How Profitable are Hospitals?

How Profitable are Hospitals?

How Profitable are Hospitals?

Hospital care represents the largest proportion of spending in the American healthcare system at nearly one-third of all costs, or a little over $1 trillion. To put that in perspective, that represents 5.5% of the United States’ GDP, or more than the total GDP of all but 15 countries in the world. Hospitals are generally viewed as being altruistic, with the mission to make sick people well, but with so much money being spent on hospital care, it’s a fair question to ask just how profitable they are. Fortunately, data from Torch Insight can help answer how profitable are hospitals.

What is hospital profitability?

“Finances are confusing,” said Captain Obvious (or at least he would say). Personal finances are challenging to understand for many, but they really have nothing on corporate finances. Hospital finances start with the general complexity of some really weedy corporate finances and then weave in healthcare-specific attributes. I don’t pretend to be an expert in the nuances, but a few things are relevant for this analysis. First, hospitals get money from two major sources: caring for patients and “other income” which includes donations, investments and non-care items like selling parking or gifts. Profitability, in the aggregate, is based on total income less total expenses divided by total expenses, but this can be subdivided into total margin or patient margin which is the margin based just on providing care.  It’s important to know that this is limited to the hospital’s costs and profit, and does not include a health system’s profits if it is made up of a hospital plus community settings. In some cases, hospitals will make a significant profit, but there may be community practices that lose money. Additionally, and very commonly, health systems will only pay attention to their aggregate performance where some hospitals may lose money but others make money, but for this I’m just going to focus on the individual hospitals.

Underlying Data

Hospitals that accept payments from Medicare – which represents nearly all hospitals – are required to submit annual reports on facility characteristics, utilization data, cost and charges, and, most importantly for this analysis, financial information. These cost reports are collected in the Healthcare Cost Report Information System (HCRIS).  While the data can be downloaded, it’s cumbersome to work with in its native format. Torch Insight includes cleaned versions of the HCRIS data which can easily be exported and matched to all the other data that

The most recent year with nearly complete financial data is 2016, so that’s what I’ll work with.  I’m also limiting this to the facilities that people typically think of as being “hospitals”, which is short-term acute care hospitals, children’s hospitals and the smaller critical access hospitals – this means facilities like psychiatric hospitals and rehabilitation hospitals are excluded. Finally, I did a quick look at data quality and dropped extreme cases by censoring hospitals that were below the 1st percentile or above the 99th percentile. This left me with a sample size of 2,689 hospitals.

 

Hospital Profitability

Mean Median Standard Deviation
Overall Margin 5.2% 3.8% 12.7%
Patient Margin -1.4% -1.7% 16.6%

 

The difference between the overall margin and the patient margin is pretty striking – on average, hospitals are losing money based on patient margins, but in the aggregate, accounting for non-patient care revenue, they are making a 5% margin. The averages and medians are interesting, but the variation is pretty extreme (note that I’ve dropped a few outliers to make the histograms clearer).

How Profitable are Hospitals? How Profitable are Hospitals?

You can see here just the amount of variability there is with some hospitals making a relative killing and others really struggling. From this, we may just assume that hospital margins are all over the place.

Drilling into the Details

Torch Insight has dozens of hospital-level characteristics that let me drill down into some more details. I’ll start by looking at the margin by the size of the hospital, using as a proxy the average daily census (ADC).

Hospital Profitability by Size

Average Daily Census Mean Overall Margin Mean Patient Margin Median Overall Margin Median Patient Margin Count
1-24 3.5% -5.1% 2.4% -4.8% 525
24-41 3.4% -5.3% 1.2% -5.2% 547
41-106 4.9% -0.3% 3.3% -1.9% 537
106-197 6.7% 2.3% 5.2% 0.6% 539
197+ 7.6% 1.5% 6.6% 1.4% 538

 

Margin tends to increase with size and, focusing on mean patient margin, the smaller hospitals that have lower than a 100 ADC are, on average, losing money based on just patient margins. Next, I’ll look at profitability by type of hospital.

Hospital Profitability by Type

Hospital Type Mean Overall Margin Mean Patient Margin Median Overall Margin Median Patient Margin Count
Children’s Hospitals 12.2% 4.0% 9.9% 0.8% 32
Critical Access Hospitals 2.4% -7.2% 2.0% -5.5% 718
Short-term Acute Care Hospitals 6.2% 0.6% 4.7% -0.4% 1,939

While it’s a common refrain that children’s hospitals are profitable, it’s interesting to see just how much of a difference there is between children’s hospitals and general hospitals. I will note that not all children’s hospitals submit cost report information, so this is just a portion of children’s hospitals in the United States.

