Interpolation, statistics and discrete distinctions
VectorPaint Simpson reversal: higher in each group but lower after aggregation
Plot fictional task successes by group and overall, checking unequal mixture weights without causal or policy recommendations.
On this page
One scheme can have a higher success rate within every difficulty group yet a lower aggregate rate. Use four fictional count cells for grouped and overall bars, adding numerators and denominators before division. This finite table evaluates nobody real; choosing a decision target requires additional goals and causal evidence.
Fix success counts, trial totals and groups
Set easy A=9 successes/10 attempts, B=80/100; hard A=30/100, B=2/10. Preserve numerator/denominator pairs. Groups are disjoint and exhaustive for this table, without duplicate or omitted attempts. A/B are fictional schemes, not institutions or measured performance. Save raw counts and separate easy, hard, aggregate and weight notes. Rate means successes divided by a positive total. Failures can be obtained by subtraction, but do not mix failure and success rates in one comparison. Failures are easy A1/B20 and hard A70/B8, totaling A71/B28. Adding successes returns 110 each, an independent count check without altering the table. Failure counts alone are not failure-rate comparisons.
Compare the same quantity within each group
Easy rates are 90% versus 80%, and hard 30% versus 20%, both A advantages of ten percentage points, not ten-percent relative increases. Retain fraction sources; 9<80 alone does not compare easy-group rates. Give within-group bars the same 100% width so different sample counts do not directly lengthen them. Label successes and totals beside each, preventing equal-width bars from hiding unequal group composition.
Aggregate numerators and denominators, not unweighted rates
A aggregates to (9+30)/(10+100)=39/110≈35.455%; B to (80+2)/(100+10)=82/110≈74.545%, reversing the within-group ordering. Averaging A’s 90% and 30% to 60% assigns equal group weights instead; B’s unweighted 50% is not its actual aggregate either. Equal aggregate denominators 110 simplify comparison but do not cause a contradiction. Different group compositions explain the reversal.
A success/total B success/total
easy 9/10 80/100
hard 30/100 2/10
aggregate 39/110 82/110
A group weights: easy=1/11, hard=10/11
B group weights: easy=10/11, hard=1/11
common 1/2,1/2 weights: A=0.60, B=0.50
Give rate bars and count cards distinct visual roles
Use a suggested 2000×1700 canvas with two group panels. Give every A/B background bar width 1000 pixels for 100%, with fills 900,800,300,200. Aggregate fills are approximately 354.545 and 745.455, retaining exact 39/110 and 82/110. Put totals 10,100,100,10 in separate count cards; one length must not simultaneously represent rates and counts. Share zero baselines and keep labels assigned to rows, avoiding an occlusion that pairs 30% with B’s 20% label.
Use weights to explain why subgroup advantages do not persist
A’s easy/hard weights are 1/11 and 10/11; B’s 10/11 and 1/11. Weighted rates (1/11)×0.9+(10/11)×0.3 and (10/11)×0.8+(1/11)×0.2 recover 39/110 and 82/110. Most A attempts belong to the lower-rate hard group, unlike B. Ten-point subgroup advantages cannot summarize away that composition. Any separate weight bars denote group shares, not success rates. Changing composition changes totals without itself changing within-group performance.
Treat common-weight comparison as a different question
As a separate stipulated target mix, give each group weight one half: standardized A=60%, B=50%. This answers a different question from the original aggregate. It neither makes the aggregate wrong nor unconditionally recommends A. Target composition, grouping role and assignment mechanism require more evidence. Arithmetic association supplies no randomization or causal effect for the same task. Label the control “different common weights”, retain originals and keep standardized rates outside the 110-attempt count ledger. For common easy weight w in [0,1], A=0.3+0.6w and B=0.2+0.6w, a constant difference 0.1. Equal weights prevent this reversal here, without selecting w for real decisions or eliminating unspecified factors.
Check denominators and groups before graphics
If aggregate A remains higher, check unweighted averaging. If easy A is lower, check counts versus rates. Wrong totals require four-cell assignments, not deleting hard cases to restore expectations. Fix bar widths from scale rather than move count cards to hide errors. Repair one copy factor and recompute subgroup rates, aggregate and weights. Preserve raw table, rate plot and standardization separately. Graphics neither infer confounding automatically nor justify real policy/performance conclusions.
Reopen and retain both orderings and causal limits
Save native work and reopen SVG/PDF, confirming A higher in both group panels but lower overall, with 39/110,82/110 and weight fractions readable. At fit-to-window size distinguish rate and weight bars; inspect rounding without changing exact counts. Reopen backgrounds, fills, counts and notes separately. Deliver all four cells and the common-weight control. Leave unperformed actions pending, without automatic statistical analysis, real trials or a resolved causal/decision problem.
Related tutorials and sources
VectorPaint bar chart: keep lengths, numbers and units consistent; VectorPaint Bayes source diagram: condition on red within sixty-four area units.
Method sources: Judea Pearl, UCLA: Understanding Simpson’s Paradox, §1.
Version and operation scope: checked on 17 November 2026 against the official VectorPaint listing and Apple public lookup: version 1.0.8, requiring iOS / iPadOS 17.0 or later. Draw manually with published shapes, pen, Bézier nodes, grids/guides, alignment, layers/groups, text and file output through visible installed controls. Prepare definitions and calculations externally; infer no built-in formula, statistics or graph solver. Keep editable native work and inspect SVG, PDF or PNG as needed. SVG/PDF containing imported bitmaps do not automatically become entirely vector. Canvases and samples are original teaching settings, without physical-device execution or real measurement; mark unperformed actions pending.