Reputations and Firm Performance: Evidence From the Dialysis Industry

Transcription

1 Reputations and Firm Performance: Evidence From the Dialysis Industry Subramaniam Ramanarayanan Anderson School of Management, UCLA Los Angeles, CA Jason Snyder Anderson School of Management, UCLA Los Angeles, CA

2 Abstract: Dialysis firm performance is publically graded using three coarse categories: above expected mortality, below expected mortality, and as expected mortality. We use the underlying performance measures that create these public categories to generate a regression discontinuity design that randomly assigns reputations to firms. We find that when a firm is assigned the grade above expected mortality this causes the firm to reduce their future mortality rates by taking better care of their patients. An above expected mortality grade also causes more informed patients to avoid the firm in the future. We do not find comparable effects for firms that are randomly assigned to the below expected mortality grade. The evidence is consistent with disappointing information being a significant motivator of firm and patient behavior. 2

3 I: Introduction Over the past fifty years a large amount of literature in economics has highlighted the central role of information in decision-making. A key implication is that improving the quality and transparency of information will lead to better social outcomes. 1 One area where these ideas have been applied has been to the implementation of firm quality disclosure programs. Numerous programs exist that provide public information on firm performance and product quality. Restaurant quality (Jin and Leslie 2003), health-plan report cards (Scanlon, Chernew, McLaughlin, and Solon 2002; Jin and Sorenson 2006; Dafny and Dranove 2008), and hospital rankings (Pope 2009) are a few areas 2 where there has been significant academic study concerning the effects of quality disclosure on firm performance and consumer choice. The empirical literature has come to two major conclusions regarding the impact of quality disclosure programs. First, at the consumer level, there is substantial evidence that quality disclosure leads to vertical sorting. All else equal consumers are more likely to choose higher quality products after learning the product quality. Secondly, there is evidence that quality disclosure leads to subsequent improvements in product quality. For example Jin and Leslie (2003) find that after Los Angeles adopts restaurant quality report cards there is a significant decline in hospital admissions for foodborne illnesses. Common to most of the previously studied disclosure programs is a quality score or an ordinal ranking 3 that is uniformly applied to all of the firms or products in a given marketplace. In practice many disclosure programs are designed quite differently. Two distinct types of programs are frequently observed, those that celebrate the exceptional and those that shame the incompetent. Consider two examples: The Blue Ribbon Schools Program honors public and private elementary, middle, and high schools that are either high performing or have improved student 1 The positive impact of information disclosure on social welfare is at times ambiguous; see Dranove, Kessler, McClellan, and Satterthwaite (2003). 2 See Dranove and Jin (2010) for a comprehensive overview. 3 For example grading restaurants on an A to F scale or ranking hospitals from 1 to

4 achievement to high levels, especially among disadvantaged students. 4 Over the past twenty-five years approximately 6,000 schools have received a blue ribbon designation out of well over 100,000 eligible schools. In contrast to the blue ribbon designation Condé Nast Portfolio has put out a list of their Toxic Ten 5, a list of ten companies that could be doing more for the environment. There are several reasons to believe that focusing disclosure policies on the extremes of the firm quality distribution could be desirable. Tournament theory (Lazear and Rosen 1981) suggests that disclosing who the very best firms are can create a powerful incentive for all firms to improve product quality. On the other hand, there are compelling arguments from behavioral economics (Kahneman and Tversky 1979) that evaluations that falls short of expectations can be a markedly more powerful motivator compared to evaluations that exceeds expectations. 6 This suggests that singling out the worst performers can provide a meaningful way to encourage improvements in product quality. Whether efforts should be focused on exposing the worst performers or praising the best is a question that is not convincingly addressed in the prior literature. Establishing a causal relationship between these program features and future behavior is difficult. The common methodology in the prior literature has been to use a difference in differences design to assess outcomes before and after the implementation of the disclosure program relative to a control group. 7 A particularly troubling possibility is that these methods can attribute the impact of the program to mean reversion. When a regulator observes distressing signals in a marketplace there could be a tendency to pursue information disclosure policies as a remedy. For example, a health plan disclosure law could be enacted because lawmakers receive 4 See the department of education website: 5 From Condé Nast Portfolio s website: 6 An example of this comes from Mas (2005) who finds that arrest rates fall when police officers receive pay raises that fall short of their expectations. A smaller sized effect is seen when police officers receive raises that exceed expectations. There is limited evidence of the role of expectations in shaping behavior in product disclosure markets. Dranove and Sefkas (2008) find that vertical sorting is dependent upon the extent to which the information that is being disclosed actually updates firm and consumer beliefs based on their prior expectations. If the information being disclosed is not surprising, it is less likely that behavior will change. 7 Pope (2009) is a notable exemption. He uses an instrumental variables approach to assess the impact of hospital rankings on future admissions. 4