Hospital Profitability by Ownership

While many hospitals are not-for-profit endeavors, there are a variety of different ownership structures for hospitals.

Hospital Ownership Mean Overall Margin Mean Patient Margin Median Overall Margin Median Patient Margin Mean Average Daily Census Count
Corporate 7.1% 3.6% 5.1% 2.2% 114 475
Government 1.7% -10.5% 1.3% -8.2% 73 561
Non-Profit 5.5% -0.2% 4.4% -1.0% 145 1589
Physician-Owned 15.7% 13.0% 16.0% 14.5% 35 63

 

Not surprisingly, for-profit hospitals tend to have better margins than non-profit or government-owned hospitals, since they have the stated objective to maximize profits. The profitability of physician-owned hospitals (which are for-profit), though, was surprisingly high. Part of this may be explained because a larger proportion of these are specialty hospitals that perform a significant amount of highly-reimbursed surgeries (like orthopedic hospitals), as can be seen by the mean ADC where they tend to have very few admitted patients at any one time.

Conclusions

This is just the tip of the iceberg of understanding hospital profitability, and maybe in the future I’ll come back to it again and do some more analysis. My key takeaway is that there are many different hospital characteristics that are important indicators of hospital profitability. Additionally, hospitals tend to be profitable, but they need non-patient revenue to be profitable. Finally, the incredibly high amount of spending on hospital care suggests that we, as a country, should really look where hospital costs can be taken out of the system – what better advice than to go where the money is.

Where does Torch Insight data come from?

Torch Insight is built on an integrated data warehouse that combines a multitude of different data sources and links the data together in a way that allows for exploration, analysis, benchmarking, and generally gaining insight into how different players in healthcare markets are connected. A key component of any data platform is understanding where the underlying Torch Insight data comes from. In this blog I’ll talk about where underlying data comes from and how it’s integrated into the platform.

Data sources

There are three major underlying sources for the data that we use: (1) publicly available data, (2) licensed data, and (3) collected data. Additionally, there is a wealth of derived data that we provide.

Publicly Available Data

There are thousands of publicly available datasets that are sponsored and maintained by a variety of different entities. The majority of these are managed by government entities and, because of the public funding that sponsors their creation, they are disseminated publicly. Others are sponsored by not-for-profit organizations, and a handful are released by for-profit entities. For Torch Insight, we work with many dozens of publicly-funded data sources including those sponsored by the Centers for Medicare & Medicaid Services (CMS), the United States Census Bureau, the Centers for Disease Control (CDC) and others.

Many of these sources are direct sources since they are the result of government-backed studies or data collection efforts. Think of these like data from the Census Bureau where the direct collection of demographic information was the intent of the programs that created them. Others, such as many of the datasets from CMS, are indirect sources since they are derivatives of other programs. Medicare beneficiary demographics, for example, are derived from the data captured which is required to run the Medicare program.

Licensed Data

A variety of data sources that we use come from proprietary sources that we license from other companies. These come from a variety of vendors that collect data for a variety of reasons. Some are in the business of collecting and selling data, while others are in a separate business and have collected large amounts of data as a result of that business and are seeking opportunities to disseminate it with business partners. An example of this data is health insurer financial and enrollment data which is reported quarterly to the National Association of Insurance Commissioners. Data that we license from others includes redistribution permissions, so we can resell that data, and the derivative work we generate from it, to our clients.

A subset of licensed data is data that we have a license to use but are only able to sell derivative work. For example, we do not sell claims data, but we do license claims data that we are able to extract information from, including standardized reports and relationships, and then we share just the extracted or generated data.

Collected Data

The third source of data is data that we manually collect. For example, our ACO Database was originally started in 2010 and we have a team that is focused on maintaining a record of all payment contracts where healthcare providers accept risk for the patients they care for. Other manually collected data is validated data where we have a team that manually insures that our other data sources are correct, such as validating addresses and phone numbers, or taking known healthcare providers that come from other sources and adding additional information such as websites and identifying which health system they are affiliated with. We have manually validated hundreds of thousands of variables and fields and continue to do so on an ongoing basis.

Derivative Data

Similar to some of the standardized reports we generate from claims data, we also create other derivative work from our base datasets. This includes making basic calculations, such as standardized financial ratios, or applying more sophisticated model derived from dozens of variables to make customized estimates.

Knowing Where the Data Comes From

Metadata Example

Understanding the underlying data sources is essential for analysts and researchers as they try to make sense of what they discover through Torch Insight. Our data is tracked with a custom-built metadata management system that keeps track of the underlying data source, the time frame it covered, the time it was released, and any calculations done to the data. This can be accessed from within many of the tools, or can also be accessed through a separate, searchable resource.

