Determinants of soft budget constraints: How public debt affects hospital performance in Austria

4 Soft budget constraints (SBCs) undermine reforms to increase hospital service efficiency 5 when hospital management can count on being bailed out by (subnational) governments in 6 case of deficits. Using cost accounting data on pub licly financed, non-profit hospitals in 7 Austria from 2002 to 2015, we analyse the associati on between SBCs and hospital efficiency 8 change in a setting with negligible risk of hospita l closure in a two-stage study design based 9 on bias-corrected non-radial input-oriented data en v lopment analysis and ordinary least 10 squares regression. We find that the European debt crisis altered the pattern of hospital 11 efficiency development: after the economic crisis, hospitals in low-debt states had a 1.1 12 percentage point lower annual efficiency change com pared to hospitals in high-debt states. 13 No such systematic difference is found before the e conomic crisis. The results suggest that 14 sudden exogenous shocks to public finances can incr ease the budgetary pressure on publicly 15 financed institutions, thereby counteracting a preexisting SBC. 16


Introduction 21
In their quest to safeguard the financial sustainability of health-care systems, policymakers 22 in several countries have implemented reforms targeting the efficiency of health-care service 23 provision in public hospitals. A prime example is the introduction of payments based on 24 prospective diagnosis-related groups (DRGs) (see Dan (2013), Kittelsen, Magnussen, and 25 Anthun (2007), but also Wagstaff and Moreno-Serra (2009) for surveys). Yet the 26 effectiveness of these reforms is undermined as long as public hospitals can expect to be 27 bailed out in times of financial distress, typically by subnational governments. Indeed, public 28 hospitals are often subject to a soft budget constraint (SBC), i.e. 'an ex ante behavioural 29 regularity, which exerts an influence on the firm's decision' (Kornai, 1979(Kornai, , 1986. Hospital 30 bailouts are often the only politically viable option at hand when policymakers want to avoid 31 snubbing their constituency. 32 Brekke, Siciliani, and Straume (2015) and Shen and Eggleston (2009) use the inverse of 33 the probability of a hospital closure as a measure of budgetary softness. In many cases, 34 however, the probability of hospital closure converges towards zero, if the probability of 35 bailout is virtually 100% in practice. Alternatively, the problem can be expressed via the 36 federal governments' commitment not to bail out additional expenditure at the local level (see 37 Bordignon and Turati (2009)). The central issue, however, remains unchanged: Why should 38 hospital management care about efficiency and not simply act as a budget-maximizing 39 bureaucrat, as outlined in Niskanen's (1968) classic model of bureaucracy? Most likely, there 40 will be an implicit upper limit on the maximum deficit that is tolerated without the 41 owners/financiers replacing the management. Hospital management hence faces the dilemma 42 of maximizing hospital budget and avoiding being laid off. Using the probability of a bailout 43 as a measure of the softness of budgetary constraints neglects this dimension. 44 3 When hospital closures are unlikely, there is an additional caveat to the approach used by 45 Brekke et al. (2015) and Shen and Eggleston (2009). If the probability of hospital closure 46 converges towards zero for all hospitals, the budgetary constraint is equally soft for all 47 hospitals and there should be no systematic differences in the efficiency changes between 48 hospital groups. However, we propose that systematic differences in efficiency change can 49 indeed be observed in countries with subnational autonomy like Austria. While the likelihood 50 of hospital closure approaches zero, the debt burden of the states, which ultimately have to 51 absorb any hospital deficits within the state, significantly influences the degree of budgetary 52 softness, leading to systematic differences in state-level hospital efficiency changes. The 53 financial crisis in 2009 and the subsequent European debt crisis constituted a strong 54 exogenous shock to Austria's public finances. EU legislation, adopted as a consequence of the 55 debt crisis, further exposed so-called 'hidden debts' in Austria, including the debt of publicly 56 owned hospitals. It is likely that revealing the 'hidden debt' of public hospitals further 57 aggravated the problem of public debt, i.e. the compliance with the Maastricht criteria, in the 58 political domain. A key hypothesis for the present analysis is that states with relatively high 59 public debt were hit hardest by this development, which considerably limited the financial 60 leeway of these state governments. The financial crisis, therefore, caused a rift in the 61 budgetary constraints of hospitals in high-debt states, tilting the dilemma of hospital 62 management towards higher budgetary discipline by making running deficits in the aftermath 63 of the financial crisis more problematic. 