5 numerous complaints about the complexity of health insurance. If subsequent improvements in the marketplace are seen, it is possible that the results could be due to mean reversion. Quality is naturally variable over time and it is possible that after the disclosure program is implemented there will be a return to the long-term average that would have occurred in the absence of the disclosure program. Mean reversion can be exacerbated when looking at the extremes of the quality distribution where difference in difference designs can substantially overstate the impact of disclosure programs. 8 To study these issues we use a natural experiment from the dialysis industry that quasi-randomly assigns quality information to firms. The most prevalent modality of treatment for kidney failure is in-clinic dialysis treatments that clean the patient s blood every 2-3 days. Approximately 20% of patients die while on dialysis each year and there is considerable dispersion in quality across facilities. 9 Up until late 2010 the only available disclosures of firm quality were three coarse grades: as expected mortality, below expected mortality, and above expected mortality. 10 Because of the coarseness of the performance categories fairly similar facilities could receive very different grades. For example, a facility that had an 90 th percentile national ranking could receive an as expected mortality grade while a facility at the 91 st percentile could be given an above expected mortality grade. Prior to the discontinuous categorization both facilities should have had essentially similar characteristics. The only difference is that the facility with the slightly higher mortality percentile rank received a much lower grade. Using recently available archival performance data we use a regression discontinuity design to isolate the causal impact of changes in the information disclosed about the facilities on future performance. We find that receiving an above expected mortality grade as opposed to an expected mortality grade leads a facility to improve their future mortality rates. The 8 These problems are prominent in other domains, Chay, et. al (2005) find that difference in difference designs can seriously inflate the estimated impact programs relative to a quasi experimental design because of the above rationales /8308/ 10 In late 2010 data on actual risk adjusted mortality rates were released for the years Prior to that year the National Kidney Foundation was extremely reluctant to disclose the underlying risk adjusted mortality data. 5

6 evidence is consistent with these changes being real as opposed to the facilities solely changing their patient profiles to manipulate the rankings. We find evidence that patients who are more informed will sort away from facilities that are ranked above expected mortality. We find no evidence to suggest vertical sorting of the nature seen in previous studies. Put differently, facilities that are rated above expected mortality see no changes in patient volumes but do treat a higher percentage of patients who were likely unaware of the ratings. Importantly, we find very limited evidence that distinguishing facilities between below expected mortality and as expected mortality has any impact on relevant outcomes. Taken together we find strong causal evidence that disclosure policies impact firm performance, but that the results are heavily skewed towards the low end of the firm quality distribution. The paper proceeds as follows: Section II will describe the dialysis and the data in greater depth, Section III discusses the methods, Section IV explains the results, and Section V offers concluding observations. II: End Stage Renal Disease, the Dialysis Industry, & the Data End stage renal disease occurs when the kidneys fail to properly clean the blood. Diabetes, hypertension, and genetics are the primary causes of kidney failure. There are two possible treatments for kidney failure: dialysis or transplantation. 11 Most physicians view transplantation as the preferred treatment. Patients live longer and healthier lives with an organ transplant. The problem is that there is a major kidney shortage; more people have kidney failure than there are available kidneys for transplantation. As a result there are almost 400,000 Americans currently being treated for kidney failure with dialysis. 12 The vast majority of US dialysis patients are treated in one of the approximately 5,000 nationwide dialysis centers 3-4 times a week. Mortality rates on dialysis are grim. Approximately 20% of patients die within their first year on dialysis and 65% die within 5 years These treatments are not exclusive. Many patients receive dialysis while waiting for a transplant Going on dialysis is similar to having stage III colon cancer. 6

