Institute for Economic Stanford Policy Research (SIEPR) THE IMPACT OF FINANCIAL ASSISTANCE PROGRAMS ON HEALTH CARE UTILIZATION Alyce Adams Jinglin Wang Stanford University New York University & Kaiser Permanente Raymond Kluender Francis Wong Harvard University NBER Neale Mahoney Wesley Yin Stanford University University of California, Los Angeles & NBER & NBER August, 2021 Working Paper No. 21-046 SIEPR | John A. and Cynthia Fry Gunn Building | siepr.stanford.edu | @siepr The Impact of Financial Assistance Programs on Health Care Utilization" t Alyce Adams Raymond Kluendert Neale MahoneyS __Jinglin Wang! Francis Wong! Wesley Yin™ August 31, 2021 Abstract Most hospitals and managed care organizations have financial assistance programs that aim to reduce financial burdens and improve health care access for low-income pa- tients. We use administrative data from Kaiser Permanente to study the effects of finan- cial assistance on health care utilization. Using a regression discontinuity design based on an income threshold for program eligibility, we find that financial assistance signif- icantly increases health care utilization initially, though effects dissipate three quarters after program receipt. Financial assistance also increases the detection of and med- ication refills for treatment-sensitive conditions, suggesting financial assistance may increase receipt of high-value care. *We thank Andrea Altschuler, Somalee Banerjee, and Lin Ma for substantial assistance during the course of this study. This work was supported by the Becker-Friedman Institute at the University of Chicago and the National Institute on Aging, Grant Number T32-AG000186 as well as the Targeted Analysis Plan sponsored by the Kaiser Permanente Delivery Science and Applied Research Program. Stanford University and Kaiser Permanente. Email: asadams@stanford.edu tHarvard University. Email: rkluender@hbs.edu SStanford University and NBER. Email: neale.mahoney@gmail.com INew York University. Email: jinglin.wang23@gmail.com '' NBER. Email: fwong@nber.org "University of California, Los Angeles and NBER. Email: email-wesyin@gmail.com I Introduction The provision of financial assistance to indigent patients is a longstanding component of the US health care system. Throughout much of the 18 and 19" centuries, hospitals were primarily charitable institutions, providing free care to people who could not afford home-based care. During the late 19 and 20" centuries, advances in medical knowledge made hospitals attractive to a broader set of patients, and financial assistance evolved to fill the gaps left by a burgeoning health insurance system (Starr, 2008). In the 1960s, federal law formally encouraged financial assistance as a way for non-profit hospitals to meet requirements for their tax-exempt status. In recent decades, federal and state legislation have increased requirements for the provision of financial assistance, both by non-profit and for-profit hospitals. Today, most US hospitals have financial assistance programs that seek to reduce fi- nancial burdens and improve health care access for low-income patients by providing a combination of debt forgiveness and reduced out-of-pocket costs (see Table 1, discussed below).! These programs target both uninsured and insured patients and are common across hospital ownership types. In 2018, hospitals provided a total of $26 billion in char- ity care, of which $20 billion was provided to uninsured patients and $6 billion provided to insured patients (Roth et al., 2021). Charity care accounted for 1.5% of total expenses for the median non-profit hospital, 1.4% of total expenses at the median for-profit hospital, and 0.9% of total expenses at the median government hospital (Bai et al., 2021). Charity care provided through financial assistance programs has been at the center of a policy debate over whether non-profit hospitals provide sufficient "community benefits" to justify their tax-exempt status." The prior literature has largely focused on the dollar value of financial assistance, drawing on hospital-level data on charity care and bad debt - collectively referred to as uncompensated care - from the American Hospital Association 1Examples of stated aims include: "improving health care access for people with limited incomes and re- sources" (Kaiser Permanente); "provide medically necessary healthcare to everyone, regardless of the ability to pay" (Community Healthcare System); and "providing quality health care services to all our patients re- gardless of their financial situation" (Mercy). See, for example, Senator Grassley's investigations discussed in https: //www.nytimes.com/2009/ 06/01/us/politics/Olhealth.html and https://www.modernhealthcare.com/government / grassley-back-it-ramping--up-scrutiny-tax--exempt-hospitals. Annual Survey and the publicly available Hospital Cost Reports. Researchers have used these datasets to document trends in uncompensated care and to examine how insurance coverage has impacted uncompensated care (e.g., Garthwaite et al., 2018; Cunningham and Tu, 1997; Dranove et al., 2016; Mann et al., 1997; Camilleri, 2018). While prior studies measure the implications of uncompensated care on hospital fi- nances, little is known about whether financial assistance programs improve health care access (i.e., increase health care utilization among those who may be deterred by costs). In principle, financial assistance could impact health care utilization through the price ef- fects of reduced out-of-pocket costs and the wealth effect of relief from previously accrued medical bills. However, identifying the causal impact of financial assistance programs is challenging because selection into these programs is typically endogenous to past or ex- pected future utilization. And because patients may seek care from multiple health care providers, estimating the full impact of a hospital's financial assistance program can be hamstrung by incomplete data on health care utilization. This study estimates the impact of Kaiser Permanente Northern California's finan- cial assistance program on health care utilization. Kaiser Permanente provides an ideal setting for studying the impacts of financial assistance programs for three reasons. First, Kaiser Permanente's program is representative of financial assistance programs at other large health care systems, offering a combination of debt write-downs for previously in- curred unpaid medical bills and the elimination of cost sharing for health care over subse- quent months. Second, eligibility is determined by a strict income cutoff rule at 350% of the Federal Poverty Level (FPL), which provides identifying variation in the form of a re- gression discontinuity (RD) design. Third, Kaiser Permanente operates a large, integrated closed-network health system, which means that patients can receive all types of care at Kaiser Permanente facilities. While our data allow us to observe some instances of care re- ceived at non-Kaiser facilities, the integrated insurer-provider system means that virtually all care occurs at Kaiser facilities, allowing us to observe complete information on health care utilization, drug prescriptions, and diagnostic tests of program applicants. We use an RD based on an eligibility cutoff at 350% of FPL to estimate the causal ef- fects of the financial assistance program at this discontinuity. We observe administrative 2 data on income, family size, and other demographics for over 25,000 program applicants. Using these data, we estimate a sharp 78.8% increase in approval for applicants below the threshold. Virtually all applicants below the threshold are approved while a small percent- age of applicants above the threshold qualify through a separate expense-based criterion. In support of the research design, we show that patient demographics and prior health care utilization trend smoothly through the discontinuity. As a placebo check, we find no effects at the discontinuity in any of the seven quarters before application. Finally, we do not detect bunching of income below the discontinuity, which suggests limited scope for income manipulation. Collectively, these results support the identification assumption that applicants immediately above and below the income eligibility threshold are similar, and that the results are not driven by confounding selection into program eligibility. We find that financial assistance increases health care utilization in the first quarter following application. Our preferred instrumental variable (IV) estimates, which scale up our estimates to account for the 78.8% increase in approval at the threshold, indicate that approval leads to a 3.6 percentage point (pp) increase in the likelihood of an inpatient encounter (relative to a mean of 6.2%), a 13.4 pp increase in the likelihood of an ambula- tory encounter (relative to a mean of 67.0%), and a 6.7 pp increase in the likelihood of an emergency department encounter (relative to a mean of 12.7%). We estimate a fairly large increase in prescription drug utilization (an increase of 32.0 in prescription drug days sup- plied relative to a mean of 136.0 days) and marginally significant increases in utilization of drugs to treat cholesterol, diabetes, and depression. The effects we observe decline in the second quarter and largely disappear by the third quarter after the date of the application decision. The first quarter impacts on utilization are large in magnitude. As a benchmark, they are proportionally similar to the effects of Medicaid on health care utilization reported in the Oregon Health Insurance Experiment (Finkelstein et al., 2012; Taubman et al., 2014) over the first year-and-a-half of the program. In other words, the initial impacts of the financial assistance program on health care utilization in our insured study population are similar in magnitude to the impacts of providing Medicaid to the uninsured in Ore- gon. However, unlike the effects of Medicaid which persist for at least the year-and-a-half study period, the effects of the financial assistance program tail off by the third quarter. The comparison highlights differences between hospital financial assistance programs and Medicaid coverage, which we discuss in Section V. In addition to the effects on utilization, we find that financial assistance increases test- ing for and detection of health conditions. Specifically, our IV estimates indicate a 4.0 pp increase in the probability of an abnormal test result (relative to a mean of 10.0%). This effect is driven by an extensive margin increase in testing; conditional on having a test, the probability of an abnormal result is unchanged. We also find increased drug utilization for treatment-sensitive conditions. Taken together, the increased detection of abnormal health conditions - a precondition for appropriate treatment - along with increased drug utilization, suggest that at least some of the increase in health care utilization caused by the financial assistance program is high value. This finding is consistent with emerging ev- idence of the detrimental effects of consumer cost-sharing on the utilization of high-value care (Chandra et al., 