RETRENCHMENT IN PUBLIC HEALTH UNEQUALLY 1 AFFECTS ACCESS TO HEALTHCARE FOR INCOME 2 GROUPS: A QUASI-EXPERIMENT DURING THE GREAT 3 RECESSION 4

Background: The Great Recession, starting in 2008, was characterized by an overall reduction in living standards. This pushed 11 several governments across Europe to restrict expenditures, also in the area of healthcare. These austerity measures are 12 known to have affected access to healthcare, probably unevenly among social groups. This study examines the unequal effects 13 of retrenchment in healthcare expenditures on access to medical care for different income groups across European countries. 14 Method: Using data of two waves (2008 and 2014) of the European Union Statistics of Income and Living Conditions survey 15 (EU-SILC), a difference-in-differences (DD) approach was used to analyse the overall change in unmet medical needs over time 16 within and between countries. By adding another interaction, the differences in the effects between income quintiles (difference- 17 in-difference-in-differences: DDD) were estimated. To do so, comparisons between two pairs of a treatment and a control case 18 were made: Iceland versus Sweden, and Ireland versus the United Kingdom. These comparisons are made between countries 19 with recessions equal in magnitude, but with different levels of healthcare cuts. This strategy allows isolating the effect of cuts, 20 net of the severity of the recession. Results: The DD-estimates show a higher increase of unmet medical needs during the Great Recession in the treatment cases 22 (Iceland vs. Sweden: +3.24 pp; Ireland vs. the United Kingdom: +1.15pp). The DDD-estimates show different results over the two models. In Iceland, the lowest income groups had a higher increase in unmet medical needs. This was not the case in 24 Ireland, where middle-class groups saw their access to healthcare deteriorate the analyse effects

Some countries made cutbacks in the provision of healthcare services, others increased 102 employee contribution rates (either for the general public or for specific population subgroups), 103 and others increased or introduced user charges for health services. Further, some countries 104 reported expanding benefits, targeting low-income groups. These are important policy 105 measures, because user charges play an important role in the threshold for healthcare access, 106 both for low-value care and for high-value care (which is cost-effective). Chaupain-Guillot and 107 Guillot 2 report a positive relationship between the height of UMN and the height of OOP citizens 108 are to pay. Rice et al. 12 also found a link between the level of OOP expenditures and the non-109 take-up of healthcare when it was considered necessary due to the high costs. They specified 110 that not only the height of the OOP was crucial, but also what citizens were accustomed to 111 paying. Forgoing medical care because of rising OOP is more likely to occur among lower-112 life, but they may also increase the net costs in the longer term 25 . 153 This article aims to study whether a reduction in health expenditures affects access to 154 healthcare differently between income-groups. 155 Using a difference-in-difference-in-differences (DDD) approach enables us to investigate the 156 extent to which budget cuts affect income groups differentially. Overall, recessions imply a 157 decrease in living standards, which affect their access to healthcare because of the increased 158 relative weight of OOP in households' budgets. This direct effect is not our central research 159 question, as this effect has been documented quite well in the existing literature. Our research 160 question looks at the indirect impact of the recession, through the ensuing healthcare budget 161 retrenchment. Our central research question is as to whether the latter affects lower-income 8 The results of this study can inform policymakers in their decisions on how to deal with future 164 challenges that require cuts in public spending while avoiding the increase of unequal barriers 165 to healthcare, which facilitate health inequity. Thus, it may help to avoid or limit large increases 166 in UMN, or at the very least to reduce inequitable effects for low-income groups. In the survey, the respondents were asked if they needed medical care but were unable to 178 take it up during the past 12 months. This indicator, "unmet medical needs" (UMN), can be 179 used as a proxy for experienced barriers in access to care. This approach has been adopted 180 in a large number of previous studies 18-20, 27, 28 . 181 Additionally, respondents are asked about the reason for not taking up medical care, choosing 182 between eight options: could not afford care, waiting times, lack of time, travel distance, fear, 183 wait and see, lack of knowledge, or others. A binary outcome variable was constructed 184 indicating the presence of UMN because of cost-related reasons (direct costs, waiting lists, or 185 travel distance). This variable, also used in previous research 18, 28 , is likely to reflect difficulties 186 in access to medical care due to situations associated with an economic crisis, such as budget 187 cuts, an insufficient supply of healthcare, higher co-payments, and lack of household economic 188 resources. Income groups were created based on the equivalized income; it is the total income To estimate whether retrenchment in healthcare affects the access to care differentially 234 according to the income-group, a DDD is set up. 235 The regression equation of the DDD is: 236 +̂8 . + 238 treatment (in 2014). The equivalized income-groups are divided into quintiles, to analyse the 241 difference between the first quintile (the 20% of the population with the lowest income) with the 242 other groups. This allows us to discover potential non-linear effects, especially to test whether 243 the middles class has evolved differently from the lowest or higher-income quintiles. The First, some scholars have recently argued estimations become more plausible if the cases are 258 similar in levels before the treatment 35 . This indeed seems to be the case in our two pairs. In 259 Sweden (2.4%) and Iceland (1.6%), the reported prevalence of UMN is not that far apart, and 260 the difference is opposite to the post-treatment situation: access to healthcare in Iceland, as 261 measured by UMN, is better before the budget cuts, and worse after. In our main comparison 262 between Ireland (1.7%) and the United Kingdom (1.1%), the pre-treatment differences are 263 particularly small. 264 The second argument supporting the parallel trends assumption is that, before 2008, EU upon in 2000, started to bear some fruit. Scholars seem to argue that the OMC helped to 268 counterbalance potential divergence in policy 36, 37 which would be some reassurance of parallel 269

