Association of Hospitalization and Mortality Among Patients Initiating Dialysis With Hemodialysis Facility Ownership and Acquisitions

Key Points Question Are acquisitions by dialysis facility chains associated with patient health outcomes? Findings In a cohort study using difference-in-differences analyses of US patients initiating hemodialysis from 2001 to 2015, hospitalization days per patient-year and mortality decreased over time in most comparison groups. However, decreases in hospitalizations were significantly slower at acquired vs nonacquired facilities; observed acquisition results were largely associated with changes at independently owned facilities that were acquired by dialysis chains, where acquired facilities had significantly slower decreases in both mortality and hospitalization rates than nonacquired facilities. Meaning Acquisition of independently owned dialysis facilities by dialysis chains suggests slower decreases in mortality and hospitalization rates than would have otherwise occurred.

 The regression coefficient exp(θ) describes the extent to which the difference-in-difference estimate varies between patients at chain-owned facilities and those not at chain-owned facilities prior to acquisitions.  The combination of exp(β + τ) describes how mortality changed in the period following acquisitions among patients at facilities that were owned by chains prior to acquisitions and that were not acquired.  The combination of exp(β + τ + γ+ θ) describes how mortality changed in the period following acquisitions among patients at facilities that were owned by chains prior to acquisitions and that were acquired by a separate chain.  The combination of exp(γ+ θ) represents a difference-in-difference estimate specific to patients at facilities that were owned by chains in the period prior to acquisitions.

3) Examining Early Facility Switches:
Our previous observation that patients are more likely to switch dialysis facilities soon after developing ESRD informed our decision to assign patients to a facility based on where they received dialysis at the start of their fourth dialysis month. However, we are unaware of specific sources where this phenomenon has been published. Consequently, we conducted an analysis to verify this assertion.
Among patients initiating in-center hemodialysis between 2001 and 2015, we examined the likelihood of changing dialysis facilities in each 90-day interval during the first 360 days of dialysis. Each probability was conditional upon receiving in-center hemodialysis at the start of the 90-day interval. In the first 90 days of dialysis, 21.5% of patients switched dialysis facilities at least once. This declined to 9.2%, 8.2%, and 7.6% among patients in the 2 nd , 3 rd , and 4 th , 90-day intervals, respectively, confirming that facility switches are more common in the first 90 days of dialysis.

4) Measuring Market Competition:
We calculated a commonly-used metric of market competition -the Herfindahl-Hirschman Index (HHI) -based on where patients lived and the facilities where they received dialysis. We calculated this metric separately for each calendar year using a point-prevalent cohort of all patients receiving in-center hemodialysis in the United States on the first day of the year. In our primary models, we defined geographic markets using hospital service areas (HSAs), although we varied this definition in sensitivity analyses. The equation used to calculate Herfindahl-Hirschman Index (HHI) is as follows: Where S i represents the proportion of patients living in an HSA receiving dialysis at the i th firm in the HSA. We calculate a measure of "observed HHI" for each hospital service area (HSA) from the following three steps: 1. Calculate a "first-stage" competition measure for each HSA (using the equation above), based on sum of squared market shares of firms where patients living in each HSA choose to dialyze. In this stage all patients residing in a given HSA define the "market" for each firm-HSA pair. Firms do not have to be located in the same HSA where patients reside, and a given dialysis facility can be included in the calculation of HHI for multiple HSAs if patients from multiple HSAs dialyze at that facility. For the purposes of this step, facilities owned by the same dialysis chain were considered to be one firm. The market share for a firm in an HSA is equal to the proportion of patients in that HSA who choose to dialyze at that firm. For example, in an HSA where half of the patients receiving dialysis went to one of four facilities owned by one firm and the other half of patients went to one of two facilities owned by a second firm, the market share would be considered to be split evenly across the two firms, with an HHI for that HSA of 0.5 2 + 0.5 2 = 0.5.
2. Calculate a dialysis-facility-level measure of competition, using a weighted average of the "first-stage" HSAlevel HHIs for patients who actually dialyze at each facility. This measure is calculated for each separate facility, regardless of which firm owns a facility. It assumes that facilities compete for patients within HSAs and can discriminate against patients living in different HSAs when competing against rival firms.
3. Calculate a "second-stage" HSA-level measure of competition from a weighted average of the facility-level-HHIs at facilities where patients residing in each HSA receive dialysis.
In summary, this index represents a weighted average of competition indices for facilities that treat patients in a given HSA, where facility competition indexes are, in turn, weighted by choices available to patients they treat.

