Assessing Catastrophic Health Expenditures Among Uninsured People Who Seek Care in US Hospital-Based Emergency Departments

Key Points Question What is the risk of a single treat-and-release emergency department (ED) visit contributing to a catastrophic health expenditure (CHE; health care costs exceeding 40% of post-subsistence income) among uninsured people? Findings In this cross-sectional study of 41.7 million ED visits from 2006 to 2017, nearly 1 in 5 (18%) uninsured treat-and-release ED patient encounters were at risk of CHE. This risk has grown over time and disproportionately burdens those with low incomes. Meaning Policies such as broadening financial risk protection for unscheduled care may help to mitigate CHE risk among uninsured people, who have few alternatives for care outside of the ED.


Conceptual Framework
The primary goal of this project was to provide national estimates of the proportion of uninsured patients who are treated and released by the ED who are at risk of a catastrophic health expenditure (CHE). The World Health Organization has defined CHE as annual out-of-pocket healthcare expenditures that exceed 40% of post-subsistence income (e.g., income that remains after accounting for food and housing costs). [1][2][3] We leverage prior methodologies that have quantified CHE risk, which is a component of financial toxicity, among disease-specific populations (e.g., trauma 4 , cardiovascular disease 5 ), to estimate this risk among uninsured ED treat-and-release patients over time.

Estimating Income
The methodology to estimate income in this analysis builds upon previous work that has applied estimated incomes for individuals living in income quartiles and then comparing this to listed charges. Specifically, this prior work 4,5 has utilized datasets included in the Agency for Healthcare Research and Quality Healthcare Cost Utilization Project, including the National Inpatient Sample (NIS)), which is similar to the primary data source utilized in this study: the Nationwide Emergency Department Sample (NEDS). 6 For these datasets, the only variable related to income is an estimated income for individuals living in a particular community ZIP code income quartile.
To inform the main parameters of shape and scale to estimate income in this analysis, we relied on Gini coefficients for the United States as provided by the World Bank for each year (2006-2017). 7 Following prior work, we defined the two key parameters of shape and scale in a microsimulation model to generate income distributions for the analytic sample. 4,5,8 For shape, we used the World Bank Gini coefficients for 2006-2017 and mapped the established Gini coefficient-based scalars by year. 7,9 For scale -this was the community quartile level income estimate provided by NEDS 6 -divided by the scalar estimated for each year.
This method requires setting a mean for each ZIP code income quartile. As has been done in prior studies 4,5,10 , we used the upper bound of the estimated median household income within a given ZIP code income quartile to inform the gamma distribution each year. For instance, the range of median incomes published by NEDS for the community income quartile 2 in 2017 was $44,000-$55,999.25 For every encounter who lived in ZIP code income quartile 2 in 2017, the value of $55,999 was set as the average median income level that informed the microsimulation model's shape and scale as defined above. The exception was for quartile 4 as this had no upper bound, thus this quartile's average median income level was set at the 80th percentile income level for each year, consistent with prior work.
As shown in eTable 1, we then mapped each Gini coefficient by year to the shape parameter from Shrime et al. 9 In eFigure 1, we display the gamma (γ) distribution by household income based on the four ZIP code community income quartiles available in the NEDS dataset (eFigure 1 Once each encounter was assigned an income in the microsimulation model, we then estimated a post-subsistence income level per encounter, as has been done in prior work. 3 Post-subsistence income was calculated as the remaining income after taking into account costs of food and housing. Estimates for food and housing costs across various income thresholds are available from the Bureau of Labor Statistics. 9 For instance, if an encounter's estimated income was <$15,000, then 91.5% of their income was estimated to have been spent on food and housing, and the remaining balance would be defined as their post-subsistence income. As another example, if an encounter's estimated income was between $70,000-$99,999, then 37.0% of their income was estimated to have been spent on food and housing.

