Housing instability data quality: eviction records and health data integration trust

Roughly 3.6 million eviction cases are filed in the United States each year. Each one generates a court record. Almost none of those records conform to a standard that health systems, Medicaid managed care organizations, or population health AI models can reliably consume.

This is the core problem with housing instability data quality. The data exists. It is scattered across thousands of county court systems, formatted inconsistently, and disconnected from every clinical data source that could give it meaning. When organizations attempt SDOH housing integration, they discover that the hardest part is not acquiring eviction data. It is trusting it enough to act on it.

What housing instability actually means in data terms

Housing instability and housing insecurity are used interchangeably in policy discussions, but they represent different data constructs. Housing insecurity is a broad category that includes cost burden, overcrowding, frequent moves, and living in substandard conditions. Housing instability is a subset that focuses on the risk of losing one's current housing, including eviction filings, lease nonrenewals, and homelessness.

For AI models and clinical decision support systems, this distinction matters. A patient who spends 55% of income on rent is housing insecure. A patient who has an active eviction filing is housing unstable. These two states produce different health risks, different intervention windows, and different data signals.

The Healthy People 2030 framework from HHS defines housing instability through five measurable indicators: high housing costs relative to income, poor housing quality, neighborhood instability, overcrowding, and homelessness. Each indicator requires a different data source, and none of them live natively inside an EHR.

Why eviction records are uniquely difficult to integrate

Eviction records are court records. They are generated by over 3,100 county court systems, each with its own filing conventions, case management software, and public access rules. Some counties digitize records and make them available through online portals. Others require physical courthouse visits to retrieve paper files.

The PMC study that currently ranks first on this topic evaluated approaches to linking eviction data with Medicaid claims and homelessness service records. The researchers found that deterministic matching on name and date of birth produced match rates as low as 30% in some jurisdictions. Probabilistic matching improved results but introduced false positive rates that degraded the reliability of downstream analyses.

Several structural problems make eviction data integration uniquely fragile:

No universal identifier. Eviction records use names and addresses. They do not contain Social Security numbers, Medicaid IDs, or Medical Record Numbers. Matching a court record to a health record requires fuzzy logic on fields that change frequently for the very population being tracked.

Filing does not equal eviction. An eviction filing is not the same as an executed eviction. Approximately 40% of eviction filings result in dismissal, withdrawal, or settlement. But the filing itself creates a permanent court record that follows the tenant. Health data systems that treat filings as binary eviction events overcount the prevalence of forced displacement.

Temporal misalignment. Court records capture a legal event at a point in time. Health records capture encounters over time. An eviction filing from March may not affect housing status until June. A Medicaid claim from April may reflect care accessed while the patient still had stable housing. Without precise date alignment, the causal inference that health AI models need collapses.

Geographic fragmentation. A patient who is evicted in one county and seeks care in another county creates a data linkage gap that no single-source integration can resolve. This is compounded by the fact that eviction-driven moves often cross county and sometimes state lines.

Housing instability and health: the data that connects them

The relationship between housing instability and health outcomes is well-documented at the population level. Adults who experience eviction have higher rates of emergency department visits, mental health diagnoses, substance use disorders, and chronic disease exacerbation. Children in evicted households show elevated rates of lead exposure, developmental delays, and food insecurity.

Housing instability statistics paint a stark picture. The Eviction Lab at Princeton University estimates that prior to the COVID-19 pandemic, landlords filed approximately 3.6 million eviction cases annually. During the pandemic moratorium period, an estimated 1.36 million excess evictions were prevented. After moratoriums expired, filing rates returned to or exceeded pre-pandemic levels in many jurisdictions.

But translating population-level evidence into individual-level risk scoring requires data that is clean, linked, and current. Most health systems do not have it.

Screening instruments like the AHC-HRSN ask patients directly about housing instability. These screens capture self-reported data at the point of care. They are valuable but suffer from response bias, inconsistent administration, and the same recency problems that plague all patient-reported data. A patient screened as housing-stable in January may face an eviction filing in February. The screening data does not update itself.

The trust gap in SDOH housing integration

Eviction record integration failure points by data quality dimension

When a health plan or Medicaid MCO attempts to integrate eviction records with claims data, the result is a data product with unknown provenance, inconsistent quality, and no validation layer. This is not a technical problem that better APIs will solve. It is a trust problem.

Consider the data flow. A third-party vendor scrapes eviction records from county court websites or purchases bulk data from court administrators. The vendor normalizes names, standardizes addresses, and applies matching algorithms to link court records to health plan enrollment files. The matched data is delivered to the health plan as a flat file or API feed.

At every step in this chain, trust degrades:

Provenance is unknown. The health plan does not know when the court record was created, when it was scraped, or how many transformations it underwent before delivery.

Consent is absent. No tenant consented to having their eviction record linked to their Medicaid enrollment. The legal basis for this linkage varies by state and is actively contested in several jurisdictions.

Recency is unreliable. Court records may be scraped weekly, monthly, or quarterly. The health plan has no way to know whether a record reflects current housing status.

Quality is unvalidated. Name-matching algorithms produce both false positives and false negatives. Neither error is visible to the end user without manual verification.

This trust gap means that even when eviction data is technically integrated with health data, no one can answer the basic question: should we act on this?

Key statistics

3.6 million eviction cases filed annually in the U.S. prior to the COVID-19 pandemic, per the Princeton Eviction Lab.

~40% of eviction filings result in dismissal, withdrawal, or settlement rather than executed eviction, making filing-level data unreliable as a binary housing instability indicator.

30% deterministic match rate reported in some jurisdictions when linking eviction records to Medicaid enrollment using name and date of birth alone.

