INTRODUCTION

Since the enactment of the Affordable Care Act, policymakers have increasingly introduced readmission measures into several of Medicare’s value-based purchasing programs.1,2,3,4,5,6 This has led to a burgeoning literature exploring the most effective strategies for reducing readmissions, which can be classified into three major types. The first set of studies focuses on improving care transition interventions that occur in health systems. This research investigates the impact of coordinating care across inpatient and outpatient providers.7,8,9 The second set of studies examines hospitals and their ability to reduce readmission rates.10,11,12,13,14,15,16 The third explores individual practices and patient-centered medical homes.17,18,19,20

Even with such a large evidence base on readmission reduction strategies, there remains an important gap, especially at the primary care practice level.17 Primary care practices play an integral role in reducing readmissions as they often assume care for patients after discharge. For example, primary care visits shortly after admission for heart failure or high-risk surgery have been demonstrated to lower readmission rates.21,22 However, we have limited knowledge of how activities conducted by practices are associated with reduced readmission rates for their patients. The few practice-level studies we do have are often based on small-scale or single interventions conducted in one, or a handful, of organizations. While these are important contributions to the literature, their findings may not be externally valid as interventions are adapted to varied local contexts.17,23 To the best of our knowledge, there are no studies examining the association of practice activities and readmission rates in a large, nationally representative sample of primary care practices.

Despite the critical importance of primary care practices in reducing readmission rates, no empirical studies have examined the association of practice activities and readmissions rates in a large number of geographically and organizationally diverse practices. In this study, we fill that gap. We use new, nationally representative data on practices to create a composite readmission reduction measure based on the existing literature on how practice-level activities influence readmission rates. We pair this data with Medicare claims to examine the association between readmission activities and readmission rates. We hypothesized that implementing more readmission activities would be associated with lower practice-level readmission rates. We based our hypothesis on the limited evidence of the effectiveness of individual readmission reduction activities to reduce readmissions along with studies demonstrating greater success of multi-component interventions.

STUDY DATA AND METHODS

We linked new survey data on physician practices with data on practice-level readmission rates to examine whether practices’ readmission reduction activities were associated with their readmission rates. We derived practices’ readmission rates from the administrative claims of individual Medicare beneficiaries receiving primary care visits from the physician practices in our survey. While our study includes both patient and practice-level data, our unit of analysis throughout the paper is at the practice level.

Primary Care Practice Data

We relied on the National Survey of Healthcare Organizations and Systems (NSHOS) to obtain practice-level activities potentially associated with readmission as well as other practice-level covariates. The NSHOS was fielded from July 2017 to August 2018.24 The survey employed a stratified-cluster sampling design to survey health systems, hospitals, and physician practices on a wide range of organizational characteristics and capabilities. In this analysis, we use the physician practice survey, which was fielded to all practices with three or more adult primary care physicians (defined as internal medicine, family medicine, geriatrics, and general practice). The sample frame was derived from IQVIA’s OneKey database, which details ownership relationships and specialty information.25,26 Further information on the survey can be found in the peer-reviewed literature.27,28,29,30,31

The NSHOS initially identified 4,976 practices that met inclusion criteria. Of these, 2,333 (47%) completed the survey. After merging 2017–2018 NSHOS data with 2015–2016 Medicare data using national provider identifier numbers, another 260 practices were excluded because no Medicare beneficiaries were assigned to them. Finally, we restricted our sample to practices with at least 25 index admissions to create reliable estimates. The final sample included 1,788 practices and 415,663 admissions from all 50 states and Washington DC.

Patient Data

Patients were included in our sample if they were 66 and older, continuously enrolled in Medicare Parts A and B 12 months prior to and including the date of the index readmission, and enrolled in Part A for at least 30 days after the date of discharge. Using an approach adapted from the Medicare Shared Savings Program,32 we assigned each patient to the practice at which they received the plurality of their evaluation and management visits delivered by primary care clinicians. If a patient did not have any primary care visits delivered by a primary care clinician during the year, we attributed them to the practice where they received the plurality of their evaluation and management visits from specialist physicians. We include comparisons of practice and patient-level characteristics for survey respondents and non-respondents in our Appendix Tables A1 and A2.

