OBJECTIVES: This study uses the healthcare costs of patients in the United States with different utilization patterns during the early COVID-19 pandemic to explore methods for improving the fit of multivariable cost modeling.

METHODS: We identified individuals with continuous claims enrollment between 03/01/2019–02/28/2022 and EHR activity during Y2019 (03/01/2019–02/29/2020) and Y2021 (03/01/2021–02/28/2022) in the Veradigm Network EHR linked to claims data. These individuals were sorted into three cohorts (skippers, delayers, and continuers) based on healthcare utilization in Y2020 (03/01/2020–02/28/2021). We evaluated six models of follow-up healthcare costs (365 days following the first Y2020 visit for delayers and continuers and Y2021 for skippers). Two-part models included a logistic regression model of having non-zero follow-up costs and a generalized linear model of follow-up costs. All models included age, region, sex, race, ethnicity, and baseline costs as covariates with Y2020 utilization cohort as the predictor of interest. Models varied on incorporation of additional covariates including Charlson Comorbidity Index (CCI), immunocompromised status, and interaction terms.

RESULTS: Before adjusting for covariates, mean (SD) unadjusted healthcare costs in the follow-up period were $10,621 ($43,647) for skippers (N = 576,030), $12,178 ($47,126) for delayers (N = 992,655), and $14,559 ($59,808) for continuers (N = 1,754,150). After adjusting for covariates, delayers had significantly higher costs than continuers in all models with average marginal effect (AME) estimates ranging from $831­ to $1,069. AME estimates for skippers versus continuers were more varied, ranging from -$1,024 to $441. In the best fit two-part model, which included a categorical CCI and interaction terms of cohort*age and cohort*baseline_costs, delayers had significantly higher costs than continuers and skippers had significantly lower costs than continuers.

CONCLUSIONS: The findings of multivariable modeling are dependent on selection of covariates and use of a two-part model improved overall fit regardless of covariates.