Conditional mean and quantile regression analysis to evaluate stroke risk factors in South Africa, trend and seasonality of stroke and its associated economic burden

Abstract

Stroke is the second largest cause of mortality and long-term disability in South Africa (SA). Stroke is a multifactorial disease regulated by modifiable and non-modifiable predictors. Little is known about modelling stroke cases in SA, particularly using modifiable and non-modifiable predictors. This study aims to address important gaps in empirical stroke literature i.e., identifying and quantifying stroke predictors through Bayesian analysis and quantile regression (QR) mea-surement error (ME) models to cater for MEs in predictors in modelling stroke cases. Analysis of stroke predictors has in the mainstream literature, concentrated on mean regression, yet mod-elling with Bayesian techniques or QR is more appropriate than using mean regression. This stem from the former’s ability to provide probability distributions for parameters, quantify uncertainty more effectively by allowing for incorporation of prior knowledge while the latter gives complete response conditional distributions. However, as a point of departure, mean regression analysis is also employed. A cross-sectional hospital-based study is used to identify and quantify stroke predictors in SA using 35 730 individual patient data retrieved from selected private and public hospitals between January 2014 and December 2018. Predictors of stroke such as hypertension, hypercholesterolemia and diabetes are usually measured with error. These measurement errors include instrumental, and observer errors, just to mention a few. If MEs are ignored, inference based on parameter estimates and confidence intervals in statistical models are often adversely affected by serious biases. QR estimates for stroke predictors and Bayesian analysis of ME correction methods using integrated nested Laplace approximation (INLA) and Markov chain Monte Carlo (MCMC) are compared. Classical logistic regression models are fitted to identify the modifiable and non-modifiable stroke predictors. The least absolute shrinkage and selection operator (LASSO) logistic regression model is employed to select predictors for modelling stroke cases in SA. A Poisson generalised linear model (GLM) for count time series data is also fitted to identify the trend and seasonality of stroke cases and associated predictors in SA. The Bayesian and QR models are further fitted to identify predictors associated with increased cost of stroke treatment/management. The LASSO logistic regression was fitted, and eight variables were significant with two vari-ables shrunk to zero/near zero, unlike the classical logistic regression that found all variables were significant. The same risk factors were found to be associated with stroke occurrence for both the Poisson generalised linear regression and Poisson QR for seasonality at conditional mean and 50ᵗʰ quantile. The LASSO logistic Bayesian QR produced the same significant predictors as the logistic QR models (two predictors were not significant). The Bayesian QR models using INLA have the smallest values of MSE across all quantiles, which indicates that the parameter estimates are more precise when compared to MCMC.