, quote percentage-point differences when subtracting success rates. Otherwise a reader may confuse relative and absolute comparisons.

Worked example 11: compare relative gains

For Easy puzzles, A’s 90% versus B’s 80% is a ten-point advantage.

Relative to B’s 80%, it is (90−80)/80=12.5% greater.

For Hard puzzles, A’s 60% versus B’s 50% is also ten points, but the relative increase is (60−50)/50=20%. Equal point differences need not mean equal relative improvements.

Weighted percentages are not ordinary means of the percentages

A has category rates 90% and 60%. The plain arithmetic mean is (90+60)/2=75%.

But the actual overall rate is 66%, because A attempted four times as many Hard puzzles as Easy puzzles.

The unweighted mean is meaningful only for a hypothetical equal-category-weight question, not as a substitute for the observed overall fraction.

Worked example 12: spot the wrong denominator

A pupil calculates A’s overall success as (90%+60%)/2=75% and labels it “observed total.”

The calculation ignored the fact that the counts were 10 Easy and 40 Hard attempts.

Correct pooled success is (9+24)/(10+40)=33/50=66%. The first erroneous step was using equal group weights without justification.

A large group can dominate an aggregate

For A, 80% of all attempts are Hard. Therefore the 60% Hard success rate strongly influences the pooled result.

For B, 90% are Easy, so the 80% Easy rate dominates its pooled result.

This principle generalises: a large subgroup can dominate a headline even when its internal rate is only one of several category rates.

Worked example 13: estimate before calculating

Without detailed arithmetic, A’s pooled rate should be closer to its Hard 60% than its Easy 90%, because most A attempts were Hard.

B’s pooled rate should be closer to its Easy 80% than its Hard 50%, because most B attempts were Easy.

The exact totals 66% and 77% match those directional estimates. This is a valuable mental check on a weighted average.

A rate must use its own denominator

When a method has 9 Easy successes out of 10 Easy attempts, its conditional rate is 9/10.

Dividing nine by the overall 50 attempts instead gives 18%, the share of all A attempts that were Easy successes, not the Easy success rate.

A mistake in denominator choice can create a false subgroup conclusion even before aggregation occurs.

Worked example 14: compare two types of fraction

For A, P(Success | Easy) in the observed frequency description is 9/10=90%.

The overall share of attempts that are both Easy and successful is 9/50=18%.

Both fractions are legitimate, but they describe different events. The vertical conditioning bar changes the denominator.

A three-way table can make the problem clearer

Every trial can be classified by method, difficulty and success outcome.

The original two-way success counts become part of a three-way frequency table. One may then condition on difficulty or ignore it.

Simpson’s paradox arises when the direction of a conditional association differs from the direction after marginalising over a third variable.

The word marginalisation has a mathematical meaning

In probability and statistics, marginalising over a variable means summing out its categories.

For method A, total success count is 9+24 and total attempts are 10+40.

Marginalising difficulty produces the overall rate 33/50. Conditioning on difficulty produces separate 9/10 and 24/40 rates. The two operations answer different questions.

Conditional probability gives another notation

Write S for success, M for method and D for difficulty.

Then P(S|M=A,D=Easy)=0.9 and P(S|M=B,D=Easy)=0.8 in the stipulated empirical table.

But P(S|M=A)=0.66 and P(S|M=B)=0.77. The signs reverse because P(D|M=A) and P(D|M=B) differ.

A formula exposes the source of reversal

The law of total probability gives

P(S|M)=P(S|M,D=Easy)P(Easy|M)+P(S|M,D=Hard)P(Hard|M).

For A, its difficulty weights are 0.2 and 0.8. For B they are 0.9 and 0.1.

Different conditional difficulty distributions produce different pooled success rates even though A leads inside each difficulty category.

Worked example 15: plug in the full conditional formula

For A: P(S|A)=0.9×0.2+0.6×0.8=0.66.

For B: P(S|B)=0.8×0.9+0.5×0.1=0.77.

All numbers lie between zero and one, and each method’s subgroup weights add to one. The reversal is mathematically consistent.

“Controlling for” a variable needs careful definition

A common phrase is “compare methods after controlling for difficulty.”

Here that means comparing within difficulty categories or computing standardised rates using a common difficulty distribution.

