Investigating the geographic disparity in quality of care: the case of hospital readmission after acute myocardial infarction in Italy

We analyse the hospital discharge data (Schede di Dimissione Ospedaliera, SDO) from the National Ministry of Health for the years 2010–2015. The data is routinely collected by all hospitals in all the regions and include not only administrative information such as diagnosis, treatment, discharge units, admission and discharge dates but also socio-demographic characteristics of the patients. Information about the hospitals in this dataset includes the type of ownership and the Local Health Authorities (LHAs) the institute belongs.

We focus on the elderly population because researches have found that patients over 65 years old are frail and at increased risk for readmission [50, 51], and that the Italian society is characterised by an ageing population suffering from a number of chronic conditions [37]. Moreover, patients were excluded from the analysis if any of the following criteria were met: In cases where patients incurred more than one admission during the first 30 days after discharge, we consider only the first readmission episode.

We select the patients diagnosed with a heart attack - Acute Myocardial Infarction (AMI) given the high volume of emergency admissions and that AMI patient unplanned readmission is commonly used as a healthcare quality indicator. We extract all patients whose main pathology is coded 410.0–410.9 under the 9th International Classification of Disease (ICD-9). Since these patients are often sent to the hospitals nearby, the potential selection bias is ameliorated when investigating the effects of geographic factors [52]. The treatments of AMI patients include Coronary artery bypass graft (CABG) or coronary bypass surgery, cardiac catheters, percutaneous transluminal coronary angioplasty (PTCA) and stent. CABG involves taking a vein or an artery from the patient’s body and using it to reroute blood from coronary arteries. A catheter is a thin, flexible tube that is inserted in a vein. PTCA is a minimally invasive procedure that uses an inflated balloon in a vessel to expand the blood vessel to improve blood flow, while the stent is a spring-shaped prosthesis used to complement PTCA. We extract the procedural codes from our dataset and control for the different interventions performed.

We also include organizational factors of the hospitals, such as the type of institution, capacity, and generic quality in the analysis. From the SDO data, we retain the hospital ownership type variable, which includes public Hospital Units (HUs or Ospedalia a Gestione Diretta), Hospital Trust (Aziende Ospedaliere), Teaching Hospitals (Clinici o Policlinici Universitari), Research Hospitals (Istituto di Ricovero e Cura a Carattere Scientifico, IRCCS), Accredited Private Hospitals (Case di Cura Accreditate) and other private providers. We calculated the volume of AMI patients per year by the provider from the SDO data. The information on the total bed counts of hospitals across the years is obtained from the Italian Ministry of Health (Ministero della Salute) website. The rationale for including the capacity information is to proxy the potential size constraints that hospitals face, which can be related to the readmission outcome. We also use the cut-off points of low, medium and high Acute Myocardial Infarction (AMI) mortality rate defined by the National Outcome Programs (Programma Nazionale Esiti) website for the broad quality categorization for the hospitals.

The study’s primary outcome measure is the risk of readmission within 30 days after discharge for elderly patients diagnosed with Acute Myocardial Infarction (AMI) during the index hospitalisation. The primary outcome measure included readmission with all causes such as infections or complications, not just those that appear related to the initial admission. This measure is in line with the established literature and the readmission measure from the US Centers for Medicare and Medicaid Services and the QualityNet reporting guideline. In addition, because comorbid elderly patients may be more likely to be readmitted to the hospitals due to different pathologies, we also consider a more restricted definition of readmission that includes only readmissions with the same Major Diagnostic Category (MDC). For the analysis on the patient level, we consider these two types of readmission as binary variables to identify the effects of other explanatory variables. In estimating the geographic variations, we treat the readmission rates of each hospital as the outcome variable. The specifications are described in the following section.

Although unplanned readmission is a widely used quality indicator [21, 53] and represents substantial social and economic burdens, we are aware of some of the limitations of this indicator. First, adjustment for patient case-mix and contextual factors need to be carried out correctly to infer risk. We used the Ontario AMI prediction rules, a disease-specific instrument, to adjust for the risk scores of the patients. Second, studies show that not all readmissions within 30 days are avoidable [54], which can potentially make the indicator inaccurate. In recognising the potential weakness of the readmission indicator, we believe that the intricate nature of hospital readmission nonetheless offers important insights on the behaviours of the providers.

