B has the higher signup rate. Why does it lose on both devices?

I'm a product manager reviewing two signup-page designs. Our team wants to ship design B to everyone because the dashboard says its signup conversion rate is higher. Can you analyze the results, tell me what you would do next, and draft the decision note I can take to the team?

This is a fictional exercise. The export covers the same 14-day enrollment period for both designs, and every included user has a complete seven-day follow-up. Each row counts distinct eligible users who first saw that design, with completed signups counted once within seven days of that exposure. A user appears in only one design and one device group; device is recorded at first exposure. Desktop and mobile are the only groups. There are no missing counts in this export. I have not yet checked how users were assigned to designs, whether allocation changed during the period, or how targeting was configured.

Design A, desktop: 300 exposed users, 90 completed signups.
Design A, mobile: 700 exposed users, 140 completed signups.
Design B, desktop: 700 exposed users, 196 completed signups.
Design B, mobile: 300 exposed users, 54 completed signups.

The proposed announcement says: "B improves signup conversion by 8.7%, so we should roll it out to all users." I need to know whether that claim and rollout decision follow from these numbers. Please show the calculations that matter and explain what evidence we still need.

The raw numbers do show a higher **overall observed conversion rate** for B, but they do **not by themselves establish that the design caused the improvement** or justify an immediate all-user rollout.

### 1. What the export actually shows

| Design |               Desktop |                Mobile |                 Overall |
| ------ | --------------------: | --------------------: | ----------------------: |
| A      |  90 / 300 = **30.0%** | 140 / 700 = **20.0%** | 230 / 1,000 = **23.0%** |
| B      | 196 / 700 = **28.0%** |  54 / 300 = **18.0%** | 250 / 1,000 = **25.0%** |

So:

* A: **230 / 1,000 = 23.0%**
* B: **250 / 1,000 = 25.0%**
* Absolute difference: **+2.0 percentage points**
* Relative conversion-rate change: **(25% − 23%) / 23% = 8.70%**

Therefore, the announcement's **“B improves signup conversion by 8.7%”** is mathematically equivalent to the observed aggregate rate difference, **if “improves” is intended descriptively as “had an 8.7% higher observed conversion rate.”**

But that wording can misleadingly imply causation.

### 2. The important complication: device mix

The designs have very different device distributions:

* A: **30% desktop / 70% mobile**
* B: **70% desktop / 30% mobile**

And desktop converts substantially better than mobile **for both designs**:

* A: 30% desktop vs. 20% mobile
* B: 28% desktop vs. 18% mobile

Within **each device group, B is actually lower**:

* Desktop: B **28%** vs. A **30%** → **−2 pp**
* Mobile: B **18%** vs. A **20%** → **−2 pp**

So the aggregate result reverses because B received a much larger share of desktop users.

This is a classic composition/mix issue. In this particular dataset, if the two designs had the **same device mix**, B would not have the observed aggregate advantage.

For example, applying A's 30% desktop / 70% mobile mix to B's device-specific rates:

**0.30 × 28% + 0.70 × 18% = 21.0%**

compared with A's **23.0%**.

Conversely, applying B's 70% desktop / 30% mobile mix to A:

**0.70 × 30% + 0.30 × 20% = 27.0%**

compared with B's **25.0%**.

That strongly illustrates why the overall 25% vs. 23% comparison should not be interpreted as a design effect without understanding assignment.

### 3. What I would do next

I would **not announce an all-user rollout yet**. First establish how users were assigned.

Specifically, check:

1. **Assignment mechanism:** Were users randomly assigned to A/B?
2. **Allocation over time:** Did the A/B split change during the 14 days?
3. **Targeting rules:** Were particular devices, countries, traffic sources, account types, or other audiences preferentially exposed to one design?
4. **Exposure counts over time:** Confirm whether the device imbalance was present throughout the period or arose from a change in allocation.
5. **Statistical uncertainty:** Calculate confidence intervals or an appropriate test for the comparison, ideally both overall and within device strata.
6. **Pre-specified primary metric:** Confirm that seven-day completed signup was the intended primary outcome and that the analysis wasn't selected after seeing the results.

The key question is whether the assignment process makes the designs comparable. If assignment was properly randomized and balanced/appropriately analyzed, the evidence can support a causal comparison. If assignment was targeted or otherwise non-random, the observed difference could reflect who received each design rather than the designs themselves.

### Decision note

Here is a version you can take to the team:

**Signup-page A/B results — decision note**

**Summary**

The export shows an observed seven-day signup conversion rate of **25.0% for Design B (250/1,000)** versus **23.0% for Design A (230/1,000)**. This is a **2.0 percentage-point absolute difference**, or an **8.7% relative difference**.

