Social Determinants of Health Data: Sources and Linkage
Social determinants of health (SDOH) data describe the conditions in which people live, work, and age, such as income, education, housing, food security, transportation, and neighborhood environment, and researchers get it from three main sources: individual screening data recorded in clinical systems, area-level indices linked by address, and linked administrative records from outside health care. Linkage usually happens by geocoding a patient’s address to a small geographic unit and joining an area index, or by privacy-preserving record linkage to non-health datasets. Because addresses and small-area data are highly identifying, the safest pattern is to perform that linkage and the analysis inside a governed environment rather than exporting joined records.
Why this matters now
The World Health Organization has long argued that social and economic conditions drive a large share of health inequality, and research has repeatedly found that neighborhood deprivation predicts outcomes such as hospital readmission, disease incidence, and mortality. Kind and Buckingham, writing in the New England Journal of Medicine in 2018, made the case for routinely using area deprivation measures in health policy and research, and the Area Deprivation Index they maintain through the University of Wisconsin Neighborhood Atlas is now widely used in US studies.
Clinical data collection has also shifted. In the United States, the Centers for Medicare & Medicaid Services introduced hospital quality measures on screening for social drivers of health, and screening tools such as PRAPARE (Protocol for Responding to and Assessing Patients’ Assets, Risks, and Experiences) and the Accountable Health Communities Health-Related Social Needs screening tool are in wider use. The Gravity Project, an HL7 FHIR accelerator, is developing consensus terminology and data standards so these data can be exchanged consistently. In Europe, the European Health Data Space regulation creates a framework for secondary use of health data that will increasingly bring social and demographic variables into research requests.
The result is more SDOH data than ever, recorded in inconsistent ways, across systems that were never designed to be joined.
Where SDOH data comes from
Individual-level clinical data
Screening questionnaires capture housing instability, food insecurity, transport barriers, and interpersonal safety at the patient level. Responses may be stored as structured answers coded in Logical Observation Identifiers Names and Codes (LOINC), as diagnoses using ICD-10-CM Z codes (the range Z55 to Z65 covers problems related to education, employment, housing, economic circumstances, and social environment), or as free text in clinical notes. Studies consistently find Z codes are underused relative to the actual prevalence of social need, so relying on them alone undercounts.
Area-level indices
Area measures assign each patient the characteristics of where they live. Common examples include the Area Deprivation Index and the Social Vulnerability Index in the United States, built from American Community Survey data, and the Indices of Multiple Deprivation in England, published for small areas known as Lower Layer Super Output Areas. They are available for whole populations and need only an address, but they describe neighborhoods, not individuals, which introduces the ecological fallacy if interpreted at the person level.
Linked administrative data
Education records, social care, housing, benefits, and employment data held by other public bodies can add individual-level social information that clinical systems lack. These datasets are powerful and heavily governed, and linking them to health records typically requires legal agreements, trusted third parties, or privacy-preserving linkage methods.
Free text and derived variables
Clinical notes often hold the richest social information: a line recording that a patient lives alone, has lost a job, or cannot afford medication. Natural language processing (NLP) can extract these mentions at scale, but the models must be validated at each site, because documentation style varies between hospitals and specialties. Notes are also among the most identifying data a health system holds. Extraction should run where the notes are stored, with only the derived, coded variable (for example, “housing instability documented: yes”) carried into the research dataset. The same principle applies to any derived score: record the algorithm and version so another site can reproduce it exactly.
Comparing SDOH data sources
| Source | Granularity | Coverage | Main strength | Main limitation |
|---|---|---|---|---|
| Screening questionnaires (LOINC-coded) | Individual | Only screened patients | Direct measure of social need | Uneven screening rates bias who is captured |
| ICD-10-CM Z codes | Individual | Patients where clinicians coded need | Already in claims and EHR data | Substantial undercoding |
| Clinical notes (via natural language processing) | Individual | Patients with documented notes | Captures context missing from codes | Needs validated extraction, identifiable text |
| Area deprivation indices | Small area | Everyone with a valid address | Complete coverage, easy to link | Neighborhood, not personal, circumstances |
| Linked administrative records | Individual | Depends on linkage success | Rich, objective social data | Heavy governance, linkage error |
The Data Harmonization angle
SDOH data is a harmonization problem as much as a collection problem. The same concept, food insecurity for example, may arrive as a screening answer, a Z code, a phrase in a note, or a neighborhood score. A study that treats these as interchangeable will produce misleading results. A study that ignores all but one will lose most of the signal. For the general principles, see our guide on what data harmonization is.
