Assessing Re-identification Risk from Household and Family Composition in HIPAA Expert Determination
Introduction
Health datasets that contain an ID linking individuals by household or family unit carry a privacy risk that is easy to underestimate. A variable that identifies which patients belong to the same household/family, whether derived from a shared insurance subscriber ID, a shared address, or another linkage mechanism, may seem innocuous in isolation. It carries no diagnosis, no date, no location. But in practice, its effect on re-identification risk can be large — even exponential.
When a dataset groups patients by household, re-identification risk is influenced not only by each person’s indirect identifiers, but also by the combined demographic profile of the household. Indirect identifiers, such as a patient’s age, sex, location, and race may not identify someone on their own but can make records more distinguishable when considered together. Because the combined characteristics of multiple household members may correspond to a much smaller population, individuals in multi-adult households can face substantially higher re-identification risk. In some cases, that risk may exceed the “very small” standard required for HIPAA Expert Determination1, even when each person’s individual demographics would not.
However, family and household IDs have several useful applications in de-identified health data, and retaining this utility without sacrificing privacy is critical. Family IDs are important for assessing the effectiveness, cost, and efficiency of health plans because they link members who share coverage. Likewise, household IDs can greatly enrich consumer data used to design or measure the impact of drug marketing campaigns.
In this blog, we explain the mechanism behind that risk, present empirical evidence quantifying it across real population data, and describe how to manage it without sacrificing analytical utility.
How re-identification risk is determined
For a dataset with no direct identifiers, such as full names and addresses, the primary risk of re-identification is through an unauthorized linkage attack. In this scenario, records from a de-identified dataset are matched against records in a separate, external dataset using characteristics that appear in both. These shared characteristics, known as indirect identifiers, might include information such as age, location, or dates of service. External datasets, which commonly include voter registries and other sources, contain both indirect identifiers and direct identifiers, including names and addresses. If the indirect identifiers match an individual between the datasets, the direct identifiers are also matched, and that individual has been re-identified..
The probability of a correct linkage depends on how distinguishable a record is, or the extent to which an individual’s combination of indirect identifiers is unique or near-unique within the population. Aligning with HHS guidance, an expert can assess the degree to which a subject’s data can be distinguished in the health information. When a combination of indirect identifiers is common, many individuals match and the risk of correct re-identification is low. When a combination is rare, the risk is high. Quantifying this requires knowledge of how indirect identifier combinations are distributed across the real-world population. In practice, this information can be constructed from sources such as the U.S. decennial census, the American Community Survey, or state voter registries. This information allows an expert to estimate the number of individuals in the population who share a given combination of indirect identifiers, which is a standard and foundational calculation in Expert Determination risk assessments.
Common demographic indirect identifiers in health data, including age, sex, three-digit ZIP code, race/ethnicity, and year of death are well understood. The challenge with household composition is that it is not a single additional indirect identifier. It is a structural variable that expands the available indirect identifier space to include the characteristics of every other member of the household simultaneously. The following example helps demonstrate how the linkage attack based on a family unit variable increases re-identification risk.
Household variables increase the risk of re-identification by combining the attributes of multiple individuals
When a dataset contains a household linkage variable, an attacker who can identify one household member gains access to the demographic signatures of all members. Consider a household of three: a 52-year-old woman, a 55-year-old man, and a 24-year-old woman, all sharing the same three-digit ZIP code. Individually, none of these records may be rare. But together, the combination of three co-resident adults with those specific age-sex profiles in that geography is likely to match very few households in publicly available information.
This is not a theoretical edge case. It is the expected outcome as a household’s size grows because the probability that another household matches all members’ profiles decreases. A dataset that contains a household linkage variable, in effect, creates the potential for this type of linkage attack, whether or not the data recipient intends it.
The risk is further compounded by what the household linkage variable reveals about household structure, independent of any individual’s demographic values. Standard de-identification often suppresses or generalizes individual attributes (e.g., rounding ages to five-year bands or dropping ZIP code to state level) on the assumption that less precise values produce larger equivalence classes. But when a household linkage variable is present, the structure of the household itself becomes an additional indirect identifier that suppression does not erase. Knowing that a patient is in a two-adult household with a member of a different race and a significantly different age narrows the population considerably, even without precise ages or ZIP codes. An attacker does not need precise values; they need the household’s demographic signature to be rare, and household composition frequently makes it so.
Re-identification risk remains high despite imperfect mapping of household variables between health data and public data
Household definitions differ between health claims and public data, raising the question of whether that misalignment meaningfully reduces re-identification risk. In insurance claims, subscriber-based household groupings link individuals who share a health plan, which may include dependents on a parent’s plan, or exclude a spouse who holds a separate employer plan. Voter registries, by contrast, construct households from shared addresses.
Voter registries are also incomplete, capturing roughly 70% of eligible voters. As a result, an attacker trying to match a household in a de-identified dataset with one in a voter registry might not find every household member, making a confident match more difficult. In general, smaller, less complete registries do produce lower risk estimates.
These limitations are real and do reduce practical attack feasibility to some degree. An attacker working from a voter registry would not always construct the same household groupings as those encoded in the health data, and some households in the health data would not be identifiable in any reasonably available external source.
