Data Investigation

Data Investigation

Use this skill to produce investigations that are fast, correct, reproducible, and communicate a clear conclusion rather than a pile of charts.

Purpose

Every investigation should be answerable in one sentence before the first SQL query is written.

Phase 1: Frame Before Querying

1. Write the one-sentence answer first

Before writing any SQL, write the sentence the conclusion is expected to be.

Example: The cohort size gap is a definition problem rather than a product behavior problem.

If the sentence cannot be written, the question is not yet understood.

2. Classify the investigation type

3. State 2-3 competing hypotheses before querying

Never investigate with a single hypothesis. That creates confirmation bias.

Order hypotheses by plausibility and note which one is currently expected to be correct and why.

Phase 2: Build Queries In Escalating Specificity

1. Establish first, explain second

Step 1 always confirms the anomaly is real and measures its magnitude. Do not jump to cause until the effect is confirmed.

select
    <time_bucket>,
    <source_a_count> as metric_a,
    <source_b_count> as metric_b,
    <source_a_count> - <source_b_count> as gap,
    round(100.0 * (<source_a_count> - <source_b_count>) / nullif(<source_a_count>, 0), 1) as pct_gap
from ...
order by 1

2. Localize by breaking one dimension at a time

After confirming the gap, break it down along one dimension per step:

• By time: when did it appear?
• By segment: who is affected?
• By signal/source: which path is missing?

3. Align timing with upstream changes

Once the anomaly start date is known, check:

• git commits to relevant models, ETL jobs, or application code
• schema migrations or source changes
• metric definition changes
• external events such as launches, campaigns, or pricing changes

A credible root cause must explain why the metric changed when it did.

4. Validate the mechanism

After identifying a candidate cause, confirm it mechanically:

• Does the affected population match the predicted population?
• What does the counterfactual look like?
• Does a second independent signal support the explanation?

Do not stop at correlation.

5. Quantify each hypothesis before concluding

For every candidate explanation, produce a number.

Examples:

• H1 accounts for 1,240 users
• H2 accounts for 1,050 users

If a hypothesis cannot be quantified, it is not yet validated.

Phase 3: SQL Hygiene Rules

Use stable time bounds for investigations

Investigation SQL should be reproducible. Favor fixed bounds over rolling windows unless the analysis is intentionally operational and evergreen.

-- Prefer fixed investigation scope
where created_at >= '2026-02-16'
  and created_at < '2026-04-04'

Add explicit grain comments in important CTEs

-- grain: one row per user
with users as (
    ...
)

Prefer named comparison outputs over raw aggregates

Bad:

select count(*) from ...

Better:

select
    count(*) as users_in_scope,
    count(distinct org_id) as orgs_in_scope
from ...

Keep diagnostic queries small and discardable

Diagnostic SQL exists to answer one question. Do not build giant reusable queries too early.

Phase 4: Communication Rules

Lead with the answer

The first sentence of the written output should state the conclusion.

Bad:

I looked at several possible explanations for the discrepancy.

Good:

The discrepancy is caused by internal users being excluded from the dashboard query but included in the warehouse baseline.

Separate evidence from interpretation

Use a structure like:

1. Conclusion
2. Evidence
3. Why alternative hypotheses were rejected
4. Recommended next action

Include the limiting assumption

Every investigation should state the biggest assumption that could change the answer.

Example:

This conclusion assumes backend event timestamps are complete for the affected week; if ingestion was delayed, the gap may be overstated.

Phase 5: Output Standard

Every completed investigation should leave behind:

1. One-sentence answer
2. Final supporting SQL or notebook
3. A note on rejected hypotheses
4. Any follow-up question or unresolved ambiguity

Common Failure Modes

• Treating a metric definition dispute as a product issue
• Comparing sources with different grain or freshness
• Accepting the first plausible explanation without quantification
• Writing one giant query too early
• Ending with charts instead of an answer

Related Skills

• Use answering-natural-language-questions-with-dbt when the goal is answering a business question rather than debugging why analysis or metrics disagree.