Forecasting the 2021 local burden of population alcohol‐related harms using Bayesian structural time–series

Abstract

Background and aims Harmful alcohol use places a significant burden on health services. Sophisticated nowcasting and forecasting methods could support service planning, but their use in public health has been limited. We aimed to use a novel analysis framework, combined with routine public health data, to improve now‐ and forecasting of alcohol‐related harms. Design We used Bayesian structural time–series models to forecast alcohol‐related hospital admissions for 2020/21 (from 2015 to 2016). Setting England. Participants We developed separate models for each English lower‐tier local authority. Measurements Our primary outcome was alcohol‐related hospital admissions. Model covariates were population size and age‐structure. Findings Nowcasting validation indicated adequate accuracy, with 5‐year nowcasts underestimating admissions by 2.2% nationally and 3.3% locally, on average. Forecasts indicated a 3.3% increase in national admissions in 2020/21, corresponding to a 0.2% reduction in the crude rate of new admissions, due to population size changes. Locally, the largest increases were forecast in urban, industrial and coastal areas and the largest decreases in university towns and ethnically diverse areas. Conclusions In 2020/21, alcohol‐related hospital admissions are expected to increase in urban and coastal areas and decrease in areas associated with inward migration of younger people, including university towns and areas with greater ethnic diversity. Bayesian structural time–series models enable investigation of the future impacts of alcohol‐related harms in population subgroups and could improve service planning and the evaluation of natural experiments on the impact of interventions to reduce the societal impacts of alcohol.

Forecasting the patterns of COVID-19 and causal impacts of lockdown in top five affected countries using Bayesian Structural Time Series Models

“…The BSTS models produced better forecasts, as compared to ARIMA, for future heath harms due to use of alcohol. They produced reasonably good 1–5 year forecast, even with only eight data points in the training dataset [15] .…”

“…These models can be efficiently used in public health for prioritizing, developing and implementing policies to tackle and avoid the adverse health situations [21] . These models have already been used to forecast the health damages due to use of alcohol [15] and to forecast the health harms and crime rates as a results of local alcohol licensing policies [6] . It is also possible to select the suitable variables via Spike and Slab priors using these models [19] .…”

“…These models select the most informative parameters and do not require the linear regression component of the model. These models provide improved representation of the certainty of estimates, and evolve (embitter their accuracy) over time [15] . However, the analytical computation of the Bayesian posterior distribution is extremely difficult.…”

Assessing the future progression of COVID-19 in Iran and its neighbors using Bayesian models

“…Although these are valuable contributions, but results of these contributions are based on quite small datasets, which can affect the reliability of the corresponding forecasts ( Moftakhar et al., 2020 ). Secondly, all of these contributions utilized the classical models which consider the model parameters as fixed quantities, which may not be suitable for dynamic systems such as this pandemic ( McQuire et al., 2019 ; Scott & Varian, 2013 ). A careful review of literature suggests that no study has involved the evolving behavior of the COVID-19 patterns.…”

“…These models consider the model parameters as random variables, which is more suitable for an evolving system, such as COVID-19. In addition, these models include the prior information and are capable to reflect the stochastic nature of temporal data more accurately ( McQuire et al., 2019 ). Using these models, the trend, seasonality and regression components of the series can be modeled separately.…”