7 End stage renal disease receives near universal coverage from Medicare regardless of age (Nissenson and Rettig 1999). Medicare pays on a fixed fee per treatment basis. 14 End stage renal disease accounted for approximately 6% of the total Medicare budget in 2010 (Fields 2010). Since Medicare was approximately 13% of the 2010 federal budget end stage renal disease expenditures made up almost 1% of the entire federal budget in Across dialysis facilities there is considerable dispersion in mortality. Figure 1 shows the risk-adjusted standardized mortality ratios for the years At the bottom 10 th percentile of the distribution a facility has a 30% lower than expected mortality rate. At the top 90 th percentile a facility has a 33% higher than expected mortality rate. 15 Garg, et. al (1999) found that differences in performance are dependent on for-profit ownership and Powe, et. al (2002) find that dialysis administrators believe that patient education, staffing ratios, and wages are crucial determinants of facility quality. In response to this considerable dispersion in quality Medicare released the dialysis facility compare ratings in This web-based tool provided consumers with information on the location, hours, and performance quality of almost all of the nation s dialysis facilities. Performance is reported in three categories: as expected mortality, below expected mortality, and above expected mortality. The National Kidney Foundation uses a four-year window as the basis for these categories. For example facility mortality between the years would be used as a basis for constructing the categories that would be released in the middle of When the program was introduced 96% of all facilities were designated as having as expected mortality. With the four-year window the thresholds switched so that approximately 80% of facilities received an as expected mortality grade. Figures 2 and 3 illustrate how these grades are determined. For data coming before the four-year window for a facility to receive an above expected mortality grade the lower confidence interval had to be greater than 1.2. These facilities were almost certainly worse than average. In the four-year window the threshold moved from 14 From the perspective of the dialysis facility competition is quality based, not price based. 15 Since approximately 20% of patients die each year these differences are large. 7

8 1.2 to 1. Now there needs to be a 95% or greater chance that a facility was worse than average to receive an above expected mortality grade. 16 We study the impact of the grades coming from the four-year window because prior to then there were too few observations to effectively use the proposed regression discontinuity design. The data from the four-year window was revealed at different times to the facilities, consumers, and researchers. Understanding the timeline is crucial to the research design of this paper. June 2008: Facilities learn what coarse category they are in based on the four-year window. The facility learns how close they are to the threshold. Facility level performance data from 2008 would be the first year to show a response to this information. December 2008: The four-year window data is released on the dialysis facility compare website and consumers learn what coarse categories all facilities fall into. Facilities also learn what coarse categories their competition falls into. Facility level consumer choice data from 2009 would be the first year to show a substantial response to this information. From this point onward in the paper as expected mortality, below expected mortality, and above expected mortality will refer to the categories being generated from the mortality data from the four-year window. December 2010: The precise mortality data from all four-year windows between became common knowledge. Before December 2010 it was impossible to obtain data on performance of facilities beyond the coarse categories on the dialysis facility compare website (Fields 2010) The same was true for for below expected mortality. 17 Robin Fields released this information in 2010 out of concern that the coarse categories were not sufficient to adequately report facility quality. It took her two years of freedom of information act requests to obtain the data. Fields (2010) effectively describes this dispersion: Innovative Renal Care and Midtown Kidney Center, clinics about two miles apart in Houston, had similar stats on Dialysis Facility Compare in 2007, including as expected survival rates. But the full data show that Innovative Renal s average annual death rate after factoring in patient demographics and complicating conditions was 34 percent higher than expected. Midtown s average rate was 15 percent lower than expected. Dialysis Facility Compare has since changed Innovative s survival rating to worse than expected, but how much worse? The unpublished 2009 data reveal that the clinic performed more poorly, versus expectations, than 92 percent of all facilities nationwide. Innovative 8