trends. 270
Finally, we test an alternative empirical strategy that has been proposed. We tested the equal 271 trends assumption by repeating the analysis on cases that are evenly treated. We, therefore, (1,4%) the lowest amount of UMN was found. In 2014 the highest amount of UMN was reported 285 in Iceland (4,4%). 286

288
To estimate the effect of budget cuts in healthcare on the accessibility of care in general a 289 difference-in-differences approach was used. Since this is not the main interest of this paper, 290 the results are presented only briefly. In both cases (Iceland compared with Sweden and 291 Ireland compared with the United Kingdom), the UMN increase more in the country that 292 introduced budget cuts compared with the countries that did not. In Iceland, the UMN increased 293 more compared to Sweden (3.24 pp), as well as in Ireland compared with the United Kingdom 294 (1.15 pp). The output of the DD is found in additional file 2. 295 The core analyses of this paper are the difference-in-difference-in-differences estimations. The 296 DDD tests the impact of reducing health expenditures on access to healthcare in depth. The 297 overall effect of a differential impact of income on UMN, was first estimated by adding 298 equivalized income to the regression as a continuous variable. These results are found in 299 additional file 2. while the difference in trend is even much more marked in quintile 5 (-3.51pp). The results are 307 quite different in Ireland, where UMN increases more in income quintiles 2, 3, and 4 compared 308 to the lowest income group. In income quintile 3 the highest increase in UMN is found 309 compared to income quintile 1 (1.60 pp) over the period 2008-2014. To test the significance of 310 this difference between the income quintiles a chi² tests were carried out. For each year, a 311 separate chi² test was carried out per quintile, using quintile 1 as a reference. This means that 312 4 chi² tests were performed per year (quintile 1-2, quintile 1-3, quintile 1-4, and quintile 1-5). In 313 2008, no significant relationship was found between the income quintile and reporting UMN 314 except from the test with quintiles 1 and 5 (people in quintile 5 were significantly less likely to 315 report UMN than those in quintile 1). In contrast, in 2014, a significant relationship was also 316 found in the chi² test with quintile 3 (χ² (1, n= 4,252) = 7.5; p= 0.006). People in income quintile 317 3 have a higher risk on UMN in 2014 than those in income quintile 1 (additional file 3). The full 318 output of the DD-and DDD-analyses is attached in additional file 2. 319