Data
The key dataset used for this analysis comes from the publicly NEDS, which represents the largest all-payer sample of hospital-based emergency departments (ED) in the United States. 6 The NEDS databased started in 2006 and the latest year of data available when this analysis was completed was 2017. The NEDS provides a discharge weight (variable = discwt) that allows for one to use a survey-weighted analysis to produce national estimates. A summary of ED encounters by year is provided in eTable 3.
A disposition variable is provided by NEDS (variable: EDEVENT) that categorizes patients as "treat-and-release", "admitted", and some additional categories. We included only treat-and-release encounters (EDEVENT=1), which is the majority of ED visits. Among this subgroup, we identified those patients that were uninsured. For this analysis and following prior work that focuses on CHE in the uninsured 4 , we defined the uninsured using the expected primary payer variable (PAY1) if they were "self-pay". Of note, some studies define the uninsured population in NEDS as encounters with the expected primary payer of both "self pay" and those listed as "no charge". However, the "no charge" group is relatively small (0.7% of the entire sample) and a heterogeneous group as these could include patients with insurance but not expected to have a charge (e.g. Medicaid patients). Since our objective was to assess CHE risk among those who were expected to not have any insurance protection among those who received a bill, this is why we included only "self-pay" in our definition of the uninsured as other studies have done. 4 Among this treat-and-release subgroup, a total of 16.4% were uninsured. Of the 48,770,796 uninsured ED treat-and-release encounters, 41,729,750 met inclusion criteria for the analysis. Further detail on how the analytic sample was derived is available in eFigure 2.
As noted in the eFigure 2, only a small number of observations were missing key variables in this analysis thus relatively few encounters were dropped due to missing data. However, the majority of dropped observations stemmed from a core variable needed for this analysis -ED charge. We provide a summary in eTable 4 of how the population demographics vary by the full analytic sample as compared to what the sample would have looked like had these observations not been dropped. As shown in the manuscript for our primary analytic sample and again in eTable 4, relatively few uninsured treat-and-release encounters were in the West as compared to other regions. This representation was further limited in that the majority of uninsured treat-and-release encounters with missing ED charges also came from the West, which NEDS has documented as a feature of these data. 6 Year Total NEDS Encounters The analytic sample is all uninsured treat-and-release ED encounters that have no encounter-level missing values. This is compared to the treat-and-release encounters that were excluded based on missing ED charges. Estimated income was not calculated for encounters outside of the analytic sample.
As shown in eTable 4, a total of 5,934,901 of treat and release encounters were missing ED charges. Once excluding those encounters with other missing values (n= 1,397,828 encounters that were missing values for age OR sex OR urban/rural designation OR ZIP code income quartile), this left a total of n=5,643,218 encounters that were missing only ED charges. The demographics of this sample that was excluded is shown in eTable 4. We also compare this to the full sample, had these observations not been excluded. Of note, we did not exclude the relatively few encounters that were missing values for hospital level variables of interest (e.g., hospital teaching status) given that this was an encounter-level analysis. Since there were no missing values for hospital region among this study's analytic sample, no observations were dropped due to missing data for this variable. For teaching status, however, in 2006, approximately 55% of the analytic sample was missing hospital teaching status. However, hospital teaching status had no missing values in the remaining years (2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017). Given this anomaly, we did not drop any of these encounters (which otherwise had all of the remaining encounter-level variables of interest, including ED charges) based solely on missing hospital teaching status. We display the distribution of teaching status over time among the analytic sample by teaching status after excluding those missing values (see eFigure 9 below), but these 1.9 million observations were not dropped from any remaining analyses.

Changes in Uninsured ED Treat-and-Release Population Over Time
The following figures (eFigures 3-9) illustrate how the uninsured population in the analytic sample (n=41,729,750) has changed over the study period (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017). In general, the distribution appeared relatively constant over time by patient sex. For age groups, there was a decline in encounters consisting of the <20 year age group over time, possibly related to health reform insurance expansion efforts targeted at younger age groups (e.g., the dependent coverage provision of the Affordable Care Act 11 ). In terms of ZIP code. This was only reflected by a slight uptick in the average age for ED uninsured treat-and release encounters over time (30.6 years in 2006 versus 33.9 years in 2017). For ZIP code income quartile, a greater proportion of uninsured treat-and-release encounters consisted of individuals living in the lowest income quartile (Quartile 1) over time (38.0% in 2006 versus 45.0% in 2006). For rurality, this was relatively constant over time, though there was a slight uptick in encounters located in urban areas in 2017. Though this study was focused on an individual-level variable (encounters), we also assessed where these encounters were located at based on some of the available NEDS hospital-level variables (region and teaching status). For hospital region, the largest share of uninsured treat-and release encounters was located in the South, and this grew over time. For hospital teaching status, a greater proportion of uninsured treat-and release encounters appeared to be from EDs affiliated with metropolitan teaching hospitals as compared to non-teaching hospitals over time.   , there appears to be a relative growth in uninsured treat-and-release encounters stemming from Eds affiliated with metro-teaching hospitals.

