Data correlation

A framework for correlating two data sources honestly — the playbook's research-agent example treats this as its own discipline, distinct from gathering the data in the first place.

Workflow

1. Establish an alignment key first. Before comparing anything, identify what the two datasets share — a time period, an entity (team, customer, region), a matching ID. Without a real alignment key, any "correlation" is coincidence.
2. Check the timing direction. Does the proposed cause precede the proposed effect? A metric that improved before the change you're crediting it to is not evidence for that change.
3. Look for confounders before concluding. Ask explicitly: what else changed at the same time that could explain this? Seasonal effects, unrelated initiatives, and small sample sizes are the usual suspects. Name at least one plausible confounder even for a strong-looking correlation.
4. Check the sample size and variance. A trend across 2 data points, or 2 teams out of 20, is not a trend — call it out as suggestive at most.
5. State correlation, not causation, unless you have a mechanism. "X and Y moved together" is a correlation claim. "X caused Y" requires a plausible causal mechanism, not just co-movement — say which one you're making.
6. Quantify when you can. "Deploys per week nearly doubled" is stronger and more checkable than "deploys increased."

Composition

Use after research-methodology has gathered the two data sources being compared. Findings from this skill feed naturally into business-intelligence for framing the "so what."

Quality bar

• An explicit alignment key (time, entity, or ID) is named, not assumed.
• At least one confounder is named and addressed, even if dismissed.
• The correlation-vs-causation distinction is explicit in the conclusion, not implied.
• Sample size or variance caveats are stated when the underlying N is small.