64 The empirical evidence from Austria is of interest for the following reasons: the Austrian 65 DRG system does not cover the entire costs of publicly and privately owned non-profit 66 hospitals providing publicly funded acute care (for simplicity referred to as 'public hospitals' 67 henceforth). It only stipulates that at least 51% of hospital costs have to be financed out of 68 8 in the behaviour of states and hospitals. Additionally, it is unlikely that a change in the 163 absolute debt ratio would have a uniform effect across the entire spectrum of debt ratios. The 164 time-invariant classification circumvents this problem. The states' debt ratios further cluster 165 the states into two groups (high-debt states and low-debt states). Keeping the number of 166 groups that are compared to each other low allows keeping the number of observations per 167 group as high as possible, which is beneficial in situations with small samples. Lastly, the 168 time-invariant classification is a safer choice, because the timing of any effects is unclear, 169 particularly as some efficiency-enhancing measures may take time to unfold (e.g. when older 170 employees are not laid off, but their position is rather left unfilled once they retire). 171 The rationale behind linking the budgetary situation with the SBC and the hospital 172 efficiency is as follows: the first relevant factor is the financial dependence of the state 173 governments. State governments cannot levy taxes. They depend on the funds allocated based 174 on negotiations with the federal government, creating a situation of vertical fiscal imbalance. 175 A high debt ratio hence increases the dependence of the state government on the federal 176 government, effectively reducing the space for political manoeuvring and the ability to handle 177 costly bailouts. Accordingly, a high debt ratio also increases the credibility of the state 178 government to commit to stricter budgetary rules and not to bail out hospital management. 179 The second relevant factor is the behaviour of hospital management. A priori, we assume 180 hospital managers want to keep their jobs, which could be jeopardized if a bailout is required. 181 However, the blame could be passed to the state government (similar to the blame game in 182 Norway in the 1990s (Tjerbo & Hagen, 2009)), claiming that deficits are due to insufficient 183 funding rather than poor management decisions. This reasoning is easier when the state's 184 resources are abound. The budgetary situation of the states thus increases the stakes 185 9 associated with a bailout and requires both state governments and hospital management to 186 adapt their behaviour. 187 We implicitly assume in our framework that a reduction in inputs does not affect the 188 quality of hospital care. This is a strong assumption, which is required by the missing 189 availability of quality indicators for Austrian hospitals. Strict budgetary discipline may come 190 at the expense of care quality and patients' well-being. But high expenditure levels in most 191 European health-care systems and professional ethos could offset this effect and prevent a 192 substantial decline in quality of care when budgets tighten. Of course, hospitals could also 193 reduce the quality of amenities, e.g. meals, which could affect patient satisfaction but not their 194 well-being. 195 Conversely, lower hospital efficiency may also just reflect higher quality of care. Overall,196 it is unclear, whether the relationship between hospital efficiency, budgetary discipline and 197 quality of care is that close. Empirical evidence suggests that higher efficiency can be realised 198 without curbing quality of care (Piacenza & Turati, 2014;Street, Gutacker, Bojke, Devlin, & 199 Daidone, 2014). 200

Sample and data 201