5) Assigning Patients to Acquisition Cohorts:
In several instances, patients appeared in multiple acquisition cohorts. This could occur, for example, if a patient started dialysis at a facility (we will call this Facility A) that was in the same hospital service area as a facility that had been acquired (we will call this Facility B In order to ensure that each patient only appeared once in our analysis cohort, we used the following set of rules to assign patients to one and only one acquisition cohort: 1) If a patient was in the acquisition group of one cohort and the "control" group of another cohort, we assigned them to the cohort where they appeared in the acquisition group. 2) If a patient was in two acquisition groups or two control groups, we randomly assigned them to one of the two groups. In sensitivity analyses, we included all patients under the assumption that they represented independent observations. Notably, our findings were not substantially different in this sensitivity analysis.

6) Examination of Proportional Hazards Assumption:
We examined our stratified Cox model for violation of the proportional hazards assumption by separately examining a "complete-case" sample and a "full-cohort" sample, In the "full-cohort" sample, we excluded variables that were multiply imputed. In both samples, we tested the null hypothesis that the slope for the scaled Schoenfeld residual on time for each model covariate was zero, and the global hypothesis that all slopes were zero.
Based on this test, the following covariates violated the proportional hazards assumption using a p-value of <0.01 as a measure of significance: immobility, Medicare coverage, employer-based group coverage, serum albumin <2, serum albumin 2-3. However, when we visualized the log-log survival curves involving these variables, there was no clear evidence of a change in the hazard rate ratio over time.

7) Replacing Nearby Non-Acquired Facilities with Alternative "Controls":
An alternative interpretation of our study findings, when considering the baseline differences in health outcomes among acquired versus non-acquired independently-owned facilities, is that acquisitions led to a narrowing in health outcomes because of responses initiated by nearby non-acquired facilities. To examine this possibility, we created a study cohort using an alternative "control" group. Specifically, we replaced our comparison groups of unaffected patients with patients initiating dialysis in hospital service areas where no-acquisition occurred. Facilities in this alternative control group would not be expected to respond to acquisitions, since there were no nearby acquisitions to encourage changes in practices. Using this alternative cohort, we ran regression models identical to those described in our primary analyses. We focus our discussion on findings related to independently-owned facilities.

8) Examining for "Regression to the Mean":
We examined the possibility that slower declines over time in hospital days (and mortality) among acquired versus non-acquired facilities can be explained by "regression to the mean". If this alternative interpretation of our primary study findings were true, and observed differences in outcomes in the pre-acquisition period were simply due to chance, we would expect to observe a narrowing of pre-acquisition differences in outcomes in the years before and after the baseline (i.e. pre-acquisition) period. Repeated sampling of patients initiating dialysis before enrollment in our study would be expected to approach population means in the same way as repeated sampling after acquisitions.
To examine this possibility, we identified patients starting dialysis at facilities included in our primary study cohort in the 2-years prior to entry into the cohort (i.e. 4th and 5th years prior to acquisitions). This required expanding our cohort to patients initiating dialysis in 1999 and 2000. We identified patients initiating dialysis in the 2-years prior to entry into the study for all 10 acquisition cohorts and included them in our primary regression models. To simplify the creation of this expanded cohort, and to ensure that adding these data would not lead us to exclude patients from our primary cohort, we deviated from our primary study selection approach in several ways when adding the two prior patient-years: 1) We did not require that facilities had "stable" ownership in the 4 th and 5 th years prior to inclusion in our study cohort. This ensured that there would not be a change in facilities categorized as acquired and nonacquired in this analysis. 2) Patients could appear more than once if the 2 nd appearance occurred in the newly-added years.
Within this model, the estimated effect of the (acquired facilities)* (before baseline) interaction characterizes whether differences in health outcomes among independently-owned acquired versus non-acquired facilities narrowed in the before-baseline period. The three-way interaction (before baseline)*(chain-ownership)*(acquired facilities) indicates whether narrowing (or widening) of the outcomes gap in the period before baseline differed at chain versus independent facilities. Because our study findings were most robust when examining independent facilities, we focused our discussion on differences between acquired and non-acquired facilities that were independently-owned in the pre-acquisition period. However, we report results involving both independent and chain-owned facilities.

