Simpson's Paradox in Ecommerce: Why Blended ROAS Rises While Every Channel Gets Worse

Why Did My Blended ROAS Go Up When Every Channel Went Down?

Because you shifted budget toward your highest-reported-ROAS channel. Blended ROAS is a spend-weighted average, so changing the weights moves the average even when every underlying channel declines. This is Simpson's paradox — an aggregate trend reversing the direction of every subgroup within it. The fix is decomposing the change into an allocation effect and a within-channel effect.

If you have ever presented a blended number that looked great while your channel leads insisted things were getting harder, you have met this. Both parties were right. What follows is the arithmetic that reconciles them, and the causal reason the reconciliation still is not the end of the story.

What Simpson's Paradox Actually Is

Simpson's paradox is the phenomenon where a trend that appears in several groups of data reverses when those groups are combined. It is a genuine mathematical result, not a data error — the aggregate and the subgroups are both computed correctly and they genuinely disagree.

In ecommerce it shows up wherever you compute a ratio across a mix that changes: blended ROAS, blended CAC, overall conversion rate, average order value, or MER. Any time the composition of the denominator shifts between periods, the aggregate can move independently of — even opposite to — every component.

The reason it is dangerous rather than merely curious is that marketers use aggregate ratios as performance signals and reallocate budget accordingly. When the signal is driven by mix rather than performance, you get a feedback loop that reinforces whatever channel is best at claiming conversions rather than causing them.

A Worked Example in Euros

Illustrative example. Figures are constructed to demonstrate the arithmetic clearly, not drawn from a specific customer account.

A Shopify brand spends €50,000 a month across two channels. In Month 1, Google's reported ROAS looks far better than Meta's, so for Month 2 the team reverses the budget split.

Month 1

Month 2 — budget flipped to 20/80

Meta's ROAS fell from 3.00 to 2.80. Google's fell from 5.00 to 4.70. Every channel got worse. Blended ROAS rose from 3.40 to 4.32 — a 27% improvement — and monthly revenue rose €46,000.

The team celebrates. The team is measuring mix.

The Mix-Shift Decomposition

Here is a diagnostic you can run in a spreadsheet in five minutes. Blended ROAS is the spend-weighted average of channel ROAS:

Blended ROAS = Σ (spend share of channel × ROAS of channel)

Verify: Month 1 = (0.80 × 3.00) + (0.20 × 5.00) = 2.40 + 1.00 = 3.40 ✓

To split the €change into its two causes, compute a counterfactual: the new spend mix applied to the old channel ROAS. That is the blended ROAS you would have achieved purely by reallocating, with no change in channel performance.

Counterfactual = (0.20 × 3.00) + (0.80 × 5.00) = 0.60 + 4.00 = 4.60

Now decompose:

The headline "+0.92 blended ROAS" is really +1.20 from reallocation, minus 0.28 from genuine performance decay. The team's entire reported win came from mix. Their underlying media efficiency deteriorated, and the aggregate hid it.

This decomposition is the core diagnostic. Run it every time a blended metric moves, before you interpret the movement as performance.

The Causal Trap Underneath the Statistical One

Fixing Simpson's paradox with decomposition is necessary but not sufficient, and this is where most treatments of the topic stop. The +1.20 allocation effect is only real gain if Google's 5.00 ROAS was causal — if those euros genuinely would not have arrived without Google spend.

They probably were not, and here is why. Google's reported ROAS in Month 1 was inflated by exactly the mechanisms that make platform-reported numbers diverge from reality:

• Branded search harvesting. A large share of Google's conversions are people typing your brand name — demand that Meta's upper-funnel spend created. Branded search is frequently non-incremental, yet it books full credit.
• Last-touch position. Google sits closest to purchase, so last-click logic hands it conversions that earlier touchpoints caused.
• Audience overlap. The same customers see both channels, and both platforms claim them.

Watch what happens in Month 2. Meta spend drops 75%. Google's ROAS drops from 5.00 to 4.70 — a small but real decline. That decline is a fingerprint: Google's efficiency depended partly on the demand Meta was generating. The channel that looked independently excellent was partly harvesting a stream that something else fed.

So the honest reading of Month 2 is worse than "mix, not performance." It is: the brand cut the channel creating demand, funnelled budget into the channel capturing it, and booked the resulting bookkeeping shift as a 27% improvement — while total incremental revenue may well have fallen. The paradox concealed a mix change; the mix change concealed a causal error.