However, the fact that external voter data is incomplete offers less protection than it might seem. The catch is that health insurance data tends to enroll entire families together, meaning that if one household member appears in a dataset, others from the same household are more likely to appear too. That correlation works against the assumption that households are randomly represented and it means the risk estimates from incomplete data are lower bounds, not reliable approximations.
Additionally, in our modelled attack scenarios, this generally did not bring risk to acceptable levels for households of three or more members. Even under conservative assumptions, where a meaningful proportion of households are mis-assigned or absent from external data, the remaining risk is still much greater than the “very small” threshold. Differences between household definitions can reduce risk, but they cannot be relied upon to provide sufficient protection.
Empirical evidence: Population data shows risk increases with household size
To quantify re-identification risk from household composition, Datavant Privacy Solutions researchers conducted analyses using voter registry data from four U.S. states: Georgia, Connecticut, Arkansas, and Oklahoma. Household groupings were constructed from address information in these registries. We estimated each person’s risk based on the number of individuals in the population who shared the same information; risk decreased as the number of matching individuals increased. We then estimated risk for the entire dataset using a metric based on the distribution of individual risk across the population. This was calculated both with and without household composition included as an indirect identifier.
The findings were consistent across all four states:
- Single-adult households showed minimal risk increase. This is expected because knowing that an adult lives alone adds limited discriminating information beyond their own demographics.
- Two-adult households showed a significant jump. Depending on the state, voter registries contained different combinations of indirect identifiers, including three-digit ZIP code, age, sex, and, in some states, race. Adding a household linkage variable was frequently sufficient for an individual to be considered at risk of re-identification.
- Three-adult households showed a dramatic increase. Even in states where only three-digit ZIP code and age (a relatively sparse indirect identifier set) were available as indirect identifiers, approximately 90% of individuals in three-adult households had risk above the “very small” risk threshold.
The rapid increase of risk with household size is important. The jump from one-adult to two-adult households is significant. The jump from two-adult to three-adult households is severe. Adding a fourth or fifth member compounds this drastically. Risk does not scale proportionally with household size; it scales exponentially.
Removing or aggregating patient geography helps reduce risk, but not enough for larger households. A natural question is whether simply removing precise location data solves the problem. It does not, at least not for households of three or more. Even when geographic detail is reduced to state only (no ZIP code, no county) roughly one in four individuals across all households were still estimated to be at risk of re-identification. For three-adult households specifically, that figure climbs to between 43% and 66%, depending on the state. Removing ZIP code is a meaningful step, but it has to be combined with other mitigation strategies for larger households.
Managing risk with a thoughtful remediation strategy
A patient-level risk assessment conducted as part of a Datavant Expert Determination will explore the risk of the specific dataset and consider use-case specific approaches rather than uniform, conservative rules that sacrifice utility unnecessarily. Here we outline four examples of mitigation strategies. In many cases, a combination of two strategies is sufficient to satisfy the “very small” risk standard required for HIPAA Expert Determination.
- Remove the household linkage variable entirely. This is the simplest approach and works when the analytical use case does not require household-level linkage. This eliminates the additional risk.
- Limit scope to households of one or two members. For datasets where household structure is analytically necessary, limiting the dataset to one or two-person households combined with geographic aggregation may allow the dataset to reach acceptable risk levels.
- Aggregate geography. Suppressing three-digit ZIP code and retaining only state-level geography can significantly reduce risk for smaller households. For three-adult or larger households, geographic aggregation alone is generally insufficient; it must be combined with additional mitigation.
- Technical and administrative controls. Safeguards governed by an enforceable data use agreement, including security measures, user access controls, and use-case limitations, can further reduce risk, in combination with data modifications.
For most datasets, the most practical path is a combination of household size restriction and geographic aggregation, with residual risk quantified at the individual level to confirm compliance. Retaining a household linkage variable alongside individual-level indirect identifiers without specifically analyzing and addressing household-level risk is not recommended.
Conclusion
Household and family composition variables introduce a structurally different kind of re-identification risk. Unlike a single indirect identifier that increases risk incrementally, a household linkage variable expands the attacker’s available set of indirect identifiers to include all members of the household simultaneously. Empirical analysis of voter registry data across multiple U.S. states confirms that this effect is not theoretical. For households of three or more individuals, the majority of members fall into small equivalence classes under realistic indirect identifier configurations, even when geographic granularity is limited to state level.
The appropriate response is not to automatically prohibit household linkage variables from de-identified datasets. It is to subject them to explicit, household-level risk quantification as part of Expert Determination — and to apply targeted mitigation strategies calibrated to the specific household size distribution and indirect identifier set in the data.
Contact us to connect with one of our privacy experts to help you perform HIPAA Expert Determination of health data that contains household or family linkage variables.
By: Leo Rogers, Sr. Data Scientist, Privacy Product R&D, James Janisse, Principal Data Scientist, Privacy Solutions, and David Williamson, Sr. Manager of Privacy Product R&D
1HIPAA does not prescribe a fixed threshold for “very small” risk; it is determined by a qualified expert using accepted statistical methods and contextual risk assessment. See The Role of HIPAA De-identification in Enhancing Health Data Connectivity.

