Impact of vertical integration on the readmission of individuals with chronic conditions

Universidade NOVA de Lisboa Escola Nacional de Saúde Pública Impact of vertical integration on the readmission of individuals with chronic conditions Master in Health Management Óscar Ricardo Brito Fernandes Lisbon, July 2016 Universidade NOVA de Lisboa Escola Nacional de Saúde Pública Impact of vertical integration on the readmission of individuals with chronic conditions Dissertation presented to Escola Nacional de Saúde Pública in partial fulfilment of the requirements for the Degree of Master in Health Management, developed under the supervision of Professor Rui Santana and Professor Sílvia Lopes. Lisbon, July 2016 The Escola Nacional de Saúde Púbica and Universidade NOVA de Lisboa are not responsible for the opinions expressed in this publication, which are the sole responsibility of its author. i ACKNOWLEDGMENTS I am grateful to my supervisors, Professor Rui Santana and Professor Sílvia Lopes, whose expertise, support, and generous guidance made it possible for me to develop this study. It was a huge privilege and pleasure to embark on this journey with both of them. I want to thank Ana Patrícia Marques, Bruno Moita and João Sarmento for all the methodological discussions and support. I am also indebted to Portuguese Central Administration for Healthcare System (ACSS) for providing the administrative datasets for this work; Centers for Medicare and Medicaid Services for answering so promptly to all my questions and requests; and to Escola Nacional de Saúde Pública and its staff for providing me the best working and learning environment possible. I would like to express my gratitude to all my teachers who put their faith in me and urged me to do better, with special thanks to Professor Adalberto Campos Fernandes, Professor Alexandre Abrantes, Professor Ana Escoval, Professor Carlos Costa, Professor João Pereira, Professor Paulo Boto, and Professor Pedro Aguiar. I want to acknowledge all my classmates who have attended this Master programme with me for the opportunity to learn from their experiences, with special thanks to Joana Fernandes and Ana Resende for being present, for their friendship and support throughout all phases. I would also like to emphasise the recent friendship with Ana Antunes, who became a strong support throughout the development of this work. I would like to thank Professor Rui Baptista-Gonçalves for his friendship and help on revising the writing of this dissertation. My deepest gratitude goes to Sónia Bernardes Mateus, whose love, friendship, and care made this entire journey possible. I will always be indebted for what you have done for me. I want to acknowledge Professor Margarida César. Despite not intervening directly on this work, the way I see myself as a research is greatly influenced by her. I also want to thank my parents and brothers for all their love and support throughout my life. Last, but definitely not least, I would like to acknowledge the love, support and sense of family that Pedro Lobo Julião and Ana Paula Lobo brought into my life, and also for believing in me and making this journey possible. For a shared vision. ii iii The only true voyage of discovery, the only fountain of Eternal Youth, would be not to visit strange lands but to possess other eyes, to behold the universe through the eyes of another, of a hundred others, to behold the hundred universes that each of them beholds, that each of them is (…). Proust, in Remembrance of things past: The prisoner (Vol. 5) (1923). x xi LIST OF ACRONYMS ACSS – Portuguese Central Administration for Healthcare System AHRQ – Agency for Healthcare Research and Quality CCI – Chronic Condition Indicator CCS – Condition Classification System CMS – Centers for Medicare and Medicaid Services DiD – Difference-in-differences ICD-9-CM – International Classification of Diseases, 9th revision, Clinical Modification LHU – Local Health Unit NHS – National Health Service OECD – Organization for Economic Co-operation and Development SRR – Risk-Standardized hospital-wide 30-day Readmission Ratio WHO – World Health Organization xii 1 INTRODUCTION Healthcare systems sustainability is a global concern, aggravated by the demographical and epidemiological changes, and with the increase burden of disease and chronic conditions creating one of the biggest strain. Thus, these threats shape the way healthcare delivery is designed[1] and how healthcare organizations are managed. Vertical integration of healthcare providers is a way to respond to these challenges, addressing differentiation and fragmentation, as well as the healthcare needs of individuals with chronic conditions and multi-comorbidities, since they require a consistent and long term care attention.[2] Furthermore, vertically integrated healthcare providers are expected to develop better transitions of care[3], with one of the estimated effects being a decrease on the readmission frequency of individuals with chronic conditions. Because readmissions are costly[4] and threaten healthcare systems’ sustainability, increase individuals’ vulnerability and expose them to several hospital level risks[5], and being individuals with chronic conditions the ones who face higher likelihood of readmission, it is urgent to investigate if and how vertical integration is addressing these concerns. In Portugal, vertical integrated care is materialized by Local Health Units (LHU) and arose from the expectation of effective coordination between primary and hospital care, in order to better respond to the needs of the population.[6,7] Since vertical integrated care approaches to healthcare are expected to decrease readmissions[8], we presume that LHU has that effect on individuals with chronic conditions. Evidence from other countries suggests a mixed impact regarding integrated healthcare and hospital readmission[9–14], thus we consider research over this topic essential for a better integrated care and centred on individuals’ points of view and needs. In addition, in Portugal there is no study, as far as we know, that evaluated LHU effects over the readmission of individuals with chronic conditions. Thus, with this research we aim to bridge this gap. This dissertation is organized in six chapters. On Chapter 1, we introduce a theoretical framework concerning integrated care, what has been done in Portugal regarding this organizational principle for healthcare delivery, the burden of chronic conditions, the importance of hospital readmissions as well as the effects of integrated care on the readmission of individuals with chronic conditions. On Chapter 2, we focus on the aims of the research, presenting the problem that configured the investigation and 2 a set of objectives proposed to be accomplished within this study. On Chapter 3, we present the methodology of the study, starting with the study design and the explanations for our choices. We also share information regarding data sources, criteria for inclusion and exclusion of observations, explanations for variables and a description of statistical analyses conducted. On Chapter 4, we present the major findings produced within this study. On Chapter 5, we provide a discussion of main findings with previous studies and theoretical framework regarding the three major vectors of this work: integrated care (with focus on vertical integration), chronic conditions and hospital-wide all-cause 30day unplanned readmissions. We also present some of the limitations we had to face while developing this study, insights over further research and recommendations for the future. Finally, on Chapter 6, we summarize our work, and present final remarks. 3 CHAPTER 1 BACKGROUND In Portugal, the population’s life expectancy has been consecutively improving over the last few years.[15] However, when compared to other European countries, Portuguese men and women aged 65 and over, live fewer healthy years.[15] Thus, ageing populations[16,17], the prevalence of multimorbidity[17], and the current increasing pressure of burden of disease challenge healthcare systems[18,19] and the way healthcare delivery is designed[1]. Some authors[20] stress that the argument is not the ageing population but the underlying health burden of chronic conditions that created one the biggest strain on healthcare systems. Throughout the years, there have been continuous political movements to reorganize the structure of healthcare delivery, with the aim of improving the quality of care and simultaneously make it more cost-effective. Thus, the National Health Service (NHS) sustainability is assumed as a difficult structural challenge, along with the transformation of the NHS towards a system where health promotion and disease prevention are the driving forces. Portuguese integrated care experiences, with horizontal and vertical integration phenomenon, are a sign of the efforts that materialize this healthcare policy. 1.1. INTEGRATED CARE AND THE PORTUGUESE EXPERIENCE Shaw and co-authors[21] define integrated care as an organizational principle for care delivery, and integration as the methods, processes and models to achieve integrated care, as a managerial response to differentiation and fragmentation[22]. On this study we assume World Health Organization (WHO) definition of integrated health services as “the management and delivery of health services such that people receive a continuum of health promotion, disease prevention, diagnosis, treatment, disease-management, rehabilitation and palliative care services, through the different levels and sites of care within the health system, and according to their needs throughout the life course”[22, p49]. Minkman[19] states that the growing relevance of integrated care is related to the reshape of healthcare systems by: i) socio-demographic and epidemiological transitions; ii) the growing number of individuals with complex care needs with chronic conditions 4 and comorbidities; iii) engagement of individuals and caregivers in the treatment decision-making; iv) new professions arising from the demanding communications between specialists and generalists and; v) the complex and fragmented healthcare systems. Integration can focus on primary and secondary care, but may also have a wider scope and involve social care[24], insurance or other human service systems as stated by Leutz[25]. The common focus of many integrated care approaches includes the support to individuals with chronic conditions to live more independently[11,26], with improvements to their care experience. Vertical integration is one of the structural dimensions of integrated care.