Outcome

We evaluated readmissions (the outcome) using 100% inpatient Medicare claims data from calendar year 2016. We used 100% inpatient and outpatient Medicare claims from calendar years 2015–2016 for risk adjustment. We used 2016 data to assess the outcome as this was the most recent Medicare claims data we possessed at the time of the analysis. While these are not identical years to the NSHOS, we do not anticipate large numbers of practices’ readmission reduction activities to significantly change in a span of 1 to 2 years nor did we witness significant differences in the mean number of readmission activities between hospitals that responded in 2017 versus 2018. If anything, we believe this difference in years of data would lead the attenuation of any associations, rendering conservative estimates of the associations between measured readmission reduction activities and readmission outcomes.

We modeled our outcome using CMS hospital-wide readmission measure methodology.33 This measure uses a hierarchical logistic regression model in order to calculate a ratio for each hospital, which represents the number of readmissions one would expect given the hospital’s case-mix over how many readmissions one would predict for an average hospital with an identical case-mix. We adapted this measure to calculate practice-level instead of hospital-level readmission rates.

A readmission was defined as any unplanned admission occurring within 30 days of discharge from a previous index admission. We excluded index admissions for patients who died during hospitalization, were discharged against medical advice, or were admitted for psychiatric, cancer, or rehabilitation treatment. We also excluded index admissions for patients no longer enrolled in Medicare Part A at least 30 days after discharge since we would be unable to determine if they were readmitted. We combined stays into a single continuous hospitalization if a patient was discharged from one hospital and admitted the same or next day to another hospital. We distinguished between planned and unplanned readmissions by applying CMS planned readmission algorithm 4.0.33

Key Exposure

Our independent variable of interest was practices’ total number of readmission reduction activities. We created this composite measure by first identifying 26 candidate variables from the survey, based on capabilities demonstrated to reduce readmissions in the literature (full list in Appendix Table A3). These include such tasks as: communicating with patients within 72 hours of discharge,15 providing home visits after discharge,10 sending discharge summaries to primary care physicians,7 establishing a standardized process to reconcile medication,17,34 ensuring patients have access to their primary care provider,7 and empowering a care manager to help patients adhere to care plans.9,12,35 While each of these individual capabilities is an important component of successful interventions, interventions with more readmission reduction activities demonstrate the most significant declines in readmission rates.10,17,34,36,37

We dichotomized Likert scale questions by recoding “never,” “sometimes,” and “often” as zero, and “always” as one to create our composite measure. We then reviewed the interitem correlations of each candidate variable, assuming tetrachoric correlations, to arrive at a list of mutually exclusive variables. We narrowed our final list to 12 variables (Table 1). Nearly all variables had interitem correlations under 0.3 and only one was above 0.5 (0.53), indicating the measures were capturing separate care components.

Table 1 Questions Included in Composite Measure of Readmission Reduction Capabilities

Next, we created a standardized score for our composite measure ranging from 0 to 1. The numerator was the total number of survey questions for which a practice responded yes or always; the denominator was the total number of survey questions without missing data for that practice. We used this method for the denominator so that practices that did not respond to all 12 survey questions were not penalized for missing data. We excluded any practices that responded to fewer than seven questions. More than 93% of practices responded to 11 or 12 of the questions and only 14 were excluded for responding to fewer than seven.

Statistical Analyses

We began by examining differences between practices with lower and higher readmission rates. We split practices into quartiles of practice-level unadjusted readmission rates and compared patient and practice characteristics. We compared means for normally distributed continuous variables using t-tests. When continuous variables were not normally distributed, we employed Mann-Whitney U tests. We used chi-square tests to compare groups on binary and categorical variables.

We calculated practices’ readmission rates using mixed effects logistic regression, modeling patient readmissions in 2016 as a dichotomous outcome. We adjusted this model for patient race, median income of a patient’s census tract, frailty, and hierarchical condition category score in 2016. We then replicated CMS method for calculating practices’ predicted over expected readmission rates.33 The predicted readmission rate is based on each practice’s patient characteristics and their practice intercept. The expected readmission rate is based on each practice’s patient characteristics and the overall mean practice intercept. This results in a ratio which we then multiplied by the observed readmission rate for all patients to arrive at practice’s readmission rates.

We subsequently estimated a practice-level linear regression model with risk-standardized readmission rates as the outcome. Our independent variable of interest was a continuous variable (0–1) representing practice readmission capability score. We included control variables for ownership status, number of physicians, Census region, urban/rural status, and percent of patients who are dual-eligible. We applied survey weights for all analyses. We conducted all analyses in Stata 16.0. This study was approved by our institution’s human subjects review board.