In more complex data, “control” can refer to regression, matching, stratification or other adjustments. Each has assumptions, and none automatically produces causal truth merely by including a variable in a formula.

Confounding can help explain a reversal

A confounder is a variable associated with the exposure or method and with the outcome, relevant to a causal comparison under a stated causal structure.

Difficulty is associated with which method receives which puzzles in the toy table, and it is associated with success rates.

This makes the difficulty composition a plausible source of confounding if one tried to treat the comparison causally. But actual causal conclusions would require a justified assignment and measurement design.

Not every grouping variable should be adjusted away

A third variable may be a confounder, a mediator or a collider in a causal model. Adjusting for it can have different consequences.

A purely mathematical Simpson reversal does not automatically tell us whether a subgroup-adjusted estimate or the overall estimate is the appropriate causal quantity.

The correct estimand, timing of variables and study design must be stated before treating a standardised calculation as causal evidence.

The raw aggregate remains a true description

In the toy dataset, B has 77 successes per hundred actual attempts while A has 33 per fifty.

It is correct to report B’s observed overall success rate is higher.

It becomes misleading only if this is used to claim B performs better on equal difficulty mixes or must cause a better outcome, despite the subgroup information.

The adjusted comparison is also conditional

The 50–50 adjusted values 75% for A and 65% for B are derived from applying the reported subgroup rates to a hypothetical shared mix.

They are not observations from a new trial where each method actually faced equal difficulty proportions.

The adjustment helps answer a comparison question, but its applicability depends on whether the observed subgroup rates would remain valid in the target setting.

Sample sizes also affect uncertainty

A’s Easy subgroup has only ten attempts, while B’s Easy subgroup has ninety.

Thus the observed 90% and 80% rates do not have equal sampling precision under a random-trial model.

The paradox can be demonstrated exactly using stipulated counts, but interpreting it as an empirical study would require uncertainty estimates and assessment of how the attempts were selected.

Worked example 16: one extra success has different impact

In A’s Easy subgroup, one additional success with the number of attempts fixed at ten would change the rate by 10 percentage points.

In B’s Easy subgroup, one additional success among ninety would change the rate by only 1/90≈1.11 percentage points.

The small subgroup’s rate is more sensitive to one outcome. This does not invalidate the arithmetic but limits strong empirical interpretation.

A statistical estimate can be exact for the sample

The fraction 9/10 is exactly 90% of the fictional listed attempts.

Whether it estimates a larger population’s true success probability accurately is a different question involving sampling uncertainty.

A mathematical explanation should distinguish exact descriptive arithmetic from uncertain population inference.

Weighted means can show Simpson’s paradox too

The phenomenon is not restricted to yes-or-no success rates.

Suppose average values are compared within two groups, and the group proportions differ between methods. A weighted overall mean can reverse the within-group ordering.

The same algebra applies: overall mean is a sum of subgroup means multiplied by their appropriate group weights.

A miniature mean-based illustration

Suppose fictional Process A has mean outcome 9 in Easy and 6 in Hard; Process B has 8 in Easy and 5 in Hard.

A leads in both categories by one unit. But assign A weights 0.2 Easy and 0.8 Hard, B weights 0.9 Easy and 0.1 Hard.

A’s overall weighted mean is 0.2×9+0.8×6=6.6; B’s is 0.9×8+0.1×5=7.7. The same reversal occurs.

Worked example 17: compare the means under common weights

Apply equal 0.5 weights to both processes.

Adjusted A mean is (9+6)/2=7.5.

Adjusted B mean is (8+5)/2=6.5. A again leads when the comparison mix is shared.

Why the paradox can disappear after stratification

Stratification examines each category separately.

Within Easy, A’s rate is higher. Within Hard, A’s rate is higher. A misleading one-number comparison becomes a more detailed description with consistent within-group directions.

However, presenting only subgroups could also hide the actual overall workload. Good communication shows both subgroup and overall results with their denominators.

This is not merely a rounding artefact

All table counts are whole numbers. The percentages 90,60,80,50,66 and 77 follow exactly from their fractions.

No rounding is needed to create the reversal.

A pupil who searches for a calculator error will miss the actual cause: the different distribution of Easy and Hard attempts across methods.

Simpson’s paradox can appear without bad intent

Aggregated data may be reported for convenience rather than deception.

