Payer provider attribution data modeling is the practice of representing which patients a provider organization is accountable for, under what contract, for which performance period, and through which assignment method. The modeling problem is not primarily algorithmic. It is a question of record ownership, temporal boundaries, and the distinction between payer-supplied assignment and locally computed attribution. The HL7 Da Vinci Member Attribution List guide treats the payer as producer and the provider as consumer, with a confirm operation that finalizes a list [1]. The Tuva Project, by contrast, computes attribution from claims, eligibility, and provider data without requiring external payer assignment values [2]. These are not competing answers to the same question. They are different layers of a data architecture, and a coherent model must keep them separate.

Record Ownership and the Producer-Consumer Boundary

The Da Vinci ATR implementation guide establishes a clear ownership model. The payer, described as the producer, publishes a Member Attribution List that the provider organization, the consumer, retrieves or requests [1]. The guide specifies that the combination of Member Identifier, Payer Identifier, Contract Identifier, and Plan Identifier is always unique [1]. This uniqueness constraint is a modeling requirement, not a suggestion: any table that stores attribution must preserve that four-part key or risk merging members across contracts and plans.

The same guide describes a negotiation workflow in which the producer may change the list periodically and the consumer may request additions or removals. The consumer invokes a confirm-attribution-list operation to indicate that no further changes are needed, and the parties agree upon a finalized list for a specific use case [1]. This means the data model needs a status or lifecycle field that distinguishes a draft list under negotiation from a confirmed list. A model that stores only the final assignment loses the audit trail of what changed during reconciliation.

Inputs and the Distinction Between Supplied and Computed Attribution

Tuva's provider attribution module documents its inputs explicitly: normalized medical claims from claims preprocessing, eligibility exploded to member months, provider data used to classify rendering NPIs, and a package-owned month spine covering January 1900 through December 2100 [2]. The module also accepts externally supplied payer and custom attribution fields from input_layer.provider_attribution, but keeps those fields separate and unchanged from Tuva-assigned attribution [5]. This separation is a useful modeling pattern: one table can hold both the payer's authoritative assignment and a locally computed alternative without conflating their provenance.

The Da Vinci guide lists the data contained within a Member Attribution List: patient demographics, attributed provider data including NPI and TIN, health plan data such as subscriber and member identifiers, attribution data including effective dates and method, and miscellaneous data such as ACO information and conditions [1]. The inclusion of attribution method as a field is significant. It means the model must record not only who is attributed, but how the assignment was produced, whether by algorithm, manual process, or patient declaration [1].

Modeling the Attribution Method as a Controlled Field

Store attribution method as a controlled vocabulary rather than free text. The Da Vinci guide lists algorithmic, manual, and patient-declared methods [1]. Tuva's five-pass methodology adds a step dimension: 12-month PCP/NPP primary-care HCPCS, 12-month specialist primary-care HCPCS, 24-month PCP/NPP primary-care HCPCS, 24-month primary-care HCPCS with any provider classification, and 24-month any rendering NPI [2]. A combined model could use one field for the source category and another for the specific pass or rule, preserving provenance without forcing payer and computed methods into a single enum.

Temporal Boundaries and Versioning

Attribution data changes over time, and the sources describe different refresh cadences. The Da Vinci guide notes that lists may be exchanged on a monthly basis and that the producer may change the list periodically based on additional data [1]. The Oregon Health Leadership Council similarly advises provider entities to work with payers to access and maintain an accurate, up-to-date list of attributed enrollees on a monthly basis [3]. A monthly refresh implies that the model needs effective start and end dates for each attribution record, not a single current snapshot.

Tuva's current output uses a data-driven as_of_date to define rolling 12- and 24-month windows. The default is the maximum claim_end_date in the attribution claim set, if it is not null and not a future date; otherwise the system date at runtime. A dbt variable can pin the as_of_date to a specific YYYY-MM-DD value [2]. The yearly output uses calendar-year windows for the performance year, with expanded windows spanning January of the prior year through December of the performance year as a fallback [2]. These are different versioning strategies: one is a rolling window anchored to a configurable date, the other is a fixed performance-year boundary.

Prospective, Retrospective, and Hybrid Assignment Semantics

The National Association of Community Health Centers action brief distinguishes three primary approaches. Prospective attribution determines patient assignments for the upcoming performance year based on claims data from a defined look-back period. Retrospective attribution determines assignments based on care and services provided in the completed performance period. Hybrid or concurrent attribution uses historical care with continuous adjustments based on care delivery patterns [4]. Each approach has different implications for when the assignment becomes fixed and whether the provider knows the accountable population in advance.

The Oregon Health Leadership Council draws a related distinction between attribution and assignment. Attribution is the statistical or administrative methodology of assigning members to providers for calculating health care costs and quality measures, often a retrospective process using claims or encounter data. Assignment is a prospective process in which a payer matches a member with a primary care provider based on criteria such as zip code, availability, or age, and some payers encourage member selection of a PCP before assignment [3]. A data model that treats these as the same field will obscure whether a roster entry reflects a member's choice, a payer's administrative match, or a claims-based calculation.

A Validation Scenario for a Combined Model

A useful hypothetical validation scenario would load a payer-supplied Member Attribution List into a table keyed by member identifier, payer identifier, contract identifier, and plan identifier, with effective start and end dates and an attribution method field. Separately, a claims-based attribution process would compute assignments using a five-pass methodology and store the assigned step, lookback window, and provider identifier. A reconciliation query would then compare the two tables for the same performance period and flag members present in one but not the other.

This scenario is a suggestion for testing model completeness, not a claim about any deployed system. The comparison would expose whether the model can represent a member who is payer-assigned but has no claims history, a member with claims history but no payer assignment, and a member whose payer assignment changed mid-period. The Da Vinci guide's confirm operation and Tuva's labeled fallback rows for persons without assignable history both suggest that these edge cases are expected, not exceptional [1][5].