Sensitivity Analyses

We constructed our independent variable based on the literature and we estimated our main regression model using this definition. However, in sensitivity analyses, we also examined several other specifications of our independent variable. We estimated the practice-level linear regression model with covariates for each individual survey question, as opposed to a composite score, to test if any of the identified activities were more significantly associated with readmission rates. We also conducted principal component analysis using all 26 original candidate measures as well as the 12 measures included in the final composite. We also revised the composite measure to only include activities that were negatively associated with risk-standardized readmission rates, using a data-driven approach as opposed to solely relying on theory. Finally, we restricted all models to practices with at least 200 admissions in order to be consistent with the Merit-based Incentive Payment System.

RESULTS

Practice and Admission Characteristics

When comparing practices by quartile of observed readmission rates, we confirmed several key differences (Table 2). Observed readmission rates increased as practices cared for greater numbers of admitted patients. Practices in the lowest quartile of observed readmission rates were more likely to be independently owned, employ 3–5 physicians, be located in urban areas, or operate in the West. We also observed increases in practices’ percentage of dual-eligible patients when moving from lower to higher quartiles. Admissions attributed to practices in each quartile were fairly similar in age, sex, and race across all four quartiles. We observed monotonically increasing values for disability, frailty, and hierarchical condition category score when moving from the first to fourth quartile. The inverse was true for median income.

Table 2 Patient and Hospital Characteristics by Observed Practice Readmission Rate Quartile

Practice-Level Model

Our main independent variable of interest, the readmission activities composite, was significantly associated (P < 0.05) with lower risk-standardized readmission rates. We observed a 0.05 percentage point decrease in risk-standardized readmission rates for each additional activity in our composite measure, holding all other factors constant (Table 3). Engaging in more readmission reduction activities was associated with lower risk-standardized readmission rates (Fig. 1). On average, risk-standardized readmission rates for practices with composite scores between 0.8 and 1.0 were one percentage point lower than practices that engaged in none of the activities in our composite measure.

Table 3 Linear Regression Results of Practice-Level 2016 Readmission Rates
Figure 1
figure 1

Association of readmissions reduction activities composite score and risk-standardized readmission rates.

Risk-standardized readmission rates for practices with more than 20 physicians were 0.37 percentage points higher than practices employing 3–5 physicians (P < .05). Compared with practices in the East, risk-standardized readmission rates for practices in the West were 0.40 percentage points lower (P < .01). Urban practices had significantly lower risk-standardized readmission rates than large rural (−0.40 percentage points; P < .05), small rural (−1.73 percentage points; P < .001), and isolated practices (−1.38; P < .001). Each percentage point increase in practices’ percentage of dual-eligible patients was associated with a 0.01 percentage point increase in risk-standardized readmission rate (P < .01). We did not observe a significant relationship between practice ownership category and risk-standardized readmission rate. We include the results of our mixed-effects model in Appendix Table A4.

Sensitivity Analyses

When we estimated our linear regression model with covariates for each individual survey question, as opposed to a composite score, we found that eight measures were negatively associated with readmissions but none of the 12 individual survey questions was independently statistically significant (P<0.05). This suggests that reductions in readmission rates were not driven by one or two activities. Instead it is the combination of multiple activities that is driving the associations in our paper. Principal component analysis using all 26 candidate measures resulted in seven factors with eigenvalues above 1.0. None of these factors by itself was significantly associated (P<0.05) with lower risk-standardized readmission rates when we jointly controlled for all seven. Principal component analysis of the 12 measures included in our final composite measure resulted in three factors with eigenvalues above 1.0 and again; none of the individual factors was significant (P<0.05). We then modeled risk-standardized readmission rates as a function of a revised composite score with only the eight tasks negatively associated with risk-standardized readmission rates (questions 1–3, 6–8, 11–12 in Table 1). As expected, the magnitude of the composite score coefficient increased, from −0.05 to −0.08. Finally, the effect size of the composite score measure nearly doubled, and remained statistically significant, when we limited our sample to practices with at least 200 admissions.

DISCUSSION

Using new, nationally representative data on primary care practices, we found that primary care practices with a greater number of readmission reduction activities experienced lower Medicare risk-standardized readmission rates. Prior studies have demonstrated this effect at the hospital-level,10 across the continuum of care (hospitals, practices, and community services),36 or within individual practices and health systems.17,34 However, our study is the first to demonstrate how readmissions activities are associated with readmission outcomes in a large, nationally representative sample of multi-physician practices.

