Back to normal? a method to test and correct a shock impact on healthcare usage frequency data

Abstract

A method based on Bayesian structural time series is proposed to predict healthcare usage trends and to test for changes in the series levels during or after an abnormal year, such as that of the 2020 COVID-19 pandemic. Our method can also serve to calculate correction factors for frequency count data that can be integrated in a preprocessing step before undertaking a cross-sectional statistical analysis, and, in this way, the impact of a shock can be eliminated. Here, adjustments are derived for a large private health insurer in Spain from estimates of average healthcare usage. Median claims rate levels in 2020 were 15% down on 2019 figures, but rose in 2021 and 2022, when the rate was 11% and 8% higher than in 2019, respectively. Once the shock correction is incorporated in the preprocessing step, our approach is shown to outperform traditional time series techniques. Healthcare insurance usage in Spain did not fully go back to normal levels (assuming that pre-pandemic values represent normality) in 2022, with the exception of some patient groups and specific medical services. Our method can be implemented in other areas of risk analysis when frequency counts are exposed to shocks and it allows estimating the difference in claims volume between real figures and those estimated, had the shock not occurred.