2010; Brot-Goldberg et al., 2017). Our research builds on two correlational studies that have examined the effects of fi- nancial assistance programs on health care utilization. Based on a survey of 308 (insured) patients, Conner et al. (2013) find that enrollment in a financial assistance program is asso- ciated with reduced health care utilization and no change in self-reported physical or mental health. Chaiyachati et al. (2020) examine the association between non-profit hospital com- munity benefit spending and hospital readmission rates for Medicare patients but finds no statistically significant relationship. In contrast to these studies, our quasi-experimental evidence indicates that financial assistance causes a substantial increase in health care uti- lization and detection of health conditions, albeit over a limited time horizon. Our study also complements the literature on hospital-level uncompensated care, men- tioned above. This includes descriptive research on patterns in uncompensated care over time and across different types of hospitals (Cunningham and Tu, 1997; Mann et al., 1997) as well as research evaluating whether the amount of charity care provided by non-profit hospitals is commensurate with the favorable tax treatment they receive (Young et al., 2013; Singh et al., 2015; Herring et al., 2018). The literature also includes research on the role 3A number of papers have examined financial assistance policies for cancer treatment and drug costs 4 that hospitals play as insurers of last resort by providing vulnerable patients medical care that is ultimately uncompensated (Garthwaite et al., 2018; Dranove et al., 2016; Camilleri, 2018). In particular, our study sample is comprised primarily of insured patients who may nevertheless face significant out-of-pocket costs. The estimated utilization effects for this population contribute to our understanding of the role that hospitals serve as safety net insurers among insured patients who increasingly face high out-of-pocket costs.* II Background A Hospital Financial Assistance Programs Financial assistance policies have a long history in the US. In 1954, the federal government added section 501(c)(3) to the Internal Revenue Code, which provided organizations ded- icated to religious, charitable, scientific, or educational purposes with an exemption from paying federal income taxes. To qualify for tax-exempt status, a hospital had to provide "to the extent of its financial ability, free or reduced-cost care to patients unable to pay for it" (James, 2016). Since 1969, hospitals have been able to use financial assistance to fulfill their "community benefits" requirement for tax-exempt status (Somerville, 2012). While the IRS has given hospitals flexibility in determining which expenditures count towards community benefits (James, 2016), recent evidence indicates that charity care and other services account for about 85% of these expenditures (Young et al., 2013). Starting in 2015, the Affordable Care Act (ACA) imposed additional requirements on non-profit hospitals, including establishing a written financial assistance policy (IRS, n.d.; James, 2016). Finan- cial assistance policies are also influenced by state-level regulations. As of 2019, ten states require both non-profit and for-profit hospitals to provide free or discounted care to eligi- ble patients, and other financial assistance requirements exist in additional states (Stark, 2020). We gathered information on the financial assistance programs offered by the 40 largest (Felder et al., 2011; Semin et al., 2020; Zullig et al., 2017). These correlative studies typically find that fi- nancial assistance offers imperfect protection against financial hardship (Paul et al., 2016) and are subject to substantial frictions that deter take-up (Spencer et al., 2018). 4See, for example, https://www.healthsystemtracker.org/brief/tracking-the-rise- in-premium-contributions-and-cost-sharing--for-families-with-large-employer- coverage/ health care systems from their websites. Table 1 displays income-based eligibility criteria and benefits for the 4 largest for-profit and 4 largest non-profit systems with available information. We also provide information for Kaiser Permanente Northern California, which is the focus of our study. Appendix Section A describes the methodology used to gather information and Appendix Table Al provides eligibility and benefits information for all 40 health systems. Among health systems with available information, eligibility cutoffs range from 200% to 400% of FPL. Applicants may also qualify if they have substantial health care expenses relative to their means. Qualifying patients typically receive some combination of forgive- ness of previously incurred medical bills and reduced out-of-pocket costs for future care, often with more generous benefits for lower-income patients. While for-profits do not face the same federal regulatory requirements as non-profits, the largest for-profit health sys- tems offer financial assistance programs that are comparable in their eligibility criteria and benefits to non-profit health systems. B Financial Assistance at Kaiser Permanente Kaiser Permanente is a large, closed-network health care system that operates 39 hospitals and over 700 medical offices across eight states, serving 12.5 million patients.° As part of the ACA requirement for non-profit health care providers, Kaiser Permanente maintains a set of financial assistance policies similar in eligibility and generosity to those offered by other large hospital systems. For this study, we use data from Kaiser Permanente North- ern California, a division within Kaiser Permanente serving 4.5 million people in the San Francisco Bay Area, Greater Sacramento, and the Central Valley. Kaiser Permanente Northern California patients qualify for financial assistance if they have family income at or below 350% of FPL. Patients can also qualify with incomes above the 350% FPL threshold if they have eligible out-of-pocket medical and pharmacy expenses exceeding 10% of household income over a 12-month period (regardless of FPL). The ma- jority of patients (71%) who enroll in financial assistance qualify based on their incomes. Patients can learn about the financial assistance program through several channels. 5Source: https://about.kaiserpermanente.org/who-we-are/fast--facts Information on the program is included on medical bills sent to patients, as well as on- line. Case workers and caregivers can also provide patients with information in person at Kaiser facilities and assist patients in applying to the program. Kaiser Permanente Northern California's financial assistance eliminates both existing debts and cost sharing for future care. Patients receive a full discount on unpaid bills previously incurred at Kaiser. Patients in a Medicare Advantage plan face no copays for 6 months after receipt, and patients in a non-Medicare Advantage plan (e.g., employer- sponsored or Marketplace) face no copays for 12 months. The program applies to virtually all health care delivered by Kaiser, including care provided at Kaiser hospitals and clinics, as well as drugs provided at Kaiser pharmacies. III Data We obtained deidentified administrative data on all applicants to the Kaiser Permanente Northern California financial assistance program between January 2016 and December 2017. For each applicant, we observe income, family size, demographics, whether the patient applied via the income or expense-based criteria, and whether the application was approved or denied. Kaiser Permanente offers an ideal setting to study the impacts of financial assistance on health care utilization. Because it is an integrated closed-network system, we observe detailed electronic medical records information for 96.3% of health care expenses incurred by program applicants.® This allows us to observe ambulatory visits, emergency depart- ment visits, and inpatient hospital stays, as well as information on prescription drug use, laboratory tests and results, and total health care costs. We observe utilization, prescrip- tion drugs, and laboratory test results from a period 24 months before to 24 months after the month of the program application decision. We observe costs for the 12-month period before the application decision. We focus our analysis on the discontinuity created by the income eligibility cutoff at 350% FPL. While the expense eligibility criterion also creates a discontinuity, data on out- On average, only 3.7% of total expenses for individuals in our baseline sample were incurred for medical services outside of Kaiser in the 12-month period before the application decision. of-pocket bills at the time of eligibility were not available so we cannot use this threshold as a research design. A total of 25,574 patients with available information on income and family size applied via the income-based criteria. We drop patients with incomes below 150% FPL or above 550%, yielding a symmetric window around the 350% FPL cutoff. Of the 18,695 applicants in the remaining sample, 13 applied more than once. We exclude these individuals because the subsequent applications may be endogenous to information concerning their likelihood of subsequent approval. The resulting baseline sample has 18,672 applicants. For each applicant in our baseline sample, we construct a monthly panel of outcomes on health care utilization for the period from 24 months before to 24 months after the month of application decision (total of 49 months). For most of the analysis, we aggre- gate the data to the quarterly frequency to reduce the noise inherent in higher-frequency measures. We define quarter 0 as months 0, -1, and -2, where month 0 is the month of the application decision; quarter 1 contains months 1, 2 and 3 after application; and so on. Our primary measures of health care utilization are indicators for whether a patient had at least one (i) ambulatory visit, (ii) inpatient visit, (iii) emergency department visit, (iv) visit of any kind (i.e., ambulatory, inpatient, or emergency department), and (v) the number of prescription drug days supplied, in the quarter. Our laboratory test data cover tests for cholesterol (HDL and total), triglycerides, and blood sugar (A1C). Cholesterol and triglycerides tests are used to assess risk and guide treatment of heart disease; A1C tests are used to manage treatment of diabetes. We construct an indicator for whether a patient had at least one laboratory test and whether the test returned at least one abnormal result.' Table 2 provides summary statistics on our baseline sample in quarter -1 (covering months -3, -4, and -5, relative to the month of application decision). Column 1 shows means for the baseline sample with incomes between 150% and 550% of FPL. On average, applicants have a family income of $40,000, have a family size of slightly more than 2, and are 58 years old. The average Elixhauser Comorbidity Index (count of chronic conditions) 7We follow clinical guidelines and define an abnormal test as less than 40 mg per deciliter for HDL choles- terol level, 240 mg per deciliter or higher for total cholesterol, 150 mg per deciliter or higher