While each of the activities in our composite measure was a component of commonly successful readmission reduction interventions, none was significantly correlated with lower risk-standardized readmission rates by itself. Instead, increasing reductions in readmission rates occurred as practices performed more activities. Average risk-standardized readmission rates for practices performing almost all or all of the activities in our composite measure were a full percentage point lower than risk-standardized readmission rates for practices engaging in none of the activities. One percentage point was the difference between being classified in the 50th or 75th percentile of risk-standardized readmission rates. One possible explanation for this finding is that successfully reducing readmissions may depend on a multitude of activities. This might explain why multi-component interventions have proven to be the most effective at reducing readmissions.38,39 However, we acknowledge the possibility that practices implementing more activities do so because of a strong commitment to reducing readmissions. Alternatively, these practices may have greater financial/human resources, enabling them to test more activities to see what works best.

The fact that none of the readmission activities in our study was independently associated with readmission rates may seem surprising, given attention in some spheres to specific interventions or readmissions reduction activities. However, systematic reviews and meta-analyses of many commonly referenced interventions demonstrate that individual interventions may not be as effective if not combined with other activities. For example, pharmacist-led medication reconciliation was not proven to reduce readmissions in meta-analyses.40,41 A systematic review of telephone follow-up after discharge found inconclusive evidence about its ability to reduce readmissions.42 Another systematic review of smaller scale interventions with fewer readmission reduction activities concluded that a majority of studies did not report significant decreases in readmission rates.43

Our findings make it challenging for practices with limited resources to prioritize what activities are the most important to implement. However, when looking at some of the most successful multi-component interventions, there are recurring activities that occur in combination. These include the use of a care-manager for complex patients, home visits after discharge for those most in need, having access to discharge summaries, continued patient education after discharge, and conducting a follow-up visit with a primary care physician after hospitalization.7,10,13,15,17,34,35

Another possible explanation for the lack of significant findings when examining individual activities is that much of the evidence on effective readmission reduction interventions has been generated from heart failure patients.10,16,22,44 However, heart failure accounts for only 4% of overall Medicare hospitalizations45 and a little over 5% in our cohort. Interestingly, a meta-analysis of many of the activities in our study found that these interventions are nearly twice as effective in reducing readmissions for heart failure patients, compared to other conditions.39 Therefore, it may be even more important to implement multiple activities when treating an all-condition population as some activities may be more effective for certain kinds of patients.

Finally, our findings should be of particular interest to large or multi-site practices. CMS has mandated that all practices with 200 or more admissions and 16 or more providers report their readmission rates if they decide to participate in the Merit-based Incentive Payment System as opposed to an Alternative Payment Model. When we limited our sample to only practices with 200 or more admissions, we found an even stronger association between our readmission composite activities score and risk-standardized readmission rates. Yet, we also found that practices employing over 20 physicians were significantly more likely to have higher risk-standardized readmission rates, even after controlling for their readmission reduction activities composite score. As such, larger practices may want to focus on what makes smaller practices more successful at reducing readmissions. It may be that smaller practices are more nimble than larger practices and can more easily adjust their activities based on their patients’ needs. It is also possible that ensuring all physicians regularly engage in these activities is easier when there are fewer clinicians, or that smaller practices are more familiar with the needs of their patients.46

We acknowledge several limitations when interpreting our findings. First, fewer than half of practices responded to the NSHOS. While practices responding to the survey were similar to non-respondents on several key observable characteristics, our sample may not be representative of the general population of multiphysician practices delivering primary care. Second, as with any survey, there is always the possibility of measurement error, which could bias our estimates. In general, this would cause attenuation bias; thus, the fact that our composite measure was significant, even in the possible presence of measurement error, is reassuring. Third, our survey is based on 2017–2018 data but we assessed readmissions for patients admitted in 2016. If practices' readmission reduction activities significantly changed from 2016 to 2017–2018, our findings may not accurately assess the relationship between readmission reduction activities and readmission rates. Fourth, the NSHOS does not include practices with fewer than three physicians, which limits generalizability. Fifth, since our study was cross-sectional, there is the possibility for omitted variable bias and our findings do not allow us to infer causality.

CONCLUSION

Routinely engaging in a greater number of readmission reduction activities was significantly associated with lower practice-level risk-standardized readmission rates. This may represent a culture-level effect in which practices engaging in more readmission reduction activities do so because they are more committed to reducing readmissions or have more resources to prevent readmissions. However, it is also possible that reducing readmissions is partly dependent on the interaction of many different activities.