[27] This process can be characterized by the transformation of the elements of a particular organizational structure, giving rise to a new structure, involving new management relations, and communication flow responsibilities. Vertical integration is therefore an organizational conceptualization of a healthcare structure, which involves creating a single management entity, and at least two units providing healthcare at different levels of care. Santana[28] points out that the cumulative conditions for recognition of vertical integration are: i) the existence of a single entity responsible for health; ii) in a given geodemographic space, well defined and limited; iii) serving a population covered by the entity; iv) with coordinated efforts among all healthcare units in the integrated care system. There is a global shift concerning integrated care[19], being given more frequently visibility to the experiences in the U.S.A. and United Kingdom[29]. Despite the many integrated care approaches across the world, there is not just a single model that best fits integrated care needs. The development of integrated care services is a non-linear and dynamic process[19], requiring the system to foster an adaptive and transformative culture, and being promoted as an approach for improving accessibility, affordability, quality of care and effectiveness, with a special focus on people with complex needs[30–32]. For recognizing the frailties in developing integration across different providers, Goodwin and co-authors[33], through case studies, argue that the starting point to integrated care should be a clinical/service model designed to improve care instead of an organizational model with a pre-determined design. Despite being a current theme in healthcare research[34], there is no widely accepted definition of integrated care[19,22,30,32,35] or a set of core factors that facilitate integrated care to be more person-centred. Hence, there is frequent confusion with 5 organizational structures and processes such as disease and case management with integrated care.[31] In addition, cultural contexts and settings configure integrated care definitions. This lack of focus of integrated care can make it difficult to implement and obtain desired outcomes, namely a decrease on the number of readmissions[36]. Portugal has been showing some legislative sensibility regarding care integration, some of it disruptive, attempting to adopt the best international practices (e.g., the first law of mental health, Law 2118/1963, 3 April).[37] The most recent Portuguese experience regarding vertically integrated care has begun in 1999, ahead of other European countries[13]. However, there is a lack of a systematic evaluation of its purpose as well as the dissemination of results and impacts.[8] The Local Health Units (LHU) arose from the expectation of effective coordination between primary and secondary care, in order to better respond to the needs of the population.[6,7] LHU are responsible for the health of a population, providing healthcare under the coordinated management of a hospital (or hospitals, in the case of a hospital centre) and primary healthcare units. In Table 1, we see that the first LHU was created in 1999 (LHU Matosinhos) by Decree-Law 207/99, 9 June[38], merging one hospital and 26 primary care providers, covering over 175 000 inhabitants. For eight more years, this was the only LHU. Only between 2007 and 2009 were created new LHU, providing integrated care for inhabitants from Alentejo, North, and central Portugal. In 2011 and 2012, two more LHU were created, one in the North another in Alentejo. Nowadays the LHU network provides healthcare for 1 145 904 inhabitants, around 11.6% of the Portuguese mainland population. Table 1: Evolution and attributes of local health units (LHU) network. LHU Year Health region Number of providers Population* Legal framework Hospital Primary care† Matosinhos 1999 North 1 26 175 321 DL 207/99, 9 June[38] Norte Alentejo 2007 Alentejo 2 94 115 663 DL 50-B/2007, 28 February[39] Alto Minho 2008 North 2 58 242 159 DL 183/2008, 4 September[40] Baixo Alentejo 2008 Alentejo 2 83 124 690 Guarda 2008 Centre 2 88 144 273 Castelo Branco 2009 Centre 1 86 105 944 DL 318/2009, 2 November[41] Nordeste 2011 North 3 98 140 440 DL 67/2011, 2 June[42] Litoral Alentejano 2012 Alentejo 1 50 97 414 DL 238/2012, 31 October[43] * Resident population by county in LHU’s catchment area was retrieved from National Statistics Institute on May 2016. Last data update by June 16, 2015. † The number of primary care providers is the sum of health centres and their extensions, as well as family health units. Information retrieved from [44] . 6 OECD[45] stresses that Portuguese healthcare needs to evolve towards a more integrated approach, increasing and deepening relations among the various levels of care, including palliative care. To prevail the success and effectiveness of these movements, a reflection period is mandatory over the existing experiences of integrated care. However, there are only a few studies about this model of organization and management of healthcare (e.g., 15, 16, and 36). There is a lack of evidence regarding economic, financial and covered population’s health outcomes, namely on individuals with complex needs of care like those with (multiple) chronic conditions. 1.2. CHRONIC CONDITIONS Chronic conditions refer to health conditions that persist across time (for at least 12 months), that require on-going healthcare interventions, either medical products, services or special equipment, and/or places limitations on self-care, independent living, and social interactions.[47] This definition also includes non-communicable diseases, mental disorders, some communicable conditions and on-going physical impairments.[48] The burden of chronic conditions has been globally underestimated by society, political and health systems regarding its transversal effects.[49,50] Chronicity is a challenge to all developed[51,52] and developing countries[50], especially because a source of the rise in cost for healthcare systems comes from the fact individuals with chronic conditions receive fragmented care.[20,53,54] Thus, international governments have committed to an integration of healthcare that specifically addresses the growing prevalence of chronic conditions in the population.[51,55,56] However the evidence of integrated interventions designed to address individuals with chronic conditions healthcare needs seems to be sparse and inconsistent.[57] In Portugal, according to the WHO[58], the population aged between 30 and 70 years have a 12% probability of dying from one of the four major chronic conditions (cancer, diabetes, cardiovascular diseases or chronic respiratory diseases). Hence, more than half of the Portuguese population (54.8%) is at risk of premature mortality due to these non-communicable diseases. The WHO[58] estimates that 86% of the Portuguese deaths are explained by one or more chronic conditions, and according to 2012 information, about 97 000 deaths were due to, at least, one chronic condition. Cardiovascular diseases were responsible for 32% of all deaths, cancers by 28%, chronic respiratory diseases accounted for 6% of all deaths and diabetes by 5%.[58] 7 According to Anderson[49], the growing trend of the presence of chronic conditions on individuals’ health illuminates less successful aspects in the evolution of health systems, historically centred in the treatment of infectious and acute diseases. Also, individuals with chronic conditions are more likely to experience hospital readmissions[53,59] since they are more vulnerable to the consequences of non-effective integrated care. 1.3. HOSPITAL READMISSIONS Hospital readmission is considered as a new inpatient episode after an index admission (i.e., an initial hospital admission) within a given period. Thus, one’s choice regarding time frame, or relations with the index admission (e.g., principal diagnosis) are crucial to define readmissions. Readmission, in the current study, is defined as a subsequent inpatient admission to any acute hospital for any condition that occurs within 30 days of the initial discharge date. Readmissions are a complex organizational and clinical challenge since not all readmissions can be prevented. Readmissions can be grouped as planned/unplanned and if related or not to the index admission diagnosis.[60] Planned readmissions are not considered a sign of poor quality of care as they are identified on the discharge process and part of the individual’s treatment. On the other hand, admission for acute conditions or complications are most likely not planned, thus a concern for healthcare systems and users[61]. Thus, reducing the volume of unplanned readmissions is a frequent strategy for improving the quality of healthcare. This research will focus on unplanned readmissions only. An all-cause readmission measure stresses the fact that, from an individual perspective, a readmission, regardless the cause, is always an adverse event[62]. Individuals are more vulnerable when they are away from their home, family and communities, exposed to several hospital level risks, and vulnerable to the experience that Krumholz[5] dubbed the post-hospital syndrome. The 30-day period for outcome measure is a much used benchmark value. It is more useful for evaluating the effectiveness of hospital discharge and post-acute period[63,64], even though some authors propose a time frame of three to seven days or a weighting scheme based on the days to readmission[65–67]. 