We estimate both the multilevel logistics model for the probability of readmission, and multilevel proportional hazard model for time-to-readmission. As the healthcare path of the patients may depend on the structures of the providers and LHAs, we allow observation within the same hospital and LHA to be correlated to each other. As such, we are accounting for the within-cluster homogeneity.

For the multilevel logistics model, we estimate the following:where \(Y_{ijk}\) is the binary variable of patient i in hospital j in LHA k, and \(Y_{ijk}=1\) if the patient is being readmitted to the hospital within 30 days of discharge. Here, each LHA cluster \(k=1\ldots n\) consists of hospital clusters \(j=1\ldots n_{i},\) and each hospital has \(i=1\ldots n_{ij}\) patient observations. \({\bf{X}}_{\textit{ijk}}\) is a row vector containing the patient-level variables including demographics, comorbidities and LOS, and \({\bf{Z}}_{\textit{jk}}\) represents a vector of hospital-level factors such as capacity and patient volume. We allow for a non-linear relationship between age and our outcome variable by including a quadratic term. As discussed in the institutional background section, providers can have different discharge incentive structures due to their payment system. We thus interact the variable \(\text {LOS}_{ijk}\) with the categorical variable of hospital types, \(\text {Type}_{jk}\), to allow for the potential heterogeneous effects. We also include a set of year and regional fixed-effects (\(\mu _t\) and \(\mu _{R}\)), as well as a regional average income variable \(\text {Inc}_{l}\) to account for the economic disparity across regions. \(\beta _x\), \(\beta _z\) and \(\beta _i\) are the fixed effects for the explanatory variables. Finally, \(e_{0k}\;\sim \;N(0,\;\theta _e^{2})\), \(\eta _{0jk}\;\sim \;N(0,\;\theta _\eta ^{2})\) and \(\;v_{0ijk}\;\sim \;N(0,\;\theta _v^{2})\) are the random error terms at the LHA, hospital and patient levels, reflecting the cluster-specific random effects. We estimate the marginal effects of the explanatory variables through maximum likelihood. However, we do not report the intra-class coefficient (ICC) in quantifying the contribution of area-level variance to total variance because the computation and interpretation of ICC are often questionable in the context of logistic regression [58, 59].

Similarly, for the multilevel survival analysis, the underlying equation is:where \(t_{ijk}\) is the observable failure (readmitted) time of the patient i nested in hospital j in LHA k and \(h(t_{ijk})\) is the hazard function of the corresponding patient. \(h_0(t)\) is the baseline hazard function. \(\beta _x\), \(\beta _z\) and \(\beta _i\) are the conditional hazard ratios, while the remaining variables are the same from Eq. (1). This more flexible model is semiparametric and thus does not have a functional form assumption imposed on the baseline hazard. We estimate the model for the time from discharge to readmission and obtain the influence of the covariates at different levels.

While it is important to identify the marginal effects of patient and hospital characteristics, we want to understand what drives the geographic variation in readmission. Since we are primarily interested in the unjustified variation generated from the providers, we aggregate the dataset to the hospital level while retaining patient variables as averages. The model consists of three geographic units—hospital, LHA and regions. As our outcome variable (hospital readmission rate) is no longer binary, we consider the multilevel mixed-effect linear model:where \(Y_{jkl}\) represents the rate of readmission in hospital j of LHA k in region l, and \(\overline{\text {LOS}}_{jkl}\) is the average LOS of the patients hospitalized in hospital j. \({\bf{Z}}_{\mathit{jkl}}\) represents the hospital ownership types, and the vector \({\overline{\bf{X}}}_{\mathit{jkl}}\) is the averaged patient-level information. \(\mu _{0l}\) is the random intercept at the regional level, \(u_{0kl}\) is the random intercept at LHA level, nested within region level, and \(\varepsilon _{0jkl}\) captures the idiosyncratic hospital factors. We assume that \(\mu _{0l}\;\sim \;N(0,\;\theta _\mu ^{2})\), \(u_{jk}\;\sim \;N(0,\;\theta _u^{2})\) and \(\varepsilon _{ojkl}\;\sim \;N(0,\;\theta _\varepsilon ^{2})\) and fit the model using restricted maximum likelihood for unbiased estimation of variances. We obtain the intra-class coefficients (ICCs) to assess the total residual variance attributable to both LHA and regional levels.