Austrian hospitals can be classified using various and partly overlapping structural features 202 (Bundesgesetzblatt I, 2017a), including, inter alia, the level of care (standard, extended, 203 maximum, and specialized, the categorization depending on the number and combination of 204 the minimum required medical specialties), type of financing (DRG-based, non-DRG-based), 205 benefit status (non-profit, for-profit) and ownership (public or private). We confine our 206 analysis to DRG-financed non-profit hospitals, because legal requirements, service level as 207 well as cost accounting and performance data are unified for this group. 208 10 In 2015, 120 hospitals were eligible for DRG financing through state health funds, 209 accounting for approximately 71% of the nationwide bed capacity. Of these 120 hospitals, 65 210 provided standard care, 23 offered extended care, 7 maximum, and 25 specialized care. Only 211 29 hospitals were privately owned, 25 thereof by religious orders. The 120 hospitals had costs 212 of around € 12.7 billion in 2015. 213 The accounting data provided by the Ministry of Health cover the years 2002 to 2015. We 214 only focus on the inpatient sector for three reasons: first, the documentation of outpatient 215 services was reformed in 2014, making a structural break in outpatient coding likely. In 216 addition, coding accuracy in outpatient departments before 2014 was not as high as in 217 inpatient departments because the level of outpatient services provided had no impact on the 218 level of funding. Lastly, distortions in outpatient data are likely following differences in the 219 hospital structure and their mapping in terms of coding algorithms (Rous, 2015). 220 The observation period is split into two subperiods (2002 to 2008, and 2009 to 2015) 221 following a major revision of the DRG system, which came into force in 2009, resulting in 222 substantially increased DRG credits per case and changes in the relative cost weights between 223 the different DRG groups. With inputs stable, this would be reflected by sudden and artificial 224 surges and drops in hospital efficiency within a DEA framework. Since the break in the time 225 series coincides with the onset of the financial crisis, we exploit this circumstance to test 226 whether there was a break in the pattern of hospital efficiency change associated with the 227 timing of the financial crisis. By performing the DEA analyses separately for the two 228 subperiods, we do not consider any efficiency changes from 2008 to 2009 that are likely to be 229 skewed by the DRG re-weighting. In contrast to the immediate re-weighting implications, the 230 impact of a budgetary constraint on the catch-up should be more gradual as hospital 231 management requires some time to take action. A gradual effect following a change in case-232 11 mix towards more lucrative DRGs is unlikely, as DRG-funded public hospitals are obliged by 233 law to admit any person in need of care so that patient selection is almost impossible. In 234 addition, hospital management cannot freely decide on the beds per speciality, as these are 235 subject to central planning by the state governments. 236 Although DRG weights are set at the national level, the monetary value of a DRG point 237 may differ between the states, since the ex-ante allocated funds per state are ex-post divided 238 by the total number of DRG points of all hospitals in the respective state. In this regard, a 239 hospitals' ability to generate additional revenue by increasing output (DRG points) are 240 limited. This implies that extra funds are needed in case of overshooting costs. The states have 241 the possibility of allocating funds beyond DRG funds to the different hospitals, not only to 242 address a hospital's specific role in the state's health-care system, but also to cover occurring 243 deficits. How generous this additional funding can be, therefore, depends crucially on the 244 state's financial situation. 245 252

Methods: data envelopment analysis 253
In the first stage, we use DEA to assess hospital efficiency changes over time. Most of the 254 analysed hospitals start from a state of inefficiency, i.e. they use more inputs than necessary to 255 provide a specific output level. By reducing inputs while keeping output stable, these 256 hospitals can improve their technical efficiency. Assuming there are no changes in the 257 production technology (i.e. shifts in the production frontier), hospitals then move closer to the 258 production frontier, i.e. they catch up. We compute the period t catch-up by: 259 with a ܿܽ‫ܿݐ‬ℎ-‫ݑ‬ ௧ < ሺ>ሻ 1 indicating deterioration (improvement) in efficiency from 260 period t-1 to period t. factor prices, information that is unavailable for public Austrian hospitals. 