Distinguishing Death from Disenrollment: Applying a Predictive Algorithm to Reduce Bias in Estimating the Risk of Rehospitalization

Young, Jessica; Dasgupta, Nabarun; Pack, Kenneth; Bloemers, Sarah; Cooper, Toska; DiPrete, Bethany; Stürmer, Til; Pate, Virginia; Lund, Jennifer; Irwin, Debra; Gibson, Teresa; Funk, Michele

Request Version for Screen Reader

Last Modified

September 28, 2022

Creator

Young, Jessica
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology
Dasgupta, Nabarun
- Affiliation: Injury Prevention Research Center
Pack, Kenneth
- Other Affiliation: IBM Watson Health
Bloemers, Sarah
- Other Affiliation: IBM Watson Health
Cooper, Toska
- Affiliation: Injury Prevention Research Center
DiPrete, Bethany
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology
Stürmer, Til
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology
Pate, Virginia
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology
Lund, Jennifer
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology
Irwin, Debra
- Other Affiliation: IBM Watson Health
Gibson, Teresa
- Other Affiliation: IBM Watson Health
Funk, Michele
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology

Abstract

Background: The inability to identify dates of death in several insurance claims data sources can result in biased estimates when death is a competing event. To address this issue, an algorithm to predict when plan disenrollment is due to death was developed and validated using the MarketScan insurance claims data. Objectives: We illustrate the bias introduced when estimating the risk of rehospitalization within 90-days of acute myocardial infarction (AMI) if death is not accounted for as a competing event. We demonstrate how this validated algorithm can be used to reduce this bias. Methods: We use a 20% sample of Medicare claims (2007–2017) to identify patients with an incident admission for AMI. Patients were required to be 66+ years of age with employer-sponsored supplemental insurance. We compare 3 methods of estimating the risk of 90-day rehospitalization. The first method uses the true death data available in the Medicare enrollment data. We used cumulative incidence functions to estimate the risk of rehospitalization, accounting for death as a competing risk. The second method mimics scenarios where death data are unavailable, and patients are disenrolled from insurance coverage shortly after death. We used Kaplan Meier curves to estimate the risk of rehospitalization, treating death as non-informative censoring at the time of disenrollment. The third method applies the validated predictive algorithm to the Medicare claims where death date has been obscured. We used a predicted probability threshold of 0.99 to distinguish between plan disenrollment and death (sensitivity = 0.92, specificity = 0.90). We estimated the risk of rehospitalization accounting for predicted death as a competing risk. Results: We identified 12 753 patients with an index hospitalization for AMI (mean age = 77.8 years). When accounting for death as a competing risk using validated death dates, the estimated 90-day risk of rehospitalization was 21.6% (20.8%, 22.3%). When mimicking a scenario where death is treated as non-informative censoring at the time of disenrollment, the estimated 90-day risk was 24.8% (23.9%, 25.6%). When using the algorithm to distinguish between death and disenrollment and accounting for predicted death as a competing risk, the estimated 90-day risk was 21.7% (21.0%, 22.4%). Conclusions: When estimating the risk of rehospitalization following AMI in a cohort of Medicare patients, applying a claims-based algorithm to predict death resulted in estimates that closely mirrored the estimates using validated death data. Alternatively, failure to account for death as a competing risk resulted in an estimate that was biased upwards.

Date of publication