for triglycerides, and 6.5% or higher for A1C for adults (Ma and Shieh, 2006; International Expert Committee, 2009). We use the corresponding thresholds from the same source for the small number of children in our sample. is 3.5 over the 12 months before application, higher than the average of 2.2 in the US pop- ulation in 2013 (Akinyemiju and Moore, 2016). The average BMI among the applicants is 29.0, which is similar to the US population mean for both men (29.1) and women (29.6) over 20 (Fryar et al., 2018). Most (88%) of applicants are enrolled in a Kaiser plan at the time of application, of which 52% are covered by a Medicare Advantage plan, and 48% are enrolled in a commercial plan (e.g., employer-sponsored or Marketplace). As expected given the circumstances, applicants had high health care costs in the months before ap- plication. In quarter -1, average costs were slightly more than $6,000, substantially higher than the $2,686 average quarterly expenditure for Americans in 2017 (National Center for Health Statistics, 2021). Applicants' healthcare utilization are also correspondingly higher than the US population average. Columns 2 and 3 of Table 2 report means for applicants with incomes between 250- 350% and 350-450% of the FPL threshold, respectively. The small differences indicate that applicants on either side of the eligibility threshold are similar along key characteristics and pre-intervention outcomes. We conduct formal tests of the validity of the RD design, described below. IV Empirical Strategy We use an RD design that exploits the program eligibility threshold at 350% of FPL to estimate the causal effect of the financial assistance program on health care utilization. Let i index applicants and f index quarters relative to application. For each quarter t € [-7,8], we estimate separate regressions of the form: yit = Bor + BisFPL; + BoeF PL} + 6:1(FPL; < 350%) + €% (1) where y; is the outcome variable, and FPL; is the applicant's income as a percentage of FPL, which serves as the running variable. The coefficient of interest, 6;, captures the effect of having an income below the eligibility threshold on the outcome in quarter t. In our baseline specification, we include a global second-order polynomial in income to capture the relationship between income and the outcome. A global polynomial is ap- propriate if the relationship between y; and income would not exhibit a structural break at 350% in the absence of the financial assistance program, and allows us to accommodate the relatively small mass of observations to the right of the threshold. We show in sensitivity analysis that our results are similar (although noisier for some outcomes) when we allow for separate polynomials above and below the threshold. Our identifying assumption is that in the absence of the discontinuity in program ap- proval, outcomes would trend smoothly through the discontinuity. Below, we present three pieces of evidence in support of this assumption. First, we show that applicant char- acteristics, such as demographics and chronic health conditions, trend smoothly through the discontinuity. Second, as a placebo test, we show there are no effects on outcomes be- fore application. Finally, we test for bunching of applicants below the eligibility threshold. The validity of the research design is also supported by the institutional environment. Patients must document their income (e.g., by submitting a pay stub) which reduces the scope for manipulation of the running variable. Applications are typically submitted fol- lowing a health event, limiting income manipulation through the retiming of applications. As we show below, while virtually all applicants with incomes below the threshold qualify for the financial assistance program, a small share of applicants with incomes above the threshold qualify via the expense-based criteria. To account for these approvals, we estimate instrumental variable (fuzzy RD) specifications. Letting FAP; be an indicator for approval, the first stage takes the form: FAP; = alY + alV FPL; + ofY FPL? + 7} 1(FPL; < 350%) + vit (2) and the second stage takes the form: yit = Bot + Big FPLi + Boy FPL} + 5," FAP; + ej (3) Since we have a single instrument, the IV coefficient of interest 5!" is numerically equiva- lent to the RD estimate 4; divided by the first stage effect 71". 10 V_ Results A Balance and First Stage To examine selection into the program, we first test for discontinuities in pre-application patient characteristics and outcomes around the eligibility cutoff. Columns 4 and 5 of Table 2 report the coefficient and associated p-value on the indicator for having income below the 350% FPL threshold from our baseline RD specification (equation 1). The top panel shows effects for patient demographics and pre-application health conditions (Elixhauser comorbidity score, BMI, smoker). The middle panel shows effects on measures of insur- ance coverage (insured and Medicare Advantage) and quarterly costs, measured in the quarter before application. Second, we implement placebo tests for differences in the outcomes before application. The bottom panel of Table 2 shows effects of our utilization outcomes, again measured in the quarter before application. The estimated effects are quantitatively small and statis- tically insignificant for every variable, supporting the assumption that outcomes would have evolved smoothly through the threshold in the absence of the financial assistance program. Appendix Figure Al displays the RD plots for the quarter -1 placebo tests. Third, we examine bunching of applications around the threshold value, which could indicate selection into the program on unobservable characteristics. Figure 1, Panel A presents a histogram of applicant income as a percentage of FPL for the unrestricted data and Panel B presents a histogram for our baseline sample of applicants with incomes be- tween 150% to 550% FPL. Neither histogram shows visual evidence of excess mass to the left of the threshold. Appendix Table A2 presents results from the manipulation test de- veloped by Cattaneo et al. (2020). The Cattaneo et al. (2020) test (henceforth, CJM) yields a p-value of 0.202, indicating the absence of manipulation. We view the CJM test as con- servative since the test over-rejects in a validation exercise. When applied to our data for placebo thresholds at 1% increments between 200% and 500% FPL, the test results in p- values of less than 0.05 for 16.6% of thresholds. See Appendix Section B for more details.® 8The problem of over-rejection appears to be even more severe for the commonly used McCrary (2008) test. In Appendix Section B, we show that the McCrary test rejects 40.5% of placebo thresholds between 200% and. 500% FPL, suggesting that this test is not well-suited to detecting manipulation of the running variable in our 11 We estimate a large first-stage effect of the cutoff on application approval. Figure 1, Panel C plots approval rates by FPL, along with fitted values from the first stage (equation 2). Applicants below the cutoff have a 78.8 pp (standard error 1.7 pp) higher likelihood of being approved for the program. Virtually all applicants with incomes below the cutoff are approved, while about one-fifth of applicants above the cutoff qualify via the expense- based criteria. B_ Utilization Impacts We next turn to estimating the effects of the financial assistance program on health care uti- lization. Figure 2 presents RD plots of the impact of financial assistance in the first quarter following application decision. For each outcome, dots show the mean of the outcome variable for 85 equal-frequency bins of income.' The solid lines show predicted values from the RD specification (equation 1) and dashed lines show the 95% confidence inter- vals. For each outcome, we also report the RD and IV estimates, their standard errors, and the mean of the outcome for applicants with an income of 350-450% of FPL (i.e., the "control group" mean). The IV estimates indicate substantial increases in utilization in the quarter after receiv- ing financial assistance. Financial assistance increases the likelihood of any ambulatory encounter by 13.4 pp (relative to a mean of 67.0% for those with income between 350-450% FPL), any inpatient encounter by 3.6 pp (relative to a mean of 6.2%), and any emergency department (ED) encounter by 6.7 pp (relative to a mean of 12.7%). The likelihood of any encounter increases by 13.0 pp (relative to a mean of 68.4%) and the number of prescription drug days supplied increases by 32.0 (relative to a mean of 136.0). All of these estimates are statistically distinguishable from 0 at the 5% level. We find that financial assistance has a large proportional effect on the likelihood of undertaking a laboratory test and on the detection of an abnormal test outcome in the quarter after program decision. Specifically, the likelihood of any test increases by 7.1 pp (relative to a mean of 19.4%) and, unconditional on having a test, the likelihood of an setting. °The number of bins was chosen to achieve bins with approximately equal observation counts on both sides of the discontinuity. Panel G has 31 bins due to its smaller sample size. 12 abnormal test increases 4.0 pp (relative to a mean of 10.0%). The increase in abnormal test results is driven by the extensive margin increase in testing, given that the conditional likelihood of an abnormal test result is unchanged (point estimate of 1.0 pp and standard error of 6.7 pp relative to a mean of 51.5%). Hence, marginal testing conducted in response to financial assistance leads to valuable diagnosis of abnormal health and is of similar diagnostic value to inframarginal testing. In Appendix Table A3, we further examine effects on abnormal test results and pre- scription drug utilization separately by chronic condition that benefit from diagnosis and management. Financial assistance almost doubles the likelihood of an abnormal choles- terol test (2.6 pp increase relative to a mean of 2.7%; p-value of 0.007) and raises utilization of cholesterol lowering drugs by a marginally significant 19% (4.9 percentage point in- crease relative to a mean of 26.2%; p-value of 0.08).!° The effect on diabetes diagnoses is imprecise, however we estimate a marginally significant 26% increase in prescriptions for diabetes (3.7 pp increase relative to a mean of 14.1%; p-value of 0.087). Financial assistance increases utilization for drugs to treat depression by about one-third (5.5 percentage points relative to a mean of 14.9%; p-value of 0.16) and has an imprecise effect on prescriptions for blood pressure. Taken together, the increased detection of abnormal health conditions - which are a precondition for appropriate treatment - along with the increased drug utilization for treatment-sensitive conditions suggest that financial assistance increases the use of high- value care for at least some patients. C Dynamics Figure 3 examines the dynamic effects of financial assistance on utilization by plotting RD estimates, and the corresponding 95% confidence intervals, for quarters t € [-7, 8]. The plots show no effect in quarters -7 to -1, supporting the validity of the research design. For some outcomes, there is a small effect in quarter 0, which is natural since program approval occurs in the last month of this quarter. Impacts of financial assistance are largest 10We cannot separate whether the increase in drug utilization is directly caused by the financial assistance or by the diagnostic testing. Numerical values of the regression discontinuity and the IV estimates for each outcome in each quarter are reported in Appendix Tables A4 and A5. 