A Five-State All-Payer Claims Database Collaboration: Onpoint Health Data’s Multi-State Behavioral Health Initiative 

According to the latest numbers from the APCD Council, 24 states so far have established allpayer claims databases (APCDs) — centralized resources that collect and integrate health care enrollment and claims data from both public and private health insurers to provide comprehensive resource for assessing health care quality, utilization, and cost. In the absence of a national APCD, these individual state-led efforts meet important policymaking and research needs by offering insights into state spending and providing opportunities for cross-state benchmarking.

Such efforts, however, can be complicated by the varying approaches adopted by state APCDs when releasing their data. For example, states may define key health care utilization measures differently, allow different types of data to be released, and distribute data in different formats, making it challenging to conduct consistent analyses across states. Additionally, data users must navigate separate data release processes with each state, further complicating efforts to access the data needed for multi-state analyses.  

Developing a Solution to Support Cross-State Analysis & Benchmarking 

As part of our pro bono work, Onpoint Heath Data recently led a cross-state APCD collaboration focused on behavioral health: the Multi-State Behavioral Health Initiative. Five states – Connecticut, Maryland, Minnesota, Vermont, and Washington — contributed their data and agreed to use standardized measures and formats to estimate behavioral health prevalence, health care utilization, costs, and chronic condition co-morbidities. This partnership, which leveraged a common data model to harmonize data across states, enabled both state-specific reporting and cross-state benchmarking. Measures included rates stratified by calendar year, payer type (commercial, Medicaid, Medicare), age, sex, and geography.

This cross-state collaboration represented more than 17.3 million covered lives annually — the largest multi-state APCD study conducted to date. Results were disseminated through a set of interactive dashboards that show the relationships between behavioral health conditions, service utilization rates and costs, chronic condition co-morbidity rates, and social drivers of health across key geographical factors.  

Prior to this collaboration, there have been several other multi-state analyses that employed a distributed research model, including the “The New England States’ All Payer Report on Primary Care Payments” (2020) from the New England States Consortium Systems Organization, the “Total Cost of Care Multi-State Analysis” (2016)  from the Network for Regional Healthcare Improvement, and the “Dartmouth Atlas of Children’s Health Care in Northern New England” (2013) from the Dartmouth Institute for Health Policy and Clinical Practice at Dartmouth College. For each of these projects, states formed a joint workgroup to develop general specifications, individual teams then developed more detailed specifications, and analysts from each state used a single, standardized methodology to generate and report comparable summary results using their respective APCD. This is similar to the analysis approach taken by other initiatives such as Academy Health’s Medicaid Outcomes Distribution Research Network (MODRN), and the Observational Health Data Sciences and Informatics (OHDSI) project. 

While the distributed research network approach eliminates the need to share raw data, it can lead to inconsistences since it places the responsibility for accurate reporting on each entity’s individual analysts. Avoiding the distributed approach by assigning the cross-state analytics work to a single entity decreases the chance for error and increases efficiency. 

Key Takeaways for States Considering Similar Collaborations 

A common data model facilitates cross-state analyses. Conducting analyses using data from APCDs is complicated by the varying file layouts that must be accurately combined before research can begin. One way to overcome this challenge involves carefully mapping incoming data fields to a common data model (Figure 1). The result is a series of standardized tables that store various data fields – spanning enrollment, claims, providers, and more – using consistent naming and table structures. This process enhances consistency in the definitions and formats for key variables used in analyses (e.g., enrollment periods, service dates, demographics, in-state residency, diagnosis and procedure codes, bill types, place of service codes, revenue codes) across years, payer types, and states.

Figure 1. Illustration of the Data Transformation Process to Harmonize Incoming Data for MultiState Analyses 

For the Multi-State Behavioral Health Initiative, we used a common data model and its series of standardized tables to run identical statistical code for each state client. The code returned results for each state that were already harmonized, allowing for valid pooled analyses and cross-state comparisons. This process is akin to the harmonization process used for distributed research network analyses; the key difference is that we used “in-house” harmonized data, skipping the difficult step of implementing identical data transformation processes across multiple analyst teams in different states.

Getting early stakeholder alignment on data and measures definitions is key to successful data sharing and buy-in. While many states are eager to collaborate on cross-state analyses and use each other’s results for benchmarking, it can be challenging to get all parties aligned on the topic, timeline, measures, and methods. Reaching consensus on these foundational elements is critical, especially when states already report on a given topic using their own definitions. For instance, a state may define behavioral health service utilization one way, while a multi-state initiative may apply a different definition to identify such use. While methodology discrepancies can be explained, states reasonably may hesitate to release data that could create results that conflict with their own established reporting. Having agreement on the cross-state methods and measures at the outset is essential to avoiding such conflicts and ensuring consistency and credibility across participating states.

Aligning on issues around data privacy and security is critical. Researchers undertaking cross-state analyses will likely be required to secure data use agreements and permissions for each state. For this project, we formalized data use agreements (DUAs) with state partners by either modifying existing agreements or creating new DUAs specific to this project. We also adhered to existing data suppression agreements such as not displaying results based on cell counts between 1 and 10 and not identifying health insurers by name. Additionally, we maintained data separately in each state’s secure environment and used aggregated data only (with data suppression rules applied) when conducting the cross-state analytics.

Even with standardization, data still varies from state to state. Even when using a shared and consistent approach, underlying data still is likely to vary across states. For example, the extent to which commercial data for self-insured state residents are missing (due to health insurers opting out of submitting data to APCDs) can affect the comparability of estimates when all payers are combined in an analysis. Also, there often are differences in the availability and timeliness of Medicare fee-for-service (FFS) data based on states’ varying approaches to acquiring these data from the U.S. Centers for Medicare & Medicaid Services (CMS). To deal with these two issues in the Multi-State Behavioral Health Initiative, Onpoint, and partnering states decided to show only estimates by payer type (with one state electing to remove Medicare FFS estimates altogether due to missing data for the study years).

Do your homework. Before conducting a cross-state analysis, it’s also essential to assess any data gaps and nuances that may affect the comparability of findings. For instance, the Multi-State Behavioral Health Initiative focused on mental health and substance use disorder (SUD) conditions. Due to differing interpretations of federal privacy regulations by health insurers, the completeness of SUD-related records in the state APCDs varied. As part of our quality assessment, we found that while many payers submit their SUD-related data to APCDs, other payers withhold such data based on their own understanding of federal rules. To mitigate this variability, we separated mental health from SUD-related diagnoses so they could be reported accurately for each state.

Reflections

The recent Multi-State Behavioral Health Initiative harnessed the power of standardized data across states and payer types using a common data model to optimize both efficiency and accuracy. This type of collaboration both extends the distributed research network approach used for other multi-state APCD analyses and aligns with important efforts to standardize definitions for robust cross-state spending analyses using APCDs. This initiative highlights the ongoing promise offered by a collaborative approach and standardized data model for future multi-state APCD analyses.