Metadata Search

Intro to Organizational Analysis: Entity-Level Facts and Relationships

Organizational Analysis on the Go

At a conference I met a hospital CEO and began chatting. I asked him about his hospital and he told me about some of its challenges and strengths. The conversation turned to physician recruitment and I asked him how many physicians were admitting patients to his hospital, and he didn’t know. I pulled up TorchInsight.com on my phone, pulled up his hospital’s dashboard and was able to list them off.

While this was more of a party trick than a practical use case, it does show one example of how organizational-level data can be accessed and use. In my last blog post, I took a look at healthcare markets and did a quick analysis of Alzheimer’s disease. Now I get to dig into the art of organizational analysis.

What is Organizational Analysis?

There is no standardized definition of organizational analysis, but it always include entity- or organization-specific data that allows the analyst to assess characteristics of named organizations.  Torch Insight includes data on hospitals, health systems, physician groups, skilled nursing facilities, health insurers, and many others.  The underlying data is derived from dozens of different sources, but the ultimate goal is to aggregate the data so that the user can get a complete perspective on the organizations.  The data can help evaluate competitors or potential partners, to identify sales targets, do an academic analysis, or countless other assessments. Once you have the data, the limit is how imaginative you can be.

Looking at Hospitals

Hospitals represent 32% of the healthcare system’s cost, and I thought it would be interesting to do some organizational-level analysis on the biggest hospitals in America.  I’m going to start by just pulling out the largest hospitals in the country and, based on 2016, there are 20 hospitals that had more than 50,000 discharges.

Locations of 20 largest hospitals in the United States

Organization facts and figures

The largest hospital, in terms of discharges, is Florida Hospital Orlando, which is part of the Adventist Health System. By clicking on the hospital I’m able to jump to its dashboard which provides data on dozens of facets of the hospital, including contact information, financial data, quality scores, health information technology platforms and more. Dozens of different origin files are incorporated into this, including publicly reported data, data calculated by Torch Insight and data that has been manually collected. The metadata management system tracks where the underlying data came from, as seen in the screenshot that shows the underlying financial data source as hospital cost reports.

Another example of factual data is the quality scores that are collected by the government for Hospital Compare reporting. Dozens of different measures exist, which can also be benchmarked against other hospitals.

Trends

 

An important part of any analysis is understanding how things are changing. The dashboard includes some default trend data – in this case net patient revenues and bed counts – but it can be customized to include dozens of longitudinal variables.

Hospital systems and relationships with physicians

 

An important component of Torch Insight is the relationships that it shows between different organizations and entities. Florida Hospital Orlando is part of Adventist, and by clicking on the connection an analyst can explore the broader system’s financial performance or look at a map of the hospitals in the system. In addition, it shows the connection of hospitals to physicians with the ability to see which physicians are admitting to the hospitals or the system. The physician groups can be explored, as well, and you can learn where the physicians are admitting.

Exporting Data

If you want to export data, you can also go through the data export tool to pull out the raw data. There are hundreds of data fields to work with. I decided to export the 148 hospitals with more than 25,000 discharges.

Once exported, you can easily evaluate whatever seems interesting with whatever tool you prefer (Excel, Stat, SAS, R, Tableau, or dozens of others).  I did a quick correlation to see if the relative market share of each of these largest hospitals was correlated with their value-based purchasing scores (they weren’t, p=0.55) and here, I made a histogram of their overall star ratings.

Conclusions about Organizational Analysis

Organizational analysis starts with data about specific entities and then allows the user to answer questions about them. In future posts I’ll provide more specific examples and case studies, including conducting some analyses that actually answer relevant policy and business questions.

Drilling Down into Alzheimer’s: A Market-Level Analysis

Introduction to Market-Level Analysis

Torch Insight is designed to help evaluate and analyze healthcare markets and organizations. There are countless analyses that can be performed, focusing on demographics, financials, performance, quality measures and many, many other pieces of information. For this demonstration, I will focus on a chronic condition that is pervasive across America: Alzheimer’s disease. The goal is to show one approach to performing a preliminary market-level analysis of a disease. 

Alzheimer’s Disease

Alzheimer’s disease is the most common form of dementia that affects over 5 million Americans, the vast majority over the age of 65. It’s a progressive illness that causes mental decline including causing difficulty with thinking, reasoning and remembering. As a country, America spends hundreds of billions of dollars a year on direct care.