272 There are some known limitations to the DEA approach: first, the units need to belong to the 273 same technological universe, using the same types of input to produce the same types of 274 output. Second, the discriminatory power of DEA depends on the total number of inputs m 275 and outputs s relative to the number of n units assessed. Our sample satisfies the rule of thumb 276 (see, e.g. Cooper, Seiford, and Tone (2007), chapters 1 and 4) requiring that 277 Third, the selection of variables has to be particularly careful since there are no tests for 278 judging statistical significance, or stability of the efficiency results. Fourth, DEA is not robust 279 to measurement errors, especially at the extreme ends of the isoquant, which can affect all 280 efficiency scores by shifting the entire isoquant. 281 We performed sensitivity analyses regarding homogeneity and the choice of variables to 282 address these issues. To increase the robustness, the data-generating process (DGP) is 283 14 simulated using the bootstrap algorithm proposed by Tone (2013), which assumes input and 284 output data to follow a triangular distribution. Since this imparts a stochastic property to the 285 efficiency scores, a second-stage analysis relating them to changes in the softness of the 286 budget constraints using regression analysis can be justified. Mitropoulos et al. (2018)  287 recently used a similar method to make use of a second-stage regression to estimate the effect 288 of hospital reforms following the financial crisis on efficiency development in Greek 289 hospitals. 290 For the DEA, we use the input-oriented non-radial efficiency and super-efficiency models 291 developed by Tone (2001Tone ( , 2002) (see appendix). The input orientation is justified by noting 292 that public hospital management has more discretionary power over inputs than outputs. The 293 use of the non-radial model has the advantage of capturing input savings beyond their 294 proportionate reduction as in the radial alternative. 295 The bias-corrected catch-up is used as the dependent variable to form a panel data set to 296 estimate the relationship between budget constraints and hospital efficiency change. As the 297 catch-up is based on the bias-corrected estimator for the unobserved efficiency obtained in the 298 first stage, we avoid the fallacy of ignoring the bias term owing to the inherent serial 299 correlation in the estimated catch-up (Simar & Wilson, 2007). 300 The DEA input-output specification (Model I) is based on the relevant literature 301 (Anonymous, 2005;Hadji, Meyer, Melikeche, Escalon, & Degoulet, 2014;Hollingsworth, 302 2008;Jakobs, Smith, & Street, 2006;O'Neill, Rauner, Heidenberger, & Kraus, 2008) and the 303 peculiarities of the Austrian hospital system (Anonymous, 2000(Anonymous, , 2014Hofmarcher, Paterson, 304 & Riedel, 2002) and avoids typical pitfalls of DEA applications as described in Dyson et al. 305 (2001). Full-time equivalents (FTEs) of physicians (PHYS), nurses (NURSE) and other staff 306 (OTHER) serve as proxies of labour inputs, imputed costs (including depreciation and 307 15 interest) as a proxy for capital input. Operating (OPER_COST) and secondary costs 308 (SEC_COST) cover the other resources to provide inpatient care. As output, we use DRG 309 credits (CREDITS), which reflect case-mix and thus severity-adjusted services. 310 To judge the stability of efficiency results, we investigate two additional input-output 311 specifications. In Model II, we decompose operating costs into medical (MED_COST) and 312 non-medical operating costs (NONMED_COST) to see if differences in resource use for 313 medical supplies and consumables affect catch-up (see, e.g. Anonymous (2015)). In Model III 314 we decompose DRG credits into credits based on major medical procedures 315 (MEL_CREDITS) and other credits (mostly credits that use the principal diagnosis to charge 316 the services) (HDG_CREDITS), because we assume specialized hospitals to be efficient in 317 the production of either MEL or HDG credits, but not necessarily in their aggregate. 318  (Banker & Chang, 2006;Hofmarcher et al., 2002) for the input-output specifications in Table  325 1. Outliers are hospitals with super efficiency higher than 1.5 times the inter-quantile range 326 (25% and 75%). As Austrian public hospitals are subject to a unified cost accounting and 327 reporting software, outliers are assumed to result from measurement error, DRG upcoding, or 328 inhomogeneous technology. In any case, hospital efficiency is likely to be distorted. 