13 in quarter 1, half as large and not statistically significant for any of our primary outcomes in quarter 2, and small and generally not statistically significant in quarters 3 to 8. Thus, while the impacts in the months immediately following program receipt are substantial, cumulative effects over longer time horizons are smaller. D_ Sensitivity Analysis We conduct several sets of robustness analyses to address standard concerns surround- ing RD designs, which complement the tests for balance and manipulation presented in Section V.A and the tests for impacts before the program presented in Section V.B. Our baseline model controls for a global second-order polynomial in income rather than separate polynomials above and below the discontinuity. This specification assumes that there is not a structural break in the underlying relationship between health care uti- lization and income that could not be captured by this global polynomial. To probe the sensitivity to this assumption, Panel A of Appendix Table A6 shows estimates that control for separate second-order polynomials above and below the cutoff, and Panel B controls for separate local linear regressions using the approach proposed by Calonico et al. (2014). The estimates are similar up to some additional noise due to the relatively lower number of observations above the income the cutoff. To provide further assurance that our results are not being driven by observations close to the income cutoff, we re-estimate our main specification using a "donut" RD de- sign where we exclude observations with incomes between 340% and 360% of FPL (see Appendix Table A6, Panel C). The estimated impacts are very similar to our baseline spec- ification. VI Discussion The goal of financial assistance programs is to provide financial relief to patients for previ- ously incurred health care expenses and to prevent financial considerations from discour- aging ongoing health care utilization to the detriment of patient health. Our setting, which offers rich data on utilization from a closed network and quasi-experimental variation in program approval, provides an ideal opportunity to study the health care utilization ef- 14 fects of a representative financial assistance program." The on-impact effects of the financial assistance program on health care utilization are proportionally similar to the effects of Medicaid on health care utilization. Appendix Ta- ble A7 summarizes the IV effects from quarter 1 to the most closely related estimates from the Oregon Health Insurance Experiment (Finkelstein et al., 2012; Taubman et al., 2014). Relative to the mean for applicants above the threshold (350-450% of FPL), the financial assistance program caused a 20.0% increase in the likelihood of an ambulatory visit, 58.1% increase in the likelihood of an inpatient encounter, 52.8% increase in the likelihood of an emergency department visit, and a 23.6% increase in prescription drug days supplied. Based on the Oregon experiment, Medicaid caused a 36.9% increase in the number of out- patient visits (comparable to ambulatory visits), 10.7% increase in the likelihood of an in- patient encounter, 20.3% increase in the likelihood of an emergency department visit, and a 15.0% increase in the likelihood of filling a prescription. A key difference is that while the impacts of Medicaid are based on averages a year and a half after Medicaid uptake, the impacts of the financial assistance program fade out by the third quarter after program approval. Thus, the cumulative impact of the financial assistance program over a year-and-a-half period are smaller than those from Medicaid. To a certain extent, the transitory impact of the financial assistance program is not surprising. The program is designed to provide one-time debt relief and elimination of cost sharing for 6 to 12 months.!2 Applicants may be patients for whom the cost-sharing reductions provided by the financial assistance program has a particularly large effect (i.e., a version of "selection on moral hazard" documented by Einav et al. (2013)). Our findings of increased abnormal test results for treatment-sensitive conditions (heart disease and diabetes), and increased prescription drug utilization for chronic health con- ditions (cholesterol, diabetes) and depression, indicate that at least some of the increase in health care utilization caused by the financial assistance program is of high value. These results are consistent with an emerging set of evidence that consumer cost-sharing has l2While our research design allows us to credibly estimate the impact of Kaiser's financial assistance pro- gram, which bundles debt relief and cost-sharing reductions; it does not allow us to separately estimate the price and wealth effects of the program. Our results are nonetheless relevant to financial assistance policies more generally, given that many hospitals' programs are similarly constructed (see Table 1). 13We do not find any differential fade-out for applicants with 6- versus 12-month reductions in cost sharing. 15 detrimental impacts on the use of high-value care (Chandra et al., 2010; Brot-Goldberg et al., 2017). Indeed, our findings of increased high-value utilization is especially notable given that 88% of our sample has insurance coverage. While the effects of financial assis- tance on the insured is of stand-alone interest - given increasing cost-sharing and the $6 billion in charity care currently provided to the insured - it is reasonable to project that the effects of financial assistance are even larger for uninsured patients. VII Conclusion This study uses a regression discontinuity design to estimate the impact of financial as- sistance programs on health care utilization. Financial assistance, in the form of debt for- giveness and reduced cost-sharing, has large on-impact effects on health care utilization, although these effects largely fade out within 3 quarters after program receipt. Finan- cial assistance also increases the likelihood of abnormal test results and drug utilization for treatment-sensitive conditions, which is suggestive of increases in high-value care that could translate into improvements in health. Given the intention of these programs to reduce barriers to care, our findings should be of interest to hospitals, which design and implement these programs, and patient ad- vocates, who help patients apply for financial assistance. Our results are also relevant to the ongoing debate surrounding the $16 billion spent on charity care by non-profit hos- pitals annually, which partly justifies the large tax exemptions granted to these hospitals (Bai et al., 2021). Our results are relevant to efforts at the state and federal level to expand financial assistance, and are broadly related to debates about expanding health insurance, for which hospital financial assistance programs are a substitute (Garthwaite et al., 2018). The analysis in our paper takes the set of applicants to the financial assistance program as given. However, hospitals may under-promote their financial assistance programs, and potential applicants may be deterred by burdensome documentation requirements.' Un- derstanding the application process and program take-up is an important area for future work. M4See, for example, https://www.wsj.com/articles/medical-debt-charity-to-buy-wipe- out-278-million--of-patients-hospital-bills-11623762001?st=pa8wazovx3vsstn& reflink=desktopwebshare_permalink 16 References Akinyemiju, Tomi, and Justin Xavier Moore. 2016. "Data on burden of comorbidities in the United States and Medicaid expansion status." Data in Brief, 8: 120-122. Bai, Ge, Hossein Zare, Matthew D. Eisenberg, Daniel Polsky, and Gerard F. Anderson. 2021. "Analysis suggests government and nonprofit hospitals' charity care is not aligned with their favorable tax treatment: Study examines government and nonprofit hospital charity care expenses compared to charity care obligations arising from the organiza- tions' favorable tax treatment." Health Affairs, 40(4): 629-636. Brot-Goldberg, Zarek C., Amitabh Chandra, Benjamin R. Handel, and Jonathan T. Kol- stad. 2017. "What does a deductible do? The impact of cost-sharing on health care prices, quantities, and spending dynamics." The Quarterly Journal of Economics, 132(3): 1261- 1318. Calonico, Sebastian, Matias D. Cattaneo, and Rocio Titiunik. 2014. "Robust nonparamet- ric confidence intervals for regression-discontinuity designs." Econometrica, 82(6): 2295- 2326. Camilleri, Susan. 2018. "The ACA Medicaid expansion, disproportionate share hospitals, and uncompensated care." Health Services Research, 53(3): 1562-1580. Cattaneo, Matias D, Michael Jansson, and Xinwei Ma. 2020. "Simple local polynomial density estimators." Journal of the American Statistical Association, 115(531): 1449-1455. Chaiyachati, Krisda H., Mingyu Qi, and Rachel M. Werner. 2020. "Nonprofit Hospi- tal community benefit spending and readmission rates." Population Health Management, 23(1): 85-91. Chandra, Amitabh, Jonathan Gruber, and Robin McKnight. 2010. "Patient cost-sharing and hospitalization offsets in the elderly." American Economic Review, 100(1): 193-213. 17 Conner, Douglas A., Arne Beck, Christina Clarke, Leslie Wright, Komal Narwaney, and Neysa W. Bermingham. 2013. "Quality and cost evaluation of a medical financial assis- tance program." The Permanente Journal, 17(1): 31. Cunningham, Peter J., and Ha T. Tu. 1997. "Trends: A changing picture of uncompensated care." Health Affairs, 16(4): 167-175. Dranove, David, Craig Garthwaite, and Christopher Ody. 2016. "Uncompensated care decreased at hospitals in Medicaid expansion states but not at hospitals in nonexpansion states." Health Affairs, 35(8): 1471-1479. Einav, Liran, Amy Finkelstein, Stephen P. Ryan, Paul Schrimpf, and Mark R. Cullen. 2013. "Selection on moral hazard in health insurance." American Economic Review, 103(1): 178-219. Felder, Tisha M., Lincy S. Lal, Charles L. Bennett, Frank Hung, and Luisa Franzini. 2011. "Cancer patients' use of pharmaceutical patient assistance programs in the outpatient pharmacy at a large tertiary cancer center." Community Oncology, 8(6): 279. Finkelstein, Amy, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gruber, Joseph P. Newhouse, Heidi Allen, Katherine Baicker, and Oregon Health Study Group. 2012. "The Oregon Health Insurance Experiment: Evidence from the first year." The Quarterly Journal of Economics, 127(3): 1057-1106. Fryar, Cheryl D., Deanna Kruszan-Moran, Qiuping Gu, and Cynthia L. Ogden. 2018. "Mean body weight, weight, waist circumference, and body mass index among adults: United States, 1999-2000 through 2015-2016." Garthwaite, Craig, Tal Gross, and Matthew J. Notowidigdo. 