The Observational Medical Outcomes Partnership (OMOP) Common Data Model gives a practical home for most of this. Screening responses and Z codes map to standard concepts in the observation and condition tables, and the location table holds geographic information that can be joined to area indices. The Observational Health Data Sciences and Informatics (OHDSI) community has a working group focused on geographic and social data, and Fast Healthcare Interoperability Resources (FHIR) profiles from the Gravity Project cover exchange. Harmonizing to these standards lets one study definition, such as “patients with documented housing instability or residence in the most deprived quintile”, run identically at every site.
The federated part matters because the joined data is so sensitive. An address plus a diagnosis plus a small-area deprivation score can re-identify people. Running the geocoding, area linkage, and analysis inside each custodian’s Trusted Research Environment (TRE) means the data never leaves the source, and only aggregate results pass an automated output check before release.
A practical linkage workflow
- Define the constructs. Decide which social domains the study needs and whether individual measures, area measures, or both are required. Write down how each will be operationalized.
- Clean and standardize addresses locally. Address quality drives linkage quality. Standardize formats and record the date of each address, since people move and deprivation indices are released for specific years.
- Geocode inside the governed environment. Convert addresses to small geographic units using a geocoder that runs locally, so addresses are not sent to external services.
- Join area indices by geography and time. Match each patient’s area and address date to the index version for that period, and keep the geographic code rather than the address for analysis.
- Link external records with privacy-preserving methods. Where individual administrative data is needed, use privacy-preserving record linkage (PPRL), which matches encoded identifiers without exposing them. Our explainer on privacy-preserving record linkage covers the main techniques.
- Map to standard concepts. Load screening answers, Z codes, and derived variables into OMOP concepts so study logic is portable across sites.
- Analyze and release aggregates only. Apply disclosure rules suited to small-area data, such as suppressing small cells and avoiding outputs at geographic levels fine enough to identify individuals.
Real-world context
Several national programs show how social and health data can be studied together under strong governance. In the United Kingdom, national statistics and research bodies have linked health records with census and administrative data inside trusted environments, following the Five Safes approach to research access. In Canada, CanPath, a Lifebit customer, combines participant questionnaires that include social and lifestyle information with health data across provincial cohorts, using federated infrastructure that keeps each cohort’s data under its own governance. In Singapore, the Ministry of Health, also a Lifebit customer, operates in a setting where national-scale health data must be analyzed under strict residency and governance rules. These are the conditions SDOH research needs: population coverage, linkage capability, and strict control of what is released.
Common pitfalls and objections
Treating area scores as personal data
An individual living in a deprived area may not be deprived. Use area indices as area-level exposures and say so in the analysis, or combine them with individual measures where available.
Ignoring who gets screened
Screening often happens in specific clinics or for specific populations. If screened patients differ from unscreened ones, missing screening data is not random, and analyses must account for it.
Linking on the wrong date
Using a current address for an event that happened years earlier misclassifies exposure. Keep address history and match indices to the period of interest.
Releasing identifying geography
Maps and tables at fine geographic levels combined with rare conditions can identify people. Agree geographic aggregation rules with the data custodian before analysis starts.
What to do next
Begin with an audit of what social data your sites already hold: screening instruments in use, Z code frequency, address completeness and history, and access to area indices or administrative linkage. That audit shows which study designs are realistic and where harmonization work is needed. For infrastructure that supports local linkage and governed analysis, see our overview of the federated Trusted Research Environment.
Frequently asked questions
What are the main categories of social determinants of health?
Commonly used frameworks group them into economic stability, education access and quality, health care access and quality, neighborhood and built environment, and social and community context. Specific measures include income, housing, food security, transportation, and social isolation.
What are ICD-10 Z codes for social determinants?
ICD-10-CM codes Z55 to Z65 record problems related to education, employment, occupational exposure, housing and economic circumstances, social environment, and other psychosocial circumstances. They are useful but underused, so they tend to undercount social need.
How is SDOH data linked to health records?
Area-level data is linked by geocoding a patient’s address to a small geographic unit and joining the index for that area and period. Individual-level external data is linked using identifiers or privacy-preserving record linkage, ideally inside a governed environment.
What is the ecological fallacy in SDOH research?
It is the error of assuming that characteristics of an area apply to every individual living there. Area deprivation scores describe neighborhoods, so conclusions about individuals require individual-level measures.
How can SDOH data be represented in OMOP?
Screening responses and Z codes map to standard concepts in the observation and condition tables, and geographic information sits in the location table, which can be joined to area indices. This lets the same study logic run across sites.
Why should SDOH linkage happen inside a Trusted Research Environment?
Addresses, small-area codes, and clinical data together are highly identifying. Performing geocoding, linkage, and analysis inside a TRE with output checking keeps identifiable data under the custodian’s control and releases only aggregate results.