329 In the case of hospital mergers, we compute efficiency scores for the respective subperiod 330 using virtual mergers between the merged hospitals in the years prior to the merger. In the 331 case of mergers of hospitals operating at different care levels, the care level specification of 332 the actual merger is used retrospectively for the virtual merger. Finally, in the case of zero 333 inputs, the hospital is dropped from the sample for the relevant subperiod, including hospitals 334 that ceased their operations during the subperiod.    We run the following time-invariant regression model separately for the pre-crisis and 357 post-crisis periods: 358 Y ୧୲ = α + ‫ܴܣܧܻ‪β‬‬ ௧ + γ‫ܤ‬ + δ X ୧୲ + ε ୧୲ (2) 359 where the dependent variable Y ୧୲ is the bias-corrected catch-up of hospital i at time t. YEAR ୲ is 360 a categorical variable capturing countrywide development at time t, B ୧ is a categorical 361 variable indicating the budgetary situation for hospital i with γ capturing the effect of interest. 362 X ୧୲ is a matrix capturing a variety of additional hospital characteristics (e.g. type, ownership, 363 case-mix, patient structure) we control for. X ୧୲ also includes a variable that decomposes 364 hospitals in efficiency quartiles based on the efficiency in (2002 and 2009, respectively) to 365 account for the heterogeneity in efficiency change due to their starting position in the two 366 18 subperiods (2002−2008 and 2009−2015).  The model is estimated with pooled OLS using White's heteroscedastic-consistent standard 371 errors, which provides consistent estimates for DEA scores in a second-stage regression 372 (Hoff, 2007;McDonald, 2009) and the catch-up as it has similar statistical properties. The 373 estimation strategy does not suffice to establish a truly causal relationship between the SBC 374 and the efficiency development. However, it can still highlight systematic differences between 375 groups of hospitals (budgetary situation) following a common shock to public finances, 376 hinting at a relationship. 377 378 20

Results 379
We report the bias-corrected catch-up based on 1,500 replications at the state level 380 separately for the two subperiods (Table 3 and Table 4)

394
The effect of the states' debt ratios on hospital efficiency changes via the channel of 395 tightening or relaxing the SBC is more thoroughly isolated through the second-stage 396 regression specified in Equation (2). Relevant results using Model I are reported in Table 6. 397 The hospital-level covariates ownership, hospital type, population density of the catchment 398 area, patient structure, case-mix and size (using the actual number of beds/100 as a proxy to 399 allow for meaningful effects of one-unit changes) do not influence efficiency changes. 400 With regard to the debt ratio, we find no systematic differences across hospitals in the pre- This effect is significant at the 5% level (p-value=0.019). Controlling for the heterogeneity in 404 the initial efficiency levels using efficiency quartiles (see columns (3) and (4)  With an R² ranging from roughly 0.01 to 0.03, the overall fit of the regression models is 417 low. This could indicate that the catch-up is not well described by the chosen covariates and 418 differences are likely to be caused by unobserved confounders, such as managerial ability, 419 informal structures of leadership, etc. These variables are, however, not available. But 420 considering the high volatility of efficiency scores obtained from DEA and that hospital-level 421 data is used -which results in a low number of not only observations but also groups for 422 comparison -the low model fit is not surprising. 423 The effect of the debt ratio is not robust to alternative DEA model specifications. In Model 424 II, the effect of the comfortable budgetary situation loses size and significance in the post-425 crisis period. In Model III, the budgetary effect misses significance at the 10% level in the 426 post-crisis period, albeit not by very much. The size of the effect remains roughly the same. 427

Conclusions 428