14 LHU and control group. Therefore, in order to address our main aim, DiD was the adequate technique. In order to develop a better understanding of the readmission phenomena in LHU and control group, we studied the associations between individuals’ risk factors and time to readmission. We conducted a Cox regression to analyse the influence of gender, age, number of chronic conditions and Elixhauser comorbidities on the risk of readmission in LHU and control group. We performed DiD analyses considering two dependent variables: i) natural logarithm of the risk-standardized readmission ratio (SRR) and; ii) occurrence of readmission (risk of readmission). We considered SRR, a hospital level measure, as dependent variable because it allowed studying the effects of vertical integration in a particular hospital’s performance to be compared to an average hospital’s performance, adjusting for discharge volume. We chose to use SRR methodology approach because it is public[95,96], there is a dense body of work of published literature regarding its utilization (e.g., 60, 65, and 66) and is still currently being discussed and challenged among researchers, policy makers, American hospitals and other stakeholders. On the other hand, the risk of readmission is an individual level measure that emphasises the probability of readmission given one’s characteristics. Since the SRR is expected to be influenced by disruptions in the continuum of care, inadequate inpatient care and discharge planning, among other factors, we decided to study if vertical integration of healthcare had an impact on diminishing these effects on readmissions, thus reducing SRR. At an individual level, vertical integration is expected to decrease one’s risk of readmission by promoting better continuum of care, considering a healthcare person-centred approach. Hence, we decided to study if vertical integration of healthcare had an impact on decreasing individuals’ risk of readmission, given one’s characteristics (age, principal diagnosis, presence of selected comorbidities, number of chronic conditions and comorbidities). 3.2. DATA SOURCE AND INCLUSION/EXCLUSION CRITERIA The Portuguese Central Administration for Healthcare System (ACSS) provided the datasets. Datasets refer to administrative data from 2002 to 2014. These datasets include inpatient claims data with individual information regarding hospitalizations within public mainland hospitals, namely a unique encoded identifier, year and hospital 15 of treatment, gender, age, principal diagnosis, secondary diagnoses, procedures, type of admission, discharge date, length of stay, and discharge status. All diagnoses and procedures were coded using International Classification of Diseases, 9th revision, Clinical Modification (ICD-9-CM). Because the number of secondary diagnoses on these datasets varied in recent years, in order to get comparability among all datasets throughout 2002 to 2014, we only used the first 20 diagnoses (one principal diagnosis and 19 secondary diagnoses). Figure 1 presents the exclusion criteria applied, being selected 9 523 432 inpatient index admissions. From these, we selected for our analyses all episodes treated at LHU (n=845 275) and at control hospitals (n=834 359), totalizing 1 679 634 index admissions. Episodes (2002 to 2014) n=20 152 283 Outpatient episodes n=7 760 877 Unknown gender n=176 Length of stay <0 or >365 days n=1 241 Age <0 or >118 years n=42 Error DRG (469 or 470) n=7 644 Error in principal diagnosis n=11 Specialized or low volume hospitals n=132 514 Discharge status different than home n=1 169 580 Episodes without at least 30 days post-discharge n=911 485 Admissions for primary psychiatric diagnoses n=154 993 Admissions for rehabilitation n=3 430 Admissions for medical treatment of cancer n=486 858 Final index admissions n=9 523 432 Figure 1: Exclusion criteria applied to all episodes from Portuguese mainland public hospitals from 2002 to 2014. 16 3.3. CONTROL GROUP In order to conduct the DiD analyses, a hospital control group was defined. Hospitals in this group were not vertically integrated, but were affected by the same external systemic effects that might have influence on readmissions. The control group was derived using ACSS’s benchmarking hospital groups (except psychiatric and public-private partnerships because these might be affected by different systemic effects). These benchmarking hospital groups were created by hierarchical clustering.[99] Hospitals from the control group had to belong to benchmark group B or C, since all LHU were clustered between those groups. After the application of the exclusion criteria for hospitals, from benchmark group B we selected the three remaining hospitals, and from group C we randomly selected three. 3.4. VARIABLES 3.4.1. READMISSION Readmissions were identified using Centers for Medicare and Medicaid Services (CMS) hospital-wide all-cause unplanned readmission measure[96] (1: readmission). First, we identified all index admissions, and selected those that could be analysed within a 30day time frame. With the final set of index admissions, we identified episodes followed by an unplanned readmission and the ones that were planned. This study focuses on the 30-day unplanned readmissions. 3.4.2. COMORBIDITIES AND CHRONIC CONDITIONS In order to better describe individuals’ comorbidities and chronic conditions we used two measures: Quan’s[100] updated version of Elixhauser comorbidity index[101] and Agency for Healthcare Research and Quality (AHRQ) chronic condition indicator[102]. Elixhauser comorbidity index is a method of categorising individuals’ comorbidities and is widely used with administrative data to measure burden of disease. Only secondary diagnoses were included to compute the index. This index was designed to be used with specific ICD coding, therefore we used ICD-9-CM 5-digit codes to ascertain the presence of comorbidities (1: comorbidity present). We used AHRQ chronic condition indicator to 17 ascertain the presence of ICD-9-CM diagnoses codes considered as chronic conditions (1: chronic condition present). We used these measures for its complementary information regarding individuals’ health status. For instance, the drug abuse Elixhauser comorbidity group is considered a comorbidity for 292.x ICD-9-CM codes, but the code 292.0 (drug withdrawal syndrome) is considered a chronic condition while 292.11 (drug paranoid state) is not. Moreover, Elixhauser comorbidity index was reported with the ability to predict length of stay[103], and was also found significantly associated with healthcare expenditures and in-hospital mortality[104]. On the other hand, the chronic condition indicator works as a proxy to assess one’s overall burden of chronic illness, engagement with the healthcare system and the total number of conditions being treated[105]. 3.4.3. INDIVIDUAL RISK OF READMISSION Individual risk of readmission was estimated using generalized linear mixed models at the specialty cohort level described elsewhere[96] (1: readmitted). For a given specialty cohort, we fitted a hierarchical logistic regression model to account for the natural clustering of observations within hospitals. From these models we obtained the risk of readmission for individual i at hospital j in year t as !!"#$ %&'()*+,-./01"# , where2is the adjusted average outcome (being readmitted within 30-day time frame) over all hospitals and 1"# a set of risk factors that are clinically relevant and strongly associated with the outcome (age, principal diagnosis and selected comorbidities). Age was treated as a continuous variable. For principal diagnosis, we used the AHRQ Condition Classification System (CCS) to group hospitalizations into clinical-coherent, mutually exclusive condition categories. Comorbid diseases were identified using the CMS Condition Category groups. To reduce bias for the risk of readmission we fitted the model considering all episodes (n=9 523 432). The model performance, assessed by the use of c-statistic, the area under the receiver operating characteristic (ROC) curve, varied between 0.60 to 0.71, similar to the work by Horwitz and colleagues[96] (0.62 to 0.67). 18 3.4.4. RISK-STANDARDIZED READMISSION RATIO The risk-standardized readmission ratio (SRR) reflects the number of readmission events at a hospital level, relative to the number of readmission events that would be expected based on average hospital performance. Therefore, this measure allows a particular hospital’s performance to be compared to an average hospital’s performance. A lower ratio (<1) stands for lower-than-expected readmission rate, and similarly a higher ratio (>1) stands for higher-than-expected readmission rate. We chose to use as dependent variable the natural logarithm of SRR for three reasons: i) to reduce outlier noise; ii) to correct positive skewness, since SRR measure is always positive and; iii) to facilitate regression coefficients interpretation, i.e., marginal changes in the explanatory variables are interpreted in terms of percentage change in the dependent variable. We provide further explanation regarding SRR in Appendix 1. 3.5. STATISTICAL ANALYSIS 3.5.1. DESCRIPTIVE STATISTICS First we used frequencies analysis to describe LHU and control hospitals regarding gender, age groups (0-19; 20-44; 45-64; 65-84; and 85+), number of chronic conditions and comorbidities (0, 1, 2, 3, 4, and 5+), for the period 2002-14. We also listed selected principal diagnosis, based on the principal diagnoses that were simultaneously most frequent, with higher readmission frequency and readmission rate: diabetes mellitus with complications, congestive heart failure, pneumonia, acute bronchitis, chronic obstructive pulmonary disease and bronchiectasis, and urinary tract disease. This selection accounted for 10% of episodes and 25% of readmissions. We computed the readmission rates for LHU and control hospitals and followed the same presentation scheme. In addition, we graphically analysed the evolution of crude readmission rates from 2002 to 2014, for LHU and control group. However, because the chronicity structure of our sample might affect the interpretation of the readmission rates, we also analysed the evolution of standardized readmission rates. For that, we applied a direct method of standardization, considering as standard population the total number of individuals distributed by the number of chronic conditions. 19 We also compared the variation of readmission rates throughout the years by number of chronic conditions, for LHU and control group, and compared readmission rate evolution with the national average readmission rate. 3.5.2. ASSOCIATIONS BETWEEN INDIVIDUALS’ RISK FACTORS AND TIME TO READMISSION We used a Cox regression (also known as proportional hazards regression analysis) to study the association between the time interval between admissions and possible covariates of a readmission. Various methodologies and techniques have been used for risk of readmission, with logistic and Cox regression being the most common methods used to identify risk factors.