Citizen Data Scientist Safe-Use Guidelines

Purpose and scope

This guide helps non-specialist analysts — product owners, operations leads, clinicians, shop foremen, marketers, educators, and others — run fast, useful analyses and experiments without creating privacy, quality, or operational risk. It explains what you can do yourself, what needs review or approval, the basic checks that stop common mistakes, and exactly when to involve a data professional.

Who should use this

If you need to answer a decision-focused question using data available to your team (for example: "Which service region has rising defect rates?" or "Does the new email subject line change conversion?") this guide is for you. It is not a substitute for qualified data scientists when work involves high risk, complex modeling, regulated data, or production automation.

Quick decision flow (use this first)

1. Clarify the question and expected decision or action.
2. Identify data sources and check whether they are approved (see Allowed Data Sources).
3. Use approved tools and follow simple statistical checks before drawing conclusions.
4. If any of the escalation triggers apply, pause and follow the escalation path.
5. Document your work using the minimal reproducibility checklist below and save artifacts where the team keeps reproducible analyses.

1) Allowed data sources and steward approval

Always prefer approved, production-ready sources. Approved sources are curated, documented, and managed with access controls. Examples:

• Enterprise reporting databases (approved schemas)
• Central data warehouse tables labeled as 'analytic' or 'trusted'
• Designated CSV extracts or APIs provided by IT/data teams

Ask for steward approval before using:

• Ad-hoc extracts from transactional systems
• Customer or employee personal data, PII, PHI, or any data classified as sensitive
• External third-party datasets whose provenance or licensing is unclear
• Data that could affect regulatory reporting, billing, safety, or compliance

Approval is typically a light steward review confirming source suitability, lineage, and any necessary access controls.

2) Acceptable tooling and reporting platforms

Use only approved tools that meet security and versioning requirements. Typical approved tiers:

• Low-risk: spreadsheet software (approved templates), BI dashboards (approved workspace)
• Medium-risk: managed notebooks (read-only access to approved datasets), versioned CSV exports stored in approved locations
• High-risk or production: any tool or workflow that automatically updates production systems, triggers actions, or uses automation must be designed and deployed with data-team collaboration

Bring a new tool into the approved list only after a security and governance review; contact the platform/data steward for the intake process.

3) Simple statistical and data-hygiene checks (run these every time)

These checks are designed to catch the majority of routine errors before you act on a result.

1. Sanity checks: Are the units, date ranges, and cohort filters what you expect? Check min/max, totals, and a few raw rows.
2. Completeness: What fraction of rows/records are missing key fields? If >5–10% missing on a critical field, investigate.
3. Duplicates: Look for unexpected duplicate keys or repeated events.
4. Outliers: Flag values >3 standard deviations from the mean or obviously impossible values (negative counts, future timestamps).
5. Aggregation sanity: Recompute a small aggregated total by hand or with a separate query to confirm summation logic.
6. Small-sample caution: Avoid firm conclusions when sample sizes are small. Prefer confidence intervals or simple bootstrap checks over binary claims.
7. Compare to baseline: Check recent historical values and known control groups — does the result fit expectations or contradict known facts?

4) When to involve an expert or resident data scientist

Pause and escalate when any of the following apply:

• Work uses sensitive personal data (PII, PHI) or regulated information
• You plan to automate decisions, trigger downstream systems, or embed results into production
• Analysis requires causal claims, advanced modeling, or forecasting beyond simple trend descriptions
• The result will affect financial reporting, billing, patient safety, legal/regulatory compliance, or major operational change
• You cannot reproduce a surprising or high-impact result using the documented steps

When in doubt, ask a resident data scientist for a quick triage — early review can save rework and risk.

5) Governance & audit expectations

Every citizen analysis that influences decisions should include minimal reproducibility and governance artifacts. Save them in the team’s approved repository.

Required metadata and artifacts:

• Analysis objective and decision/use-case (one paragraph)
• Data sources used (table name, steward, extract time, snapshot ID if available)
• Tooling and software version (e.g., BI tool, notebook kernel version)
• Query or transformation scripts (saved and versioned)
• Key assumptions and known limitations
• Reviewer name and date (if review gate applied)
• Result artifacts: charts, CSV outputs, and a small reproducible example that rebuilds the headline number

Retention: keep these artifacts for the period defined by your data steward (commonly 6–24 months depending on sensitivity and business needs).

6) Quick-reference cheat sheet: common pitfalls & hygiene

• Always check timezones: misaligned timestamps create phantom trends.
• Beware backfill: infrastructure jobs may change counts historically — confirm using snapshot IDs.
• Label cohorts clearly: don’t mix "users" and "accounts" without noting the difference.
• Watch for implicit filters in dashboards (date picker, segments) before exporting.
• Prefer counts and rates over raw percentages when denominators change.
• Annotate tests and experiments with start/end dates and assignment method.
• Document the reproducible query — a screenshot is not sufficient.

Templates & minimal reproducibility example

Use the provided one-page template (Objective, Data sources, Steps to reproduce, Key outputs, Assumptions & limitations) and attach your query or notebook. A short reproducible example often consists of:

1. A saved query that selects the minimal columns needed
2. A small, dated CSV or snapshot stored in the approved repository
3. A short script/notebook that produces the headline chart or number from that CSV

Escalation & review path

Follow this path when your analysis meets any escalation triggers:

1. Notify your immediate data steward or analytics guardian with the template artifacts.
2. Request a light review (goal: reproducibility and risk check). Typical SLAs: 24–72 hours for non-urgent items; faster for high-impact items.
3. If the review finds issues or the work must be productionized, schedule a handoff to the analytics team with agreed acceptance criteria.

Example: lightweight submission template

Title: [one-line objective]

• Decision needed: [what decision will this analysis inform?]
• Data sources: [list tables, extracts, steward names]
• Tools used: [BI tool, notebook, spreadsheet]
• Key checks performed: [sanity, missingness, duplicates, baseline compare]
• Recommended next step: [act, run A/B test, escalate for modeling]

Next steps and where to get help

Keep this guide handy. For steward approval, new tool requests, or a quick review, contact your team’s data steward or the resident analytics hub. If your organization has a citizen-analytics toolkit or template collection, consider adopting it as your working baseline.

Closing note

Citizen analysts accelerate decision-making across the organization. These guidelines are designed to let you move quickly while keeping the enterprise safe and learnable. When you follow the checks and documentation above, your work can be trustworthy, reusable, and valuable to others.