The Average Hid Two Very Different Segments

A textbook example of Simpson's Paradox where a flat aggregate performance line concealed massive internal shifts between power users and new recruits.

AnalystLisa Wong
Case Published2026-08-01
TypeSegment Analysis

The Signal

The core business metric for this platform was 'Average Session Duration'. For three consecutive reporting periods, the average remained remarkably stable at approximately 4.5 minutes. From a high-level executive perspective, this signaled a healthy, predictable product-market fit. The internal analytics team viewed the flat line as a successful baseline, leading to a period of reduced focus on engagement optimization and a shift in resources toward new feature development.

The Conflict

The conflict arose when the marketing department complained of lower conversion rates for new sign-ups despite the 'stable' engagement. Upon applying tenure-based segmentation, we discovered that the population was splitting. 'Power Users' (accounts older than 6 months) had increased their engagement by over 40% due to advanced feature adoption. Meanwhile, 'New Users' had seen their average session time collapse by nearly 50%. The overall average remained stable only because these two powerful, opposing trends were canceling each other out perfectly.

The Resolution

The team shifted from monitoring a single aggregate mean to tracking segmented cohort engagement. This revealed that a recent onboarding update had removed vital introductory tutorials, thinking they were redundant. By restoring localized onboarding for new users while maintaining the high-efficiency environment for veterans, the engagement for new sign-ups returned to historical levels. Total platform usage increased by 22% within one month of resolving the segment conflict.

  • Primary Segment Shift+40.2%
  • Secondary Deviation-50.5%
  • Overall Baseline Impact0.04% Variance

A static average is often an indicator of high internal variance. In heterogeneous systems, you must assume that the mean is a lie until you have validated that the variance within each key segment is also stable.

Case Discussion

Tom H.
Tom H.
07/28/2026

Simpson's paradox in action!

Rachel M.
Rachel M.
07/30/2026

Always segment your data.

Post an Analysis Reply