9 We have data on all 4,665 firms that received performance evaluations on the dialysis facility compare website in December of We are able to link these firms to their past and future mortality rates, organizational form, ownership information, and patient characteristics on a yearly basis from All of our analysis is performed at the facility/year level. III: Methods Our primary methodological approach is to use a regression discontinuity design to estimate the causal effect of information on future outcomes. Imbens and Lemieux (2007) describe, In regression discontinuity designs for evaluating causal effects of interventions, assignment to a treatment is determined at least partly by the value of an observed covariate lying on either side of a fixed threshold. The key idea is that close to a threshold of interest all other characteristics and choices that could influence an outcome will be orthogonal to the treatment being studied. Regression discontinuity design is a well-established methodology with a straightforward causal interpretation. Figure 4 helps to understand how this design is applied to our context. We plot the relationship between the 95 th percent lower confidence interval for the risk-adjusted mortality ratio and the 2008 risk-adjusted mortality percentile ranking. Any facility with a lower confidence interval just above 1 receives an above expected mortality grade and any facility just below 1 receives an as expected mortality grade. On the y-axis the 2008 risk-adjusted mortality percentile ranking provides a national percentile position of all facilities based solely on their 2008 riskadjusted mortality ratio. 18 For example a facility with a percentile ranking of 1 would be in the top 1% of all facilities in terms of survival. In this graphic there is strong evidence of a discontinuity precisely at the point where a firm moves from as expected mortality to above expected mortality. A firm that just barely falls in the above expected mortality category is much more likely to improve subsequent performance relative to a firm that just barely falls into the as expected mortality category. Visually it appears Renal s administrator, Scott Sullivan, said the clinic had a difficult patient pool, but its most recent results have shown improvement. We ve put things in place to make sure those numbers are corrected, he said. 18 This percentile ranking only uses data from

10 unlikely that the change in the 2008 risk-adjusted mortality percentile ranking is due to anything other than just barely being rated as having above expected mortality. Our empirical approach is to estimate the size of that jump for various outcome measures when firms receive info that they are above or below expected mortality outcomes. In our analysis we separately estimate the difference between as expected mortality and above expected mortality from the difference between as expected mortality and below expected mortality. Our principal specification is a facility level regression: Outcome! = α + β Threshold + β N!h degree polynomial of CI! + ε! Threshold is an indicator variable equal to 1 if the firm receives an above expected mortality grade or a below expected mortality grade. 19 The parameter estimate on Threshold is the size of the jump being estimated. Figure 5 provides a graphical representation of what this estimation process yields when applied to figure 4. Threshold is the difference between the two estimated lines when the lower confidence interval equals to 1. The figure also clearly illustrates the role of the polynomial controls. The polynomial controls for the underlying trend and allows Threshold to identify the discontinuous break in the data. The more degrees on the polynomial controls the greater the fit to the underlying trend. IV: Results Table 1 provides summary statistics. Importantly there are some inconsistencies in the number of recorded observations. This occurs for two reasons. First 153 of the 4,665 facilities reported on the December 2008 dialysis facility compare website closed in This has the potential to introduce survivorship bias into our results. In unreported results we find that firm survival is not related to the coarse categorical grades. This reduces the concern that our results could be driven by survivorship bias. Secondly for some of the variables the recording was not uniform across all firms. For example we observe 4,512 firms with data in 2009 but only 4,464 with information on whether a new patient visited a nephrologist in the 19 In none of our specifications will an above expected mortality firm and a below expected mortality below be in the same regression. 10

11 prior year. These missing values appear to be a relatively small portion of the overall data. The identifying assumption of the regression discontinuity design is that all potential confounding variables are orthogonal to the treatment. While this assumption cannot ever be comprehensively tested it is possible to find evidence that is consistent with the assumption. The conventional approach is to show that potentially confounding covariates are not systematically distributed on either side of the threshold. Tables 2 and 3 perform this test. In table 2 we find no evidence of a jump in any of the 2007 variables in response to barely exceeding the above expected mortality threshold. 20 This implies that there is no evidence of strategic sorting around the threshold. In table 3 we have similar findings. There is some evidence to suggest that chain owned firms are more likely to be just barely inside the below expected mortality threshold. However this result only holds at the 10% confidence level. It is not surprising that with 14 different variables one of them is significant at the 10% confidence level. Importantly column (1) of tables 2 & 3 shows that risk-adjusted mortality ratio from is not related to the thresholds. 21 This is a precisely estimated zero. Taken together there is little evidence to suggest that the assumption that the threshold treatments violate random assignment. The first four columns of table 4 show results consistent with figure 5. All of the columns show that just barely falling into the above expected mortality grade leads to a percentile improvement in the 2008 risk-adjusted mortality percentile ranking. Importantly we show that this is robust to the two major factors that influence the results of regression discontinuity designs: the size of the sample window and the degree of the control polynomial. Column (1) uses a 3 rd degree polynomial as a control and a sample window of all observations with lower confidence intervals between.9 and In column (2) we use a 10 th degree polynomial as opposed to a 3 rd degree polynomial and find similar results. In column 20 This is not surprising since the categories from the four-year window were not known until Since the thresholds come from the confidence intervals there is no reason to believe this is necessarily true or false. 22 This corresponds to a sample window range of.2. For below expected mortality threshold regressions a sample window range of.2 means that the upper confidence interval would fall between.9 and

