Variation Among Primary Care Physicians in 30-Day Readmissions

Abstract

Readmissions for Medicare beneficiaries are costly, and some may represent poor quality of care (1, 2). Prompted by the high cost, poor quality, and hospital variation associated with readmissions (3), the Centers for Medicare & Medicaid Services (CMS) launched the Hospital Readmissions Reduction Program, a pay-for-performance program that encourages hospitals to reduce readmissions by decreasing payments to hospitals with excess readmissions (4). Evidence suggests that the program may have succeeded in reducing readmissions (5). Nevertheless, rates remain high, and further improvements will require a better understanding of other factors that may influence readmissions. Risk for readmission might be influenced by the physicians providing care. This includes care during the initial hospitalization by the inpatient physician, follow-up care after discharge by the primary care physician (PCP), and emergency care after discharge by the emergency department (ED) physician. Readmission rates are higher in patients receiving inpatient care from hospitalists than in those receiving it from their PCP (6). However, risk for readmission does not vary by individual hospitalist (7). Early follow-up after hospital discharge is sometimes associated with lower readmission rates (8, 9). We previously reported that risk for readmission varies moderately but significantly by ED physician, identifying an opportunity for improvement (10). However, whether PCP care influences readmissions is uncertain, although CMS has implemented a policy incentivizing PCPs to reduce readmissions (11). We undertook this study to determine whether risk for readmission varies by PCP. We also assessed variation among PCPs in the rate at which they see patients in outpatient follow-up within a week of discharge. Methods Data Source We used 100% of data from Texas Medicare claims for 1 January 2007 through 31 December 2015. These include the Medicare denominator file for demographic and enrollment information, the carrier file for physician claims, the outpatient statistical analysis file for outpatient claims, and the Medicare Provider Analysis and Review (MedPAR) file for inpatient claims. Cohort Selection The study used 2 different cohorts, 1 to study 30-day readmission rates and the other to study whether patients were seen by their PCP within 7 days of hospital discharge. To develop the readmission cohort, we started with all hospitalized patients who were discharged between 1 January 2012 and 30 November 2015 (Appendix Table 1). For beneficiaries with multiple hospitalizations in the same year, the first hospitalization was kept. We then identified patients who were discharged alive and not transferred to other acute care hospitals. We included only hospitalizations for beneficiaries aged 66 years or older with continuous coverage from Medicare Parts A and B and no HMO in the 12 months before and 1 month after hospitalization. We excluded those who died without a readmission within 30 days after hospital discharge. We selected beneficiaries who had an identifiable PCP in the year before hospitalization and excluded those whose PCPs had fewer than 50 hospital admissions in the data. We defined a PCP as a generalist (general practitioner, family physician, internist, or geriatrician) who saw a given patient on 2 or more occasions in an outpatient setting in the year before the hospitalization of interest (12). We used CPT (Current Procedural Terminology) codes 99201 to 99205 (new patient encounters), 99211 to 99215 (established patient encounters), and G0463 to identify outpatient visits from carrier files. The generalist with the most visits was assigned as the PCP. When 2 generalists had an equal number of visits, the most recently visited provider was assigned. Appendix Table 1. Steps in Cohort Selection for Analyzing Readmission Within 30 Days After Discharge The follow-up cohort differed from the readmission cohort in that it included only patients discharged directly to the community and did not exclude those who died in the 30 days after discharge (Appendix Table 2). Appendix Table 2. Steps in Cohort Selection for Analyzing 7-Day Follow-up Visit* Measurements Patient age, sex, ethnicity, and Medicaid eligibility were obtained from Medicare beneficiary summary files. We used the Medicaid eligibility indicator as a proxy for low socioeconomic status. Admission type (emergency vs. nonemergency and weekday vs. weekend) and diagnosis-related group (DRG) codes were obtained from MedPAR files. We determined residence in a nursing facility in the 3 months before the hospitalization of interest using the MedPAR files and evaluation and management codes 99304 to 99318 (nursing facility services) from carrier files (13). Total numbers of hospitalizations and outpatient visits in the prior year were identified from MedPAR files and carrier files, respectively. Education level at the ZIP code of residence was