anobel / riskadjustedoutcomes

Beyond Classical Risk Adjustment

Socioeconomic Status and Hospital Performance in Urologic Oncology

Anobel Y. Odisho MD MPH, Ruth Etzioni PhD, John L. Gore, MD MS

Department of Urology, University of Washington; Fred Hutchinson Cancer Research Center

Abstract

Importance Safety-net hospitals take care of a more patients from lower socioeconomic status (SES) than non-safety-net hospitals and may be disproportionately punished under readmission risk adjustment models that do not incorporate SES. Adequate risk adjustment is essential to ensure that reimbursement aligns with performance and is not influenced by patient selection.

Objective To develop a readmission risk adjustment framework that incorporates SES and to assess impact of SES on safety-net hospital readmission rate rankings.

Design Quasi-experimental design using California Office of Statewide Health and Planning Data from 2007-2011. Unadjusted hospital rankings and predicted rankings under a base model, which simulated the Medicare Hospital Readmissions Reduction Program model (age, sex, comorbidity), were compared to predicted rankings under models incorporating SES and hospital factors. SES (multifactorial neighborhood score) at ZIP code level calculated from US Census data and hospital data was obtained from Medicare.

Setting All acute-care hospitals in California.

Participants All patients undergoing radical cystectomy (RC) for bladder cancer (n = 3,771), partial nephrectomy (PN) (n = 5,556) and radical nephrectomy (RN) (n = 13,136) for kidney cancer.

Main Outcomes and Measures Hospital rankings based on thirty-day all-cause acute-care readmission rate and differences between model predicted rankings.

Results The thirty-day readmission rate was 26.1% for RC, 8.3% in RN, and 9.5% in PN. For all procedures, the addition of SES, geographic, and hospital factors changed the overall hospital rankings significantly compared with the base model (p < 0.01), with the exception of SES in RC (p = 0.07) and SES and rural factors in PN (p = 0.12). For both RN and PN, the addition of SES and hospital factors significantly changed the mean ranking of SNHs. The mean ranking of SNHs increased when adding these factors, improving the ratio of observed relative to expected rankings (p < 0.01). For RC there was no significant change in rankings of SNHs with the addition of either SES, rural status, or hospital factors.

Conclusions and Relevance Adding SES to existing Medicare readmission risk adjustment models leads to significant changes in hospital rankings, with a differential impact on safety-net hospitals.

Data Sources

All data necessary to reproduce our results, except for private patient data, has been made available here both in raw and tidy formats. We do not have permission to make patient data publicly available, limiting our ability to make this work exactly reproducible. However, we have made every effort to make this work easy to replicate with your own patient data. All other data that was used is publicly available and the code to generate the tidy data from raw sources is available here. The data, in raw and tidy formats, is available in this GitHub repository in the data directory. In addition, at the time of manuscript submission, the data was deposited at to the Open Science Framework, frozen to prevent further editing, and assigned a Digital Object Identifier (DOI): 10.17605/OSF.IO/D643S.

Description of Files

To re-create the entire tidy data set, the files should be executed in the following order.

rao_sessioninfo.txt: Detailed info regarding all software versions used at the time of final analysis.

rao_processing_zip.R: Create a ZIP to ZCTA crosswalk. This should be run *first8, as the crosswalk is used in many subsequent steps. This will create data/tidy/zcta.csv and data/tidy/zctalatlon.csv.

rao_processing_oshpd.R: This file imports raw OSHPD patient data (not available publicly), and tidies it in preparation for merging with additional datasets and final analysis. This will create data/patient/tidy/pt.rds.

rao_processing_oshpdfinancial.R: Download California Office of Statewide Health and Planning data to create a crosswalk from OSHPD ID to Medicare ID, used for merging OSHPD and Medicare data. This creates data/tidy/oshpdxwalk.csv.

rao_processing_DSH.R: Download and tidy Medicare data for hospital factors, Disproportionate Share Index, Case Mix Index. This creates data/tidy/dsh.csv.