This is the difference between correlational and causal attribution. Decomposition tells you the aggregate moved because of mix. Only causal attribution tells you whether the mix change was a good idea. Methods that estimate incremental contribution with explicit uncertainty — Bayesian causal models or geo lift tests — answer the question the decomposition raises but cannot settle.

Where Else This Bites in Ecommerce

The blended ROAS case is the most common, but the same mechanism operates anywhere a ratio spans a shifting mix:

The unifying rule: never interpret a ratio without checking whether its denominator changed composition.

The Four-Step Mix-Shift Audit

1. Pull spend and revenue by channel for both periods. Compute channel-level ROAS and spend shares. If the shares moved more than about five percentage points, you have a mix-shift candidate.
2. Compute the counterfactual. Apply the new spend shares to the old channel ROAS figures. This is your allocation-only blended ROAS.
3. Decompose. Allocation effect = counterfactual − old blended. Within-channel effect = new blended − counterfactual. Report both numbers. Never report the total alone.
4. Interrogate the allocation effect causally. Ask whether the channel you shifted toward has a genuinely higher incremental return, not merely a higher reported one. Check whether the channel you shifted away from was feeding it. If the receiving channel's own ROAS declined as you scaled it, treat that as evidence of dependency.

Step 4 is the one that requires more than a spreadsheet — it is a jump in attribution maturity, not just in analytical rigour — and it is the one that determines whether your reallocation created value or destroyed it.

Common Mistakes

• Reporting only the blended number. If your weekly dashboard shows blended ROAS without spend shares beside it, the number is uninterpretable.
• Treating decomposition as the finish line. Splitting allocation from performance is arithmetic. Deciding whether the allocation was correct is causal inference.
• Assuming the paradox only runs one direction. It equally hides improvements: channels can all get better while blended declines, causing teams to kill working strategies.
• Comparing periods with different channel counts. Launching a new channel mid-period makes period-over-period blended comparison meaningless.
• Segmenting until something looks good. Slicing repeatedly until you find a favourable cut manufactures spurious patterns. Decide your segmentation before you look.
• Trusting platform ROAS as the input. Decomposition built on inflated channel ROAS produces a precisely-calculated wrong answer. Garbage in, decomposed garbage out.
• Ignoring the scale relationship. ROAS is not constant as you scale spend. Moving €30,000 into a channel changes that channel's marginal return, so applying its old ROAS to a much larger budget overstates the allocation effect.

Checklist: Before You Act on a Blended Metric

• I have channel-level spend shares for both periods, not just the aggregate.
• Spend shares moved by less than five points — or I have run the decomposition.
• I have reported allocation effect and within-channel effect separately.
• I have checked whether channel ROAS figures are platform-reported or incrementality-validated.
• I have asked whether the channel I scaled depends on the channel I cut.
• I have checked that no channel launched or paused mid-period.
• I have confirmed the comparison periods are seasonally comparable.
• My segmentation was decided before I looked at results.

Key Takeaways

• Simpson's paradox is arithmetic, not error. Aggregate ratios can move opposite to every component when the mix shifts.
• Blended ROAS is a spend-weighted average. Changing the weights changes the average without any performance change whatsoever.
• Always decompose: allocation effect = (new mix × old ROAS) − old blended; within-channel effect = new blended − that counterfactual.
• Decomposition is necessary but not sufficient. The allocation effect is only real if the receiving channel's reported ROAS is genuinely incremental.
• A declining ROAS in a channel you are scaling is a dependency signal — it suggests the channel was harvesting demand something else created.

The deeper lesson is that this paradox is not a reporting bug to be patched. It is a structural consequence of using correlational metrics to make allocation decisions. A rules-based attribution model assigns credit by position in a click path; it has no concept of what would have happened otherwise, so it cannot distinguish a channel that creates demand from one that captures it. Feed those numbers into a blended average and you get a metric that rewards demand-harvesting and penalises demand-creation — which is precisely how brands end up cutting their upper funnel and calling it optimisation. A causal attribution approach estimates each channel's incremental contribution directly, which makes both the paradox and its underlying trap visible at once. That shift sits at the centre of any serious ecommerce analytics stack, and it is what separates genuine measurement from dashboard decoration. It is also the first thing worth checking when evaluating any attribution tool: ask whether it reports incremental contribution or merely redistributes credit across a click path.

If you want to sanity-check the decomposition on your own numbers, the blended ROAS calculator will give you the aggregate, and proper test design will tell you whether you have enough volume to validate the channel ROAS figures feeding it.

For €99, upload any historical GA4 period and get causal attribution for every channel in 5–10 minutes — no pixel, no migration. Go Pro at €299/mo for continuous attribution, an AI chatbot for your data, and a developer API.