[106] The hazard function in this case refers to the probability that an individual will experience readmission within a 30-day time frame (risk of readmission at time t). Covariates were gender, age group, number of chronic conditions, and number of Elixhauser comorbidities. All covariates association with time interval between admissions were assessed by a preliminary univariate Cox regression. Any covariate with 34567 was used after in a multivariate Cox regression analysis. The multivariate model used a backward likelihood ratio stepwise data selection method. A cut-off value of 34 5658 was applied to remove covariates from the final model. Verification of the proportional hazards assumption was based on a visual inspection of Kaplan-Meier plots. These analyses were conducted separately for LHU and control group index admissions. 3.5.3. IMPACT OF VERTICAL INTEGRATION 3.5.3.1. Impact on the risk-standardized readmission ratio The effect of vertical integration on the risk-standardized readmission ratio can be estimated by comparing the treatment and control group before and after integration. Let &9 :!! #$ be the natural logarithm of the risk-standardized readmission ratio for hospital j at year t. Let ; #$ be a binary variable taking on value 1 if hospital j at year t is LHU and 0 otherwise; <$ a set of year dummies capturing period specific effects; =>2is a constant term and ? #$ an idiosyncratic term. We considered the following model estimated by a pooled OLS: Model 1: &9 :!!#$ %=>.@; #$ .<$2.A#$ 20 The estimated coefficient @ stands for the DiD. To account for possible unobservable characteristics that might affect SRR over time, we adjusted standard errors by clustering at the hospital level. This mitigates bias over @. In order to account for time-invariant unobservable at hospital level we considered a set of hospital dummies B#$ 62We also considered the average number of chronic conditions CDEFGH'9#$ 2and Elixhauser comorbidity index CDEI&)JGKLMNH #$ 2 for all individuals at hospital j as covariates thought to potentially influence readmission frequency, thus influencing SRR. Hence, the new model can be estimated as a fixed effects model: OPQRS2TU2&9 :!!#$ %=>.@; #$ .<$.B#$ .CDEFGH'9#$ .CDEI&)JGKLMNH #$ .A#$ In order to relax the parallel trend assumption we used a DiD model with a differential trend model.[107] Incorporating the assumption we obtain the following fixed effects model: OPQRS2VU2&9 :!!#$ %=>.@; #$ .B#$ .CDEFGH'9#$ .CDEI&)JGKLMNH #$ .W#XYNKH X Z [\] .^ #XYNKH X; #$ . Z [\] A#$ The parallel trend assumption can be verified by testing the non-linear restriction _`a b cde f`a b cde %56 Non-rejection of the null hypothesis of Wald’s statistical test provides evidence in favour of the parallel trend assumption. 3.5.3.2. Impact on the risk of readmission We used DiD technique to analyse the effects of vertical integration on the risk of readmission. We considered an unconditional logit model with fixed effects using dummy variables. There is still discussion regarding the use of dummy variables similarly as fixed effects. Greene[108] states that bias in estimators is large when number of years of data is small. Because we have a large number of observations per hospital and the number of years considered in this analysis for each LHU is T=8, according to Coupé[109], bias in the unconditional estimator is small and, for the purpose of our research we find it acceptable. Let the dependent variable Y "#$ be a binary variable for episode i in hospital j at year t, assuming value 1 if it is a readmission episode and 0 otherwise. Let ; #$ be a binary 21 variable taking on value 1 if hospital j at year t is LHU and 0 otherwise; g'M*"#$2 takes value 1 if admission of individual i occurred in hospital j in the year post-integration and 0 otherwise; hHNK* #"$ takes value 1 if admission happened in a LHU; <$ is a set of year dummies capturing period specific effects;22!!)i* is the individual risk of readmission for individual i admitted in hospital j in year t, computed through a logistic generalized linear mixed model as discussed by Horwitz and colleagues[96]; =>2is a constant term and ?"#$ an idiosyncratic term. The Model 1 estimated @ coefficient stands for the DiD. To account for possible unobservable characteristics that might affect risk of readmission over time, we adjusted standard errors by clustering at the hospital level to mitigate bias over @. Model 1: &'()* gH'j Y "#$ %7%=>.@;"#$ .kg'M*"#$ ./hHNK*#"$ .<#$ .!!"#$ 2.A"#$ We considered a set of hospital dummies B#$ 2to account for time-invariant unobservables at hospital level.2We also considered the number of chronic conditions FGH'9"#$ 2and the Elixhauser comorbidity index I&)JGKLMNH "#$ for each index admission at hospital j as covariates thought to potentially influence risk of readmission. Hence, the new model can be estimated as a “fixed effects” unconditional logit model: OPQRS2TU2&'()* gH'j Y "#$ %7%=>.@;"#$ .kg'M*"#$ ./hHNK*#"$ .<#$ .B#$ .FGH'9"#$ .I&)JGKLMNH "#$ .!!"#$ .A"#$ We used Wagstaff and Moreno-Serra’s differential trend model[107] in order to relax the parallel trend assumption. Incorporating this assumption in the model, we obtained the following: OPQRS2VU2&'()* gH'j Y "#$ %7%=>.@;"#$ .kg'M*"#$ ./hHNK*#"$ .B#$ .FGH'9"#$ . .2I&)JGKLMNH "#$ .!!"#$ .W#XYNKH X Z [\] .^ #XYNKH X;"#$ . Z [\] A"#$ The parallel trend assumption can be verified by testing the non-linear restriction _`a b cde f`a b cde %56 Non-rejection of the null hypothesis of Wald’s statistical test provides evidence in favour of the parallel trend assumption. We also developed a model that specifies each year’s impact. In this case, because we have the effect of a treatment over different periods, Autor[110] states that it is suitable 22 to interact the treatment variable with time dummies. Let l #$ be a treated/control dummy equal to 1 if hospital j is in the treatment group and 0 otherwise. Then we construct Model 4 as: OPQRS2mU2&'()* gH'j Y "#$ %7%=>.@nl #$ ] n\+o .<$.B#$ .FGH'9#$ .I&)JGKLMNH #$ .2!!"#$ .A"#$ All time dummies interacted with l #$ are expressed relatively to the omitted time period which serves as baseline (year I-5), thus only years I-4 to I+2 will be presented (being I the year when LHU was created). It is possible to see how the effect varies over time, if it stays constant, decreases or increases by analysing2@>, @,, and @]. 3.5.4. SENSITIVITY ANALYSIS We recalculated DiD for risk of readmission under an alternative constitution of the control group to determine the impact of control group choice. This new control group accounted for hospital’s organizational evolution between the period of 2002-14. We excluded all specialized and teaching hospitals. We compared the odds ratios results for each LHU against the new control group with those obtained with the initial control group. We considered that the previous result was stable when there was less than 5% difference between odds ratios results for the same period. The generalized linear mixed models to compute individual risk of readmission were run using SAS University Edition. The Cox regressions were performed using IBM SPSS software version 23 and the DiD analyses were performed using Stata software (v. 13). 23 CHAPTER 4 RESULTS 4.1. CHARACTERISTICS OF THE SAMPLE Our sample had 1 679 634 index admissions (Table 2). The sample was evenly distributed, with LHU accounting for 50.3% of index admissions. Index admissions from female individuals were more frequent on both LHU and control group (55.8% and 57.1%). Individuals with admission at LHU were older compared to those from control group. The average age for LHU individuals was 51.3±28.4 years and for control group 48.5±28.9. The distribution of number of chronic conditions and comorbidities was similar in LHU and control hospitals. Pneumonia was the most prevalent principal diagnosis in LHU and control group hospitals (4.4% in LHU and 4.7% in control group). Table 2: Characteristics of the sample by gender, age group, number of chronic conditions, Elixhauser comorbidity index, and selected principal diagnosis, in the period 2002-2014, for treatment (LHU) and control group. Total LHU Control group N % N % N % 1 679 634 100.0% 845 275 100.0% 834 359 100.0% Gender Female 947 826 56.4% 471 566 55.8% 476 260 57.1% Male 731 808 43.6% 373 709 44.2% 358 099 42.9% Age 0 - 19 362 884 21.6% 171 594 20.3% 191 290 22.9% 20 - 44 308 598 18.4% 148 511 17.6% 160 087 19.2% 45 - 64 319 657 19.0% 159 725 18.9% 159 932 19.2% 65 - 84 555 524 33.1% 295 788 35.0% 259 736 31.1% 85+ 132 971 7.9% 69 657 8.2% 63 314 7.6% Number of chronic conditions 0 997 634 59.4% 498 858 59.0% 498 776 59.8% 1 282 529 16.8% 143 726 17.0% 138 803 16.6% 2 196 736 11.7% 99 850 11.8% 96 886 11.6% 3 120 637 7.2% 60 780 7.2% 59 857 7.2% 4 52 686 3.1% 26 437 3.1% 26 249 3.1% 5+ 29 412 1.8% 15 624 1.8% 13 788 1.7% 30 – 1.465). The increased number of comorbidities represented increasing risk of readmission, both at univariate and multivariate models. Individuals with five or more comorbidities in LHU faced a likelihood of readmission 2.509 times higher than those that had no comorbidities. In the control group, the likelihood of readmission of these individuals with multiple comorbidities were 2.403 times higher than reference category. 