obtained from the 2011 American Community Survey estimates from the U.S. Census Bureau. Elixhauser comorbidity indicators were identified using claims from the MedPAR, carrier, and outpatient statistical analysis files in the year before the hospitalization of interest (14). Study Outcomes The primary outcome was any readmission within 30 days of discharge. A secondary outcome was whether patients were evaluated in an outpatient setting by their PCP within 7 days of hospital discharge. Overview of Analytic Approach Our aim was to describe the degree of variation in readmission rates that is attributable to PCPs. We explored many analytic approaches, which yielded near-identical results. Our main results were generated from a marginal logistic model. We repeated the analyses using a multilevel logistic regression model and a conditional model. To plot risk-standardized rates of 30-day readmissions for each PCP, we used bootstrapping with the multilevel regression model to estimate the 95% CI for each PCP. We estimated the stability of the PCP-adjusted readmission rates by comparing the results based on admissions from 2012 to 2015 with those based on data from 2008 to 2011 for the 3408 PCPs with data in both time periods. Finally, we calculated the 95% CIs for different postulated rates of readmission for a PCP at different sample sizes (numbers of admissions) to explore the feasibility of generating robust information on PCP performance using the readmission rates of their patients. Statistical Analysis Descriptive analyses were used to summarize the association between patient characteristics and readmission rates within 30 days of hospital discharge. We used generalized estimating equations (population average or marginal) models in the GENMOD package in SAS to obtain predicted 30-day readmission rates and population-averaged odds ratios associated with patient characteristics. We obtained the predicted 30-day readmission rate and 95% CI for each patient characteristic from a multilevel logistic model and a generalized estimating equations model using the margins command after running a logistic regression in Stata (15). For continuous variables, the predicted readmission rates were calculated at the median point. The model was adjusted for patient age, ethnicity, sex, Medicaid eligibility, education, emergency admission, weekend admission, DRG weight, major DRG diagnostic class, nursing home residence in the 90 days before admission, the 31 Elixhauser comorbid conditions entered separately, number of acute hospitalizations in the 12 months before the admission, and number of outpatient visits in the prior year. As an alternative, we implemented a multilevel logistic regression model (patient and PCP) using the same covariates. The model was implemented using the GLIMMIX package in SAS, and the QUAD method allowed variables to be estimated by quadrature (16, 17). We estimated the risk-standardized rate of 30-day readmissions for each PCP with the method used by CMS (18). Risk-standardized rates were calculated as the ratio of the predicted to expected number of readmissions, multiplied by the national unadjusted rate of readmission. For each PCP, the numerator of the ratio is the number of readmissions predicted within 30 days based on the PCP s observed case mix. The denominator is the number of readmissions expected based on the nation s performance with that PCP s case mix. The model was repeated 1000 times, and the means and 95% CIs were calculated on the basis of these 1000 values for each PCP. We then plotted the adjusted rate and 95% CI for each PCP and ranked them from low to high. We considered a PCP s readmission rate to be statistically significantly higher or lower than the mean if the 95% CI excluded the mean risk-standardized rate of readmission for all PCPs. We evaluated the stability of the PCP-level profiling by comparing the risk-standardized rate of 30-day readmissions generated from the bootstrapping method in 2 time periods (1 January 2008 to 30 November 2011 and 1 January 2012 to 30 November 2015) for the 3408 PCPs with at least 50 admissions in each time period. For each time period, we categorized each PCP as having adjusted rates that were significantly higher than the mean, significantly lower than the mean, or not significantly different from the mean; we then compared the categories in a 23 table. To explore the robustness of the estimates of PCP readmission rates, we calculated the 95% CIs for different postulated readmission rates at different postulated sample sizes using the following formula: We set the expected readmission rate at 13% and calculated the minimum sample sizes that could be used to detect different postulated observed rates (starting at 14.0% and increasing by 0.25percentage point intervals) that would be significantly different from the expected rate (that is, their 95% CIs excluded 13%). The analyses of rates of follow-up visits with the PCP were similar to those of readmission rates and had the