(Header: President Barack Obama speaks at an Affordable Care Act event in Dallas TX. The prominence of “narrow network” plans on ACA state marketplaces have translated into considerable savings at the risk of decreased access to care for some patients. Photo Credit: Obama White House Archives)

How well does the
American healthcare system work? There are many components to this pertinent
question of healthcare “efficacy” — quality of care, insurance coverage, costs,
and access to providers. In the United States, while federal law ensures a minimal
standard for all health insurance plans, it is
largely up to states to regulate insurance
plans along these dimensions. States like California strictly dictate that all
in-state plans provide users with a minimal level of geographic access to
primary care, while regulations in other states remain less defined.

To compete in state-level health insurance markets (implemented by the Affordable Care Act), some insurance providers offer what have become known as “narrow networks” — plans that drive down insurance premiums by providing fewer provider options. While potentially cost-saving, narrow networks run the risk of stranding many users with inconvenient or infeasible geographic access to basic health services. Clunky, opaque healthcare evaluation software has stalled most efforts to measure whether, and in which localities, narrow networks are bottlenecking or even cutting off physical access to healthcare services for users.

That’s where Bayes Impact steps in. Bayes Impact, a technology nonprofit with a mission to build the data-driven social services of the future, is using network algorithms and data visualization to tackle this problem of narrow networks. Their end goal is threefold: 1) provide policy-makers with a tool to evaluate insurance networks; 2) enable insurance companies to create healthcare plans with adequate provider networks; and, 3) empower the public to make informed decisions about insurance plans and providers.

Bayes Impact’s Executive Director Mehdi Jamei presenting at MIT.

One of Bayes Impact’s most promising tools in development is the Network Adequacy Explorer, a dashboard that measures whether primary care providers across different insurance plan networks are within accessible proximity to users in accordance with state law. Although they expect to roll this tool out in all states, Bayes Impact deployed an initial pilot in California after initial work with the Department of Managed Health Care.

How do they know where healthcare users are? U.S. Postal Service to the rescue: using anonymized public mailing addresses, they generate more than 100,000 geo-coded points to represent users. Additionally, they acquire than 70,000 provider locations belonging to over 100 networks. According to California law, an insurance plan with an “adequate” network must include a primary care physician within 30 minutes or 15 miles of each covered person’s residence or workplace. Using a number of fast distance-calculating algorithms, the tool finds that in Southern California, some 2% of the points (or roughly 20,000 addresses) cannot access Health Service Delivery Centers (HSDC) within 15 miles or 30 minutes. With a municipal population of 43 million, 2% overall corresponds to nearly a million residents without adequate geographic access to care. [NOTE: HSDCs are federally funded facilities and thus not directly mandated to meet the California minimum distance rule which is applicable only to to commercial and managed care insurance plans. However, this failure to meet the minimum distance rule poses a problem for those in Southern California who depend on HSDCs as their primary source of care: Medicaid beneficiaries and the uninsured.]


A demo of the Network Adequacy Explorer. The tool displays a municipality in Southern California, where nearly a million potential residents may be without adequate geographic access to a health service delivery center based on California’s “minimum distance rule” of service within 15 miles / 30 minutes of any user.

Below are three key takeaways from Mehdi’s talk on Bayes Impact’s work.

Vertical and Horizontal: Two Different Kinds of Research Scalability

Seminar participants identified two directions of expansion for the Network Adequacy Explorer tool: layering additional state-level healthcare access indicators on top of the existing network adequacy data and adapting the tool to consider legal constraints in other states. The first direction is undoubtedly crucial: clinic-level data points and non-geographic metrics are needed to capture access to increasingly popular community health clinics and telehealth platforms in provider networks. The second line of expansion is needed to usefully capture insurance efficacy in other states, such as in Texas, where abortion access laws might prompt certain healthcare users to seek out-of-state care, which is not currently accounted for in the tool.

Borrowing from database terminology, one might consider the first type of expansion as “vertical scalability”, increasing the resolution on existing measurements, and the latter type as “horizontal scalability”, or adapting measurement tools to other contexts. In scaling or expanding any sort of social science research, it is important to consider which “direction” — vertical (resolution) or horizontal (context) — to prioritize.

Public and Private Data Science

Releasing tools like the Network Adequacy Explorer to the public may be the best way for other stakeholders (citizens, data scientists, advocacy groups, lawmakers) to continue the work of “solving” the “social problems” headlined in our series. While this can certainly close the feedback loop, open source tools also run the risk of exacerbating the very problems they set out to fix. For example, in the case of partisanship gerrymandering in U.S. legislative districts, computational redistricting tools and simulators may be increasing partisan gerrymandering rather than neutralizing it. However, Jamei suggested that private insurance companies may already have their own troves of data enabling them to back out of high-risk provider areas, and the current system already demonstrates worsening health care access disparities along racial lines. Open-sourcing this same demographic-augmented data can allow the public to mobilize it for social good and level the playing field.

Making Generalizable Tools in Data Science

Finally, one of the most exciting things about Bayes Impact’s work is its applicability in evaluating citizen access to other types of geographically pertinent services such as schools, correctional facilities, and public service agencies. As such, this tool could easily find a place in political science research. Such is the power of analytical methodologies that are developed for a specific domain problem but are flexible, integrated with well-designed interfaces, and thus easily re-usable in other problem areas. Data scientists, take note.

Bayes Impact plans to further develop other tools in healthcare policy evaluation, some that measure other dimensions of “efficacy” such as price transparency and patient quality of care. We’ll be watching their work closely to see how state governments respond and whether they take up Bayes Impact’s open source tools to assess and address critical health care access gaps.