9) Assessing Differences in Comorbidities among Comparison Groups:
It is not obvious how differences in reported patient characteristics over time at independently-owned facilities and chain-owned facilities would be expected to influence the measurement of health outcomes. To examine this issue, we used estimates from our stratified regression models to compare predicted probabilities hospitalization days and death among patient subgroups. In order to retain all patients used in the primary analyses, and to avoid unnecessarily complicating the analyses with multiple imputation, we excluded parameters with high numbers of missing values: laboratory values, BMI, and eGFR. When modeling the probability of death, we assumed a parametric survival function with an exponential distribution (unlike all other analyses, which used Cox regression).

10) Sensitivity Analyses:
We tested the sensitivity of our stratified analyses to a number of alternative model specifications. We focus our discussion of these sensitivity analyses on regression coefficients pertaining to independently-owned facilities but also include a coefficient identifying whether the difference-in-differences estimate varied among independentlyowned versus chain-owned facilities (eFigures5-6). In all sensitivity analyses, we used the full regression model described in Equation 1 and supplemented with additional covariates for stratified analyses.
A) We conducted sensitivity analyses where we included dummy variables representing the calendar year of dialysis initiation as additional model covariates. B) It is possible that hospital service areas do not accurately represent a local market for dialysis facilities. To examine this potential inaccuracy, we adapted our method of calculating HHI to calculate a market concentration index at the zip-code level. This involved substituting the hospital service area (HSA) of patient residences with their zip-codes in steps 1 and 3 of the algorithm described above in "measuring market competition." Within each HSA, we then compared zip-code level HHIs among acquired and nonacquired facilities both before and after acquisitions. We did this using the following linear regression model with fixed-HSA effects: In this analysis, we found small but statistically-significant differences in the change in zip-code level HHIs (ranging from 0 to 0.02) among acquired and nearby non-acquired facilities following acquisitions. Because of this slight variation, we performed an additional sensitivity analysis where we controlled for market competition prior to and following acquisitions based on patient zip-codes. C) We also conducted a sensitivity analysis where we used counties, rather than HSAs to define markets. This involved calculating county-level market competition indices in steps 1 and 3 described in "measuring market competition." In this analysis we also defined geographic proximity of facilities according to counties (rather than HSAs), meaning that the control group consisted of non-acquired facilities in the same county as acquired facilities. D) We conducted a sensitivity analysis where we excluded Gambro facilities that were divested in the process of DaVita's acquisition of Gambro and Renal Care Group (RCG) facilities that were divested in the process of Fresenius's acquisition of RCG. We identified these facilities in instances where they changed ownership at the time of DaVita's acquisition of Gambro to a chain other than DaVita or where they changed ownership at the time of Fresenius's acquisition of RCG to a chain other than RCG. E) We conducted sensitivity analyses where we excluded patient comorbidities where reporting on the Medical Evidence Report may be less certain. Specifically, we excluded cancer, heart failure, cerebrovascular disease, and coronary disease. F) We conducted a sensitivity analysis where we only considered facilities to have been acquired if the acquisition was by a large dialysis organizations (LDOs). We used the following definition of LDO adopted by the Centers for Medicare and Medicaid services: an organization with at least 200 facilities in a given year. G) We conducted a sensitivity analysis where we did not exclude patients starting dialysis in the 6-months prior to and following a dialysis acquisition. In this analysis, we assumed that the post-acquisition periods began on January 1 st in the year of ownership change. H) We conducted a sensitivity analysis where we did not require that patients only appear in the combined cohort one time. We did this by assuming that each patient contributed an independent observation, irrespective of the number of times that their record appeared in our combined cohort. I) When examining mortality, we performed an additional analysis where we included all patients who started dialysis and who lived to the first day of the fourth dialysis month. This "expanded cohort" included patients who did not qualify for Medicare after 3 months of dialysis. J) When examining hospital days, we performed two additional sensitivity analyses where we included: 1) hospital service area fixed effects, and; 2) dialysis facility fixed effects. We used linear probability models rather than negative binomial models in order to facilitate computation. K) We also examined whether the change in hospital days and mortality from acquisitions at independentlyowned facilities differed in the years following the 2011 end-stage renal disease (ESRD) payment reform. We did this in a model that only included patients at independently-owned facilities. In this model, we tested for a three-way interaction between the difference-in-difference estimate -(after acquisitions) * (at an acquired facility) * (initiating of dialysis on or after 2011). L) We examine whether our analysis of mortality was sensitive to inclusion of patients at the start of dialysis.
To do this we assigned patients to the facility where they initiated dialysis. Unlike our primary analyses, we did not exclude patients who died in the first 90 days of ESRD. We only conducted this sensitivity analysis for mortality, since hospitalizations required Medicare claims, which >40% of patients in our cohort did not have until after 90 days of ESRD.
We examined for a potential time-varying effect of acquisitions on mortality and hospitalizations among patients at independent facilities. This was done within the Cox and negative binomial models by creating binary indicator variables representing each of the three years following acquisitions, and by interacting these variables with an indicator of whether or not a patient was at a facility that had recently been acquired. These interaction terms identified potential variation in the acquisition effect over time. By testing the joint hypotheses that the "(post-acquisition year)*(acquisition group)" interaction terms were zero, we were able to assess whether the effect of acquisitions varied over time. In these analyses, we did not find evidence that the associations between acquisitions and each outcome varied over time in the post-acquisition period. In the case of mortality, the p-value for the test for heterogeneity in interaction terms was 0.9. In the case of hospitalization days, the p-value for the test of heterogeneity in interaction terms was 0.7.

