Root Cause Analysis Framework
When to use formal RCA
RCA is warranted when a metric moves materially and the cause is not immediately obvious. "Material" depends on context:
- Revenue: > 3% unexpected deviation
- Engagement: > 5% week-over-week change not explained by a known product change
- Error rate: any unexpected increase
Do not invest full RCA effort in movements explained by known causes (holiday, planned outage, product launch).
The structured investigation sequence
Step 1 — Confirm the signal
Before investigating, verify the movement is real:
- Is the data pipeline fresh? Check for ETL delays.
- Is the metric definition consistent? Has a query or schema changed?
- Is it one dashboard/report or does it appear in multiple independent sources?
A false alarm wastes team time. Confirm first.
Step 2 — Scope the movement
Answer: when did it start, how large is it, which populations are affected?
- Plot a time series back 12+ weeks — is this a sudden drop or a gradual trend?
- Identify the exact start date (± 1–2 days)
- Check if it affects all users or a specific segment
Step 3 — Correlate with known events
Check your event log:
- Deployments / feature releases
- Marketing campaigns or pricing changes
- Seasonal patterns (compare to same period last year)
- External events (competitor outage, news event, platform change)
If an event aligns with the movement onset, it is a strong candidate.
Step 4 — Dimension breakdown
Use the drill-down tool to break the metric by:
- Platform / device
- Geography / region
- User segment (plan, cohort, acquisition channel)
- Product area / feature
The goal is to isolate whether the movement is concentrated in a specific slice.
Step 5 — Generate and rank hypotheses
For each hypothesis, record:
- Description of the proposed cause
- Evidence for (what data supports it)
- Evidence against (what data contradicts it)
- Test that would confirm or refute it
Rank by: (likelihood × ease of testing). Test the most likely and easiest first.
Step 6 — Validate the root cause
The root cause is confirmed when:
- The hypothesis explains the timing, magnitude, and affected population
- Removing or reversing the cause stops the movement (if possible to test)
- No alternative hypothesis explains the data as well
Step 7 — Document and close
Record the finding in the RCA report template. Include:
- Confirmed cause
- Impact quantification
- Corrective action and owner
- Preventive measure (how to catch this earlier next time)
Contribution analysis
When a metric is a weighted sum (e.g., total revenue = sum across segments), use contribution analysis to find which segments drove the change:
segment contribution = (value_b - value_a) / total_change
A segment with large absolute change AND large contribution % is the primary driver.
Also decompose into:
- Volume effect: the segment had more/fewer users
- Rate effect: the metric per user changed
- Mix effect: the segment's share of total changed
Common root causes by metric type
Revenue drop
- Pricing change (contraction/churn)
- Payment failure spike
- High-value customer churn
- Seasonal / calendar effect
Conversion rate drop
- UX bug in checkout / onboarding flow
- Traffic mix shift (lower-intent source)
- Page load regression
Engagement drop
- Feature removed or degraded
- Notification or email suppressed
- Competitive displacement
- Cohort aging effect