Hospital readmission rates are a standard indicator of hospital performance. There are concerns that current risk-adjustment methods need to account for the numerous external factors that can influence readmission rates. If these external factors are not considered, hospitals may be unfairly penalized when they discharge patients to communities that are less capable of supporting care transitions and disease management. While incorporating adjustments for the myriad of social and economic factors outside of the hospital setting could improve the accuracy of readmission rates as a performance measure, the number of potential variables and the scarcity of data to measure them limit the feasibility of doing so. This paper evaluates a practical solution to this problem: using mixed-effect regression models to estimate case-mix adjusted risk of readmission by a community of patients’ residences (community risk of readmission) as a performance indicator in addition to hospital readmission rates.
We examine whether case-mix adjusted community risk of readmission can be used as a quality indicator for community-based care using hospital discharge data and mixed-effect regression models with a random intercept for the community. An unplanned repeat hospitalization was our outcome of interest. Our primary exposure came from our neighborhood.
The case-mix adjusted risk of unplanned repeat hospitalization is associated with the community of residence. The risk of readmission in a community can be estimated and mapped as an indicator of the community’s ability to support care transitions and long-term disease management.
Contextualizing readmission rates through a community lens can assist hospitals and policymakers in improving discharge planning, reducing hospital penalties, and, most importantly, providing higher-quality care to the people they serve.
Health services research, pay for performance, performance measures, quality improvement, and health policy are some of the terms used.
Risk-adjusted hospital readmission rates are a standard indicator of hospital performance and quality of care worldwide.
1–4 Many governments (for example, Denmark and Canada) routinely monitor hospital readmissions, which the Canadian Institute of Health Information (CIHI) describes as common, costly, and often avoidable. 5–7 Other countries, such as Germany, England, and the United States, have taken more concrete steps to reduce hospital readmission rates by instituting readmission reduction programs that include financial penalties for hospitals that fail to meet predetermined thresholds. 8
There are concerns that the risk adjustments used in the models that inform these readmission reduction programs do not account for factors that can affect readmission rates outside the hospital setting.
9 10 Researchers discovered that poverty,11 12 social isolation,13 transportation access14, and access to primary care15 are associated with readmission rates, even after adjusting for diagnostic, procedural, and demographic factors. 16–18 Additionally, there is evidence that investing in community-based services and resources can reduce hospital readmissions among local populations. 19–22 A variety of interacting elements outside the hospital setting can influence readmissions,23 and putting the entire burden of reducing readmissions on hospitals alone may unfairly penalize hospitals discharging patients to communities with less capacity to address these elements.
While adjusting for the myriad of social and economic factors outside of the hospital setting could improve the accuracy of readmission rates as a measure of hospital performance, the feasibility of doing so is limited. Data on these factors are frequently not routinely collected, the number of potential variables is significant, and managing and measuring many important and interacting variables would be difficult. While expanding risk-adjustment models to include factors outside of the hospital setting is a worthy goal, significant gaps will remain, raising concerns about fairness and bias.
Rather than attempting to capture and adjust for every potential variable and interaction between factors outside the hospital system that could affect hospital readmission rates, a more practical strategy is developing risk-adjusted readmission performance indicators for communities. The risk of readmission associated with being discharged to a specific community (community risk of readmission) could also be estimated using similar modeling approaches as those used to calculate hospital readmission rates. Estimates of community risk of readmission would consider many downstream factors influencing readmission risk and the complex interactions between them. These estimates could then be combined with traditional hospital performance measures to identify underserved communities. While this would not solve the problem of risk under adjustment for hospital performance indicators, it would provide another lens through which readmission rates at a hospital could be contextualized.
Measuring readmission rates for communities as a performance measure has only been attempted in relatively large populations. For example, the National Health Service (NHS) reports indirectly standardized emergency readmission rates for geographically defined Clinical Commissioning Groups (CCGs) in the United Kingdom. 24 CCGs serve populations ranging from 96 564 to 1 860 111 people, many of which include multiple cities. 25 As a result, reported rates may obscure significant geographical risk heterogeneity. Furthermore, the methods used to generate the NHS estimates could be more suitable for estimating rates in areas with fewer people. The indirect standardization limits the number of factors incorporated into case-mix adjustment and, mainly when applied to smaller populations, needs to distinguish between natural and chance variation. 26 27 The use of mixed-effect regression models for case-mix adjustment, combined with empirical Bayes estimation of adjusted rates, addresses these limitations and has thus become the industry standard for small area estimation and hospital performance measurement. 27–29
ORDER WITH US AND GET FULL ASSIGNMENT HELP FOR THIS QUESTION AND ANY OTHER ASSIGNMENTS (PLAGIARISM FREE)
Furthermore, the NHS reporting emphasizes readmission within 30 days. While the initial transition from hospital to community settings is critical to high-quality care,30-32 is only the first step in long-term management. Once patients are discharged from the hospital, it is up to community-based health and wellness services, informal supports, and patient self-management to facilitate care and reduce readmission rates. As a performance indicator, community risk of readmission can extend beyond a 30-day window following discharge, shifting the emphasis away from the transition of care efforts and toward broader disease management. Understanding the risk of readmissions associated with communities and hospitals allows for developing and implementing integrated hospital and community quality improvement plans where they are most needed.
This study had two goals as a first step toward developing a measure of community risk of readmission. The first was to determine the magnitude of community variation in case-mix adjusted readmission rates for adults aged 30 and older in Nova Scotia, Canada, between 2010 and 2014. The second goal was to see if differences in estimated risk-adjusted readmission rates for specific communities were significant enough and precise enough to be helpful as a quality indicator for community-based care after hospital discharge.