12 (3) we expand the sample window to include all facilities with a lower confidence interval between.75 and While the results drop in magnitude the economic significance is still quite large. In column (4) we include controls for for-profit status, whether a chain owned the facility, total number of dialysis stations, and the riskadjusted standardized mortality ratio from The results are unchanged. This suggests that the treatment is randomly assigned. Columns (5)-(8) of table 4 perform the same test looking at the 2008 risk-adjusted mortality percentile ranking of firms who just barely pass the below expected mortality threshold. The results are much smaller and less significant compared to the first four columns. This suggests that poor performing firms respond to negative disclosures by improving quality. The response of exceptional firms to positive disclosures is more limited and difficult to distinguish from zero. How does this mortality improvement occur? An important component of high quality dialysis care is how well a patient s anemia is managed. Most patients on dialysis do not have an ideal red blood cell count and are therefore anemic. Having poorly managed anemia increases mortality risk. A blood test of hemoglobin levels given to patients undergoing dialysis provides a measure of how well anemia is being managed. Ideal hemoglobin levels are between 10 g/dl and 12 g/dl. While all centers are required to frequently check and report hemoglobin levels they are not required to have all of their patients between 10 g/dl and 12 g/dl. To achieve this ideal level a facility must actively adjust anemia medication. A major problem is that incentives are not aligned. A facility can over-prescribe Medicare reimbursed anemia medication to patients to increase revenues. When too much medication is used hemoglobin levels can fall below 10 g/dl. Well-managed anemia means that facilities are actively engaged with patients in medication management and avoiding the revenue related temptations to over-prescribe. From the dialysis facility reports we have data on what percentage of patients treated at the facility have well managed anemia. 23 From this we construct two variables, whether a facility is above the median in anemia management and whether the facility was in the top 10 percent of anemia management in the US during In column (1) of table 5 we find no 23 Unlike the mortality data this is not risk adjusted. These results should be cautiously interpreted. 12

13 evidence that being just barely within the above expected mortality threshold leads to above median anemia management. 24 However in columns (2) & (3) we find the probability that 90% of the facility s patients will have well managed anemia increases substantial if the facility falls just inside the above expected mortality threshold. This finding is robust to changing the sample window range and the degree of the polynomial control variables. This suggests that the improvement in care is not uniform across treated firms. A limited number of firms exhibit dramatic performance improvements. Columns (4) (6) show no result for facilities being just within the below expected mortality threshold. Table 6 shows results regarding the characteristics of new patients at the facility. Since patients do not see the coarse performance data until December of 2008 the first time we would expect to see results on the number of new patients or the composition of new patients is in Column (1) shows results on the number of new patients at the facility in The result is an imprecisely estimated zero effect. From column (1) it could be the case that just barely being in the above expected mortality category has no effect on the number of new patients. It is also possible that it could reduce the number of new patients by 25%, which would be an economically significant effect. 25 From table (1) approximately 30% of patients did not visit a nephrologist prior to starting dialysis. Nephrologists are kidney specialists who help patients manage the course of end stage renal disease. The advantage of visiting a nephrologist is that the nephrologist is more likely to be aware of the dialysis facility compare website and the quality of local facilities when referring a patient to a dialysis center. Our assumption is that if a patient did not see a nephrologist before starting dialysis they are less likely to be aware of the facility performance rankings. In columns (2) & (3) of table 6 we find strong evidence that when a facility is just barely within the above expected mortality threshold 10% more of the new patients in 24 We use linear probability models with these discrete dependent variables because regression discontinuity designs quite often lead non-linear maximum likelihood estimators such as probits or logits to not converge. 25 This number comes from the fact that for the parameter and standard error the 95% lower confidence interval is approximately equal to -5. Since the average firm takes on 20 new patients a year this is an approximately 25% reduction. 13