Market-Level Analysis of Alzheimer’s

The impact of the disease is not uniform around the country. Using Torch Insight, we can start to visualize the differences. Torch Insight has both market-level data and organizational-level data (such as information about physicians, hospitals, insurance companies, etc.). For this example, I’m focusing on the market-level prevalence of Alzheimer’s.

Since Alzheimer’s primarily affects those that are on Medicare (those aged 65 and older or those that with a disability – including Alzheimer’s disease – that qualifies them for supplemental security income), Medicare data is a great resource to assess the prevalence of the disease. Medicare is a government-backed health insurance program and collects massive amounts of information based on enrollment information and the claims it pays on behalf of beneficiaries. Some of this data is then released in a variety of public use files (PUFs) and limited use files. The prevalence of Alzheimer’s in the Medicare population comes from the 2015 Medicare Geographic Variation PUF, which is prepared by the Centers for Medicare & Medicaid Services (CMS). This file provides a variety of pre-tabulated fields, including data at the state- and county-level.  While I’m demonstrating this data based on the pre-tabulated data, we’ve also calculated it directly from Medicare claims data.

In the following map you see the percent of Medicare beneficiaries that have Alzheimer’s or a related disorder, by state.  The map is interactive and you can zoom, scroll and hover over the individual states to see values.

You can see that the percent of the Medicare population with these dementia-related illnesses ranges from a low of around 7% to a high of around 12%.  For states that are providing services for these patients, either through social services or through Medicaid for those that are dually eligible for Medicare and Medicaid, the prevalence can make a significant impact to the state budget.

Asking questions about Alzheimer’s

When I start to study a topic, I like to look at the data and start to ask questions, and the map is a great way to do this. In this example, after I looked at the distribution of the prevalence of the disease, a few questions occurred to me:

(1) The disease visually appears to be more prevalent in states with larger populations (Florida, Texas, New York, California, etc.) and I wonder if there’s a correlation?

(2) I know that Alzheimer’s is an expensive disease, so I wonder if an increased prevalence of Alzheimer’s is associated with higher Medicare costs?

(3) I wonder if the variation at the state level is similar to the variation at smaller geographies.

Alzheimer’s correlation with population

The first two questions will require a little more data. Luckily, Torch Insight has hundreds of pre-built, healthcare-related variables so I can select a few that are interesting.

Rather than overlaying the variables, I generally prefer to work with the raw data, so I exported the file and brought the variables into Stata.  It turns out that Alzheimer’s is positively correlated with absolute population size (corr=0.4467, p=.001), and I can plot it:


The R2 is nearly 0.2, which, is modestly explanatory for a single variable looking at a linear relationship.  There are a few states that are outliers with population, so I transformed the population and took the log of it and the R2 went up to 0.31 – decently explanatory.

 

This means that not only are states with higher populations more likely to have more total Alzheimer’s patients, but they are more likely to have a higher prevalence of Alzheimer’s patients.

Alzheimer’s and Medicare Costs

When comparing costs, it’s important to think through what costs are important. Medicare provides a variety of cost numbers in the PUF file, including total and per-beneficiary costs, as well as costs that are standardized (meaning they adjust for different prices that are paid in different regions of the country) and risk-adjusted. Since risk-adjusted costs account for the prevalence of different diseases, I will use state-average, standardized annual costs per Medicare beneficiary. This time the correlation is really strong (corr=0.781,p<.0001).

State-level market-analysis

The only outlier with low costs and high prevalence of Alzheimer’s is Hawaii, which may partly be explained by how costs are standardized with Hawaii. I can confidently say that a high prevalence of Alzheimer’s disease within a state is associated with higher total Medicare costs.

Alzheimer’s at the County Level

While state-level analyses are interesting, it’s often valuable to drill down to a smaller geographic level. Torch Insight includes states, hospital referral regions, metropolitan areas, congressional districts and counties as options to dig deeper within the visual interface.  I just selected the state with the highest prevalence of Alzheimer’s – Florida – and included the county-level map.

The prevalence of the disease ranges from 7% (not much higher than the best states) to 21% in Miami-Dade county. Not only is that a huge difference, but the highest prevalence is in the most populous county, which drives up the state average. Similar to the previous analyses, the prevalence of the disease at the county-level in Florida is strongly associated with population (corr=0.655, p<.0001) and costs (corr=0.627, p<.0001).