2018. "Hospitals as insurers of last resort." American Economic Journal: Applied Economics, 10(1): 1-39. Herring, Bradley, Darrell Gaskin, Hossein Zare, and Gerard Anderson. 2018. "Com- paring the value of nonprofit hospitals' tax exemption to their community ben- efits." INQUIRY: The Journal of Health Care Organization, Provision, and Financing, 55: 0046958017751970. 18 International Expert Committee. 2009. "International Expert Committee report on the role of the A1C assay in the diagnosis of diabetes." Diabetes Care, 32(7): 1327-1334. IRS. n.d.. "Charitable Hospitals - General requirements for tax-exemption under Section 501(c)(3)." James, Julia. 2016. "Health policy brief: Nonprofit Hospitals' community benefit require- ments." Health Affairs. Ma, Hongbao, and Kuan-Jiunn Shieh. 2006. "Cholesterol and human health." The Journal of American Science, 2(1): 46-50. Mann, Joyce M., Glenn A. Melnick, Anil Bamezai, and Jack Zwanziger. 1997. "A profile of uncompensated hospital care, 1983-1995." Health Affairs, 16(4): 223-232. McCrary, Justin. 2008. "Manipulation of the running variable in the regression disconti- nuity design: A density test." Journal of Econometrics, 142(2): 698-714. National Center for Health Statistics. 2021. "National Center for Health Statistics. Health, United States, 2019: Table 44." Paul, Christine, Allison Boyes, Alix Hall, Alessandra Bisquera, Annie Miller, and Lorna O'Brien. 2016. "The impact of cancer diagnosis and treatment on employment, income, treatment decisions and financial assistance and their relationship to socioeconomic and disease factors." Supportive Care in Cancer, 24(11): 4739-4746. Roth, Luke, Jessica Naber, Luke Metz, and Nina Nikolova. 2021. "Hospital care for the uninsured in the United States: An analysis of national data sources." Semin, Jessica N., David Palm, Lynette M. Smith, and Sarah Ruttle. 2020. "Understand- ing breast cancer survivors' financial burden and distress after financial assistance." Sup- portive Care in Cancer, 1-8. Singh, Simone R., Gary J. Young, Shoou-Yih Daniel Lee, Paula H. Song, and Jeffrey A. Alexander. 2015. "Analysis of hospital community benefit expenditures' alignment with 19 community health needs: Evidence from a national investigation of tax-exempt hospi- tals." American Journal of Public Health, 105(5): 914-921. Somerville, Martha H. 2012. "Community benefit in context: Origins and Evolution-ACA Section 9007." Spencer, Jennifer C., Cleo A. Samuel, Donald L. Rosenstein, Katherine E. Reeder-Hayes, Michelle L. Manning, Jean B. Sellers, and Stephanie B. Wheeler. 2018. "Oncology navigators' perceptions of cancer-related financial burden and financial assistance re- sources." Supportive Care in Cancer, 26(4): 1315-1321. Stark, Andrea Bopp. 2020. "An ounce of prevention: A review of hospital financial assis- tance policies in the States." Starr, Paul. 2008. The social transformation of American medicine: The rise of a sovereign profes- sion and the making of a vast industry. New York:Basic Books. Taubman, Sarah L., Heidi L. Allen, Bill J. Wright, Katherine Baicker, and Amy N Finkel- stein. 2014. "Medicaid increases emergency-department use: Evidence from Oregon's Health Insurance Experiment." Science, 343(6168): 263-268. Young, Gary J., Chia-Hung Chou, Jeffrey Alexander, Shoou-Yih Daniel Lee, and Eli Raver. 2013. "Provision of community benefits by tax-exempt US hospitals." New Eng- land Journal of Medicine, 368(16): 1519-1527. Zullig, Leah L., Steven Wolf, Lisa Vlastelica, Veena Shankaran, and S.Yousuf Zafar. 2017. "The role of patient financial assistance programs in reducing costs for cancer pa- tients." Journal of Managed Care & Specialty Pharmacy, 23(A4): 407-411. 20 Table 1: Hospital Financial Assistance Policies Hospitals* Eligibility Criteriat Benefits Panel A. Largest For-Profit Health Systems Health System. Number of HCA 185 Healthcare Community 105 Health Systems Tenet Healthcare 65 Universal 26 Health Services Income < 200% FPL Income between 200 and 400% FPL Income < 200% FPL Income between 201 and 301% FPL Income < 200% FPL Income < 200% FPL Income between 201 and 250% FPL Income between 251 and 300% FPL Panel B. Largest Non-Profit Health Systems Ascension 151 Health? Trinity Health 92 Providence 51 Health Atrium 50 Health Kaiser Permanente 39 § Income <= 250% FPL Income between 250 and 350% FPL Income between 351 and 400% FPL Income below 250% FPL Income <= 300% FPL Income between 301 and 350% FPL Income <= 200% FPL Income between 201 and 300% FPL Income between 301 and 400% FPL Income <= 350% FPL 100% write-off of emergency services costs Cap out-of-pocket balances at 4% of annual income using a sliding scale Receive care for free Receive care discounted to the amount generally billed to Medicare 100% charity care discount 100% discount off gross charges 83.5% discount off gross charges 67% discount off gross charges 100% discount off patient responsibility amounts 75% discount off patient responsibility amounts 67% discount off patient responsibility amounts 100% discount on patient financial obligations 100% write-off on patient responsibility amounts 75% discount off patient responsibility amounts 100% discount on eligible services for 180 days 75% discount on eligible services for 180 days 50% discount on eligible services for 180 days 100% discount on patient responsibility bills; may also include an eligibility period for follow up services * Based on information as of July 2019. See Appendix A for details for an expanded list and details on the construction of this table. t The table lists income-based eligibility criteria only. The exact terms for each financial assistance program may include other eligibility requirements such as patient insurance status, medical expenditure incurred, and asset level, which in turn might be associated with different benefit terms. $ Eligibility criteria vary by state and by hospital. The policy here pertains to Lourdes Hospital in Binghamton, NY. § Eligibility criteria vary by region. The policy here pertains to Kaiser Permenente Northern California. 21 Table 2: Summary Statistics and Covariate Balance Sample Mean Discontinuity 150-550% FPL 250-350% FPL 350-450% FPL RD Estimate p-value Q) 2) B) (4) 6) Demographics at Application Family Income ($) 40299.29 49355.72 65714.77 -197.67 0.82 Family Size 2.15 2.19 2.17 -0.00 0.97 White (%) 50.65 50.99 48.24 -0.00 0.85 Male (%) 42.65 44.38 48.43 --0.04 0.13 Age 57.77 56.57 55.13 -0.42 0.71 Elixhauser Comorbidity Index* 3.52 3.45 3.36 0.13 0.43 Body Mass Index (BMI) 29.05 29.27 29.23 0.34 0.35 Ever Smoked (%) 42.34 41.63 41.80 -0.02 0.46 Insurance and Cost in Quarter -1! Insured (%) 88.12 89.48 90.47 -0.01 0.73 Medicare Advantage (%) 52.40 49.76 46.23 -0.01 0.67 Total Cost ($) 6094.89 6461.90 6884.35 -132.03 0.92 Key Outcomes in Quarter -1 Any Encounter (%) 67.05 67.98 69.11 -0.00 0.93 Any Ambulatory Encounter (%) 65.36 66.43 67.78 -0.00 0.93 Any Inpatient Encounter (%) 7.30 7.45 7.72 -0.00 0.92 Any Emergency Department Encounter (%) 15.24 14.90 14.01 0.01 0.37 Prescription Drug Days Supplied § 132.50 133.83 132.60 -3.77 0.64 Any Test Record (%) 21.70 21.88 24.02 -0.02 0.26 Any Abnormal Test Result (Unconditional) (%) 11.61 12.18 12.39 0.00 0.98 Any Abnormal Test Result Conditional on Test (%) 53.48 55.66 51.59 0.06 0.28 Note: Quarter -1 corresponds to event months -3, -4, and -5 relative to the quarter of application decision. The income eligibility cutoff is 350% of FPL. Column 1 shows means for the baseline sample with income between 150% and 550% of FPL; columns 2 and 3 show means for applicants within 100 percentage points of the 350% FPL threshold. Columns 4 and 5 report coefficient estimates and p-values on an indicator for income below the 350% of FPL threshold from our baseline regression discontinuity specification (equation 1). *Calculated for the 12 months prior to program application. + The insurance coverage variables are indicators for being insured in all months of quarter -1. § Winsorized at the 95th percentile. 22 Figure 1: Distribution of Applicant Income and First Stage A. Distribution of Applicant Income 1200 @ 800 c oO > oO oO 2 400 0 2 0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 Income (% FPL) B. Distribution of Applicant Income in Baseline Sample 500 : : T-statistic: 1.277 400 p-value: 0.202 & 300 | oO a oO ® 200 We 100 0 : 150 200 250 300 350 400 450 500 550 Income (% FPL) C. First Stage 1.00 ances | RD Estimate (S.E.) = 0.788 (0.017) B © 0.75 © Q < = 0.50 2 oO © Le 0.25 . 150 200 250 300 350 400 450 500 550 Income (% FPL) Note: Panel A shows the distribution of applicant income, with the gray region showing applicants outside of our baseline sample (150-550% of FPL). Panel B shows the distribution of income for the baseline sample and reports the t-statistic and p-value from the manipulation test proposed by Cattaneo, Jansson, and Ma (2020). Panel C shows the first stage effect on approval at 350% of FPL. Dots show the mean approval rate within equal-frequency income bins. Solid lines are fitted values from a second-order polynomial; gray bands show a 95% confidence interval. 23 Figure 2: RD Estimates for Quarter 1 A. Any Ambulatory Encounter B. Any Inpatient Encounter I RF: 0.106 (0.023) 15 RF: 0.029 (0.012) IV: 0.134 (0.029) IV: 0.036 (0.016) Mean: 0.670 Mean: 0.062 Z A - Z . - s | a 05 oo ~~ e a 0 T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) C. Any Emergency Department Encounter D. Any Encounter (Ambulatory, Inpatient, ED) I 25 I RF: 0.053 (0.017) 85 RF: 0.102 (0.023) . . | IV: 0.067 (0.022) IV: 0.130 (0.029) 2 * 7 Mean: 0.127 8 Mean: 0.684 2 | / 7 e .75 15 . 6 it s e A 65 T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) E. Prescription Drug Days Supplied F. Any Lab Test I RF: 21.674 (8.299) 3 I RF: 0.056 (0.020) IV: 27.519 (10.589) ° «| IV: 0.071 (0.025) Mean:0.194 Mean:131.203 9 T T T T T T T T T T T T T T T 1 50 25 50 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) G. Any Abnormal Lab Result (Unconditional) Income (% of FPL) H. Any Abnormal Lab Result (Conditional on Test) 2 | RF: 0.031 (0.015) 7 I RF: 0.008 (0.055) | IV: 0.040 (0.019) | IV: 0.010-(0.067) PS Mean: 0.100 6 .« * i sma __ Mean: 0.515 15 * | " ae" e e : s | ° . * 8, o ° . e vise all 5 a ee el " en ' ." . " | *_ -~ - 4 e e e -_- ~ ~ os ® . - ee 4 Ss ° | ~~ | * .05 | 3 T T T T T T T T T T 1 T T T T 1 50 200 250 300 350 400 450 500 £550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) Note: Figure shows regression discontinuity plots of the impact of financial assistance in the first quarter after the application decision. Dots show the mean of the outcome for 85 equal-frequency bins (220 applicants per bin, except for Panel G where there are 130 applicants per bin). Solid lines show fitted values from a second-order polynomial; dashed lines show 95% confidence intervals. For each outcome, we also report the RD and IV estimates, their standard errors, and the mean of the outcome for applicants with an income of 350-450% of FPL (i.e., the "control group" mean). Prescription Drug Days Supplied is winsorized at the 95th percentile. 