14 2009 will not have seen a nephrologist prior to dialysis. This suggests that informed patients are shying away from facilities with an above expected mortality grade. Columns (4) (6) show no impact of just barely being within the below expected mortality threshold on new patients or the percentage of new patients who previously did not see a nephrologist. V: Conclusions The evidence in this paper is consistent with information disclosure programs being most effective when the incompetent are shamed. One important caveat when trying to generalize this result beyond the dialysis industry is that firms in the dialysis industry are capacity constrained. The best quality firms can only treat a limited number of patients. In industries where firms are not capacity constrained it is likely that information disclosure programs that praise the exceptional are as effective if not more so. The lack of a result on the grade changing the number of new patients is puzzling since it stands in contrast to the prior art. It raises the possibility that mean reversion is driving the results in the existing literature. It points to the need for more compelling research designs in the information disclosure literature to adjudicate this debate. The result on nephrologist visits points to a practical implication for the design of the dialysis facility compare program. It seems that the information on the website is not being utilized by all of the patients searching for dialysis treatment. Finding a means to make the information more salient to consumers could lead to important welfare improvements. 14

15 VI: Bibliography Dafny, L, and D Dranove Do report cards tell consumers anything they don't already know? The case of Medicare HMOs. The RAND Journal of Economics 39(3): Dranove, D et al Is More Information Better? The Effects of Report Cards on Health Care Providers. Journal of Political Economy 111(3): Dranove, D, and A Sfekas Start spreading the news: A structural estimate of the effects of New York hospital report cards. Journal of Health Economics 27(1): Dranove, D, and G Jin Quality disclosure and certification: theory and practice. Journal of Economic Literature 48(4): Fields, R God Help You. You're on Dialysis. Atlantic Monthly: Garg, P et al Effect of the ownership of dialysis facilities on patients' survival and referral for transplantation. New England Journal of Medicine 341(22): Imbens, G, and T Lemieux Regression discontinuity designs: A guide to practice. Journal of Econometrics 142(2): Jin, G, and A Sorenson Information and consumer choice: The value of publicized health plan ratings. Journal of Health Economics 25(2): Jin, G, and P Leslie The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards. The Quarterly Journal of Economics 118(2):

16 Kahneman, D, and A Tversky Prospect Theory: An Analysis of Decision under Risk. Econometrica 47(2): Lazear, E, and S Rosen Rank Order Tournaments as Optimum Labor Contracts. Journal of Political Economy 89(5): Mas, A Pay, reference points, and police performance. Quarterly Journal of Economics 121(3): Nissenson, A, and R Rettig Medicare's end-stage renal disease program: current status and future prospects. Health Affairs 18(1): Pope, D Reacting to rankings: Evidence from. Journal of Health Economics 28(1): Powe, N et al Cost-quality trade-offs in dialysis care: A national survey of dialysis facility administrators. American Journal of Kidney Diseases 39(1): Scanlon, D, M Chernew, and C McLaughlin The impact of health plan report cards on managed care enrollment. Journal of Health Economics 21(1):

17 Figure 1: Kernel density of risk-adjusted mortality ratio Risk-adjusted mortality ratio

18 Figure 2: Kernel density of lower confidence interval of the risk-adjusted mortality ratio New Threshold Above expected mortality Old Threshold Lower confidence interval Below expected mortality Old Threshold Figure 3: Kernel density of upper confidence interval of the risk-adjusted mortality ratio New Threshold Upper confidence interval

19 Figure 4: Above expected mortality threshold and future risk-adjusted mortality percentile rank 2008 Risk-adjusted mortality percentile rank Lower confidence interval Note: Lower rank indicates improved performance Figure 5: Above expected mortality threshold and future risk-adjusted mortality percentile rank 2008 Risk-adjusted mortality percentile rank Lower confidence interval Note: Lower rank indicates improved performance. Regression follows specification (1) in table 4 performed on the 8 data points in the graph. 19