The paradox is a reminder that a headline can conceal meaningful composition effects even when every reported count was recorded accurately.

The best response is to ask what groups exist, how they are weighted and which question the comparison is meant to answer.

Simpson’s paradox and correlation differ conceptually

A correlation coefficient summarises association between numerical variables under a particular definition.

Simpson’s paradox describes a reversal of association after aggregation or conditioning on groups.

Both concern how data are organised and summarised, but neither is a universal proof of causation. A dataset can have an apparent trend that changes when subgroup structure is considered.

Worked example 18: a direction reversal in signs

Let the aggregate A-minus-B rate difference be R_A−R_B=0.66−0.77=−0.11.

Within Easy, difference is 0.90−0.80=+0.10.

Within Hard, difference is 0.60−0.50=+0.10. Thus the aggregate sign is opposite to both subgroup signs.

A bar chart can conceal denominators

Two bars labelled “66%” and “77%” show only the pooled rates.

They do not reveal that A had 50 attempts and B had 100, nor their radically different difficulty mixes.

A clearer visualisation groups bars by difficulty and includes counts as well as rates. Data design should help readers understand the mechanism, not hide it.

Count labels matter as much as percentages

A subgroup with two successes out of two has 100% observed success, but very little information about a wider population.

A subgroup with ninety successes out of one hundred has 90% observed success with a different sample size.

Comparing the percentages alone discards precision and sampling context. Always retain numerator and denominator in a serious analysis.

A histogram is not the same as a stratified success table

A histogram typically represents the distribution of a numerical variable across intervals, with conventions for frequency density and bin width.

Our table separates categorical difficulty groups and binary outcomes. It should not be interpreted through histogram area rules without a distinct numerical binning problem.

This is a reminder to select the correct statistical representation for the type of data.

Simpson’s paradox depends on the unit of analysis

A table could count attempts, people, sessions or devices. Those denominators are not interchangeable.

In our example, the unit is one attempted puzzle. If one participant attempted many puzzles, treating all attempts as independent would require additional justification for inferential statistics.

For the exact descriptive proportions, the attempt unit is well defined. For causal or population conclusions, the data-generating process matters further.

Repeated attempts can create dependence

Suppose each fictional participant tries many puzzles. Their outcomes may be correlated because of ability or shared conditions.

The table of aggregate attempts can still be counted, but confidence intervals that assume all attempts are independent might understate uncertainty.

This is a separate statistical issue from Simpson’s paradox, and it illustrates why one correct calculation cannot validate every possible conclusion.

Standardisation requires overlapping groups

A common-weight comparison needs meaningful subgroup rates for each method in the reference categories.

If one method never appears in a particular category, its rate for that group may be unknown from the observed data.

Then direct standardisation cannot simply insert an invented value. The missing subgroup requires evidence, a model with assumptions or a narrower comparison population.

Worked example 19: recognise missing support

Suppose method A has no Hard attempts at all.

Its observed Easy rate may be known, but its Hard success rate cannot be calculated from the sample because the denominator is zero.

A hypothetical 50–50 adjusted A rate therefore cannot be determined from that table alone. Treating missing Hard attempts as a 0% success rate would be an unjustified assumption.

Changing the target population changes the standard

A future workload mostly composed of Hard puzzles makes the Hard category more influential in the relevant standardised comparison.

A future workload mostly Easy makes the Easy category more influential.

Under our particular table, A’s advantage remains ten percentage points for every common mix because its subgroup advantages are equal. In other datasets, the adjusted advantage can vary with the chosen common weights.

Worked example 20: a varying advantage example

Suppose a hypothetical table instead has A rates 90% Easy and 55% Hard, B rates 80% Easy and 50% Hard.

A leads by ten points Easy and five points Hard.

Under a 50–50 mix, A’s adjusted rate is 72.5% and B’s is 65%, a 7.5-point lead. Under 90% Easy, the lead is 9.5 points. The target mix matters when subgroup advantages differ.

Standardisation does not automatically solve causality

Even after applying common weights, the methods may differ in other hidden ways: puzzle format, timing, selection mechanism or repeated observations.

Difficulty adjustment addresses the difficulty composition in the stipulated model and nothing else.

It is appropriate to say “the adjusted descriptive comparison under these rates favours A.” It would be inappropriate to declare a causal treatment effect without a defensible design.