31 Table 5: Multivariate Cox regression for LHU and control group index admissions. LOCAL HEALTH UNITS CONTROL GROUP Covariate Scoring Unadjusted hazard ratio (95% CI) Adjusted hazard ratio (95% CI) Unadjusted hazard ratio (95% CI) Adjusted hazard ratio (95% CI) Gender 0: Male (Reference) (Reference) (Reference) (Reference) 1: Female 0.863 (0.846 - 0.880) 0.906 (0.889 - 0.924) 0.797 (0.782 - 0.811) 0.861 (0.845 - 0.878) Age 0: 0 - 19 (Reference) (Reference) (Reference) (Reference) 1: 20 - 44 0.957 (0.920 - 0.995) 0.928 (0.893 - 0.966) 0.706 (0.680 - 0.733) 0.683 (0.658 - 0.710) 2: 45 - 64 1.096 (1.056 - 1.137) 0.839 (0.807 - 0.872) 1.007 (0.973 - 1.042) 0.713 (0.688 - 0.740) 3: 65 - 84 1.970 (1.912 - 2.030) 1.281 (1.238 - 1.325) 2.059 (2.004 - 2.116) 1.197 (1.158 - 1.236) 4: 85+ 2.784 (2.685 - 2.887) 1.716 (1.647 - 1.787) 3.167 (3.065 - 3.272) 1.755 (1.689 - 1.823) Number of Chronic conditions 0 (Reference) (Reference) (Reference) (Reference) 1 1.792 (1.745 - 1.840) 1.298 (1.251 - 1.347) 2.001 (1.950 - 2.053) 1.398 (1.347 - 1.452) 2 2.255 (2.193 - 2.318) 1.287 (1.230 - 1.347) 2.699 (2.629 - 2.771) 1.456 (1.391 - 1.524) 3 2.578 (2.498 - 2.661) 1.266 (1.203 - 1.333) 3.215 (3.122 - 3.310) 1.472 (1.398 - 1.549) 4 2.787 (2.669 - 2.910) 1.233 (1.160 - 1.312) 3.385 (3.252 - 3.522) 1.396 (1.314 - 1.483) 5+ 3.174 (3.014 - 3.343) 1.201 (1.118 - 1.291) 3.759 (3.573 - 3.954) 1.362 (1.267 - 1.465) Elixhauser comorbidity index 0 (Reference) (Reference) (Reference) (Reference) 1 1.792 (1.748 - 1.838) 1.280 (1.233 - 1.330) 2.044 (1.994 - 2.095) 1.285 (1.237 - 1.335) 2 2.368 (2.304 - 2.433) 1.604 (1.532 - 1.679) 2.771 (2.701 - 2.843) 1.583 (1.512 - 1.657) 3 2.833 (2.737 - 2.933) 1.896 (1.797 - 2.001) 3.490 (3.382 - 3.602) 1.935 (1.836 - 2.040) 4 3.454 (3.290 - 3.626) 2.296 (2.150 - 2.452) 3.946 (3.767 - 4.133) 2.192 (2.053 - 2.339) 5+ 3.800 (3.569 - 4.046) 2.509 (2.318 - 2.716) 4.306 (4.039 - 4.590) 2.403 (2.215 - 2.608) 32 4.3. IMPACT OF VERTICAL INTEGRATION 4.3.1. READMISSION RATES BEFORE AND AFTER VERTICAL INTEGRATION We analysed crude readmission rates for LHU and control group, before and after vertical integration, for the group of individuals with no chronic conditions, with one to three chronic conditions, and with more than three chronic conditions (Figure 4). Overall, data suggests the same decreasing pattern for LHU and control group for individuals with no chronic conditions. For individuals with chronic conditions, we obtained mixed results, specifically for the individuals with more than three chronic conditions. For individuals with no chronic conditions, readmission rates faced a reduction in all LHU after vertical integration, being more expressive in LHU 1 (-0.8%). The same decreasing pattern occurred in the control group. For individuals with one to three chronic conditions, most of LHU decreased readmission rates after vertical integration, with LHU 5 decreasing by -0.9% and LHU1, LHU 6, and LHU 7 decreasing -0.5%. Only LHU 4 experienced an increase on readmission rates (1.4%). The control group faced an increase in readmission rates for these individuals. For the individuals with more than three chronic conditions, data suggests mixed results. While LHU 1, LHU 4, and LHU 6 increased the readmission rate after vertical integration, the others were able to decrease. The decrease was most expressive in LHU 3 and LHU 5 (-1.8%). The control group in the period after the creation of LHU 3, LHU 4, and LHU 7 also experienced an effect of reduction on the readmission rates. 33 No chronic conditions One to three chronic conditions More than three chronic conditions Figure 4: Average readmission rates (%) for each LHU and control group, before and after vertical integration, by number of chronic conditions. 34 4.3.2. IMPACT OF VERTICAL INTEGRATION ON THE RISK-STANDARDIZED READMISSION RATIO We computed the SRR for LHU and control group in the period 2002 to 2014 (Figure 5). For all years, the control group presented a higher SRR than LHU. From 2007 to 2010, LHU experienced a high decrease of SRR. Later in the control group, from 2009 to 2011, we observed the same effect. Figure 5: Risk-standardized readmission ratio (SRR) for LHU and control group in the period 2002-2014. We performed DiD models with different specifications with natural logarithm of risk-standardized readmission ratio as dependent variable. In the last model we also ran a Wald test to check the parallel trend assumption for DiD (Table 6). For LHU 1, vertical integration has promoted a decrease in SRR, consistent with all three models. The result for LHU 2 shows a decrease in SRR in model 1 and 2, but an increase in model 3. However, the coefficient from model 3 belongs to the 95% confidence interval of model 2. The Wald test was significant, therefore parallel trend assumption cannot be assured, and one cannot attribute these results solely to vertical integration. There is evidence that vertical integration promoted a 9.4% increase in SRR for LHU 3, but it was not significantly different from the control group. For LHU 4, LHU 5 and LHU 6, we observed a reduction of SRR varying from -4.8% to -7.2%, but without statistical significance, meaning there is no significant differences in risk-standardized readmission ratio between the LHU and the control group in the period preand postintegration. For LHU 7, the results show an increase of 23.4% of SRR in model 3, but this was not statistically significant. 35 Table 6: Difference-in-differences models for risk-standardized readmission ratio for each LHU compared to the control group, in the period I-5 to I+2. Model 1 Model 2 Model 3 Specification test Coefficient (95% CI) R2 Coefficient (95% CI) R2 Coefficient (95% CI) R2 Wald test p-value LHU 1 - 29.3% 0.319 - 19.6% 0.798 - 30.8% 0.817 0.00 0.983 (-41.0% ; -17.6%) (-31.8% ; -7.4%) (-45.1% ; -16.4%) LHU 2 - 17.3% 0.193 - 4.7% 0.787 4.7% 0.794 5.30 0.021 (-29.8% ; -4.8%) (-15.5% ; 6.0%) (-16.5% ; 25.9%) LHU 3 12.8% 0.232 6.9% 0.727 9.4% 0.743 0.32 0.571 (1.4% ; 24.3%) (-4.3% ; 18.1%) (-9.8% ; 28.6%) LHU 4 10.3% 0.269 1.2% 0.792 - 4.8% 0.800 0.69 0.406 (-0.1% ; 20.7%) (-4.9% ; 7.3%) (-18.1% ; 8.6%) LHU 5 - 32.0% 0.297 - 7.8% 0.829 - 3.4% 0.831 0.11 0.742 (-43.8% ; -20.3%) (-17.3% ; 1.7%) (-14.5% ; 7.8%) LHU 6 3.0% 0.145 - 2.6% 0.694 - 7.2% 0.718 0.00 0.964 (-8.7% ; 14.7%) (-17.3% ; 12.1%) (-24.2% ; 9.9%) LHU 7 - 0.2% 0.194 14.6% 0.677 23.4% 0.684 0.57 0.450 (-13.9% ; 13.5%) (-7.3% ; 36.6%) (-3.2% ; 50.0%) 4.3.3. IMPACT OF VERTICAL INTEGRATION ON THE RISK OF READMISSION We performed a logistic regression to study DiD of risk of readmission in LHU and in the control group. In Table 7, we summarize the results for the three different models specified. In the last model we conducted a Wald test to check for the DiD parallel trend assumption. For LHU 2 and LHU 4, the Wald test was significant, meaning that the effects observed in model 3 (despite being significant or not) cannot be solely attributed to vertical integration. Other factors could have happened that contributed to these results, and were not captured by the model specification. The LHU 1 and control hospitals had a total of 600 086 index admissions, being 69 725 from LHU 1. Model 3 suggested a higher risk of readmission for LHU 1 when compared to the control group (OR 1.017; 95% CI: 0.940 – 1.101). For LHU 2, model 3 suggested a decrease in the risk of readmission (OR 0.991; 95% CI: 0.952 – 1.032). Although the result was not significant for model 3, model 2 suggested the same decrease in the risk of readmission but with statistical significance. For LHU 3, all models suggested that vertical integration promoted a decrease in the risk of readmission when compared to the control group. Model 3 suggested an odds ratio of 0.911 (95% CI: 0.837 – 0.991). For LHU 4, all models indicated a higher risk of readmission despite vertical 36 integration, but only model 3 was statistically significant (OR 1.240; 95% CI: 1.149 – 1.338). However, the parallel trend assumption was not verified, thus results should be interpreted with caution. LHU 5 performed similarly in all three models. Results suggest that integration promoted a significant decrease in the risk of readmission (OR 0.860; 95% CI: 0.790 – 0.936). For LHU 6, model 1 and model 2 showed a decrease in the risk of readmission, but none with statistical significance. Model 3 suggested an odds ratio for risk of readmission of 1.076 (0.992 – 1.166). Lastly, LHU 7 performed similarly in the three models, all suggesting a decreased risk of readmission. The parallel trend assumption was confirmed by Wald’s test of significance, thus we can state that there is no statistical evidence that vertical integration promoted a decrease in the risk of readmission, with LHU 7 performing better than the control group but without statistical significance. 37 Table 7: Difference-in-differences models for risk of readmission for each LHU compared to the control group, in the period I-5 to I+2. Model 1 Model 2 Model 3 Specification test Sample size (LHU cases) Odds Ratio (95% CI) R2 Odds Ratio (95% CI) R2 Odds Ratio (95% CI) R2 Wald test p - value LHU 1 600 086 0.872 0.078 0.875 0.079 1.017 0.079 2.15 0.142 (69 725) (0.807 – 0.941) (0.810 – 0 .946) (0.940 – 1.101) LHU 2 594 776 0.899 0.074 0.902 0.075 0.991 0.075 5.03 0.025 (62 738) (0.846 – 0.956) (0.846 – 0.961) (0.952 – 1.032) LHU 3 603 417 0.949 0.083 0.949 0.083 0.911 0.083 1.06 0.304 (94 839) (0.859 – 1.049) (0.856 – 1.052) (0.837 – 0.991) LHU 4 589 376 1.042 0.079 1.046 0.079 1.240 0.079 10.74 0.001 (64 711) (0.976 – 1.111) (0.980 – 1.117) (1.149 – 1.338) LHU 5 614 334 0.844 0.079 0.846 0.080 0.860 0.080 3.73 0.05 (83 973) (0.779 – 0.914) (0.779 – 0.919) (0.790 – 0.936) LHU 6 667 178 0.976 0.079 0.980 0.080 1.076 0.080 0.94 0.331 (136 817) (0.923 – 1.057) (0.905 – 1.061) (0.992 – 1.166) LHU 7 527 353 0.947 0.087 0.946 0.088 0.937 0.088 0.05 0.820 (28 246) (0.832 – 1.079) (0.826 – 1.083) (0.866 – 1.013) 38 4.3.3.1. Analysis of the evolution of risk of readmission before and after integration We performed a logistic regression to analyse DiD in the evolution of risk of readmission in the period pre-integration (I-4 to I-1) and post-integration (I to I+2) in LHU compared with the control group (Figure 6). An overall look at Figure 6 shows that there was no clear pattern regarding the evolution of risk of readmission. LHU 1 in the pre-integration period presented an increased risk of readmission; in the year I+1 and I+2, the risk of readmission decreased (OR 0.892 and 0.779), presenting statistical significance in year I+2. For LHU 2, the risk of readmission decreased between year I-4 and I-2, then increased until year of integration. In the period after integration, the risk of readmission in LHU 2 was significantly lower when compared to the control group (OR 0.863 and 0.875). For LHU 3, the risk of readmission decreased until year I-2, and then increased in year I-1. Since vertical integration, the risk of readmission decreased in this LHU. In LHU 4 and LHU 6, the risk of readmission did not present a clear pattern in the period prior to vertical integration. Nevertheless, since integration the risk of readmission had been decreasing. We observed a decreased risk of readmission for LHU 5, being statistically significant since year I-1. On the other hand, LHU 7 presented in the same time frame increased risk of readmission, even though without statistical significance. The risk of readmission presented a decrease tendency, but not significantly. 