The total residual variance attributable to the LHA level is:And the total residual variance attributable to the regional level is:Larger values of ICC indicate that a considerable proportion of the residual variance in readmission rate is attributable to these levels. Visually, we compare the plots that rank the LHA and regional residuals for both the empty and the full models to assess the variation explained by the observed variables qualitatively.

We want to understand how much readmission variance is explained by differential hospital behaviours, here proxied by LOS and the different surgical procedures. This can be achieved by comparing the increase in explained variance after including the predictors. In the multilevel analysis, the presence of multiple variance components challenges the reporting of \(R^{2}\) [60, 61] and we, therefore, calculate the proportional reduction in total variance after incorporating these predictors [62] using the following formula:We can argue that while, for instance, ownership-driven variation reflects organizational or structural disparities that are beyond the control of hospitals, inequalities that are driven by differential risk-adjusted LOS and the use of surgical procedures are arguably mitigable. If we observe a substantial increase in \({R^{2}}\) after the inclusion of LOS and innovative procedure, this implies the importance of discharge behaviour in driving the regional differences.

Before looking into the marginal effects of patient and hospital factors on readmission risks, we first present a descriptive graph of the readmission Nelson–Aalen cumulative hazard estimates as a function of days after discharge across the selected large regions, as seen in Fig. 2. We observe that the baseline readmission risk for patients who are admitted to hospitals in the Southern regions of Apulia and Sicily are significantly higher throughout the days after discharge than those admitted in Lombardy and Lazio.

Table 2 reports the coefficients of the patient and hospital-level variables from Eqs. (1) and (2). We analyse 30 days readmission for all causes as the primary dependent variable, as well as readmission with the same MDC as a secondary indicator. For all-cause readmission, we observe from the coefficients of the interaction term for LOS that, for both the multilevel Logit and hazard models, the probability to be readmitted decreases with LOS for patients admitted to all types of hospitals. The magnitude of this negative effect is higher for patients admitted to Hospital Units and Private Clinics than that of other hospitals. Moreover, the coefficients for hospital types show that independent public hospitals have significantly lower readmission probabilities than the Hospital Units and Private Clinics. This finding is particularly interesting as it partially relates to hospital incentive structures. For hospitals under a global budget as in the case of Hospital Units, there is little incentive to save costs, and thus the index hospitalisation LOS is more effective in preventing future unplanned readmission. Whereas for hospitals under a PPS with some budget allocations such as Hospital Trust, Teaching and Research Hospitals, the LOS is relatively less effective in reducing the probability of all-cause readmission given their incentive to improve efficiency. However, other than the effects of LOS, these independent public hospitals have significantly lower readmission risks than Hospital Units, indicating that other mechanisms other than payment systems are also driving the differences in readmission. Finally, for the profit-making private hospitals that operate solely under PPS, the effect of LOS in reducing the probability of readmission is the strongest, but they have the highest overall hospital readmission. For the more restricted outcome indicator, readmission with the same MDC, we observe a similar effect for LOS in terms of the directions of the coefficients. However, the coefficients are only significant for hospitalisations in Hospital Units and Private Clinics but not for the independent public hospitals.

For the demographic factors, the probability of readmission increases with age, but the effect diminishes with age. Males are more likely to be readmitted than females, and foreigners are less likely to be readmitted. Patients who underwent PTCA and Stent, CABG and Catheter all have less risk of all-cause readmission than patients with no operation performed. However, for readmission with the same MDC, the CABG procedure does not reduce the probability of readmission. The fact that patients were previously discharged to home hospitalisation, rehabilitation institution or other types of integrated home care does not affect the probability to be readmitted.