A causal diagram could clarify assumptions

A simple causal question asks whether difficulty affects both method assignment and success. If it does, difficulty can create confounding when comparing methods.

One may draw arrows from difficulty to method selection and from difficulty to success under a hypothetical design.

But causal diagrams describe assumptions about mechanisms. They are not automatically established by the same frequency table used to illustrate a statistical reversal.

Simpson’s paradox is a warning, not an excuse to discard averages

Weighted averages are mathematically correct and often useful.

The danger is failing to say which weights were used, or comparing rates from unlike populations as though the mixtures matched.

Good analysis preserves the overall average while making the subgroup structure and intended question visible.

Worked example 21: compute the grand total across both methods

Add all attempts: 50+100=150.

Add all successes: 33+77=110.

The combined success rate is 110/150=73⅓%. This is a third descriptive statistic covering the entire fictional table, not a direct comparison favouring A or B.

The combined total cannot replace a method comparison

The 73⅓% rate combines results from A and B.

It answers “What share of all 150 listed attempts succeeded?” not “Which method performs better?”

Using it to evaluate either method alone would blend the very categories under comparison and hide method-specific rates.

One extra category can create more complexity

Real datasets may be grouped by difficulty, time period, equipment type and task length simultaneously.

Different marginalisations can produce different summaries. A reversal over one grouping may disappear or change when more variables are included.

The mathematical response is to define the relevant target and model, not to search endlessly for the partition giving the preferred headline.

Why disaggregation should follow a question

If the goal is fair comparison at similar difficulty, stratifying by difficulty is meaningful.

If the goal is reporting actual throughput across the workloads assigned, the overall rates are meaningful.

A transparent analysis makes the intended question explicit and does not quietly substitute one estimand for the other.

Common error 1: average the subgroup percentages equally

A’s Easy and Hard rates are 90% and 60%, but it attempted unequal counts.

The plain mean 75% is not A’s observed pooled success rate. The true overall rate is 33/50=66%.

The first fix is to use subgroup attempt counts as weights, rather than assuming equal categories.

Common error 2: overlook the different mixes

A did mostly Hard puzzles; B did mostly Easy puzzles.

A headline graph with only 66% versus 77% conceals this difference.

Inspect the denominators before concluding the pooled ranking is a fair equal-difficulty comparison.

Common error 3: think the subgroup and aggregate figures contradict

The rates are 90% versus 80%, 60% versus 50%, and 66% versus 77%.

All are exactly compatible because they describe differently weighted populations.

Write the weighted-average formula to reveal how each total arises.

Common error 4: interpret subgroup superiority as proven causation

A’s listed success fraction exceeds B’s in each difficulty category.

That descriptive result does not prove giving a particular puzzle to A causes a higher chance of success. Assignment may have been nonrandom within categories.

Use language such as “higher observed rate in the stipulated table,” not “guaranteed better result.”

Common error 5: assume any adjustment is inherently correct

A 50–50 adjustment and a pooled-mix adjustment answer somewhat different hypothetical workload questions.

Both can be reasonable when clearly defined. Neither automatically supplies an empirically verified causal effect.

Choose the standard from the real question and disclose it.

Common error 6: ignore subgroup uncertainty

The Easy category for A has only ten attempts; B has ninety.

Even with exact fractional calculations, population inference would have different levels of uncertainty.

A successful algebraic explanation of Simpson’s paradox does not remove sampling variation.

Common error 7: treat a zero-count subgroup as zero performance

If a method has no observations in one subgroup, its observed rate there is undefined, not automatically zero.

A missing category cannot be silently filled in to support an adjusted comparison.

The correct response is to state the missing information and the assumption needed for any further model.

Primary preparation: weighted averages from unequal groups

Primary pupils can compare two groups containing different numbers of counters and calculate an overall fraction.

For example, one group may have nine successes in ten and another 24 in forty.

The early skill is adding numerators and denominators correctly rather than averaging percentages that describe unequal group sizes.

Lower Secondary: percentages and data tables

Secondary pupils can calculate subgroup proportions, percentage-point differences and overall rates.

They can learn to ask “How many attempts are represented?” before interpreting a headline.

The full conditional probability and confounding discussion can follow once weighted averages and simple probability notation are secure.