39 Figure 6: Evolution of risk of readmission (odds ratio) in the period I-4 to I+2 (95% CI) for all LHU compared to the control group. !"# !"$ !"% !"& '"! '"' '"( '") '"* !"#!"$!"%!"&!!'& !'% ()*+& !"# !"$ !"% !"& '"! '"' '"( '") !"#!"$!"%!"&!!'& !'% ()*+% !"# !"$ !"% !"& '"! '"' '"( '") !"#!"$!"%!"&!!'& !'% ()*+$ !"# !"$ !"% !"& '"! '"' '"( '") '"* '"+ !"#!"$!"%!"&!!'& !'% ()*+# !"# !"$ !"% !"& '"! '"' '"( !"#!"$!"%!"&!!'& !'% ()*+, !"# !"$ !"% !"& '"! '"' '"( '") '"* !"#!"$!"%!"&!!'& !'% ()*+- !"# !"$ !"% !"& '"! '"' '"( '") '"* '"+ !"#!"$!"%!"&!!'& !'% ()*+. 46 Much work has to be done at a clinical level to develop a more integrated approach, but also at the community-level. Increased number of chronic conditions are more associated with economic difficulties[130] or low levels of health literacy, thus producing a denser and more complex problem to address. Since readmission for these individuals is more frequent than desired, and care for individuals with chronic conditions is still much focused on an acute care model, different and multiple approaches have to be developed to minimize this problem and to promote a paradigmatic shift. Burke and colleagues[3] found evidence that monitoring and managing symptoms after discharge, enrolling help from individuals’ communities, and promoting conditions for self-management through health literacy are associated with decreased risk of readmission, but so is a more effective use of information and communications technology across healthcare providers[78]. Additionally, what individuals and carers perceive as the reasons for – and preventability of – readmissions[62], have indeed an effect on avoiding readmissions[76] and on configuring healthcare professionals’ practices. 5.2. LIMITATIONS OF THE STUDY This study presents a set of limitations that deserve some attention. Our research relies on administrative data and used a retrospective study design, therefore is limited in its ability to prove causation. Another limitation is due to the model we selected to identify readmissions, as well as to predict individual risk of readmission for each index admission. One might assume stated readmission rates as conservative in the sense that, and as referred by another study[131], after a person being discharged alive, one might have died outside the hospital due to the care received. Thus, readmission was prevented and was not considered in this study. This situation is more feasible to happen for individuals with more complex conditions. There are many models, with varying complexity, to predict readmissions[132], but we chose to use solely the CMS methodology[95,96], that could be considered as analytical bias. The selected chronic condition indicator sets another limitation. The identification of chronic conditions would be different if other aggregator was used (e.g., Tonelli and colleagues[133]). However, we chose AHRQ because its development is well documented, it assesses each ICD-9-CM code to check if the diagnosis is clinically considered to be chronic or not, and offers the possibility to analyse chronicity through body systems. 47 Another limitation stands out because this is an observational study rather than a controlled trial. For that reason, we might have incurred in some selection bias on choosing control groups (for the main analyses, and for the sensibility analysis) to develop the DiD analyses. Also, the DiD analysis using a logistic regression is slightly biased as previously assumed, but since bias is residual it represents no major prejudice to our study aims. We also point another possible bias, this time related to selection of LHU index admissions. We did not account for the area of residence of individuals treated at these units. It would be very difficult to do so because of the evolution of hospitals distribution across mainland Portugal throughout the years, and the intense hospital horizontal integration phenomena that occurred within that period. Despite these limitations, and because we did not limit the population under study to a particular age group or to a set of specific conditions like most studies, it can add new information to debate. In addition, this study brings to the fore new information regarding an under-evaluated policy measure (the creation of LHU), and provides more information on the evolution of readmission rates, prevalence of chronicity and comorbidities and their association with readmission. 5.3. FURTHER RESEARCH AND RECOMMENDATIONS Our main goal was to study the impact of LHU on the readmission of individuals with chronic conditions. Our findings suggest mixed results, so further research is required to develop a deeper understanding of the impact of LHU on individuals’ health outcomes and therefore adjust integrated care policy to more effective paths. The understanding of the effects of social and organizational mechanisms, and their interaction over readmission, is already complex. However, it seems vital to a better understanding of the phenomena to measure how socio-economic factors increase the risk of readmission, specially on those with most complex care needs. Or even to establish casual pathways between socio-economic factors with readmission. To measure these variables and incorporate them into risk-assessment models is complex, but most likely necessary. There is also a research gap regarding how LHU are addressing more complex individuals’ healthcare needs, and how readmissions are being avoided. Our mixed findings over integrated care effects on decreasing risk of readmission for individuals with chronic conditions suggest the existence of different integrated care approaches, and 48 most likely different levels of integration intensity. Therefore, it seems crucial to understand how some LHU developed positive results in decreasing readmissions over the years, while other did not. Crossing our findings with what has been the Portuguese integrated care policy framework, we would like to suggest some recommendations. First, it is imperative the development of an integrated care policy, with specialized working teams with the ability and competence to address the organizational challenges that integrated care practices demand, and also address the training needs of LHU professionals. The services of these teams could also be made available for any public hospital who wished to develop integrated care practices, despite not being formally constituted as LHU. The effectiveness and effects of integrated care policy and specialized working teams for a better integrated care has to be measured. The first LHU was created in 1999, and from 2007, a new expansion movement occurred. However, there is still a lack of evidence of this network’s effects on delivering better health outcomes (namely on decreasing readmission). A consistent evaluation throughout time facilitates a deeper knowledge of integrated care impacts and stresses the possibility to readapt this policy measure to current healthcare needs, namely addressing the complex needs of people with chronic conditions and multi-comorbidities, and promote an organizational cultural change among healthcare providers of different levels. Second, person-centeredness is at the core of integrated care. However, individuals’ voices (and carers’ voices) are quite silent on the construction of a more integrated care approach, with deeper connections on a community-level. So, a national strategy to address chronic conditions should be configured by the ones who will benefit from it the most. This strategy should be intrinsically linked to the recent national education program for health, literacy and self-care[134], empowering individuals with healthcare needs and carers to a better control over their health, hence decreasing acute admissions and readmissions. Third, addressing individuals with chronic conditions needs is a complex task, because it involves different levels of care, each with its own specificities. Therefore, in order to overcome a hospital-centric care model for these individuals, and thus promoting better and more effective integrated care, payment schemes should considerer adjustedreadmission criteria as a financial incentive to organizations’ budgets. And fourth, health regions should have a more proactive role engaging healthcare organizational structures and populations in the promotion of better health outcomes.[45] 49 On the one hand, health regions should develop a better understanding of the evolution of populations’ health. Cumulatively, health regions should act as dialogue engaging partners for a more integrated care, despite formal constitution of healthcare providers as LHU. Primary and secondary healthcare should develop regional specific disease management programmes or a wider chronic condition care model, involving other levels of healthcare, namely palliative care. A more integrated approach with palliative care can act as key anchor on decreasing readmissions[80], specifically on those with multimorbidities. 