Using Generalized Structured Additive Regression Models to Determine Factors Associated with and Clusters for COVID-19 Hospital Deaths in South Africa (preprint)

ABSTRACT

Background: The first case of COVID-19 in South Africa was reported in March 2020 and the country has since recorded over 3.6 million laboratory-confirmed cases and 100 000 deaths as of March 2022. Transmission and infection of SARS-CoV-2 virus as well as deaths in general due to COVID-19 have been shown to be spatially associated but spatial patterns in hospital deaths have not fully been investigated in South Africa. This study uses national COVID-19 hospitalization data to investigate the spatial effects on hospital deaths after adjusting for known mortality risk factors. Methods: COVID-19 hospitalization data and deaths were obtained from the National Institute for Communicable Diseases (NICD), who together with the South African National Department of Health (SANDoH) collected hospital admissions data through DATCOV, an active electronic hospital surveillance system for COVID-19. We used the generalized structured additive logistic regression model that allows for modelling spatial correlation to realistically estimate risk factors for hospital COVID-19 deaths. The model included patient demographic and clinical factors as well as time in months which accounted for different waves. Continuous covariates were modelled by assuming second-order random walk priors, while spatial autocorrelation was specified with Markov random field prior and fixed effects with vague priors respectively. The inference was fully Bayesian. Results: The risk of COVID-19 in-hospital mortality increased with patient age as well as with admission to intensive care unit (ICU) (aOR=4.16; 95% Credible Interval 4.05-4.27), being on oxygen (aOR=1.49; 95% Credible Interval 1.46-1.51) and on invasive mechanical ventilation (aOR=3.74; 95% Credible Interval 3.61-3.87). Being admitted in a public hospital (aOR= 3.16; 95% Credible Interval 3.10-3.21) was also a significant risk factor for mortality. Risk of deaths also increased in months following a surge in infections and dropped after months of successive low infections highlighting crest and troughs lagging the epidemic curve. After controlling for these factors, districts such as Vhembe, Capricorn and Mopani in Limpopo province, and Buffalo City, O.R. Tambo, Joe Gqabi and Chris Hani in Eastern Cape province remained with significantly higher odds of COVID-19 hospital deaths suggesting possible health systems challenges in those districts. Conclusion: The results show substantial COVID-19 in-hospital mortality variation across the 52 districts. This highlights the importance of modelling spatial patterns simultaneously with fixed and nonlinear effects of continuous covariates to identify clusters at high risk of health outcome. The flexible approach to modelling data that has spatial patterns helps to account for possible loss of efficiency due to spatial correlation that spatial patterns can induce in data. Our analysis suggests notable COVID-19 hospital deaths clustering in some districts in Limpopo and Eastern Cape provinces and this information can be important for strengthening health policies and the public health system for the benefit of the whole South African population. Understanding differences in in-hospital COVID-19 mortality across space could guide interventions to achieve better health outcomes.