eTable1. Number of Acquired Facilities and Nearby Non-acquired Facilities by Year
Year  Acquired  113  331  483  377  58  88  79  85  203  58  1,875  Nearby not-acquired *  226  327  586  401  147  330  206  194  568  274 3,259 * Not-acquired facilities are in the same hospital service area as acquired facilities. All facilities listed above had the same owner in the three years before and after the acquisition year.

eTable2. Dialysis Facility Characteristics Prior to Acquisitions
Distance between home and facility -Median(IQR) 4.7 (9.0) 3.6 (5.5) 4.3 (8.7) 3.5 (5.4) 0.28 IQR is inter-quartile ratio. * Indicates >10% standardized difference in characteristics in the pre-acquisition period. † represents the statistical significance of the interaction term in a model where each characteristic of interest is a function of case vs. control, pre-acquisition vs. post-acquisition, and the interaction between case and postacquisition. ± based on zip-code level data. Logistic and linear regression was used for binary and continuous outcomes, respectively. When comparing the hazards of death using the alternative control, acquisitions were no longer independently associated with mortality in the pre-acquisition (i.e. baseline) period at independently-owned facilities (HR 0.99; 95% CI 0.95 to 1.04). Declines over time in mortality at independently-owned acquired facilities continued to be slower than declines in mortality at independently-owned facilities that were not acquired. Acquisitions were associated with a (relatively) higher rate of death (diff-in-diff estimate 1.06; 95% CI 0.99 to 1.15). This 6% difference is less than the 15% difference observed in our primary analysis, and was of marginal statistical significance (p=0.09). Specifically, the magnitude of the difference-in-differences estimate for independently-owned facilities was 60% smaller in the alternative control compared to the primary model of survival.

eTable 4. Baseline Characteristics among Chain-owned Facilities
A similar pattern emerged when comparing changes over time in hospitalization days at independently-owned facilities using the alternative control group. Among independently-owned facilities, acquisitions continued to be independently associated with a slower decline in hospital days (diff-in-diff estimate 2.1; 95% CI 0.8 to 3.3. This was only slightly smaller in magnitude than the difference of 2.7 (95% CI 1.4 to 4.0) observed in the primary analysis. Specifically, the magnitude of the difference-in-differences estimate for independently-owned facilities was 22% smaller in the alternative control compared to the primary model of hospitalizations.