Design of the study, data, and setting
This was a population-based, descriptive study of community variation in the case-mix adjusted risk of unplanned readmission to the hospital after discharge to a community setting (rather than a long-term care or assisted living facility, other nonacute care institutions, or a psychiatric hospital). To account for exposure time and right censoring, the risk of unplanned hospital readmissions was modeled using time-to-event regression models with risk-adjustment variables as fixed effects and a random intercept for a community of residence as a random intercept. Following that, empirical Bayes estimation and mapping of case-mix adjusted risk of readmission for specific communities were performed. The study population included all people aged 30 and up in Nova Scotia, Canada, who were discharged from a hospital between 2010 and 2014. In the analysis, we only considered the first eligible discharge for each person. From 2010 to 2014, the average population of Nova Scotia was around 937 000 people, and it is served by a regionally organized network of 43 local, regional, and tertiary hospitals of varying sizes and capacities.
Our data came from provincial health registry eligibility files linked to the CIHI Discharge Abstract Database, which we accessed through Health Data Nova Scotia (DAD). The DAD data set includes all discharges from all acute care facilities that care for the study population. DAD data are coded by certified health records technicians using hospital charts, and electronic data entry systems per CIHI-established validated standardized protocols. 33
Our primary exposure was a community of residence, and our outcome of interest was hospital readmission, operationalized as an unplanned repeat hospitalization (URH) after an index discharge from the hospital. The number of days someone spent in a community between index discharge and URH or censoring was used to calculate an exposure time. URH was defined as any hospitalization after a valid index hospitalization (see online supplemental appendix 1) coded as ‘urgent’ or ’emergent’ in the data’s ‘Admit Type’ field. This does not include planned hospitalizations, such as hip or knee replacements due to joint deterioration. Index cases were followed up until a URH, death, or the end of the study period.
The forward sortation area (FSA) was defined as the first three letters of Canadian postal codes. FSAs in Nova Scotia vary significantly in size and population. Generally, they delineate areas that are small enough to avoid excessive aggregation while ensuring a large enough sample size for community-level estimation. FSAs also offer a reasonable approximation to meaningful communities and neighborhoods without relying on geocoding postal codes to other geographical entities, such as census geography, which has been shown to result in significant allocation errors. 34
Models included fixed effects for age-sex groups, Elixhauser comorbidity index score, percentage of time spent in the community during the last year of life, indicator variables for 25 different individual health conditions (see online supplemental appendix 2), and the total number of health conditions to account for community variation in patient case-mix. Persons in their final year of life are at an increased risk of unplanned hospitalizations,35, so our model took this into account. To accomplish this, we calculated the percentage of each person’s exposure interval that fell within 365 days of their death date and included that figure in the regression model. Age was measured in 5-year increments, with a single group for people aged 30-54 years and another for those aged 85 and up. The Elixhauser comorbidity index score calculated multimorbidity and a simple count of health conditions. The Elixhauser score is a good predictor of rehospitalization and mortality in people with chronic conditions36 37. In contrast, a simple disease count is one of the best comorbidity measures for predicting hospitalization. 38
The expected time to a URH for a given community was defined as the community risk of readmission. We used a mixed-effect accelerated failure time (AFT) regression model with case-mix adjustment variables as fixed effects and a random intercept for a community to estimate community variation in the case-mix adjusted time to URH. We chose an AFT model over the more commonly used Cox-proportional hazards model for two reasons. First, the AFT model expresses community effects as the relative expected time to a URH, which policymakers and the general public find more intuitive than relative risks. Second, the AFT model estimates failure distribution effects rather than hazard rate effects. It is more robust to model assumptions such as the proportionality of hazards than the Cox hazards model. 39 The estimated effects of covariates are referred to as ‘acceleration factors’ in the AFT model. 40 An acceleration factor more significant than one indicates that the community’s expected time to a URH is longer than the provincial average (lower community risk of readmission). In contrast, an acceleration factor of less than one indicates that the community’s expected time to a URH is shorter than the provincial average (higher community risk of readmission).
Case-mix adjusted intercepts for each community and 95% confidence intervals (CIs) were estimated to identify specific communities with a higher or lower time to URH. Estimated best linear unbiased predictors were used, as is common practice in small area estimation empirical Bayes estimators. 41 We exponentiated the estimated community intercepts and confidence bounds to obtain community-level estimates of each community’s acceleration factors relative to the average community (i.e., a community with a random intercept of zero). Then, community estimates were mapped using ArcGIS software to reveal spatial patterns.
Following the completion of the initial modeling, additional analyses were performed to assess the sensitivity of the results to which health conditions were included in the case-mix adjustment and the length of follow-up. The first sensitivity analysis excluded index hospitalizations with one of the four conditions in the primary analytical model that had the most significant influence on time to a URH: cancer, stroke, injury, or diabetes. The second sensitivity analysis used only the last three years of our data to assess the impact of follow-up length.
Patient and public participation
The Maritime SPOR SUPPORT Unit (http://www.spor-maritime-srap.ca/) was formed through a collaboration of researchers, managers, patients, and policymakers. Two 1-hour teleconferences with hospital staff from across the province directly involved in discharging patients and supporting transitions to home inspired the study objectives. Three major topics were discussed: community factors that increase the likelihood of repeat hospitalization, community factors that influence discharge planning and coordination of care between hospital and community settings, and relevant community factors that differ from one location to the next. There was widespread agreement that discharging patients to some communities was much more complex than others due to differences in formal and informal supports, remoteness, and socioeconomic and demographic characteristics.
To ensure that the findings of this study were available to as many stakeholders as possible, we shared our results with various academic and policy audiences. Health authorities have incorporated the findings into community profiles to aid primary care planning.