24 Figure 3: RD Estimates for Each Quarter A. Any Ambulatory Encounter B. Any Inpatient Encounter 0.15 | 0.06 0.10 0.04 g g £ 0.05 - 0.02 a ii 0.00 0.00 -0.05 | -0.02 l T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T 7 6 5 -4 -3-2-1012 3 4 5 67 8 7 6 5 4 -3-2-10123 4 5 67 8 Event Quarter Event Quarter C. Any Emergency Department Encounter D. Any Encounter (Ambulatory, Inpatient, ED) 0.10 015 | 0.10 @ 0.05 2 oO oO £ £ 0.05 i fi 0.00 - 0.00 -0.05 | 0.05 | T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T 7 6 5 -4-3-2-1012 3 4 5 67 8 7 6 5 4 -3-2-10123 45 67 8 Event Quarter Event Quarter E. Prescription Drug Days Supplied F. Any Lab Test 40.00 | 0.10 | 0.05 @ 20.00 o z z s £ 0.004 wu uw 0.00 - - -0.05 | -20.00 | -0.10 | T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T 7 6-5 4 -3-2-10123 45 6 7 8 7 6 5 -4-3-2-1012 3 45 67 8 Event Quarter Event Quarter G. Any Abnormal Lab Test (Unconditional) H. Any Abnormal Lab Test (Conditional on Any Test) 0.06 | 0.25 | O04 0.15 2 0.02 2 £- E 0.05 iy 0.00 - rt -0.02 -0.05 | -0.04 | -0.15 | T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T 7 6-5 4-3-2 -10%12 3 45 67 8 7 6 5 -4-3-2-10123 45 67 8 Event Quarter Event Quarter Note: Plots show regression discontinuity estimates, and the corresponding 95% confidence intervals, of the impact of financial assistance for quarters t € [-7,8]. Event quarter 0 corresponds to event months 0, -1, and -2 relative to the month of the application decision. Prescription Drug Days Supplied is winsorized at the 95th percentile. 25 The Impact of Financial Assistance Programs on Health Care Utilization Online Appendix Alyce Adams Raymond Kluender Neale Mahoney Jinglin Wang Francis Wong Wesley Yin A Hospital Financial Assistance Policies We focus on the 40 largest health systems by number of hospitals as of July 2019, com- piled by Becker's Hospital Review (www. beckershospitalreview.com/largest- hospitals--and-health-systems--in-america-2019). To determine whether a health system has a financial assistance program, we search on the health system's organization website using keywords such as financial assistance and charity care. For eligibility crite- ria and benefits, we refer to the most recent financial assistance/charity care policy doc- uments available on the organization's website. We record only income-based eligibility criteria and use the organization's own language to describe the benefits (with small modi- fications for succinctness). To determine whether a health system is not-for-profit, we refer primarily to the organization's website (or other sources found via internet search if such information is not available on the organization's website). B Manipulation Tests Appendix Table A2 reports results from manipulation tests of the density of applicants around the 350% FPL threshold. For reference, the first column reports the coefficient on an indicator for income less than the 350% FPL threshold from the first stage regression (equation 2). The second column reports results from the manipulation test proposed by Cattaneo et al. (2020, henceforth CJM) using the recommended second-order polynomial 26 with bandwidths of 31.48 pp and 40.25 pp below and above the discontinuity, respectively. The p-value for the test statistic of 0.202 fails to reject the null of no manipulation. The third column reports results from the manipulation test proposed in McCrary (2008) using the recommended bin size (1.05 pp) and bandwidth (80.31 pp). The p-value of for this test rejects the null of no manipulation. Because the result of the McCrary test conflicts with that from CJM, and because ex- cess mass below the cutoff is not evident in visual inspection of the density (Figure 1), we assess the performance of both methods by implementing these tests at placebo thresh- olds throughout the distribution of income in our sample (i.e. at various points that do not correspond to any relevant program cutoff). Our baseline sample is comprised of ap- plicants with an income of +/- 200% FPL around the 350% FPL threshold. We construct placebo thresholds at 1% intervals for the 301 points between 200% of FPL and 500% of FPL, and implement the CJM and McCrary tests on samples restricted to applicants +/- 200% FPL from these placebo cutoffs. As we do above, we use the recommended bin sizes and bandwidths for all of these exercises. Appendix Figure A3 plots the resulting p-values of the test statistics against the placebo thresholds from this exercise. The CJM test (Panel A) is moderately prone to over-rejecting the null of no manipulation, with p-values of less than 0.05 for 16.6% of placebo thresholds. In comparison, the McCrary test (Panel B) is much more biased towards over-rejection, re- jecting the null with a p-value below 0.05 in 40.5% for placebo thresholds. Based on this simulation, we conclude that the McCrary test is not well-suited to our environment. We view the fact that the CJM moderately over-rejects on average but fails to reject at the true 350% threshold as fairly strong evidence in support of the research design. 27 adod yxau uo panujzuoy S[]}q [ROTPAUT Jo YO-s}1IM %OOT "Tdi %00Z => euroouy os WeeHweapy -IT. sAep (ST 10} SBdTAras a]qISI[a UO yUNODSTP %0G Idi %00P PUR LOE U2EMj9q eUTODUT sAePp (ST 10} S8dTAIOS B[qIST[S UO JUNODSIP %G/, Idi %00E PUR TOZ U2EMjaq SUTODUT sAep (ST 10} SadTAIAS a[qIST[a UO JUNODSIP %OOT "Td JO %00Z => auI0SUT og wey uminy =-OL syunoure AyqIqtsuodsar yueyed uo saSreyp [eursiio Woy yUNODSTP %GZ Idi %OSGE PUB [OE Usemjoq auTOsUT syunoure AyqTIqisuodsaz yuatyed Uo JJO-3}LLM % OT Td %O0E => auTooUy Is yeep souepraolg 6 yUNOosIp ared AyTeYD %OOT Tdi %007 "oyeq autoouy g9 STRUTS] OUST, Z suoyesy[qo [epueuy Jusyed uo yMODSIp %0OT Tdi %0G7 > awODUT 26 yay Ayu, S *Sa0LAJas ONS IO} sya} -ed areorpayy 0} pariq Ajjersue3 yunoure ay} 0} payunodsip ared aATe0ayy Idd %LOE Pue % LOZ Usemyeq suT0sUT Sura} daly JOJ ald BATA09x{ Tdi %00Z > auosuyT ZL -sAg upeayy Aqrunumut0> P syunoure Ayqiqisuodsaz yuaryed Jyo yuNODsIp %79 Tdi %00F Pue [ge usamjaq awosuy syunoure Ayqiqisuodsaz yuaryed yo yuNOdsIp % GZ Tdi %0GE pue 0¢z Usamjaq aUIOdUT syunoure Ay Iqisuodsaz yuaned Jyo UNODSTIP %OOT Td %0GZ => aurosuy ISL WTeeyY uotsusssy Z "apes SUIpT[s B SuIsn suIOoUT [enuUR jo %F ye padded are sadueTeq yoyood-jo-jnO "Id %00F Pue 0OZ Ueemjoq auTOoUT sadtAlas ADUaZIIUIa O} P3}e[AI $}S09 JO JJO-SILIM %,QOT. Tdi %007% > auWODU] SSI aIeyTePeH VOH T sjeqidsopy yyaueg Aypqisye urero0rg Jo Jaquinn| woayskg yyeoH = yURY SITIO] BOULISISSY [PIOULUL] 'LV e1qeL 28 ased Jxau uo panuizuoy syunoure Ayyiqisuodsax yuayed jyo yuNoosIp ares Aye %0OL Idd %00€ Mojeq auzosuy ge WH sMSHYyD =-s «IZ saolAras Aressadau-AT][eoIpaul IOJ ares val] "Id %0SE => autosuT 6€ auQArIeW 6L saotaras dn Moyo 103 porsad AyTiqisya ue apnypout ose Aeur 'AyqIqisuodsaz yuayed uo yUMOdISIP Y%OOT Td %0SE => auTosUT 6E DJUSULULIAY Jasieyy gL 9}PI PUMOSSIP % 07 72 aTed aATODIOY %OOF-LOE USeaMjaq sUTOoUT a}eI YUNOSSIP %g Je Aled aATaDay %00E-LEZ Usamjaq auloouy areo AyTeYD PayUNodsIp %OOTL PATPRY "Tdi %0S7 > awOoUy OF OWdnN ZI yunoosip ueprshyd % Og pue yunoosip [eydsoy %F %O0E-1EZ UBEMJaq SUIODU] yunoosip ueprshyd %0/ pue yunoosip Teydsoy %0g %OGT-1GZ UBEaMjaq aUODUT yunoostp uepishyd pur feydsoy %001 Tdi %00Z > awWOoU] Ip ARAL peTiq ATje1ua3 yunowre ay} ueY) Ja}ea13 Ou aq [LM souRTeq BuTUTeUreI ay} yey) Yons Burpue}syno aouLTeq 3y} JO yUNOUIL ayy Jo UOHONpar TenIeg Tdi %GZE pue 97z usamjaq awosuT aouryeq anp yuated Jo ssauaars10j ajyo[dutop "Idi %G7Z => BUTODUT FP UIeaL] proyures SI siseq [euUOTsai b UO oTeOs SUIpT|s B UO paseq ale> payUNOosIp aAta0oyy Idd %00F Puke LOZ UseMJaq BUTODUT WaeeH DOURISISSE [PDULUTT %OOT. Tdi %00Z => BuTOUT gp ABA sanoveg uog €1 peTig ATTerauesy yunouy au} uey} Joyear3 ou sa8reyp ssoi3 s,juaT -ed ay} Jo Y%CT JO aouepeq yUNODDe s,jUaTed ay} Jo Jessa] oY} SaMO JOnEG Tdi %00¢ pue 00z usemjeq suosUT WarPeH saoueyeq yuNOoe yueHed Surpue}s}no Uo yUNOISTP % OT Tdi %007Z > aUOSUT 8h aM % Hoos JojAeg ZI sjeqidsoyy qWeusg Aypqrsye weisorg jo Jaquiny washs yyeaH = UR aod snowadd wolf panu1jzuod - TY 21qe], 29 ased Jxau uo panuizuoy sadIeyD SsoId FO JUNOISIP %G'Eg "Idd %0S¢ PUP % LOZ USeMjoq SUTO SUT SODTA sadieyp sso13 Jyo yUNOOSTP %QOT. Tdi %007 > eu0DUT 9% -FeS Wee éjesieatup 0€ syuaped pamsut Joj 0OSZ$ JO Sse0xa Ul sdUBINSUT Jaye aduLTeq Jo syueyed pamsumm Joj GOV Jo yUMoosIp %QgG Td %O0F pue %00¢ Usemjeq auTOoUT sjuayed pamsut 1oJ 99¢Z$ JO ssaoxa UT sdUBINSUT Jaye aouLTeq JO syuayed pamsumm Joy gyYy Ho yMoosIp % Gz Tdi %00E pue % 007 Usemjeq suTOoUT syuaned pamsut 107 0Ogz$ JO SS80xe UT BoURINSUT Jaye aoueypeq 10 syuayed pamsutun 10; yunoo0e yuayed Yo yMOdsIp % OT. Tdi %00Z > awosuy 87 wpTeay rauueg Zt yuaunjsn{pe aourjstsse [epueuy enieg Td %009 pue %0¢z Usamjaq autoouy yunowre Ayyiqrsuodsar yuaryed uo yuauysnipe aourjsisse [epueuy %QOT "Tdi %0SZ => autosuy 8Z - WATeaLT Bsorny ayeooapy 9% Ked 0} a[qe paweep st yuaned ayy yunoure Aue ssaj soueyjeq anp ayy 0} jenba yunoosiq, "Idi %00Z => auT0sUT 6% - SaoTNOSaYY UATPOEY Sexa, GZ ajeos SurIpys e uo paseq soBieyp Jo ssauaars10j %G7 0} dq) Tdi %0SP Pue % [Gz Uaemjaq suIOSUT aoueqy] sa8reyo JO JJO-OFLIM [NY Tdi %067 => auodu] 6% -IV UNROE SureTg way HZ aTeos SUTpT]s B UO paseq ares Aressadau AT[eoIpsul Jo yUSSIaUta IO; saBreyp Jo ssousalB10j % 6 0} dE) Tdi %00F Pue %OST Usemjoq ouTOoUT area Aressavau AT[eoTpaul Io yUaSIaUra 10J saBIeyo Jo ssauaars1oJ %,QOT. Tdi %0ST Mopeq auroouy €€ WpTeay essay 7Z areoIpayl 0} payig AyTestaues yUnoury 9y} ye peddeo sa8reup ssoi3 yuanEg Tdi %LOF Pue %00E Usemjeq suTOoUT sjeqidsoyy qWeusg Aypqrsye weisorg jo Jaquiny washs yyeaH = UR aod snowadd wolf panu1jzuod - TY 21qe], 30 ased Jxau uo panuizuoy ATuo syuayed painsut 0} pepig Ayjeiouesy yunoury ayeos SuIpTs eB uO paseq sjuaned pamsut 0} payig AjTetauey yUNoury au} JJo yUNODsTp [eIeg "Idi %009 Pue TOP UeeMyoq SUTO SUT "Idi %00P PUP TOZ Use Meg sULOSUT JUNODSTP %OT Tdi %00Z > 2wOsUT zz WeoH wiogAyuy) = BE Ayes Aq sore, Tdi %00P < ewoouy ajeos SUIPT[s ve UO paseq aoursISse [eIDUeUy [enieg Idi %00F PUB LOZ UsaMyjeq aUTODUT yuNoosTp sourjsisse [eOURUTy %0OT Tdi %00Z > eWOsU] 2 yndeg = ZE a[eos BuIpTs ve UO peseq aourjsisse [eDuUeUg [enIeg Idi %00F PUC %LOZ UeeMjoq auOdUT yuNOosTp aourjsisse [eUeUTy %00T Tdi %002 > eur0DU] €@ WIRPEH NSS =--9E suosied painsut 0} perig AT[eIeues yuNouTY ay] ye peddeo yusged ay} 0} a[qei[Iq yUNOUTe ay] YIM aouR\sISsy [eDURULY [enIeg Tdi %00G PUP %LOL Ueemjaq aurosuy aoueysisse [ePueULy [IN Td %00T MOaq aut0suy 4 WTeeH TPMYMON = GE aoueyeq Aed-yjas ay} yo yuousn{ipe %0¢ Td %00F PUe %00Z USEMJaq SUTODUT ula} aouryeq Aed-yjas ay} jo yuauysn{pe %00T Tdi %00Z > aut0uT €z -Shg YAH STUD ode bE ares Ayreyp [Ne Tdi %00F > AWODU] bz yeeH onng = ZE ayeos SUIPTS B UO paseq soUR]sISse [eDURUy TenIeg Idi %00¢ pue %007 UseMjoq aUIODUT o1e9 aired Jo aposida sad Ayyiqisuodsaz yuoyed TeurMou ke snutu 'asouRAsIsse [[NJ Idi %007 MOjaq aulosuy FO -yeeH WeyuMowley] Te sasIeyp SsoIs JJO JUNODSTP %/9 Idi %00€ Pue %IGzZ UsaMjaq aUIODUT sjeqidsoyy qWeusg Aypqrsye weisorg jo Jaquiny washs yyeaH = UR a8od snowwadd woLf panurjzuod - TY 21qe], 31 '3[qe} ST} JO UOTN|suOD ay} UO s[IeJep JoJ Y UOT}Dag xIpuaddy sag :ajon (SIOUTTT]) Td-I %009 10 (uIsuODsT(\) a[eos SUTIPT[s B UO paseq JUNODoR JUaTed Jo yMOoSIP Teed Td %00P PUC Tdi % LOT UaeMjeq aUTODUT uaayshg yunoooe JusTyed yo JUNODSIP %OOT "Idi %00¢ > SuOdU] GI YNeaH sioysig jerdsoy OF ayeos SuIpys & Uo paseq seSreyp paliig AT[eteuesy yUNOWY uO yunoosIp TenIeg Tdi %OSF pue % G7 Usamjeq auTOoUT aourjsisse [eueUyy %00T "Idd %S7 > SUIODSUT 1? UHTeeH Peed 6€ sjetidsopzy yyaueg Ayqisye wreig0i1g jo Jaqunn wiayshs WeoF] = yureyy a8nd snotoasd wolf panuiquod - LY e1qe], Table A2: First Stage and Manipulation Tests First Stage CJM Test McCrary Test Coef. 0.7876 0.0003 0.4318 Std. Err. 0.0169 0.0002 0.0749 Test Statistic 46.6977 1.2767 5.7627 P-value 0.0000 0.2017 0.0000 Obs. 18672 18672 18672 Note: Column 1 reports the coefficient on an indicator for income less than the 350% FPL threshold from the first stage regression (equation 2). Column 2 reports results from the Cattaneo, Jansson, and Ma (2020) manipulation test using the recommend second-order polynomial with bandwidths of 31.05 pp and 39.30 pp below and above the discontinuity, respectively. Coef. is the difference between the local quadratic density estimators to either side of the cutoff; test statistic is the t-score. Column 3 reports results from a McCrary (2008) manipulation test using the recommended bin size (1.04 pp) and band- width (81.14 pp). 