50 51 CHAPTER 6 FINAL REMARKS With this study, our main aim was to understand the impact of vertical integration on the readmission of individuals with chronic conditions. In order to address our aim, we undertook an outcome research, with a longitudinal and retrospective observational design. We compared 30-day readmissions before and after the creation of seven LHU. We considered an eight-year time frame for each studied LHU, five years before integration and three years after, and used difference-in-differences to address our main aim. In order to understand the associations of time to readmission with individuals’ risk factors we developed a Cox regression model for LHU and control group. Cox regression results suggest that for LHU and control group hospitals female individuals are less at risk of readmission, the risk increases with increasing age and number of comorbidities. At LHU, we observed a decreased risk of readmission with increasing number of chronic conditions, suggesting LHU could manage more effectively readmissions. Difference-in-differences results suggest that vertical integration promoted a decreased of risk-standardized readmission ratio in four LHU. Also, when analysed the individual risk of readmission we observed that it was reduced for four LHU, but only significantly for two (LHU 3 and LHU 5). When we performed a sensitivity analysis, annual evolution of odds ratio of risk of readmission was stable for most years. Integrated care was supposed to address more effectively individuals with chronic conditions and complex healthcare needs, reducing the effects of fragmentation and differentiation across different levels of providers, with a decreasing effect over readmissions. Addressing this problem is a concern to healthcare systems because of ageing populations, increasing prevalence of multiple chronic conditions and burden of disease, hence pressuring healthcare systems sustainability. Thus, reducing the volume of unplanned readmissions is a frequent strategy to promote a more sustainable system, and with better quality for its users. However, our findings suggest mixed results, with some LHU decreasing their risk of readmission for individuals with chronic conditions, but others did not. This is a sign that even for integrated care, addressing solutions to such a challenge as readmissions is complicated, mainly due to the many factors that contribute to their understanding and decreasing (e.g., inpatient care quality, after discharge care, comorbidities, and social 52 determinants). Still, reducing hospital readmissions is a priority for health stakeholders. Therefore, much has to be done to develop a more integrated approach, but also at the community-level. However, the success of Portuguese vertical integration experiences cannot be assessed solely by its effects on the readmissions of individuals with chronic conditions. Instead, a wider set of indicators measuring different dimensions should be considered, as well as an analysis to the efforts developed to mitigate evidence-based known barriers to vertical integration of care. So, in order to promote a better healthcare to individuals with chronic conditions, namely protecting them from readmission, healthcare organizations, despite being formally constituted as LHU, should: develop integrated care pathways for the most prevalent chronic conditions on their catchment area; revise discharge processes; continuously evaluate health outcomes; and share best practices of integration involving community and other levels of care (namely palliative care). These challenges upon LHU, but also to all healthcare organizations, require a new culture of sharing and openness from healthcare providers. Within this paradigmatic shift, healthcare providers could better address healthcare fragmentation and develop healthcare approaches more person-centred, with its basis on health education and promotion. In a time where citizens are given unrestricted movement across all hospitals within Portuguese NHS context, empowering citizens’ decision-making over their health and well-being, this research adds a space-time of discussion over the design of vertically integrated care approaches and their impacts on health outcomes and system sustainability. Consequently, with this study we also expected to bridge the gap of little evidence regarding both the Portuguese experiences of vertical integration of healthcare and readmission frequency of individuals with chronic conditions, inspiring different level providers to move towards a more integrated healthcare approach. 53 REFERENCES 1. Mohrman SA, Kanter MH. Designing for health: learning from Kaiser Permanente. In: Organizing for sustainable health care [Internet]. Emerald Group Publishing Ltd.; 2012. p. 77–111. Available from: http://tinyurl.com/hxncvt3 2. Strandberg-Larsen M, Krasnik A. Measurement of integrated healthcare delivery: a systematic review of methods and future research directions. Int J Integr Care [Internet]. 2009;9(Jan-Mar):e01. Available from: http://tinyurl.com/z3jp486 3. Burke RE, Guo R, Prochazka A V, Misky GJ. Identifying keys to success in reducing readmissions using the ideal transitions in care framework. BMC Health Serv Res [Internet]. BioMed Central; 2014;14:423. Available from: http://tinyurl.com/heupula 4. Golmohammadi D, Radnia N. Prediction modeling and pattern recognition for patient readmission. Int J Prod Econ [Internet]. Elsevier; 2016;171:151–61. Available from: http://tinyurl.com/jkmsc77 5. Krumholz HM. Post-hospital syndrome: an acquired, transient condition of generalized risk. N Engl J Med [Internet]. 2013;368(2):100–2. Available from: http://tinyurl.com/glazncf 6. Fernandes J, Barros P, Fernandes A. Transformar para servir melhor [Transform to better serve]. In: Três olhares sobre o futuro da saúde em Portugal [Three points of view about the future of healthcare in Portugal]. Cascais: Princípia; 2011. p. 51–81. 7. Santana R, Marques P, Lopes S. Definição de um modelo de acompanhamento da actividade desenvolvida pelas Unidades Locais de Saúde e monitorização da modalidade de pagamento aplicada às Unidades Locais de Saúde em 2009: relatório II - diagnóstico de situação [Definition of a model of activity monitoring developed by the Local Health Units and monitoring the payment method applied to the Local Health Units in 2009: report II - diagnostic] [Internet]. Lisbon; 2009. Available from: http://tinyurl.com/hjupcwc 8. Santana R, Marques AP, Barba A, Rua T, Sarmento J, Funenga I. Uma alteração de paradigma [A paradigm shift]. In: Santana R, Marques AP, editors. A integração vertical de cuidados de saúde [Vertical integration of healthcare]. Lisboa: Escola Nacional de Saúde Pública; 2016. p. 19–69. 9. Nolte E, Pitchforth E. What is the evidence on the economic impacts of integrated care? [Internet]. Permanand G, editor. WHO Regional Office for Europe and European Observatory on Health Systems and Policies; 2014. Available from: http://tinyurl.com/jbjlfff 10. Walston SL, Kimberly JR, Burns LR. Owned vertical integration and health care: promise and performance. Health Care Manage Rev [Internet]. 1996;21:83–92. Available from: http://tinyurl.com/gtfvxny 11. Dorling G, Fountaine T, McKenna S, Suresh B. The evidence for integrated care [Internet]. 2015. Available from: http://tinyurl.com/zr76jqc 12. White B, Carney PA, Flynn J, Marino M, Fields S. Reducing hospital readmissions through primary care practice transformation. J Fam Pract [Internet]. 2014;63(2):67–74. Available from: http://tinyurl.com/jut6umb 13. Polanco NT, Zabalegui IB, Irazusta IP, Solinís RN, Del Río Cámara M. Building integrated care systems: a case study of Bidasoa Integrated Health Organisation. Int J Integr Care [Internet]. 2015;15(Apr-Jun):e026. Available from: http://tinyurl.com/gv2lpd7 14. Kodner DL. Managing high-risk patients: the Mass General care management programme. Int J Integr Care [Internet]. 2015;15(September):1–15. Available from: http://tinyurl.com/gkrxhfk 15. Crisp N, Berwick D, Kickbusch I, Bos W, Antunes JL, Barros PP, et al. Um futuro para a saúde: todos temos um papel a desempenhar [A future for health: we all have a role to play] [Internet]. Lisboa: Fundação Calouste Gulbenkian; 2014. Available from: http://tinyurl.com/grpf8ow 16. Lee KH, Low LL, Allen J, Barbier S, Ng L, Ng M, et al. Transitional care for the highest risk patients: findings of a randomised control study. Int J Integr Care [Internet]. 2015;15(Oct-Dec):1–10. Available from: http://tinyurl.com/hs4uw4w 17. Valentijn PP, Schepman SM, Opheij W, Bruijnzeels MA. Understanding integrated care: a comprehensive conceptual framework based on the integrative functions of primary care. Int J Integr Care [Internet]. 2013;13(4):e010. Available from: http://tinyurl.com/h2ftvvm 18. Rachev BT. The economics of health service transformation: a business model for care coordination for chronic condition patients in the UK and US. Clin Gov [Internet]. 2015;20(3):113–22. Available from: http://tinyurl.com/gwwtmsk 19. Minkman MMN. The current state of integrated care: an overview. J Integr Care [Internet]. Emerald Group Publishing Limited; 2012;20(6):346–58. Available from: http://tinyurl.com/jj6v7dk 54 20. Navickas R, Visockienė Ž, Puronaitė R, Rukšėnienė M, Kasiulevičius V, Jurevičienė E. Prevalence and structure of multiple chronic conditions in Lithuanian population and the distribution of the associated healthcare resources. Eur J Intern Med [Internet]. 2015;26(3):160–8. Available from: http://tinyurl.com/zjdy8g8 21. Shaw S, Rosen R, Rumbold B. What is integrated care? [Internet]. Nuffield Trust. London; 2011. Available from: http://tinyurl.com/zvlz8vk 22. Lillrank P. Integration and coordination in healthcare: an operations management view. J Integr Care [Internet]. Emerald Group Publishing Limited; 2012;20:6–12. Available from: http://tinyurl.com/zzwvyrk 23. World Health Organization. People-centred and integrated health services: an overview of evidence [Internet]. 2015. Available from: http://tinyurl.com/gtvjq6k 24. Ham C, Curry N. Integrated care. What is it? Does it work? What does it mean for the NHS? 2011;1– 8. Available from: http://tinyurl.com/j7g982m 25. Leutz WN. Five laws for integrating medical and social services: lessons from the United States and the United Kingdom. Milbank Q [Internet]. 1999;77:77–110. Available from: http://tinyurl.com/zwftwyk 26. OECD. Health reform: meeting the challenge of ageing and multiple morbidities [Internet]. OECD Publishing. 2011. Available from: http://tinyurl.com/jkkdtpl 27. Santana R, Costa C. A integração vertical de cuidados de saúde: aspectos conceptuais e organizacionais [Vertical integration of healthcare: conceptual and organizational aspects]. Rev Port Saúde Pública [Internet]. 2008;7:29–56. Available from: http://tinyurl.com/jblt37p 28. Santana R. O financiamento por capitação ajustada pelo risco em contexto de integração vertical de cuidados de saúde: a utilização dos consumos com medicamentos como proxy da carga de doença em ambulatório [Risk adjusted capitation funding in the context of vertical integration of healthcare: medical consumption as a proxy for burden of disease in outpatient setting] [PhD thesis] [Internet]. Universidade NOVA de Lisboa; 2011. Available from: https://run.unl.pt/handle/10362/6049 29. Verhaegh KJ, MacNeil-Vroomen JL, Eslami S, Geerlings SE, de Rooij SE, Buurman BM. Transitional care interventions prevent hospital readmissions for adults with chronic illnesses. Health Aff [Internet]. 