20 Table 1: Summary statistics Variable Mean Std. Dev. Min. Max. N Risk-adjusted mortality percentile rank Risk-adjusted mortality ratio Above expected mortality Below expected mortality Total stations For profit status Chain facility Number of new patients Number of new patients No nephrologist in prior year No nephrologist in prior year Left market

21 Table 2: Observables are balanced at the above expected mortality threshold (1) (2) (3) (4) (5) (6) (7) Risk-adjusted Total For profit Chain Above median Number of No nephrologist mortality ratio stations status ownership patient anemia new patients in prior year VARIABLES management Above expected mortality (0.016) (1.559) (0.055) (0.072) (0.105) (4.172) (0.034) Lower CI polynomial degree Sample size Note: Stars denote significance levels: 99 percent confidence level (***),95 percent confidence level (**), and 90 percent confidence level (*) Standard errors clustered at the state level. The Sample window range contains observations lower confidence intervals between.9 and.1. Table 3: Observables are balanced at the below expected mortality threshold (1) (2) (3) (4) (5) (6) (7) Risk-adjusted Total For profit Chain Above median Number of No nephrologist mortality ratio stations status ownership patient anemia new patients in prior year VARIABLES management Below expected mortality * (0.016) (1.697) (0.083) (0.081) (0.084) (2.664) (0.039) Upper CI polynomial degree Sample size Note: Stars denote significance levels: 99 percent confidence level (***),95 percent confidence level (**), and 90 percent confidence level (*) Standard errors clustered at the state level. The Sample window range contains observations upper confidence intervals between.9 and.1. 21

22 Table 4: Threshold status and future risk-adjusted mortality percentile ranking (1) (2) (3) (4) (5) (6) (7) (8) Risk-adjusted Risk-adjusted Risk-adjusted Risk-adjusted Risk-adjusted Risk-adjusted Risk-adjusted Risk-adjusted mortality mortality mortality mortality mortality mortality mortality mortality percentile percentile percentile percentile percentile percentile percentile percentile VARIABLES rank 2008 rank 2008 rank 2008 rank 2008 rank 2008 rank 2008 rank 2008 rank 2008 Above expected mortality *** *** ** ** (4.587) (4.682) (4.462) (4.381) Below expected mortality (4.112) (4.225) (2.753) (2.797) Lower CI polynomial degree Upper CI polynomial degree Control variables No No No Yes No No No Yes Sample window range Sample size Note: Stars denote significance levels: 99 percent confidence level (***),95 percent confidence level (**), and 90 percent confidence level (*) Standard errors clustered at the state level. 22

23 Table 5: Threshold status and anemia management (1) (2) (3) (4) (5) (6) Above median in Top 10 percentile Top 10 percentile Above median in Top 10 percentile Top 10 percentile patient anemia in patient anemia in patient anemia patient anemia in patient anemia in patient anemia VARIABLES management 2008 management 2008 management 2008 management 2008 management 2008 management 2008 Above expected mortality ** 0.084* (0.073) (0.064) (0.044) Below expected mortality (0.070) (0.061) (0.042) Lower CI polynomial degree Upper CI polynomial degree Sample window range Sample size Note: Stars denote significance levels: 99 percent confidence level (***),95 percent confidence level (**), and 90 percent confidence level (*) Standard errors clustered at the state level. Anemia is well managed if a patient has a Hemoglobin between 10 and 12 g/dl. 23

24 Table 6: Threshold status and the number and characteristics of new patients (1) (2) (3) (4) (5) (6) Number of No nephrologist No nephrologist Number of No nephrologist No nephrologist new patients prior to prior to new patients prior to prior to VARIABLES 2009 dialysis 2009 dialysis dialysis 2009 dialysis 2009 Above expected mortality ** 0.087*** (2.816) (0.044) (0.031) Below expected mortality (1.673) (0.041) (0.029) Lower CI polynomial degree Upper CI polynomial degree Sample window range Sample size Note: Stars denote significance levels: 99 percent confidence level (***),95 percent confidence level (**), and 90 percent confidence level (*) Standard errors clustered at the state level. 24