rao_processing_icd9cm.R: create labels for ICD9 codes from Medicare data. This creates data/tidy/icd9cm.csv.

rao_processing_medicarepos.R: Download Medicare Provider of Service files and obtain number of FTE residents per facility. This creates data/tidy/residents.csv.

rao_processing_drivingdistances.R: This uses ggmap to calculate driving times and distances from patient to hospital zip centroids. This creates data/tidy/distances.csv.

rao_processing_rural.R: Assign RUCA codes at the zip code level for each patient. This creates data/tidy/rural.csv.

rao_processing_spatial.R: Merge data with TIGER/LINE shape files for creation of interactive maps. This creates data/tidy/spatial.rda.

rao_processing_acs.R: Using the acs package, data is downloaded directly from the US Census American Community Survey for California and Diez-Roux Neighborhood scores are calculated. This is then saved as data/tidy/acs.csv for analysis. This creates data/tidy/acs.csv.

rao_processing_all.R: This file merges tidy patient data with additional datasets to create the final analytic cohorts. This should be executed LAST. This creates data/tidy/ptcombined.rds data/tidy/ptgu.rds.

Results

floating.css is necessary for the left-sided navigation menu used for all markdown reports.

rao_describe_oshpd.Rmd: Basic decriptive analyses of all variables from mulitple data sources and detailed output of all statistical models for each cohort. To create results for each cohort, name the desired cohort (Cystectomy, PartialNx, or RadNx) as a paramater in the YAML.

rao_tables.Rmd Generates Tables 1 and 2 in the manuscript, which include all basic descriptive analyses by cohort and univariate tests of significance.

rao_oecalculations.Rmd Generates Tables 3 and 4 in the manuscript. This includes detailed model output for all comparisons in all three cohorts, along with tests of significance (Wilcoxon rank-sum and permutation testing).

riskadjustedoutcomes's Issues

convert all imports to read_csv from readr

trends in readmission: plot readmissions over time by cohort

SES models with all components

Calculate distance from patient zip to treatment hospital

Stratify sampling for training set

when making training set, instead of taking random sample across entire dataset, group by hospital and sample randomly from within that hospital to ensure all hospitals present in training AND testing data

create extended LOS flag

based on some metric from other data (SEER medicare, NCDB, NIS), not our own data

Comment OE calculations file

add Rural/Urban classification

create models for hospital level factors

add hospital characteristics into random effects model

procedure specific volume (continuous and quintile)
Volume of all GU procedures (continuous and quintile)
hospital bed number
hospital type (academic/not)

Create new OE table of wilcoxon results and AUC

O/E calculations: combine models

Combine single factor models with multi factor grouped models into single markdown file

Comment tables markdown table (tables 1 & 2)

merge hospital chars data from DSH data

logistic models replicating medicare TKA

create chart of readmission rates by year by disease

consider looking at subsets of hospitals for predictions

Incorporate resident data

import medicare Provider of Services data, # of resident FTE per hospital, academic yes/no
https://www.cms.gov/Research-Statistics-Data-and-Systems/Downloadable-Public-Use-Files/Provider-of-Services/POS2007.html

Calculate O/E ratios for models

calculate hospital volume for each procedure, with quintiles

Account for in-hospital mortality for each cohort

incorporate mortality data

recently obtained from OSHPD: patient mortality data

Classify Safety Net Hospitals

safety net as defined by medicare DSH as highest quartile
See: Matlin Gilman, E. Kathleen Adams, Jason M. Hockenberry, Ira B. Wilson, Arnold S. Milstein and Edmund R. Becker. California Safety-Net Hospitals Likely To Be Penalized By ACA Value, Readmission,

And Meaningful-Use Programs Health Affairs, 33, no.8 (2014):1314-1322)

Add combined models to AUC calculations

table 1 showing frequency of various model variables, in readmission yes/no groups

Add SES Diez-Roux neighborhood Z-score (continuous and quintile)

assign weekday to discharge date

Center, scale (linear) ACS variables

will make interpreting regression results easier