# Data qualityData qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → scorecards for government datasets
A caseworker in a state Medicaid office pulls a report: 40,000 households flagged as "over income" for benefit renewal. Before anyone sends a termination notice, someone should ask: how many of those 40,000 rows have stale addresses, missing income fields, or duplicate case IDs from a system migration three years ago? In 2023, several US states purged Medicaid rolls using automated eligibility checks after pandemic-era protections ended, and investigative reporting (see KFF's tracking of Medicaid unwinding) found many terminations were due to paperwork and data errors, not actual ineligibility. That is a data qualitydata quality failure with real human consequences.
This lesson builds a practical tool for that exact moment: a scorecard that tells you whether a government dataset is trustworthy enough to act on.
In government, data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → failures don't just cost money. They can deny someone a benefit, misdirect emergency resources, or produce a policy built on a fiction.
Data qualityData qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → refers to whether a dataset is fit for its intended use. It is usually assessed across four to six dimensions. We'll focus on the four most actionable for public sector analysts:
These mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → closely to the framework used by the US Government Accountability Office (GAO), the federal audit agency, in its data reliability assessments for program audits.
Imagine a state's Supplemental Nutrition Assistance Program (SNAP, the US federal food assistance program) eligibility file, 100,000 rows, one row per household, used to decide who gets renewed automatically.
Percentage of required fields that are non-null across critical variables (income, household size, address, last verification date).
Worked example:
Completeness score = average of critical fields = (96 + 99.5 + 88) / 3 = 94.5%
A common public sector benchmark: fields used directly in eligibility determination should exceed 98% completeness before automated decisions are made on them. Below that, flag for manual review.
Harder to measure directly; usually estimated via a sample audit against a source of truth (pay stubs, employer records, a cross-match with wage databases).
Worked example: Auditors pull a random sample of 500 records and verify income against the state's wage reporting system.
For high-stakes eligibility data, many audit standards (echoing GAO guidance) treat anything below 95% accuracy as insufficiently reliable for automated adverse actions (denials, terminations) without human review.
How current is the data relative to when it's used?
Worked example: Households are supposed to be re-verified every 12 months.
Do values agree across systems or within the file? Common check: does household size in the eligibility system match household size in the linked tax or wage system?
Worked example:
| Dimension | Score | Threshold for automated decisions | Pass/Fail |
|---|---|---|---|
| Completeness | 94.5% | 98% | Fail |
| Accuracy | 93% | 95% | Fail |
| Timeliness | 82% | 90% | Fail |
| Consistency | 91% | 95% | Fail |
Overall verdict: This dataset fails three of four dimensions against reasonable thresholds for high-stakes automated action. It may still be usable for aggregate policy analysis (estimating total caseload trends) but should not drive individual termination decisions without a human-in-the-loop review layer.
This is the core judgment call of this lesson: the same dataset can be "good enough" for one use and dangerous for another. A scorecard doesn't give you a single verdict, it gives you a use-case-specific verdict.
import pandas as pd
def completeness(df, cols):
return df[cols].notna().mean().mean() * 100
def timeliness(df, date_col, months_allowed=12, as_of=pd.Timestamp("2026-01-01")):
cutoff = as_of - pd.DateOffset(months=months_allowed)
return (df[date_col] >= cutoff).mean() * 100
critical_fields = ["income", "household_size", "last_verified"]
score_completeness = completeness(df, critical_fields)
score_timeliness = timeliness(df, "last_verified")
print(f"Completeness: {score_completeness:.1f}%")
print(f"Timeliness: {score_timeliness:.1f}%")Run this monthly, plot the trend, and you catch degradation before it becomes a scandal.
Knowledge check
1. In the Medicaid unwinding example, why is this framed as a data quality failure rather than simply a policy outcome?
2. A caseworker finds a household's income field is populated but reflects a job the person left two years ago. Which data quality dimension is most directly violated?
3. Why does the lesson frame data quality as a governance issue rather than purely an IT issue?
4. Select ALL correct answers about the four core data quality dimensions described in the lesson.
Select all the correct answers.
5. Select ALL correct answers about why an analyst should build a data quality scorecard before acting on a government dataset like a benefits eligibility file.
Select all the correct answers.
Thresholds (98% completeness, 95% accuracy, etc.) aren't universal laws, they're policy choices, and someone has to own them.
In the US, agencies increasingly formalize this through data governance boards and are shaped by the Foundations for Evidence-Based Policymaking Act (2018), which requires federal agencies to appoint Chief Data Officers and build data inventories and quality plans. The OMB's Federal Data Strategy sets expectations for data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → practices across agencies.
In the EU, the Data Governance Act (2022) and national open data laws push public bodies toward documented data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → standards, particularly for datasets shared for reuse. Eurostat, the EU's statistical office, publishes its own quality assurance framework with dimensions nearly identical to the four above, plus "coherence" and "accessibility."
The practical lesson: before you trust a scorecard, ask who set the pass/fail thresholds, and whether they were calibrated to the decision at hand (a benefits termination) versus a lower-stakes one (a public dashboard).
A dataset can pass every technical threshold and still mislead you if it's systematically biased, for example, if outreach and re-verification happen more often in urban counties with better broadband, rural households will show artificially poor timeliness scores that reflect infrastructure gaps, not household behavior. Always pair the scorecard with a subgroup breakdown (by region, language, age) before using it for policy.
🎬 [VIDEO: "Data QualityData QualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → Dimensions Explained" - https://www.youtube.com/results?search_query=data+quality+dimensions+explained - search results for accessible explainer videos covering completeness, accuracy, timeliness, and consistency frameworks used across public and private sectors]