However, these aggregate results do not currently establish that Design B caused the improvement or support an immediate rollout to all users.

**Why**

The device mix differs substantially between the designs:

* Design A: 300 desktop and 700 mobile users
* Design B: 700 desktop and 300 mobile users

Within both device groups, B has a lower observed conversion rate than A:

* Desktop: A 30.0% vs. B 28.0%
* Mobile: A 20.0% vs. B 18.0%

Thus, the aggregate advantage for B is driven by its much higher proportion of desktop users, whose conversion rate is higher under both designs. The aggregate 25.0% vs. 23.0% result should therefore not be treated as a clean estimate of the design effect without examining assignment and device composition.

**What we should verify before rollout**

1. How users were assigned to A or B, including whether assignment was randomized.
2. Whether the allocation between designs changed during the 14-day period.
3. Whether targeting rules caused different user populations to receive the two designs.
4. Whether the device mix and conversion differences were consistent over time.
5. Statistical uncertainty around the observed differences, including device-level comparisons.

**Current decision**

Do not characterize the result as evidence that B “improves” conversion by 8.7% in a causal sense yet, and do not use these aggregate numbers alone as the basis for an all-user rollout.

The verified descriptive result is: **B had a 25.0% observed seven-day signup conversion rate versus 23.0% for A during this 14-day export, a relative difference of 8.7%.**

A rollout decision should follow the assignment/targeting review and an appropriately stratified statistical analysis.

The most important takeaway for the team is that **the 8.7% calculation is correct as an aggregate relative difference, but the proposed causal interpretation and rollout do not follow from these numbers alone.**

I'm a product manager reviewing two signup-page designs. Our team wants to ship design B to everyone because the dashboard says its signup conversion rate is higher. Can you analyze the results, tell me what you would do next, and draft the decision note I can take to the team?

This is a fictional exercise. The export covers the same 14-day enrollment period for both designs, and every included user has a complete seven-day follow-up. Each row counts distinct eligible users who first saw that design, with completed signups counted once within seven days of that exposure. A user appears in only one design and one device group; device is recorded at first exposure. Desktop and mobile are the only groups. There are no missing counts in this export. I have not yet checked how users were assigned to designs, whether allocation changed during the period, or how targeting was configured.

Design A, desktop: 300 exposed users, 90 completed signups.
Design A, mobile: 700 exposed users, 140 completed signups.
Design B, desktop: 700 exposed users, 196 completed signups.
Design B, mobile: 300 exposed users, 54 completed signups.

The proposed announcement says: "B improves signup conversion by 8.7%, so we should roll it out to all users." I need to know whether that claim and rollout decision follow from these numbers. Please show the calculations that matter and explain what evidence we still need.

The proposed claim and rollout decision are **incorrect** and based on a classic statistical illusion: **Simpson's Paradox**.

While Design B appears to win overall, Design A is actually equal or superior on both desktop and mobile when analyzed separately.

---

### Step 1: The Calculations That Matter

#### 1. Segment-by-Segment Conversion Rates

* **Desktop:**
* Design A: $90 / 300 =$ **30.0%**
* Design B: $196 / 700 =$ **28.0%**
* *Verdict:* **Design A wins** by 2.0 percentage points.


* **Mobile:**
* Design A: $140 / 700 =$ **20.0%**
* Design B: $54 / 300 =$ **18.0%**
* *Verdict:* **Design A wins** by 2.0 percentage points.



#### 2. Aggregate Conversion Rates

* **Design A Overall:** $(90 + 140) / (300 + 700) = 230 / 1,000 =$ **23.0%**
* **Design B Overall:** $(196 + 54) / (700 + 300) = 250 / 1,000 =$ **25.0%**
* *Difference:* $25.0\% - 23.0\% = +2.0$ percentage points ($+8.7\%$ relative improvement).

#### Why is this happening?

The aggregate metric is distorted because the **traffic mix (device allocation) is heavily unbalanced**:

* Desktop users convert at a much higher baseline rate (~28–30%) than mobile users (~18–20%).
* **Design B received 70% desktop traffic**, while **Design A received only 30% desktop traffic**.
* Design B looks better overall simply because it was shown to significantly more high-converting desktop users, not because the design itself performs better.