In this paper, we analyse the effect of a tightening of budget constraints on hospital 429 efficiency change of Austrian DRG-financed hospitals resulting from an exogenous shock to 430 public finances. We use an input-oriented slacks-based DEA efficiency model to compute the 431 annual catch-up over a period of 13 years. In the second stage, we analyse the impact of the 432 budgetary situation of the states on hospital efficiency change using a pooled OLS regression. 433 The main motivation to investigate the association of the budgetary situation of the financing 434 body and the hospital efficiency is that any policy measures to increase the efficiency of 435 hospital service provision are undermined whenever hospitals are subject to a SBC. 436 We consider the peculiarities of the Austrian health-care system as we neither abstract 437 from the problem that, in practice, hospital closure might be virtually impossible, nor do we 438 assume the existence of SBCs to be exogenously given. We rather argue that a tightening or 439 24 further softening of an SBC is closely related to the financial situation of the financing 440 government body, i.e. the state government. We thereby explore the possibility of using the 441 state government's public debt ratio as a proxy for the changes in the SBC. 442 Using cost-accounting data from Austrian DRG-financed hospitals from 2002 to 2015, we 443 find that hospitals with low initial levels of efficiency have successfully improved efficiency. 444 Hospitals with high initial levels of efficiency seem to face less pressure to further improve 445 efficiency so that their catch-up is considerably lower. The results could reflect that it is 446 probably simpler to reap higher efficiency gains from low initial levels compared to higher 447 efficiency levels. It could also be argued that the results only show that reforms aimed at 448 improving efficiency -mostly targeted at low-performance hospitals -were successful. But 449 this is not the entire story. We find a change in the pattern of nationwide hospital efficiency There are some limitations to the study design, which may impact the result. First, there are 464 only nine Austrian states. As the debt ratio is defined at the state level, only nine groups are 465 available for comparison, which makes it more difficult to obtain significant results, 466 particularly when the effect is not very strong. This is also a possible explanation for why the 467 results somewhat depend on the DEA model specifications (in addition to the generally low 468 number of observations, and the varying sample sizes due to different outliers). We further 469 stress that the high volatility of the efficiency scores impedes a causal interpretation of our 470 regression results. As our estimates could be subject to omitted variable bias, they should be 471 interpreted as explorative. A second limitation is that there is still no quality indicator 472 available for hospital services. A good opportunity for future research would be to use a 473 broader definition of efficiency that also includes a quality dimension, allowing for the 474 possibility that decreases in cost can come at the expense of the quality of the health-care 475 services provided to patients. And lastly, a possible shortcoming of measuring the state debt 476 ratio as financial debt to overall budget is that the debt ratio does not include information 477 about the assets of states vis-à-vis their financial liabilities, as the relevant information was 478 not available for a sufficiently long period.

Data envelopment analysis: method
Assume that = , … , hospitals produce = , … , outputs, using = , … , inputs . The production possibility set is assumed to satisfy the axioms stated in Charnes, Cooper, and Rhodes (1978) and Banker, Charnes, and Cooper (1984). The input-oriented nonradial efficiency for hospital k, which belongs to the set ‫ݐ‬ of units to be analysed ('reviewed') and uses , input quantities to produce , output quantities, is evaluated with respect to the reference set ݂ as follows, The super-efficiency measure δ ᇱ satisfies δ ᇱ ≥ 1, with δ ᇱ > 1, indicating the minimum average expansion rate of inputs, which still guarantees that the pertinent unit is located on the frontier of reference set ݂. As only non-oriented models ensure feasibility of super-efficiency models, infeasible solutions to (8) to (12) may also occur.
To increase the robustness of the efficiency scores, we use the bootstrapping technique proposed by Tone (2013) to simulate the data-generating process (DGP) as follows: 36 Step 1 Compute the input-oriented non-radial efficiency ߜ based on the actual input and output data Step 2 Repeat the following substeps ܾ = 1, … , ‫ܤ‬ times: i. Simulate the input/output data assuming a triangular distribution for the inputoutput data with data variations being taken from historical data ii. Compute the input-oriented non-radial efficiency ߜ መ based on the simulated input and output data We then derive a bias-corrected efficiency ߜ ሚ by