eTable10. Examining for "Regression to the Mean"
Mortality

Summary of Findings from Examination of "Regression to the Mean"
In an examination of mortality, there was a 10% lower hazard of death in independently-owned acquired facilities compared to independently-owned non-acquired facilities in the baseline period prior to acquisitions. This is similar in magnitude to the 10.4% "gap" in mortality observed at independently-owned facilities in our primary analysis. Similar to findings from our primary analysis, this mortality difference narrowed following acquisitions, as acquired facilities experienced smaller declines over time in adjusted mortality (difference-in-difference estimate for independently-owned facilities: 1.15; 95% CI 1.05 to 1.27). In contrast to the post-acquisition period, baseline differences in mortality were similar when looking back to the 2-years prior to entry into each cohort. Compared to patients at non-acquired facilities who initiated dialysis in the baseline period (who served as the "index"), the estimated hazard of death at acquired facilities was 10% lower in the baseline period and 8% lower in the 2-years prior to baseline. The p-value testing whether differences in adjusted mortality between acquired and non-acquired facilities changed in magnitude in the 2-years prior to acquisitions was 0.92.
When examining hospitalization days in the baseline period, patients at independently-owned facilities that were acquired had 2.5 (95% CI -3.3 to -1.6) fewer hospitalization days per patient-year compared to those at non-acquired facilities. Similar to findings from our primary analysis, this difference narrowed to 0.6 in the period following acquisitions (p-value for difference-in-difference estimate of <0.001). Yet, the difference in hospitalization days between acquired and non-acquired independently-owned facilities remained virtually unchanged (at 2.4 fewer hospital days) in the 2-years prior to acquisitions. The p-value testing the significance of this difference was 0.84.
Together, these findings suggest that slower declines in health outcomes following acquisitions at acquired versus non-acquired facilities are unlikely to be explained by random chance. Differences in health outcomes between acquired and non-acquired facilities persisted when looking back in time, and only narrowed following acquisitions.

eTable11. Assessing Differences in Comorbidities among Comparison Groups
Outcomes predicted from observed changes in case-mix and geography Variation in in 1-year mortality and hospital days predicted from observed differences among patient groups in geographic and patient characteristics was relatively small, ranging from 0 to 0.4 in hospital days per patient-year and 0.0 to 0.2 in the 1-year probability of death. While these findings do not directly address unobserved differences in case mix, they suggest that differences in patient and geographic characteristics did not lead to substantial changes in health outcomes.

eFigure2. Unadjusted One-Year Probabilities of Death
Note: Death during study period does not include death following censoring. P-values for differences in probabilities of death in pre-acquisition periods among acquired vs. non acquired facilities were <0.001 at All Facilities and at Chain-owned facilities. Probability of death at independently-owned facilities prior to acquisitions was 20.2% at acquired facilities and 19.5% at non-acquired facilities (p-value for difference-0.21) Note: Unadjusted improvements in rate of death were obtained from the difference of the cumulative probability of survival (using Kaplan Meier method) at each day among patients starting dialysis before versus after acquisitions.

eFigure4. Unadjusted Hospital Days per Patient-Year
Note: Unadjusted hospital days at independently-owned facilities that were acquired declined by 0.2 days following acquisitions. This change was not statistically significant (p=0.3). Among chain-owned facilities prior to acquisitions, hospitalization days per patient-year were 17.3 at acquired and 18.5 at non-acquired facilities. Following acquisitions, hospital days declined by 1.1 (p<0.001) at acquired facilities versus 2.0 (p<0.001) at nonacquired facilities.

eFigure5. Sensitivity Analyses Focusing on Mortality
Detailed Results of Sensitivity Analyses Involving Mortality: Note: All models control for covariates listed in eTable7. "Modification of diff-in-diff at chain-owned facilities" is the estimated coefficient of 'θ' in the stratified regression model. Note: All models control for covariates listed in eTable8. "Modification of diff-in-diff at chain-owned facilities" is the estimated coefficient of 'θ' in the stratified regression model.

Summary of Sensitivity Analysis Results:
Our primary study findings were not sensitive to the alternative model specifications listed above. In particular, the difference-in-differences estimates involving patients at independently-owned facilities remained close to the values observed in our primary cohorts. Acquisitions did not consistently predict changes in health outcomes at chainowned facilities in our primary analyses, and the predicted effects of acquisitions on chain-owned facilities varied somewhat across different model specifications.
In an additional sensitivity analysis (Results not shown), acquisitions of Gambro by DaVita and Renal Care Group by Fresenius were not associated with mortality or hospitalizations. This is consistent with the broader finding that acquisitions of chain-owned facilities were not associated with health outcomes.
We did not find evidence that the estimated effect of acquisitions changed following enactment of ESRD payment reform and we did not observe heterogeneity over time in the estimated effect of acquisitions on hospital days and mortality. (Results not shown)