Comparison of ED Charges and Income Over Time
All values were adjusted for inflation and displayed in 2017 US Dollar terms. Any encounters that were missing ED charges (variable = TOTCHG_ED) were dropped from the analysis. To eliminate erroneous outliers in ED charges, we computed a modified ED charge that replaced those charges exceeding the 99 th percentile of charges with the dollar amount of the ED charge listed at the 99 th percentile.
In eTable 5 and eFigure 10, we show that, even after accounting for inflation, ED charges have risen over time while the estimated median household income (one estimate was randomly drawn among the 1000 simulations to illustrate a representative example of income distributions for the analytic sample) slightly decreased. Given the makeup of the outpatient uninsured ED population in terms of ZIP code community income quartile, this relatively stable to decreasing incomes is not unexpected since fewer members living in the highest income quartile make up the analytic sample over time (eFigure 7, eFigure 11).  . The estimated income is the average of the estimated median household income after applying a microsimulation model to estimate income for each encounter based on ZIP code income quartile levels. The estimated income data shown above comes from one randomly selected draw from the 1,000 simulations, which allowed us to provide discrete estimates by year.

CHE Risk by Definition
For the purposes of this study, a Catastrophic Health Expenditure (CHE) is calculated by comparing the listed ED charge for each encounter and comparing this to the estimated income for each encounter (which was calculated from the microsimulation model based on each encounter's ZIP code income quartile, as described in detail above). Though there remains a number of ways in how CHE has been defined and operationalized in prior studies, we employ the World Health Organization's (WHO) definition of CHE which has been described as out-of-pocket spending on healthcare that exceeds 40% of one's post-subsistence income (e.g., income that has accounted for estimated costs of housing and food). 2,3 For this study, we define "CHE 40 Risk" as an ED charge for a single encounter that exceeds 40% of the encounter's estimated annual post-subsistence income.
An alternative definition that has been used in CHE/financial hardship literature is out-of-pocket healthcare spending that exceeds 10% of annual overall income. 4,[12][13][14] For this alternative definition, we defined what we call "CHE 10 Risk" as the proportion of encounters that had an ED charge that exceeded 10% of their estimated annual household income.
For this study, CHE 40 Risk is at a similar level to that of CHE 10 Risk in 2006 (13.6% and 13.4%, respectively), but CHE 10 Risk grows to an absolute higher level over time as compared to CHE 40 Risk. Specifically, 31.0% of the analytic sample met criteria for CHE 10 Risk in 2017 as compared to only 22.6% of the sample meeting CHE 40 Risk that same year (eTable 6).    Note: CHE=Catastrophic Health Expenditure (CHE). CHE 40 Risk = ED charge that exceeds 40% of estimated annual post-subsistence household income (i.e., income that remains after accounting for food and housing costs). CHE 10 Risk = ED charge the exceeds 10% of estimated annual household income.

CHE Risk by Diagnosis Category
The primary focus of this project was to showcase CHE risk nationally and over time. However, we also conducted an exploratory analysis to see if and how CHE risk may vary by diagnosis categories. We limited this analysis to our most recent year of data (2017). We mapped each encounter to one of the twenty-one categories (chapters) of ICD-10 CM codes as defined in 2017 by the primary diagnosis (variable dx1). 15 Of the 3,421,828 observations from our analytic sample in 2017, only 1783 were missing a primary diagnosis code and an additional 32 had an invalid value for the listed code. This equates to a total of 3,420,013 observations (99.9%) were matched to one of the twenty-one ICD10-CM categories.
In eTable 7, we present the risk of CHE based on both the primary definition used in this analysis (CHE 40 Risk) as well as an alternative definition (CHE 10 Risk), as defined above. We display these results sorted by the prevalence of the sample that was mapped to a particular ICD-10 category. For instance, 20.2% of the 2017 uninsured treat-andrelease encounters in our analytic sample had a primary diagnosis code that mapped to the ICD-10 Category (Chapter) of Injury, poisoning, and certain other consequences of external causes (S00-T88

Additional Analysis of CHE Results
In eTable 9, we provide the unadjusted results for CHE risk for each covariate in this encounter-level study. The results come from a series of survey weighted linear regression models with CHE risk as the dependent variable and one covariate of interest (e.g., sex, age groups, income quartile, etc.) as the independent variable. After computing each regression model, the margins function was used to provide CHE risk (as opposed to the coefficient from the model output). A corresponding 95% confidence interval (CI) and p-value was provided. A p-value <0.05 suggests that there was a statistical difference in the predicted CHE risk between a group of the covariate as compared to the reference group within that variable (i.e., CHE risk for female versus male). Given the extremely large sample size available to us, it is not surprising that even small differences that may not be substantively different were statistically significant.
As shown in the primary analysis, we also provide the same results for each of the separate survey-weighed linear regression models containing both the covariate of interest as well as year fixed effects to account for secular trends. As before, the margins command was used to provide the estimated CHE risk for each variable that accounts for year fixed effects. As an example, encounters from the highest income quartile (Q4) had a predicted CHE risk of 6.6% as compared to 22.9% among the lowest income quartile (p-value <0.001), holding year constant. Age Groups