Advanced statistical literacy and scope

Simpson’s paradox connects descriptive statistics, conditional probabilities, causal inference and study design.

The SEAB 2027 G3 Mathematics syllabus directory offers school-level statistics context. The full causal-analysis discussion here is enrichment rather than a claim of universal SEC G3 examination content.

The valuable long-term skill is resisting misleading aggregate comparisons and choosing the right denominator.

A twenty-minute family diagnostic

Minutes 1–5: Calculate A’s Easy and Hard rates: 90% and 60%. Calculate B’s: 80% and 50%.

Minutes 6–10: Compute pooled A=33/50=66% and pooled B=77/100=77%. Explain why the reversal is not a calculator error.

Minutes 11–15: Reweight both methods to 50% Easy, 50% Hard. Obtain A=75%, B=65%.

Minutes 16–20: Ask which comparison answers the observed-workload question, which answers a common-difficulty-mix question, and why neither alone proves causation.

A seven-day learning sequence

Day one practises fractions with different denominators. Day two computes category-specific rates.

Day three calculates pooled rates by summing successes and attempts. Day four derives weighted-average equations.

Day five applies a shared reference mix. Day six discusses uncertainty and observational assignment. Day seven diagnoses a new apparent ranking reversal and writes an honest interpretation of both subgroup and aggregate results.

Practice questions with checked answers

A. A Easy rate 9/10? Answer: 90%.

B. A Hard rate 24/40? Answer: 60%.

C. A pooled rate 33/50? Answer: 66%.

D. B pooled rate 77/100? Answer: 77%.

E. A’s 50–50 standardised rate? Answer: 75%.

F. B’s 50–50 standardised rate? Answer: 65%.

G. A’s adjusted rate using pooled two-thirds Easy and one-third Hard weights? Answer: 80%.

H. Can the ranking reverse with identical subgroup weights when A leads strictly in every subgroup? Answer: no.

Frequently asked questions

What is Simpson’s paradox?

A reversal of an association or comparison when data are aggregated across groups instead of examined conditionally, often because subgroup proportions differ.

Is the paradox a mathematical contradiction?

No. The aggregate and subgroup rates are different weighted summaries of the same data.

Why can a method win every subgroup but lose overall?

It may be assigned a much larger share of difficult observations, while the other method receives a larger share of easier ones.

What is a weighted average?

An average in which each subgroup rate is multiplied by its proportion of the relevant total.

Can I simply average two percentages?

Only if that equal weighting answers the question. Actual pooled rates use subgroup denominators as weights.

What does standardisation do?

It applies the same reference group proportions to both sets of subgroup rates, creating a common-mix comparison.

Does standardisation prove causation?

No. It adjusts specified descriptive group differences; a causal interpretation needs justified design and assumptions.

Is the overall success rate still correct?

Yes. It describes the actual mixture of attempts, provided the counts and definitions are accurate.

Are these actual Bukit Timah school results?

No. Every task, attempt and success count is a fictional mathematical teaching example.

How can parents recognise understanding?

The learner can reproduce every total, state the group weights, explain the reversal and distinguish descriptive comparison from causal claims.

Official context and connected Mathematics in Bukit Timah

Consult the SEAB 2027 G3 Mathematics syllabus directory for Singapore school statistical foundations. All values here are invented and are not assessments of real students or tutors.

Read Mean, Median and Outliers, Sampling Bias and Population Estimates, Scatter Plots, Correlation and Causation, Bayes’ Theorem and Base Rates and Confidence Intervals and Sampling Uncertainty for related statistical ideas.

The immutable Clementi Secondary 1 Mathematics tutoring reference anchors first-principles working, checks, careful diagnosis and independent reasoning.

The lasting habit is to keep every numerator with its correct denominator, expose the group mix, compare under suitable shared weights when appropriate and never convert a descriptive reversal into an unsupported claim about real educational outcomes.

Companion Mathematics in Bukit Timah articles from this four-part batch: Hamming Distance, Parity Checks and Error-Correcting Codes · Hypothesis Testing, P-Values, Errors and Statistical Power · Fourier Series, Harmonics and Signal Reconstruction. Together these original guides explore reliable digital coding, statistical evidence, subgroup-weighted interpretation and Fourier signal representation, with worked checks and clearly fictional examples.