33 Table A3: RD Estimates for Quarter 1, Clinical Outcomes Reduced Form Instrumental Variables Control Mean Coef (SE) 95% CI P-value Coef (SE) 95% CI P-value A. Cholesterol Abnormal Cholesterol t 0.027 0.021 [ 0.006, 0.036] 0.007 0.026 [ 0.007, 0.046] 0.007 (0.008) (0.010) Any Abnormal Cholesterol Drugs 0.262 0.038 - [-0.004, 0.081] 0.079 0.049 -[-0.006, 0.103] 0.080 (0.022) (0.028) Days Supplied for Abnormal Cholesterol Drugs i 24.874 3.353 [-0.717, 7.423] 0.106 4,257 [-0.922, 9.436] 0.107 (2.077) (2.642) B. Diabetes AIC Level > 6.5 0.070 0.011 [-0.014, 0.036] 0.399 0.014 [-0.018, 0.046] 0.399 (0.013) (0.016) Any Diabetes Drugs 0.141 0.029 -[-0.004, 0.063] 0.087 0.037 [-0.005, 0.080] 0.087 (0.017) (0.022) Days Supplied for Diabetes Drugs t 18.437 5.146 [0.290,10.002] 0.038 6.533 [ 0.359,12.708] 0.038 (2477) (3.150) C. Depression Any Antidepressants 0.149 0.044 [0,.008, 0.079] 0.015 0.055 [0.011, 0.100] 0.016 (0.018) (0.023) Days Supplied for Antidepressants t 14,211 3.793 [0.311, 7.275] 0.033 4816 [ 0.384, 9.248] 0.083 (1.776) (2.261) D. Blood Pressure Any Blood Pressure Drugs 0.398 0.030 = [-0.018, 0.077] 0.219 0.038 -[-0.023, 0.098] 0.220 (0.024) (0.031) Days Supplied for Blood Pressure Drugs + 66.062 9.691 [0.358,19.025] 0.042 12.305 [ 0.410,24.199] 0.043 (4.762) (6.069) Note: Table reports regression discontinuity estimates for quarter 1 with standard errors in parentheses. + Abnormal Cholesterol is defined as having either high total cholesterol or low HDL test results at any point in the given quarter. A high total cholesterol level is defined as 240 mg per deciliter or higher for adults (age 18+) and 170 mg per deciliter or higher for non-adults. A low HDL cholesterol level is defined as less than 40 mg per deciliter for adults or less than 45 mg per deciliter for non-adults. $ Winsorized at the 95th percentile. Control mean is the mean for applicants with incomes between 350% and 450% of FPL. N = 18,672 observations. 34 *SUONPAIASO Z29'8T = N YIM suorsserSar UO paseq are Jajzenb yea Jo} seyeurysy *prooar qry e Suraey UO [eUOTIpUOD J[Nsez qui [euUIoUge AUY = qe] [eEUOYIpUOD 'prodar qe| e SuTAey Uo TeuonIpuUOOUN j[NSaI qeT [ewOUge Auy = qe] [eUOyTpuOCsU:] 'p10daI 489} qeT Auy = qey Auy '(epquaored yyg6 aug ye poziiosutm) parddns sAep 3nip uondtiosarg = xy "eyuNoousa (qq Jo 'yuayedut 'Aroyepnqure Auy = royunosuq Auy '1ayun0.u9 yueuryredap Aoua8iewe Auy = Gy Jeyunooua yuepedut Auy = yuapeduy Jayunooua Aroye~nqure Auy = Aroye_nqury "uorsPpep uopeordde Jo yyOW ay} 01 aapeyar Z- pure 'T- '9 syyuour yuaAe 0} spuodsaiioo (9 JoyIeNf) 'sasoyjuared ul SIOLIO prepurys YIM Joyenb yore 10; sayeuryjsa AyInuUOSstp uoTssersaI sy1odar afqey, :2j0N (8¢0°0) (ST0'0) (610'0) (sre'Z) (<z0°0) (FI0'0) (Z00°0) (<Z0'0) 6990 _-s FEO L260 Too'0- «Ss 6Sh'0-s«FTO'O- Ss FISO)=-ssLOZ'F)-SCFFT = 9EN'O-~-SsC«dT9FO)=-s«éOT'0-Ciés«#STGHO«Ss«TOO'O--SCés«C*CdT'TZ'O~-=riTEO*O 8 (Fg0'0) (9100) (0z0'0) (€98"Z) (¥Z0'0) (S100) (600°0) (Sz0"0) Zg¢0 = Z€0°0 Sge'0 FIOO }§6=6 9FG'0-sZ10'0-'«é"OT'0s«i«'«é9GTZTSsC%S9'D-«Cs«COTO--Ss GOTO-Ci-irOTOSCisi'iZSTOCZOOsCiCSLW'D.-s« STOO Z (9¢0°0) (sT0'0) (0z0°0) (et0'8) (7200) (710'0) (600°0) (<z0'0) r190 = 800 940°0 9700 «= gh0'0-s«étOO'D-sis'éO'N-Cé='éireZE'ST=SCHO'N-Ciéi'éiso'd:Ci'éi'éiOT:Cié'éTO:«CGZZ-«C'iONsCi'éi'éBOID-SCs«éHOV'D 9 (7¢0°0) (9T0°0) (0Z0'0) (801'8) (¥20'0) (ST0'0) (300°0) (¥Z0'0) zzg0 = SEDO 813°0 vOO0 9640 9000- 8670 S/T6 P60 Z000- 210 0700 49890 1000 98cg0 F000- ¢ (700) (910°0) (0z0'0) (Z¥0°8) (¥20'0) (ST0'0) (600°0) (¥20'0) 6800 1600 €L¥'0 1100 §=-s99F'0-s-s« STO'O- =Ss«OZT'0-s-sSTS'ZLSCOOFD-s-C'<*dTT'O:-Ci'é'éZTZ'O'-CtiC«"iSTO'D'-«C(€§s«C#CT/'-Cs«éFO/'O-s-«sCZI80~=- 9000 F (g¢0'0) (sT0'0) (0z0'0) (v0E's) (7Z0'0) (910'0) (0T0'0) (7Z0'0) 080°0 960°0 919°0 g000 = s«S6Z'0--s-«*dTZO'O--S'-«OOST'0--sCéSB'OT: «=COTS'0-s-CéSTO'N--Ss"«éBO'O-SCs«CZZO'00'-«dS$G9'0-«sS0D'0--'--«s469'0-S «6000 € (s¢0°0) (9100) (0z0'0) (7978) (¥Z0'0) (9100) (OLO'0) (¥z0'0) PeZ0 ~©=- G6 T0'0 L¥T0 sI00 }§=-s« PZT ZZO0--sCO9LD0-sCsGENFT )«=CSMT'0siéE'N--i'iT'CiC'*éTZO'O-Cié=i'éET':«SCsC«CTO'|OD--SCés«C'=dTZ'OD-Cs«LED'OD rd (s¢0'0) (ST0'0) (0z0'0) (667'8) (€z0'0) (410'0) (Z10'0) (€z0'0) 6280 - 800°0 ZE0'0 Ts0'0 }§=-ss G00'0--«ss«é#FPO'D--sis'«éiwODs«s«é='éTZ™~=Ss«iOs«'«'éZOT'-C(i'«ON--Ciés'éeGO'OD-C('és«éTZO'O-Ciés«iZ'-«C(€és«éFHKT!H'0.-=S-=«é="PP-OT'0 I (sr0'0) (610°0) (€Z0'0) (cre's) (0z0'0) (zz0'0) (0z0'0) (1Z0'0) @wz70 86700 scr'0 rIoo }§8=-s S80 ss OOO Ss OG Ss GONG )=s«ZT'0-«édTS0O «= OD|D-sCi'iDSCiC'éTEE'O=SsCé@N-Cié'éiTTO-Cé«éEOV'D 0 (zg0°0) (9100) (1z0'0) (TST'8) (€z0°0) (£100) (E100) (€z0°0) 0sz0 ~=--_- ZS0°0 €86°0 0000 §=-s«09Z0.-«édEZO'O--'-«sd9'0-Ss-sLZ'E- «=H s«ZOOO- Ss GZEO=-s STO'|D-s-s«C«é8T60-Ss«TOD'O- = S@G0- Z00°0- T- (g¢0°0) (910'0) (0z0'0) (000°8) (7200) (sT0'0) (600°0) (7Z0'0) 7960 -- €00°0- s19°0 s00'0 «so HFFO-SsCSTO'"O-sCé#RSB'IO.SCsCEZO'T)=-sC10Z'0-s«édH00-"'é'étZO=3s«é#OSTOND-Ciés«éFNT'-Cié="«éET'OD-Ciés«G'OD-Cs=«*GTOV'OD t (sg0°0) (9100) (0z0'0) (396°2) (sz0"0) (F10'0) (600°0) (sz0"0) Isv0 = 00 7920 sooo =-s«sTeZ0-'wLO'O---s-«"iLD-siCiHE'E = SZETO-CiZO'-Ciés«éMHTON-«s«s«CdOO'D-Ci'éirH' «= «000--'s-«s@EO~-s«EZO'N ¢- (890°0) (S100) (610°0) (1¥8°Z) ($z0"0) (€10'0) (g00°0) (<z0"0) 9980 O10 €16°0 Z000 }©=6 7Z6'0--s-«sTOO'O--'s-«s8GZ0~-s«éLW'Z- «= 97TO-~-s 87ND ~=-s«éT9N'0 Ss GZS 8Z0-sOO'0- Ss ZHED =: 1200 r- (090°0) (st0'0) (610'0) (222) (<z0°0) (€10'0) (z00°0) (<z0'0) 9070 ~-s- S00 €7E0 roo )}83-sesc0--Cis'étT'|?7D- «ss PH'O-iéicaSG = g990--s«éaTTO';O~Sssr8G0 = ZOO'0--Sss66TOS 600'0-_-s--s«s7Z9'0~-S-s«zT TOO ¢- (9¢0°0) (9100) (6100) (628°2) ($z0°0) (Z10'0) (800°0) (<z0"0) ooeo = S00 ctV'0 zlO0 )§=-ss«a686'0-s-«éOOW0-s«s«s«éGO-SsC«édTS6O- «=-949'0--is«éOT'0-i'é'é@Z0-SsCODOS CaSO 0'0---s«OF'!0-«ZT' 9 (690°0) (STOO) (610'0) (zgz°Z) ($z0"0) (€10'0) (z00°0) (Sz0'0) 770 0 900s eZZ0-iéTOTC'iéi'iTD:C(<«si«iTD-Ci'édeEO CEL SCGEO SC EWOSCE@™ZO-«-«*"TOO Ss SHO = S0010-",--«--«sS6Z0--"«9ZON L£ aNeAd jeep anyeA d poop aneAd POD anrAd FPO ond jo TNeRAd OD aNrAd jfoD anreAd fod JenengH qe] [euorpuod = qe'] [euonrpuocouy) qe] uy xa qayunoouq Ay da quapeduy Aoyenquy Jaren?) yoey Oj sayewysy CM PV FQeL 35 *SUONPAIASO 79'S = N YIM suorsserSar uO paseq are Jajzenb yore Joy sayeurysy *procar qe] & SULARY UO [eUOIpUOD s}[NsaI qeT [eWsOUge AUY = qe] [eUOyIpUO? 'piodar qe] e SUTARY UO TeuOHTpUOOUN s}[Nsar qe] TewIOUge Auy = qe] [euonIpucoUuA) *piooal 3sa} qey Auy = qe] Auy '(e[queorad u3¢6 ayy ye peztiosum) parddns shep 38nip uoydiosarg = xy 'C/q 10 'yueqedut 'Aroye;nquae Surpnpour reyuNooua Auy = Jayunooug Auy 'Jayumooue yueuyredap Aouasiawie Auy = Cy Jeyunoous yuayedut Auy = yueyeduy 'aymooue Aloyenqure Auy = Aroyetnqury 'uorspep uoyeoydde JO YJUOU oY} 0} BARTAI Z- pu 'T- 'Q SyYJUOW yUAaAd 0} spuodsaii0d (C JayeNg) 'sesoyyuared ul sIOLIe prepueys YIM Joyenb yee 10; soyeurysa A] soda a]qey, :ajoN (9Z0°0) (610°0) (6Z0'0) (Te€'6) (T€0'0) (810°) (600°0) (TE0'0) 1990 #00 260 z0O'0- ««6SP0)=-s STO'O---s«sFISO = 0609 PIO }8=: 90'0--="-s-édTOW'"~S-s«ETO'O~S-s«8T6'0~Ss«TOO'O~=-s«saT'TZ/O~--6€0°0 8 (890°0) (0Z0°0) (9z0°0) (<Z0'01) (T€0°0) (6100) (T10'0) (T¢0°0) 9690 OF0'0 Gge'0 sloo §=-s«oFSO0-Ss«CT'D-"'édCOT'OCisCSHOD«=s7B9'DSsCELOO- Ss GOZO) PO'DSsCZBTO SC STOO SZPO Z2'0- Z (0Z0°0) (610°0) (¢z0'0) (Z1z01) (1€0°0) (8100) (T10°0) (TE0"0) rI90 ~=-- SEO 9200 ccoo }§6=-s €PO'0--s-C'*dTSO'O.-Ciés«éO'O-Ci'éPEEZT©=«C§THOT-CiésC'ie90-'ésS=EOCD:«=CEOCOSSz@O--CSTND-«Cié=«éSOD-C"'é'éVGV'ID 9 (z90°0) (0z0°0) (¢z0°0) (FOE 01) (T€0'0) (6100) (TT0°0) (TE0°0) Tzs0 = € #00 sIs°0 Goo0 §=-s-«98SZ'0-"«OO- S's BGZTO0 GO TIL)=SEGO = 00'0-- ss Z6T'0 Ss STOO Z9B0 Ss ZODS-s«BO «9000 S (390°0) (0z0°0) (9Z0°0) (sz 01) (T¢0°0) (6100) (TT0°0) (TE0°0) 680°0 = STT'0 €Z¥'0 PIOO }§=-:99F'0-s-«CGSTO'O--Ss«OZT'0-ié'HGSGTD. «=sCO']D-Cié'DTO'D-Ci=iéATZON-CsC*#ZO'ON-Cié'iCF'O':«Ss«0DN-«ézT'O =O ZO ¥ (7Z0°0) (0Z0°0) (¢z0'0) (19¢°0T) (0¢0'0) (0Z0'0) (€T0'0) (T€0'0) 180°0 ~-s-: O€T'0 919°0 0100 §=-s TO-«C'iZO'O™-<C'é'édTCT'D-Cé«CST'ET~=s«éOTS'D-sC«éOZOTO0-Ciés'iSROCD--Cié'éO'OsSCsCéG9'0--«ébOO'0--S ss Z69'-s-«s ATO' € (g90°0) (0Z0'0) (gz0'0) (e001) (0¢0°0) (0z0"0) (€10'0) (0¢0"0) €7Z0 ~-- €Z0'0 Zb70 €z700 86 9ZT'0-sCHEO'O «= ZZO'O--sCZSBG'BTCEZT'O0 Ss FOD-iéiST'sCss«Z@O|N-Cié-"'iHT'D-Cis«ZTO'D-Cté=«'dTZT'9'-s«ZO'D Zz (z90°0) (610°0) (¢z0'0) (68°01) (6Z0°0) (2Z0'0) (910°0) (6Z0°0) 6280 ~=-Oto'd Z€0'°0 oroo0 3=-ss S0010--«s"«édTZO'0.-Cié'iOO'D:-C'éST'ZZ@~=s«éOD--«C'é'éETD:s«CiC'OIN-s«:«d9000-ié'éTZO'|:-«C(iéiétK'N-s«és'ét-SC*-*'VECT.'D T (¢¢0'0) (€Z0'0) (620'0) (009°0T) (SZ0°0) (220°0) (9Z0°0) (970°0) Trz70 ©=- $900 scr0 ZI00 »§=6 SS8'0 ss S00'0- ss HOS C*ONTT'Z 9710 = 6000S s-«sPO0'0--i'"wo4O'N-Ci'éEET'!|O:~=Os«SZO-sC*T'O-CiéPO'D 0 (390°0) (0z0°0) (9z0°0) (Fe 01) (6Z0°0) (1z0'0) (9T0'0) (0€0"0) 0sz0 =--_- ¥Z00 €86'0 0000 »§=-s«OSw0sCéOO- Ss sEP'D-i'édTSZ'F---«CEG'O «= C0N'|0- Ss FZEO )=s« GTO'0-ié'é#P3TGHO = ZOO'0-- ss SZHO- 000 I- (g90°0) (0Z0°0) (¢z0'0) (ZST01) (o¢0'0) (0Z0°0) (Z10°0) (TE0'0) 7960 -- €00°0- S190 0100 »§=-s HFOCsCéCSTO'N-séB68N CTT 10Z0 6s Z@10.0---«édT/#Z/O~=3--_-s«dTZO'!OW Ss Z9T'0-s-sCé*STO'D-Ci'éEIN':C*'«CGTU'D t (ZZ0°0) (0Z0°0) (¢z0'0) (IZT'01) (T¢€0°0) (810°0) (T10'0) (T¢0'0) Ish'0 = F400 7920 9000 +~=«édTeZ0~=-s«00'0--s-«OZ'0-s-siLOSTHCC'«GSETN-siés«SZO'D-<ié=«iBZGN-«s-«ON-siés-"'im'sCs«NDD-C'é'éE'C:C*'«EOV'D €- (0Z0°0) (610°0) (¥z0'0) (096°6) (1€0°0) (Z10°0) (Ot0'0) (TE0"0) 9980 €10°0 £160 z000 86 @Z60--s:«OO'0--s-s«BGZ0CS9E- ss G9ZO---CiseO'DSsCdT9D--Ciéw'éE'N-Ciés«BZ'Ns«C«SOO'O- «Ss ZHEO = ZZ0°0 a (020°0) (810°0) (¥z0°0) (S18°6) (T¢0'0) (Z10°0) (600°0) (TE0°0) 6070 --6S0°0 €Ze0 sioo }3=-s ess0-sCTOOSCs«CéDOV'. TSU $990 e100 7890 6000- 6610 Z100- 7290 £100 ¢- (z90°0) (0z0°0) (¢z0'0) (100°0T) (T¢0°0) (SL0°0) (OT0'0) (ZE0"0) 8670 0200 Str'0 SI0';0 }§=-s:« 686'0.