Conclusions

My goal with this was not to explain what’s driving the spread of Alzheimer’s Disease and other forms of dementia or to do any sort of a causal assessment. Hopefully, you were able to see the power of being able to quickly combine market-level characteristics and see how you can dig into the variation that exists. Hopefully, also, this raises questions in your mind about Alzheimer’s disease – it certainly does for me and there are multiple avenues of further inquiry I could perform (How does its prevalence relate to other chronic diseases? What other diseases are as predictive of total Medicare costs? Why might there be a higher prevalence of Alzheimer’s disease in higher population areas and larger cities? And many more). The joy of data exploration is that there is always more to learn and Torch Insight can accelerate that learning.

 

 

 

Understanding Healthcare Markets: Introduction to the Torch Insight Blog

Torch Insight Blog

Blogs solve a couple of really important challenges: (1) quick dissemination of information to a broad audience and (2) a lasting reference source.  Often, the quickest way to share information is to speak to someone or send an email, but that doesn’t go to a broad audience, while lasting references, like books, take a long time to get to the market.  The Torch Insight Blog exists to quickly share interesting data-driven findings and conclusions about healthcare markets that can be referenced by anyone who is interested in healthcare.

As the inaugural blog, I thought I would share a little bit about the history of Torch Insight and explain what it is and why we built it.  In the future I will focus more on interesting quantitative findings and commentary around what’s happening in the healthcare system.

A Brief History of Torch Insight

Since its inception, Leavitt Partners has sought to help clients understand how the healthcare system is changing and create strategic plans.  A key to understanding the system is accurate data about the state of healthcare.  Back in 2011, when I first started doing quantitative analysis for clients, I learned that there was no available data set that brought in comprehensive data about the healthcare system and linked it together in a meaningful fashion.  To understand a healthcare market you need to know who is in a market and how the different organizations and stakeholders interact, so we began a process of linking together disparate databases to provide a comprehensive view of healthcare markets.  During these early years we identified many different data sources and also began collecting some of our own data, particularly our ACO tracking data, and figured out how to bring the data sources together for each project we did.

After five years of manually connecting databases, we recognized that there was an opportunity to standardize the process of linking data and that some of the customized work that we then performed could be automated.  So, in the summer of 2016 we kicked off the development of the tool that would become Torch Insight.  The process was relatively straightforward:

  1. Take the knowledge we had gained about which data sources were needed to provide a 360-degree perspective of a healthcare market and build out the process and system to keep hundreds of disparate data sources updated, linked and accessible.
  2. Build a platform to access the data, visualize answers and gain insights about market-level relationships.
  3. Develop algorithms and analytic approaches to derive meaningful insights from the data that are not available anywhere else.

Working from this plan, we have taken Leavitt Partners’ expertise of data and markets and built a platform that allows users to explore the relationships between healthcare providers, payers and populations.  While there are definite use cases where clients are already using the platform, the amazing part of an interlinked platform is the ability to come up with novel uses for it on a daily basis.

Healthcare Markets

While healthcare is a 3.3 trillion dollar industry, it’s far from a uniform industry with each market having developed from its own little accident of history where thousands of disparate decisions and strategies have turned the market into what it is today.

Each market has its own makeup of providers, payers and patients.  For example, some markets are dominated by a single not-for-profit health system that primarily employs many of the physicians, while others have many competing health systems or independent physician practices or a high number of for-profit systems, or countless other dynamics.  There are no national health systems, like we see with national chain restaurants or retail stores, and each organization has its own structure.

Torch Insight shows how these markets are structured and also brings in relationships between providers so you can see which hospitals are working with which physicians and value-based arrangements, like ACOs.

It also shows how patients interact with the system and create visualizations to show where patients naturally move between providers, provides financial and enrollment data for health insurers, population demographics (including a wealth of disease metrics), competition metrics, and tools for visualizing and exploring the relationships between each of these factors.

One of my favorite things to do is to overlay different factors to start to see how markets have formed.  This could be as simple as overlaying disease burden with the location of different types of healthcare facilities, or integrating dozens of different variables into a composite metric which tracks some interesting facet of the market.  Since the data can always be exported on demand, it doesn’t take much effort to take the raw numbers and run some statistical tests on them to enhance certainty about conclusions.

Torch Insight Map of Healthcare Markets
Map showing pinpoints overlaid on demographic data

Analysis of Healthcare Markets

My objective with this blog is to show how Torch Insight can help answer questions about healthcare markets.  I’ll do this by applying the data to questions that are relevant to current issues, answering questions that I have, and highlighting interesting points that I discover.  Some of the posts will focus on using the tool and showing practical examples of how to answer a question, but others will focus more on showing what I find when I take the data out of the tool and look at it in a statistical package.  At heart, I’m a data guy and a researcher which means I like to ask, and answer, questions with data.  I hope you’ll join me on this process as we explore the American healthcare system.