2014;33(9):1531–9. Available from: http://tinyurl.com/zx7p47m 30. Valentijn PP, Boesveld IC, van der Klauw DM, Ruwaard D, Struijs JN, Molema JJW, et al. Towards a taxonomy for integrated care: a mixed-methods study. Int J Integr Care [Internet]. 2015;15(JanMar):e003. Available from: http://tinyurl.com/jejzz3o 31. van der Vlegel-Brouwer W. Integrated healthcare for chronically ill. Reflections on the gap between science and practice and how to bridge the gap. Int J Integr Care [Internet]. 2013;13(AprJun):e019. Available from: http://tinyurl.com/jejtg6c 32. Armitage GD, Suter E, Oelke ND, Adair CE. Health systems integration: state of the evidence. Int J Integr Care [Internet]. 2009;9(Apr-Jun):e82. Available from: http://tinyurl.com/hnqoxoy 33. Goodwin N, Dixon A, Anderson G, Wodchis W. Providing integrated care for older people with complex needs Lessons from seven international case studies [Internet]. 2014. Available from: http://tinyurl.com/oygprwf 34. Sun X, Tang W, Ye T, Zhang Y, Wen B, Zhang L. Integrated care: a comprehensive bibliometric analysis and literature review. Int J Integr Care [Internet]. 2014;14(Apr-Jun):e017. Available from: http://tinyurl.com/jxy3sbl 35. Martínez-González NA, Berchtold P, Ullman K, Busato A, Egger M. Integrated care programmes for adults with chronic conditions: a meta-review. Int J Qual Heal Care [Internet]. 2014;26(5):561– 70. Available from: http://tinyurl.com/j8fqfxx 36. Kansagara D, Chiovaro JC, Kagen D, Jencks S, Rhyne K, O’Neil M, et al. Transitions of care from hospital to home: an overview of systematic reviews and recommendations for improving transitional care in the Veterans Health Administration [Internet]. 2015. Available from: http://tinyurl.com/h52xjlj 37. Santana S, Szczygiel N, Redondo P. Integration of care systems in Portugal: anatomy of recent reforms. Int J Integr Care [Internet]. 2014;14(Jul-Sep):e014. Available from: http://tinyurl.com/hw9p28m 38. Decreto-Lei n.o 207/99. Diário da República. 1a série - N.o 133 - 9 de junho de 1999. 39. Decreto-Lei n.o 50-B/2007. Diário da República. 1a série - N.o 42 - 28 de fevereiro de 2007. 40. Decreto-Lei n.o 183/2008. Diário da República. 1a série - N.o 171 - 4 de setembro de 2008. 41. Decreto-Lei n.o 318/2009. Diário da República. 1a série - N.o 212 - 2 de novembro de 2009. 42. Decreto-Lei n.o 67/2011. Diário da República. 1a série - N.o 107 - 2 de Junho de 2011. 43. Decreto-Lei n.o 238/2012. Diário da República. 1.a série — N.o 211 — 31 de outubro de 2012. 55 44. Portugal. Ministério da Saúde. Entidade Reguladora da Saúde. Estudo sobre o desempenho das Unidades Locais de Saúde [Study on the performance of Local Health Units] [Internet]. Porto; 2015. Available from: http://tinyurl.com/hureou5 45. OECD. OECD reviews of health care quality: Portugal 2015. Raising standards. [Internet]. OECD Publishing; 2015. (OECD Reviews of Health Care Quality). Available from: http://tinyurl.com/gpbv99e 46. Santana R, Marques AP, Lopes S, Barreto X. Projecto de definição de um modelo de acompanhamento da actividade desenvolvida pelas Unidades Locais de Saúde. Relatório III-Estudo sobre o grau de integração de organizações de saúde (EGIOS) [Project for the definition of an activity monitoring model developed for the Local Health Units. Report IIIStudy about the degree of integration of healthcare organizations] [Internet]. Lisboa; 2010. Available from: http://tinyurl.com/jmpsdeq 47. Perrin EC, Newacheck P, Pless IB, Drotar D, Gortmaker SL, Leventhal J, et al. Issues involved in the definition and classification of chronic health conditions. Pediatrics [Internet]. 1993 Apr;91(4):787–93. Available from: http://tinyurl.com/zmg8k7o 48. World Health Organization. Innovative care for chronic conditions: building blocks for action. Global report [Internet]. Geneva: World Health Organization; 2002. 1-99 p. Available from: http://tinyurl.com/jm4l755 49. Anderson G. Chronic Care: making the case for ongoing care [Internet]. 2010. Available from: http://tinyurl.com/hlcxu6t 50. Halpin HA, Morales-Suárez-Varela MM, Martin-Moreno JM. Chronic disease prevention and the new public health. Public Health Rev [Internet]. 2010;32(1):120–54. Available from: http://tinyurl.com/jqov564 51. Contel JC, Prevention C, Ledesma A, Social I, Blay C, Prevention C, et al. Chronic and integrated care in Catalonia. Int J Integr Care [Internet]. 2015;15(Apr-Jun):e025. Available from: http://tinyurl.com/j5tg9w5 52. Cassel CK, Johnston-Fleece M, Reddy S. Aging: Adding complexity, requiring skills. Stud Health Technol Inform [Internet]. 2010;153:47–69. Available from: http://tinyurl.com/hvdum5t 53. Jackson CT, Trygstad TK, DeWalt DA, DuBard CA. Transitional care cut hospital readmissions for North Carolina medicaid patients with complex chronic conditions. Health Aff [Internet]. 2013;32(8):1407–15. Available from: http://tinyurl.com/jcp26tm 54. Goodman R a, Posner SF, Huang ES, Parekh AK, Koh HK. Defining and measuring chronic conditions: imperatives for research, policy, program, and practice. Prev Chronic Dis [Internet]. 2013;10. Available from: http://tinyurl.com/z6ljlvv 55. Mitchell GK, Burridge L, Zhang J, Donald M, Scott IA, Dart J, et al. Systematic review of integrated models of health care delivered at the primary – secondary interface: how effective is it and what determines effectiveness ? Aust J Prim Health [Internet]. 2015;21(4):391–408. Available from: http://tinyurl.com/hsvhf25 56. Cheah J. Integrate now, create health. Int J Integr Care [Internet]. 2011;11(Apr-Jun):e022. Available from: http://tinyurl.com/h6q2qrb 57. Nolte E, Pitchforth E. What we know: a brief review of the evidence of approaches to chronic care. In: Nolte E, Knai C, Saltman RB, editors. Assessing chronic disease management in European health systems: concepts and approaches [Internet]. WHO Regional Office for Europe and European Observatory on Health Systems and Policies; 2014. p. 9–22. Available from: http://tinyurl.com/hukxzhd 58. World Health Organization. Global status report on noncommunicable diseases 2014 [Internet]. Geneva: World Health Organization; 2014. 176 p. Available from: http://tinyurl.com/laj7nou 59. Gruneir A, Bronskill SE, Maxwell CJ, Bai YQ, Kone AJ, Thavorn K, et al. The association between multimorbidity and hospitalization is modified by individual demographics and physician continuity of care: a retrospective cohort study. BMC Health Serv Res [Internet]. 2016;16:154. Available from: http://tinyurl.com/grg3fjj 60. American Hospital Association. Rethinking the hospital readmissions reduction program. Trendwatch [Internet]. 2015;(March). Available from: http://tinyurl.com/nv3qzf2 61. Kristensen SR, Bech M, Quentin W. A roadmap for comparing readmission policies with application to Denmark, England, Germany and the United States. Health Policy (New York) [Internet]. Elsevier Ireland Ltd; 2015;119(3):264–73. Available from: http://tinyurl.com/zw6u8j2 62. Jeffs L, Dhalla I, Cardoso R, Bell CM. The perspectives of patients, family members and healthcare professionals on readmissions: preventable or inevitable? J Interprof Care [Internet]. 2014;28(6):507–12. Available from: http://tinyurl.com/hp5npjy 62 63 Appendix 1 Risk-standardized readmission ratio We estimated the hospital-specific SRR using generalized linear mixed models at the specialty cohort level, modelling data at the individual and hospital levels to account for outcome variance within and between hospitals. For a given specialty cohort, we fitted a hierarchical logistic regression model to account for the natural clustering of observations within hospitals. Let # $% denote the outcome (1: individual i is readmitted within 30 days) for an individual in one of five mutually exclusive cohort &'()*)+ at hospital j. Let ,$- . ,/),0)*)," denote a set of k risk factors. Let M denote the total number of hospital and 1the number of index individual stays in hospital j. We used the linear relationship between outcome and covariates through a logit function with dispersion 23456 7839 # $.(.:;<-;=>,$- ;?$-, with <-@AB)C0. In the model equation, : is the adjusted average outcome over all hospitals; <- denotes for the hospital-specific intercept, also called the empirical Bayes estimator and; C0 is the between hospital variance component. For each year and each specialty cohort a model was ran, in a total of 65 models. Thus, individual risk of occurring a readmission for individual i in hospital j at year t is computed by DD$-% .23456E/:;=>,$- . With the result of each generalized linear mixed model, we computed the predicted and expected number of readmissions at each hospital. The predicted number of readmission is computed as the sum of the predicted probability of readmission for each individual, including the hospital-specific intercept. The expected number of readmission is computed in a similar way, but ignoring the hospital-specific intercept. Using similar notation to the previous model, for the predicted number of admissions 78FGHI for index admissions in cohort &J()*)+ at hospital j we use 78FGHI.23456E/:;<-;=>,$- where the sum is over all index admissions in cohort C at hospital j. For the expected number of index admissions KLMHI we use KLMHI.23456E/:;=>,$- . Thus, the risk-standardized readmission ratio in cohort C at hospital j NDDHI is computed as 78FGHIKLMHI. The risk-standardized readmission ratio at the hospital level is a combination of all specialties cohort NDDHI 64 pooled for each hospital using a volume-weighted logarithmic mean to create a hospitalwide SRR composite: NDD -.FLM 1HI2O NDDHI1HI. The use of a geometric mean is a more appropriate indicator to summarize benchmark results.[135] Thus, given a hospital case mix and service mix, we are able to compare with an average hospital’s performance with the same case mix and service mix. A lower ratio (<1) is a sign of lower-than-expected readmission rates or, in other words, better quality; on the other hand, a ratio higher than 1 denotes a higher-than-expected readmission rate.