---

### Step 2: What Evidence We Still Need

Before making any rollout decision, you need to investigate **why the allocation was skewed**:

1. **Traffic Split Verification:** Was this an unstratified 50/50 split that accidentally suffered from a bad random assignment, or was targeting misconfigured (e.g., routing desktop users to Design B)?
2. **Allocation Over Time:** Did the split change mid-experiment? (e.g., ramping up Design B on desktop midway through).
3. **Statistical Significance:** With $N = 1,000$ per variant, running a chi-squared test or two-proportion z-test on each segment will reveal whether Design A’s 2 percentage point lead on mobile and desktop is statistically significant or just noise.
4. **Behavioral Metrics Post-Signup:** Since conversion rates within device types favor Design A, check down-funnel retention or activation to ensure user quality isn't compromised.

---

### Step 3: What To Do Next

1. **Do NOT ship Design B to everyone.** Shipping Design B globally would likely decrease your overall conversion rate due to its lower segment performance.
2. **Pause or fix the experiment setup** to ensure an even 50/50 randomized split across both device types (stratified sampling).
3. **Re-run or continue the test** with balanced allocation until statistical significance is reached within each device segment.

---

### Step 4: Draft Decision Note for the Team

```text
SUBJECT: Hold on Design B Rollout – Re-evaluating A/B Test Results

Team,

I am recommending that we pause the full rollout of Design B. While the top-line dashboard shows Design B with a 25.0% overall conversion rate vs. Design A's 23.0% (+8.7% relative lift), a segment breakdown reveals that this is caused by Simpson's Paradox due to an uneven traffic split across devices.

Key Findings:
1. Desktop Conversion: Design A (30.0%) beats Design B (28.0%).
2. Mobile Conversion: Design A (20.0%) beats Design B (18.0%).
3. Traffic Imbalance: Design B was exposed to 70% desktop traffic (which converts at a higher baseline rate), while Design A was exposed to 70% mobile traffic. 

Because Design A actually outperforms Design B on both desktop and mobile individually, rolling out Design B to 100% of users would likely cause our total signups to drop.

Next Steps:
- Engineering/Data Science to review experiment configuration to understand why traffic allocation was skewed (70/30 instead of 50/50 per device).
- Calculate statistical significance on the per-device segment results.
- If necessary, run a clean follow-up test with proper stratified 50/50 randomization before making a final decision.

Let's discuss during our sync today.

```

I'm a product manager reviewing two signup-page designs. Our team wants to ship design B to everyone because the dashboard says its signup conversion rate is higher. Can you analyze the results, tell me what you would do next, and draft the decision note I can take to the team?

This is a fictional exercise. The export covers the same 14-day enrollment period for both designs, and every included user has a complete seven-day follow-up. Each row counts distinct eligible users who first saw that design, with completed signups counted once within seven days of that exposure. A user appears in only one design and one device group; device is recorded at first exposure. Desktop and mobile are the only groups. There are no missing counts in this export. I have not yet checked how users were assigned to designs, whether allocation changed during the period, or how targeting was configured.

Design A, desktop: 300 exposed users, 90 completed signups.
Design A, mobile: 700 exposed users, 140 completed signups.
Design B, desktop: 700 exposed users, 196 completed signups.
Design B, mobile: 300 exposed users, 54 completed signups.

The proposed announcement says: "B improves signup conversion by 8.7%, so we should roll it out to all users." I need to know whether that claim and rollout decision follow from these numbers. Please show the calculations that matter and explain what evidence we still need.

# Bottom line

**The claim and the rollout decision do not follow from these numbers.** B's higher overall rate comes from a difference in device mix between the arms, not from B performing better. Within each device group, B converts 2.0 points lower than A. The data is also too small to show B is worse.

## 1. Overall result (what the dashboard shows)

| Design | Exposed | Signups | Rate |
|---|---|---|---|
| A | 1,000 | 230 | 23.0% |
| B | 1,000 | 250 | 25.0% |

- Absolute difference: +2.0 pp. Relative lift: 25.0 / 23.0 − 1 = **8.7%**, which is where the announcement's figure comes from.
- The claim calls this "conversion improves by 8.7%", but it is a relative lift, and the absolute change is 2.0 pp.
- Even taken at face value, this is not statistically distinguishable from zero. Pooled z ≈ 1.05, p ≈ 0.30, 95% CI for the difference ≈ −1.7 to +5.7 pp.