321
This study aims to estimate whether the effects of budget cuts in the healthcare sector on the 322 accessibility of medical care are different between income groups. To estimate these effects, 323 this study evaluates the change in access to healthcare between income-groups over the 324 period 2008-2014, between countries with similar shocks in the business cycle, but differences 325 in the decrease in health expenditures. This approach allows us to analyse whether the effects 326 of budget cuts on access to healthcare are different between income-groups, isolated from the 327 effects of the recession on living standards. The core contribution of this study deals with the 328 expected differential impact of budget cuts by income group. In this respect, the outcome is 329 dramatically different between Iceland, where no specific measures were taken to protect the 330 lowest income groups, and Ireland, which introduced measures intending to protect low-331 income groups. In Iceland, the lowest income groups were hit especially hard. In contrast, in 332 Ireland, some are holders of a medical card, which gives them free access to a general 358 practitioner (GP) and hospital care, and to prescribed medication at a reduced cost (category 359 I). Others have to pay these costs themselves (category II) 38 . In 2014, the number of medical 360 cardholders was nearly 40% 39, 40 . As entitlement to such a medical card depends on the level 361 of income, most beneficiaries are in income quintile 1 and, to a lesser extent, in income quintile 362 2. As the policy measures implemented by Ireland affected people exclusively from Category 363 II (e.g., increase in user charges and the abolition of automatic entitlement to medical cards 364 for people above 70 years), they were able to guarantee access for low-income groups that 365 were medical cardholders 41 . Our findings confirm the concerns expressed in earlier research 366 about unequal access to care in Ireland, especially for those who are just above the income 367 level to be entitled to a medical card 42 . Also, Schneider and Devitt 21 report a higher increase 368 in UMN during the Great Recession in the higher income groups but no increase in the lower-369 income groups in Ireland. 370 These findings demonstrate the importance of research on the impact of healthcare cuts on 371 access to care, isolating these cuts from changes in living standards. In addition, in-depth research is important to understand which groups of the population are most affected by the 373 measures. 374

375
The use of a difference-in-differences (DD) approach makes it possible to set up a stronger 376 causal design since the effect of budget trends can be isolated, net of differences that already 377 existed before the crisis. This is, even more, the case for the difference-in-difference-in-378 differences design. This allows us to adequately test inequities in healthcare access for 379 different income groups. Thereby the careful selection of treatment and control cases 380 effectively isolates changes in health expenditures, net of the impact of the crisis on living 381 standards. Moreover, the use of multiple cases makes it possible to distinguish between 382 different effects among countries, depending on the type of measures taken by a country. To 383 the best of our knowledge, this more sophisticated set of techniques has not been used before 384 to investigate the differential effect of austerity measures in healthcare on access to care for 385 different income groups. 386 Despite several strengths, this study has some limitations. 387 First, difference-in-differences designs are based on a common trend assumption: in the 388 absence of the treatment, the change in both groups remains the same. It is not possible to 389 test this directly. Therefore, this is typically tested for a period where the treatment did not take reporting may lead to variation in bias between income groups and over time, amongst others, 411 due to adaptation. The subjective relative deprivation theory hypothesizes that, in economic 412 downturns, people adapt their preferences so that those who lose economic resources switch 413 their opinion from 'cannot afford' to 'do not want' to shield themselves from unrealistic goals 43 . 414 Although this theory has only been tested on consumption goods, it might have an impact on 415 a person's perception of healthcare needs. This could lead to an underestimation of the results 416 found. A possible solution would be to use objective data in addition to subjective data, such 417 as real take-up of healthcare (GP consults, number of admissions). Since both subjective and 418 objective data have their limitations, combining both data can increase the quality of research, 419 which is in line with the proposals of Thompson et al. 44 . Moreover, the variable UMN is taken 420 up in the cross-sectional part of the EU-SILC survey. Because cross-sectional data have 421 certain limitations, the use of panel data is recommended for this purpose 44