-«is-i«éiO0-<i'i'ikOSC«*dTS'T--'-i-i'é'-Cié=-=*STO'D-CiéiéZD-«s"«CGDN--«ié'='éiwsSG'SCsC«C'O--Ss«aO'O--s STON o- (z90°0) (610°0) (¥Z0'0) (798°6) (z€0'0) (9L0°0) (800°0) (Z€0'0) ozo 60 F400 '-sZZO--C*dT@ONsCsC'éZTD-SC'd'OC'"ETO 9G CGE 6ZO'Ns«éE@ZO-«C"-Sé=-«CGHO«(0010----«#TO-SséEDI L anfeAd ja0p anyeaA d pod aneAd jeoD oneAg jf0OD omeAd POD MRAd PO omMNPAgd pod aneAd fo wyHenH Qe [euopHIpuoD qe] euonrpucouy) qe Auy eI qayumooug Suy aa quspeduy Aroyenqury Joven?) yoeyq oj soyeuyjsy AT :SV FGeL 36 Table A6: RD Estimates, Alternative Specifications Reduced Form. Instrumental Variables Control Mean Coef (SE) 95% CI P-value Coef (SE) 95% CI P-value Panel A. Separate Polynomials on Either Side of Cutoff Any Ambulatory Encounter 0.670 0.115 -[ 0,052, 0.178] 0.000 0.146 [ 0,066, 0.226] 0.000 (0.032) (0.041) Any Inpatient Encounter 0.062 0.033 - [-0.000, 0.067] 0.052 0.042 -_[-0.000, 0.085] 0.052 (0.017) (0.022) Any Emergency Department Encounter 0.127 0.065 [ 0,020, 0.110] 0.005 0.083 = [ 0.025, 0.141] 0.005 (0.023) (0.030) Any Encounter (Ambulatory, Inpatient, or ED) 0.684 0.116 = [ 0,054, 0.179] 0.000 0.148 = [ 0.069, 0.227] 0.000 (0.032) (0.040) Prescription Drug Days Supplied * 131.203 35.255 - [13.416,57.095] 0.002 44.807 [16.723,72.891] 0.002 (11.142) (14.329) Any Lab Test 0.194 0.064 [0.011,0.117] 0.018 0.081 [0.014, 0.149] 0.018 (0.027) (0.034) Any Abnormal Test Result (Unconditional) 0.100 0.046 [ 0.006, 0.086] 0.023 0.058 = [ 0.008, 0.109] 0.023 (0.020) (0.026) Any Abnormal Test Result (Conditional on Test) 0.515 0.078 - [-0.076, 0.232] 0.321 0.096 -_[-0.095, 0.287] 0.323 (0.079) (0.097) Panel B. Locally Linear Polynomials on Either Side of Cutoff Any Ambulatory Encounter 0.670 0.091 -'[-0.010, 0.191] 0.077 0.118 = [ 0.003, 0.233] 0.044 (0.051) (0.059) Any Inpatient Encounter 0.062 0.023 - [-0,034, 0.080] 0.427 0.031 --[-0.041, 0.102] 0.403 (0.029) (0.037) Any Emergency Department Encounter 0.127 0.057 -'[-0.017, 0.130] 0.132 0.069 --[-0.027, 0.165] 0.158 (0.038) (0.049) Any Encounter (Ambulatory, Inpatient, or ED) 0.684 0.103 -[ 0.009, 0.197] 0.031 0.131 [0.019, 0.243] 0.022 (0.048) (0.057) Prescription Drug Days Supplied * 131.203 1.442 [-35.936,38.821] 0.940 18.397 [-24.848,61.642] 0.404 (19.071) (22.064) Any Lab Test 0.194 0.079 - [-0.007, 0.166] 0.072 0.103 -_-[-0.009, 0.215] 0.071 (0.044) (0.057) Any Abnormal Test Result (Unconditional) 0.100 0.019 -'[-0,041, 0.079] 0.537 0.044 = [-0.024, 0.112] 0.202 (0.031) (0.035) Any Abnormal Test Result (Conditional on Test) 0.515 0.086 [-0.326, 0.154] (0.483 0.127 [-0.442,0.187] 0.427 (0.122) (0.160) Panel C. Donut RD Any Ambulatory Encounter 0.678 0.105 = [ 0.054, 0.157] 0.000 0.133 --[ 0.067, 0.199] 0.000 (0.026) (0.033) Any Inpatient Encounter 0.058 0.036 [ 0,009, 0.063] 0.008 0.046 [0.012, 0.080] 0.008 (0.014) (0.017) Any Emergency Department Encounter 0.128 0.054 [ 0.015, 0.092] 0.006 0.068 [0.019, 0.117] 0.007 (0.020) (0.025) Any Encounter (Ambulatory, Inpatient, or ED) 0.693 0.097 [ 0.046, 0.148] 0.000 0.123 [ 0.058, 0.188] 0.000 (0.026) (0.033) Prescription Drug Days Supplied * 136.203 21.482 [ 2.283,40.682] 0.028 27.123 = [ 2.786,51.460] 0.029 (9.795) (12.417) Any Lab Test 0.200 0.054 [ 0.008, 0.100] 0.022 0.068 [ 0.010, 0.126] 0.022 (0.023) (0.030) Any Abnormal Test Result (Unconditional) 0.106 0.029 -[-0,006, 0.064] 0.102 0.037 -[-0.007, 0.081] 0.102 (0.018) (0.023) Any Abnormal Test Result (Conditional on Test) 0.531 0.003 - [-0.120, 0.125] 0.965 0.003 -_[-0.140, 0.146] 0.965 (0.062) (0.073) Panel D. Count Gutcomes Number of Ambulatory Encounters * 3.813 0.516 [ 0.038, 0.994] 0.034 0.655 [0.047, 1.264] 0.035 (0.244) (0.310) Number of Inpatient Encounters + 0.062 0.029 [ 0.004, 0.053] 0.021 0.036 - [ 0.006, 0.067] 0.021 (0.012) (0.016) Number of Emergency Department Encounters * 0.166 0.073 [ 0.027, 0.120] 0.002 0.093 = 0.034, 0.152] 0.002 (0.024) (0.030) Total Number of Encounters (Ambulatory, Inpatient, ED) * 4.129 0.636 [ 0.113, 1.159] 0.017 0.807 [ 0.142, 1.473] 0.017 (0.267) (0.340) Note: Table reports alternative specifications of the regression discontinuity estimates for quarter 1 with standard errors in parentheses. Panel A reports estimates that control for separate second-order polynomials in income on either side of the threshold. Panel B shows estimates that control for local linear polynomials using the optimal bandwidth proposed by Calonico et. al (2014). Panel C reports estimates that control for a global second-order polynomial, as we do in our baseline specification, but excludes applicants with incomes + 10% FPL from the cutoff (340-360% FPL). Panel D shows estimates that control for a global second-order polynomial, as we do in our baseline specification, but with count outcomes as the dependent variables. t Winsorized at the 95th percentile. Control mean is the mean for applicants with incomes between 350% and 450% of FPL. N = 18,672 observations. 37 'poriad Apnjs yUOUI-gT Ue JaAO eyep SAeIISTUMUpE Wo Ss] aINsvaur aUIOIINEC "(FL0Z) 'Te Je URUIGNR], 'Zz afqey, :aomost 'ayep UoQeoyOu Jaye syyuoUl GT ynoge sind90 asuodsar AdAIns a8eseAer oy} BIO '(SUOTSSTUIpe yuaryedut pure syista yuaTedyno Joy potad yoeq-yoo] yUOUT-9 B UWI) sesuodsal AsaIns WO are saInsead aUIODING *(ZI0Z) 'Te J UlaIS[OYULJ 'A BTQR], :aaIMo0s} fssniq uoydtosarg porddns sheq %O'ST €0 €7 yuaamy jo IoquUNN %0'TZ TTEL GZ sniq uonduoserg ft s}Ist, yuouIed Jamnoouq Jueunsred %EOT GPE %0L -0q Aouasiourg Auy %8'7G WL TL %L9 -aq AouaSiewg Auy {SUOTSSIUIpy Jayunod %LOL EL %LL0 eytdsopyyueneduy Auy = MT "9g %79 %9'E -uq juegjeduy Auy jayunosuyq %69E WLS %OTTZ +SHSIA UaQedjnO Auy %0'0Z %0°LZ9 WHET Aioyenquy Auy dno (Tdi %0SF-0S€) (AD poya joquod wt (qLV7) pea dnorsjoquod ur ayeurysa Jeuonsodorg anjea ueay = Pay awlodInNQ jeuoyJodo1g anjea ueaw 1%) Gy IO aul0DjnNc_ yueutiedxy sourmsuy] yeep] UOZaI_O weISOLJ sURISISSY [eIURULY [eoIpey] Jasiey yuourtiadxy sduemsuy yyyeep{ UosIIO YIM UOsTIedwOZD poyq Peuonsodorg zy ajqey 38 Figure A1: RD Estimates for Quarter -1 A. Any Ambulatory Encounter B. Any Inpatient Encounter Income (% of FPL) C. Any Emergency Department Encounter 75 I RF: -0.002 (0.023) 12 ' I RF: -0.001 (0.013) °* oe L- IV: -0.003 (0.030) . > | IV: -0.002 (0.016) 7 *e a> o oe «- ~Mean: 0.678 a "eee : ee 3 | : -Mean: 0.077 e o s ¢ 08 "2. "wre 0" * Je 68 os . es -" e ee ° 2 : . ° .06 os e * oe | 6 oe % © e ? s .04 8 | s s 55 02 e | T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) D. Any Encounter (Ambulatory, Inpatient, ED) Income (% of FPL) G. Any Abnormal Lab Result (Unconditional) I 25 I RF: 0.015 (0.017) 75 .° I RF: -0.002 (0.023) | IV: 0.019 (0.021) ; .o* bY IV: -0.003 (0.029) * Mean: 0.140 ° e« We Mean: 0.691 2 e ® ° | 7 7 . ° he Ss. . . _- ° _- 65 @ & te 15 0 ® 6 & = - e ™ A | ~ 55 T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) E. Prescription Drug Days Supplied F. Any Lab Test ° : 3.774 (8.151 3 : -0.023 (0.021 160 | RP (8.151) : | RF (0.021) ee ° "ef, | IV: -4.791 (10.343) ° . | IV: -0.030 (0.026) eos © F © . Mean: 432.601 ee° ee °*_ _Mean:-0.240 14075 Me, "eae - 2, . . - . .25 2 e - oa i oe ° tr, ' oe eee % 1204 "sen, eee, ' I S 27 wg, ene. te J ~_ °* ° se | ° ~ . * e e | % too. °° ~\ 15 * . ! . | 80 | 4 | T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) H. Any Abnormal Lab Result (Conditional on Test) Income (% of FPL) e . RF: 0.000 (0.016) 7 I RF: 0.057 (0.052) 15 . - | IV: 0.000 (0.020) IV: 0.074 (0068) wore. gg ee Mean:0.124 6 _Mean: 0.516 A 6 | ok 4 _ ™ 05 3 T T T T | T T T T T T T T ' T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Note: Figure shows regression discontinuity plots of the impact of financial assistance in quarter -1, which corresponds to event months -3, -4, and -5 relative to the month of application decision. Dots show mean of the outcome for 85 equal-frequency bins (220 applicants per bin, except for Panel G where there are 130 applicants per bin). Solid lines show fitted values from a second-order polynomial; dashed lines show 95% confidence intervals. For each outcome, we also report the RD and IV estimates, their standard errors, and the mean of the outcome for applicants with an income of 350-450% of FPL (i.e., the "control group" mean). N = 18,672 observations. 39 Figure A2: RD Estimates for Quarter 0 A. Any Ambulatory Encounter B. Any Inpatient Encounter RF: 0.033 (0.021) 35 I RF: 0.020 (0.020) IV: 0.042 (0.026) ' | IV: 0.025 (0.026) Mean:0.766 3 ° . Mean: 0.207 ng *, 7 1 yr e e " ~ 2 15 T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) C. Any Emergency Department Encounter D. Any Encounter (Ambulatory, Inpatient, ED) a4ee | RF: 0.062 (0.022) 9 I RF: 0.031 (0.020) . ea | IV: 0.079 (0.027) -.° a | IV: 0.039 (0.025) ® aaj ee Mean;0.236 ee er Mean: 0.787 . te oes @ | J a. a *ou as od _ - a = 3 | 8 wien en . 9 SS nee rmeemargretgs | ee e ' e | ° : 25 ° _- 75 ' ; e ° s : | 2 7 ° | T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) E. Prescription Drug Days Supplied F. Any Lab Test 1801 ° I RF: 5.605 (8.345) 4 . : I RF: -0.004 (0.023) we "8 . | IV: wars 80m . | IV: - % Se & coe lean: B:} °,* oe jean: 0. 160 ._° e ae a | * wt 35 . o*ts .* ° e | s 9 e ° & e 9 ¥ e§ 2. 140 3 ewe ° 2 ibe "ee, *, | . 120 257 come % @ . | s 100 2 | T T T T T T T T T T T T T T T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) G. Any Abnormal Lab Result (Unconditional) H. Any Abnormal Lab Result (Conditional on Test) 25 I RF: 0.014 (0.019) 8 I RF: 0.052 (0.045) | IV: 0.017 (0-023) | IV: 0.065 (0.055) ® Mean: 0.153 7 Mean: 0524 * 2 A . 2 ° » @ . e | aan | 5 Pee eo* . s oe * "6 | " eo put % 2 J ' 6 15-4 wee (ee eo FS ° e gee ® . ° 5 e a a ee * ® s ' ° le ® e 4 | 4 | T T T T T T T T T T T T T 1 T T T T 150 200 250 300 350 400 450 500 550 150 200 250 300 350 400 450 500 550 Income (% of FPL) Income (% of FPL) Note: Figure shows regression discontinuity plots of the impact of financial assistance in quarter 0, which corresponds to event months 0, -1, and -2 relative to the month of application decision. Dots show mean of the outcome for 85 equal-frequency bins (220 applicants per bin, except for Panel G where there are 130 applicants per bin). Solid lines show fitted values from a second-order polynomial; dashed lines show 95% confidence intervals. For each outcome, we also report the RD and IV estimates, their standard errors, and the mean of the outcome for applicants with an income of 350-450% of FPL (i.e., the "control group" mean). N = 18,672 observations. 40 Figure A3: Distribution of P-values for Placebo Manipulation Tests A: CJM Test B: McCrary Test 16.61% of p-values are below 0.05 40.53% of p-values are below 0.05 ' ' 1.00 1 : 1.00 1 I . I - l . ! ! we ! 0.75 : . 1 . 0.75 1 : 1° 1 © ! . © ! 3 . 3 3 . 1 a . ! > . . > * ee 40.50 : . . *L & 0.50 : ow ! B * ! B . . ! oO oO FB ! . FB e ! .t ° ! 0.25 aa . i of. . 0.25 ! ' oe fe i 1 . 1 . 0.05} oe awe -af-- Sy - t-te 4 Scncns a 2 0.05 os 7a fet ee ge weimee & Ba See ee ee : ee a ee ee ool «. * F % ¥ ae Pay . O00, tere eee eh we dy Se a | 1 200 250 300 350 400 450 500 200 250 300 350 400 450 500 Cutoff (% FPL) Cutoff (% FPL) Note: Panels A shows the p-values from placebo CJM tests conducted at 1% increments for the 301 points between 200% and 500% FPL. Panels B shows the p-values from 301 placebo McCrary manipulation tests conducted at the same increments. The vertical dashed lines show the actual 350% FPL cutoff for the financial assistance program. The horizontal dashed lines show the conventional 0.05 p-value threshold for rejecting the null of no manipulation. 41