Informative censoring by health plan disenrollment among commercially insured adults

Abstract

Purpose Health plan disenrollment occurs frequently in commercial insurance claims databases. If individuals who disenroll are different from those who remain enrolled, informative censoring may bias descriptive statistics as well as estimates of causal effect. We explored whether patterns of disenrollment varied by patient or health plan characteristics. Methods In a large cohort of commercially insured adults (2007‐2013), we examined two primary outcomes: (a) within‐year disenrollment between January 1 and December 30, which was considered to occur due to patient disenrollment from the health plan, and (b) end‐of‐year disenrollment on December 31, which was considered to occur due to either patient disenrollment from the health plan or withdrawal of the entire health plan from the commercial insurance database. In yearly cohorts, we identified factors independently associated with disenrollment by using log‐binomial regression models to estimate risk ratios (RR) and 95% confidence intervals (CI). Results Among 2 053 100 unique patient years, the annual proportion of within‐year disenrollment remained steady across years (range, 13% to 14%) whereas the annual proportion of end‐of‐year disenrollment varied widely (range, 8% to 26%). Independent predictors of within‐year disenrollment were related to health status, including age, comorbidities, frailty, hospitalization, emergency room visits, use of durable medical equipment, use of preventive care, and use of prescription medications. In contrast, independent predictors of end‐of‐year disenrollment were related to health plan characteristics including insurance plan type and geographic characteristics. Conclusions Differential risk of disenrollment suggests that analytic approaches to address selection bias should be considered in studies using commercial insurance databases.

Toward Evaluation of Disseminated Effects of Medications for Opioid Use Disorder within Provider-Based Clusters Using Routinely-Collected Health Data

“…Standard estimation methods assume that there is no bias due study drop out. If this assumption is questionable, censoring weights could be employed in the analysis; however, selection bias in routinely-collected health data may complicate application of existing methods due to different censoring mechanisms [75, 76].…”

Toward evaluation of disseminated effects of medications for opioid use disorder within provider‐based clusters using routinely‐collected health data

“…Standard estimation methods assume that there is no selection bias due study drop out. If this assumption is questionable, inverse probability of censoring weights could be employed in the analysis; however, selection bias in routinely‐collected health data may complicate application of existing methods due to different censoring mechanisms 75,76 …”

“…If this assumption is questionable, inverse probability of censoring weights could be employed in the analysis; however, selection bias in routinely-collected health data may complicate application of existing methods due to different censoring mechanisms. 75,76 Based on our simulation study, we recommend extending these estimators for alternative distributions of the random effects to better match distributions observed in the data, 62,63,77 as well as possibly a generalized estimating equation (GEE) approach to quantify the cluster-level exposure weights. 78 A GEE model would be a way to avoid the issue with the normality of the random effects observed with linear mixed models in this setting; however, this approach would make different assumptions about modeling the cluster-level treatment.…”

Commonly used definitions in real-world studies may underestimate the prevalence of renal disease among nonvalvular atrial fibrillation patients

“…This would have less of an impact for more severe patients treated in secondary care, as the relevant ICD-10 codes are re-entered in HES. For renal disease, it is also feasible that the patients with mild disease are seen relatively infrequently, as current guidelines only recommend frequent monitoring for more severe patients [26]. Other comorbidities, which were particularly affected by the increased look back window, were malignancies, cerebrovascular disease and chronic pulmonary disease.…”