3,100+ county court systems generate eviction records independently, with no shared data standard, identifier, or reporting cadence.

95% time reduction achieved by SuperTruth's DTI Engine in standardizing 105,000 diagnostic records for imaware, demonstrating the operational impact of trust-scored data pipelines on fragmented record systems.

What housing insecurity vs instability means for AI model design

AI models that predict health outcomes from SDOH data need to distinguish between housing insecurity and housing instability because the two conditions produce different risk profiles and different intervention timelines.

A housing insecurity model might incorporate census-tract-level data on median rent burden, housing quality indices, and neighborhood change indicators. These are area-level measures that can be geocoded to a patient's address. They are relatively stable over time and available from public sources like the American Community Survey.

A housing instability model needs individual-level data: eviction filings, utility shutoff notices, shelter intake records, rapid rehousing program enrollment. These data sources are volatile, fragmented, and privacy-sensitive. They require entity resolution across systems that were never designed to interoperate.

The distinction matters for data trust scoring. Area-level housing data has strong provenance (U.S. Census Bureau), reasonable recency (updated annually or every five years), and high concordance across sources. Individual-level eviction data has weak provenance (county courts with no data governance), poor recency (scraped at unknown intervals), and low concordance (matching rates below 50% in many cases).

Any AI model that treats these two data types as equivalent will produce predictions that cannot be validated, explained, or trusted.

How eviction records interact with homelessness data

Eviction is the most common precursor to homelessness, but the data systems that track each condition are entirely separate. Eviction records live in court systems. Homelessness data lives in Homeless Management Information Systems (HMIS), which are administered by local Continuums of Care.

Linking eviction records to HMIS data could identify individuals on the pathway from housing instability to homelessness, creating an intervention window. But HMIS data has its own quality problems: inconsistent data entry, variable system implementations across communities, and coverage gaps for people who do not access shelter services.

The integration challenge is multiplicative. Each data source has independent quality, recency, and provenance problems. When combined without a trust framework, the errors compound rather than cancel.

What trust-scored housing data looks like

DTI trust dimensions applied to eviction record scoring

A trust framework for housing instability data would score each record across dimensions that matter for clinical and operational use. Using the Data Trust Index framework, an eviction record would be evaluated on:

Provenance (25% weight): Can the record be traced to a specific court filing, with a docket number, filing date, and court jurisdiction? Or was it scraped from a secondary aggregator with no chain of custody?

Consent (20% weight): Is there a legal basis for linking this record to health data? Has the patient been informed? Does the linkage comply with state-specific tenant protection laws?

Recency (15% weight): When was the record last verified? Does it reflect current housing status or a historical event? Is the scraping cadence documented?

Quality (10% weight): Was the name match deterministic or probabilistic? What was the confidence score? Were address fields standardized before matching?

Concordance (10% weight): Does the eviction record align with other available data, such as address changes in claims data, shelter intake records, or self-reported screening responses?

Without scoring across these dimensions, an eviction record integrated into a health data system is just a liability. It looks like information, but it cannot support a clinical decision, a risk adjustment calculation, or an AI prediction.

The consent problem specific to eviction data

Eviction records are public court records in most jurisdictions. This creates a false assumption that they can be freely used for any purpose, including health data integration. The legal reality is more complicated.

Several states have enacted or are considering tenant protection laws that restrict the use of eviction records in housing decisions. Some jurisdictions seal eviction records after a period of time or when cases are dismissed. The patchwork of state laws means that a national eviction data product may include records that are legally restricted in some jurisdictions but not others.

From a health data trust perspective, the question is not whether the data is legally accessible. The question is whether the patient would reasonably expect their court record to be linked to their medical history. For populations already experiencing housing instability, unexpected data linkages can erode trust in health systems and reduce willingness to seek care or participate in screenings.

This is why consent governance cannot be an afterthought in SDOH housing integration. It must be embedded in the data pipeline from the point of ingestion.

What health systems and Medicaid plans should do now

Organizations working to integrate housing instability data into clinical and operational workflows need to address the trust problem before scaling the technical integration. Three steps matter most:

1. Score before you integrate. Every eviction record entering a health data system should receive a trust score that reflects its provenance, recency, match confidence, and consent status. Records below a defined trust threshold should be flagged, not suppressed, so that analysts understand data limitations.

2. Distinguish filing from outcome. Data pipelines must capture eviction case disposition, not just filing status. A dismissed filing and an executed eviction are fundamentally different events with different health implications. Treating them identically inflates instability prevalence and degrades model accuracy.

3. Document the chain of custody. From court system to vendor to health plan to AI model, every transformation, match, and inference applied to an eviction record should be logged. When a regulator, auditor, or patient asks how their housing data was used, the answer needs to be specific and complete.

The organizations that build trust into their housing data pipelines now will be the ones whose risk models, care coordination programs, and value-based contracts survive regulatory scrutiny in 2026 and beyond.

The DTI Engine scores every health data record 0 to 100 across 8 trust dimensions before your AI model sees it. For organizations integrating housing instability data, eviction records, or any SDOH data source into clinical workflows, this is the layer that turns fragmented court records into something a model can actually trust. If your team is evaluating SDOH data for Medicaid risk adjustment, population health AI, or value-based care programs, contact Louis Simeonidis at louis@supertruth.ai or (215) 918-4140.

Further reading:

DTI™ Engine

Health systems solution

Address-level SDOH data trust: geocoding accuracy and the census tract mismatch problem

Community health organizations and SDOH data quality: the trust gap

Social risk factor screening data trust: AHC-HRSN instrument quality requirements