## 2. By device (the reversal)

| Device | A rate | B rate | B − A | 95% CI for B − A |
|---|---|---|---|---|
| Desktop | 90/300 = 30.0% | 196/700 = 28.0% | −2.0 pp | −8.2 to +4.2 pp |
| Mobile | 140/700 = 20.0% | 54/300 = 18.0% | −2.0 pp | −7.3 to +3.3 pp |

B is lower in both groups yet higher overall. This is Simpson's paradox, driven by traffic mix:

| Design | Desktop share | Mobile share |
|---|---|---|
| A | 30% | 70% |
| B | 70% | 30% |

Desktop converts better for both designs, and B's traffic is 70% desktop while A's is 70% mobile. B's overall number is lifted by who saw it, not by the design.

## 3. Same-mix comparison

Weighting both designs to the same device mix removes the composition effect:

| Standard mix | A | B | B − A |
|---|---|---|---|
| A's mix (30% desktop) | 23.0% | 21.0% | −2.0 pp |
| B's mix (70% desktop) | 27.0% | 25.0% | −2.0 pp |
| 50/50 | 25.0% | 23.0% | −2.0 pp |

The device-adjusted estimate is **−2.0 pp** (B lower), with a 95% CI of roughly −6.1 to +2.1 pp (z ≈ −1.0). So the data does not show B is worse either. It is inconclusive.

## 4. A second red flag: the allocation itself

Each design has 1,000 users, but the device split is 300/700 versus 700/300. If users were randomly assigned to A or B regardless of device, the splits would be near-identical. A chi-square test on device × design gives χ² = 320, which is far beyond chance. **Assignment was not independent of device**, so this does not look like a clean randomized test. Possible causes are device-based targeting, a change in allocation partway through the period, or a rollout by platform.

This matters because the device-adjusted comparison in section 3 is only trustworthy if assignment within each device group was random. That is unverified.

## 5. Evidence still needed

1. **Assignment mechanism:** who or what decided A vs B, and whether it was random within each device group.
2. **Allocation history:** daily exposure counts by design and device, to see whether allocation or targeting changed mid-period. If it did, time is another confounder.
3. **Targeting rules:** any device, geography, traffic source, or new/returning filters that differed by design.
4. **Other mix differences:** acquisition channel, country, browser, and campaign timing by design, since device may not be the only imbalance.
5. **Sample ratio check:** the intended split versus the observed split, overall and by device.
6. **Guardrails:** signup quality, activation, retention, error rates, and page speed. A signup lift that doesn't carry through to downstream metrics isn't a win.
7. **Power:** at a ~23% baseline, detecting a 2 pp difference needs roughly 7,000 users per arm (80% power, α = 0.05). This export has 1,000 per arm, and per-device analysis needs more.

## 6. What I would do next

1. Do not announce or roll out B based on this dashboard.
2. Get the assignment and allocation answers above, since they decide whether this data can be salvaged.
3. Rerun as a properly randomized test, stratified by device, with a fixed sample size and a pre-declared primary metric (device-adjusted signup rate) and guardrails.
4. If the two designs need to be compared sooner, use the device-adjusted comparison as the primary readout, not the pooled dashboard number.

## Draft decision note

**Title:** Signup page B vs A: rollout decision

**Status:** Not ready to roll out. Recommend a controlled rerun.

**Rationale:**
- The dashboard shows B at 25.0% vs A at 23.0% (+2.0 pp, +8.7% relative). This gap is not statistically significant (p ≈ 0.30).
- The overall gap is a device-mix artifact. B's traffic was 70% desktop and A's was 70% mobile, and desktop converts better for both designs.
- Within each device, B is lower: desktop 28.0% vs 30.0%, mobile 18.0% vs 20.0%. Adjusted for device mix, B − A is about −2.0 pp (95% CI ≈ −6.1 to +2.1 pp), which is inconclusive.
- The device split between designs is far from what random assignment would produce, so assignment and targeting are unverified.

**Cost:** [Engineering and design cost of rollout: to be filled in]. Cost of delay: one more test cycle. Cost of a wrong rollout: possible lower conversion and an announcement we would have to retract.

**Expected effect:** Unknown. The data supports neither a B improvement nor a B decline; the plausible range for B's effect spans roughly a 6 pp decrease to a 2 pp increase.

**Response:**
1. Hold the "8.7% improvement" announcement and the full rollout.
2. Confirm the assignment mechanism, allocation history, and targeting rules.
3. Rerun with device-stratified randomization and a pre-set sample size (about 7,000+ per arm for a 2 pp effect), tracking signup quality and retention as guardrails.
4. Decide on the device-adjusted result.

**Assumptions:** I treated the export as accurate and complete as you described it. The confidence intervals use normal approximations, and the device-adjusted estimate assumes assignment was random within each device group, which is exactly what we still need to verify.