At the hospital level, we have discussed that the independent public hospitals such as Hospital Trust and Teaching Hospitals have significantly lower risks of all-cause readmission than LHA-managed Hospital Units. However, the same effect is not observed for readmission with the same MDC. The volume of AMI patients reduces the probability of both types of readmission, indicating some degrees of learning effect. Furthermore, hospitals with higher capacity have lower probabilities of readmission. This effect is expected, as bed constraints may contribute to early patient discharges and in turn, result in unplanned readmission. Finally, patients admitted to hospitals with low and medium in-hospital AMI mortality (according to the National Outcome Programs) have lower likelihoods of all-cause readmission. The coefficients for comorbidities, education, years and regional fixed effects can be found in Appendix Tables 5 and 6. Overall, the probability to be readmitted has decreased with comorbid patients with Shock, cerebrovascular disease and Cardiac Dysrhythmias are less likely to be readmitted, while diabetic patients are more likely to be readmitted. This correlation for the above conditions can be explained by more considerable attention offered by the providers for patients with these severe cardiovascular comorbidities. However, the medical interpretation of the conditions is beyond the scope of this paper. In Appendix Tables 5 and 6, we also observed that all-cause readmission decreases over the years, but same -MDC readmission increases. Many of the Central and Southern regions have positive and significant coefficients, indicating higher general readmission risks.

Hospital-level variation in the readmission outcome is estimated in terms of variance and intra-class correlation. We first present the coefficient estimates for the variables collapsed at the hospital level in Table 3. Although most coefficients have the same signs as in Table 2, some of them cease to be significant. Notably, for all-cause readmission rates, the coefficients for average LOS are significant for all types of hospitals except for Research hospitals, while for the same MDC readmission rate the coefficient is only significant for Private Clinics. Moreover, the percentage of patients who underwent PTCA and stent procedures have significantly negative coefficients for both types of readmission rate, while the percentage of CABG procedure only reduces all-cause readmission rate. These variables represent the underlying hospital behaviours and are robust to the aggregation (Table 3).

We graphically represent the residuals from the empty and the full models at both the LHA and the regional levels. As seen in Figs. 3 and 4, we order the residuals by the unadjusted and adjusted LHA and regional averages of readmission rate and plot the 95% confidence intervals around each residual estimate. The adjusted residuals represent the unexplained variation after accounting for the differences across patient and hospital factors. We observe that, without accounting for the explanatory variables, some LHAs and regions exhibit significantly different levels of variation for both types of readmission rates. The variation is more pronounced for all-cause readmission, as we observe LHA and regions both significantly below or above the average readmission rates. In particular, we see in Fig. 4 Marche, Piedmont, Veneto and Lombardy have significantly lower-than-average regional all-cause readmission rate, while Emilia Romagna and Sicily has a significantly high readmission rate. After adjusting for patient and hospital characteristics, these variations diminished considerably.

Since we are interested in how hospital behaviours, here represented by LOS and different surgical procedures, explain the variation, we investigate the variance components in two separate models for both types of readmission rates. We further compute the intra-class correlation (ICC) and the explained variance as represented by Eqs. (4)–(6). The intra-class correlations (ICC) estimate the proportion of overall variation in outcomes explained by the variation between geographic units. As seen in Table 4, for the model excluding LOS and surgical procedures for all-cause readmission, we observe around 11.69% of the total variation is attributed to higher geographic levels, with about 7.89% at the LHA level and 3.8% at the regional level. After incorporating LOS and surgical procedures into the specification, the ICC decreased by 2.72% at the LHA-level and by around 0.4% at the regional level. At the same time the \(R^{2}\) almost doubled from 0.1172 to 0.2175. This result indicates how LOS and the use of surgical procedures played an indispensable role in driving the geographic variation in all-cause unplanned readmission. Although the scale of results for same-MDC readmission rate is much smaller, we do observe that these hospital factors explained a considerable proportion of the overall variance at the LHA and the regional levels.