Decomposing Metric Movements
A decomposition answers “which terms of an identity changed?” It does not answer “what intervention caused the change?” This distinction matters because mathematically exact contributions can still reflect selection, measurement drift, seasonality, or common causes.
Prerequisites
Complete monetization and unit economics. Be able to distinguish an accounting identity from an estimated relationship.
Learning goals
- Write the aggregate as an explicit identity before decomposing it.
- Use a symmetric two-factor decomposition whose contributions add exactly.
- Distinguish volume, rate, mix, and interaction terms.
- Preserve causal language for designs that identify effects.
Decision contract
For adoption, the identity is : adopters equal eligible active actors times adoption rate. From baseline to restated current , adopter count rises from two to four. A midpoint decomposition assigns zero to volume and two to rate:
The identity closes exactly, but “rate contribution” remains descriptive. It does not say the rollout caused two additional adopters.
For multiple segments, also report starting weights, within-segment rates, and changing mix. Choose a declared ordering or a symmetric method; otherwise interaction allocation can depend on analyst preference.
Worked example
active_0, rate_0 = 4, 0.5
active_1, rate_1 = 4, 1.0
change = active_1 * rate_1 - active_0 * rate_0
volume = (active_1 - active_0) * (rate_0 + rate_1) / 2
rate = (rate_1 - rate_0) * (active_0 + active_1) / 2
print(f"adopters_change={change:.1f} volume_effect={volume:.1f} rate_effect={rate:.1f} identity_match={change == volume + rate}")
print("description_only=True causal_explanation=False")
Expected output:
adopters_change=2.0 volume_effect=0.0 rate_effect=2.0 identity_match=True
description_only=True causal_explanation=False
Limitations and common mistakes
- Decomposition inherits metric-definition, eligibility, identity, and missingness errors.
- A changing average can reflect within-group change, composition change, or both.
- Contributions can depend on factor order unless the method treats interaction explicitly.
- Exact closure is arithmetic validation, not causal validation.
- Percent change is unstable near zero; retain levels and denominators.
Practice
Add two actor segments with different baseline rates. Decompose the aggregate change into mix and within-segment terms, state the chosen interaction rule, and list three causal stories consistent with the same arithmetic result.
Next steps
Observational product analysis and confounding explains why descriptive contrasts do not identify intervention effects.
Primary sources
Checked 2026-10-19: NIST, Process Modeling; Eurostat, Methodology of short-term business statistics; Hernán and Robins, Causal Inference: What If.