A Comparative Analysis of Regression Models for Predicting COVID-19 Mortality

The rapid proliferation of COVID-19 data necessitated robust machine learning models for forecasting and analysis. This study aimed to identify the most suitable regression model for predicting COVID-19 mortality by comparing Simple Linear Regression (SLR) and Multiple Linear Regression (MLR) models on a comprehensive dataset. Data preprocessing involved removing features with over 30% missing values and imputing the remainder using the sample median. SLR analysis demonstrated that models based on aggregated continental data were consistently non-significant (p > 0.05), whereas all models based on country-level attributes were statistically significant (p < 0.05). The strongest SLR model utilized total cases per million as a predictor, explaining 35.01% of the variance. However, subsequent diagnostic checks on the fitted MLR model revealed critical violations of classical linear regression assumptions, specifically the failure of the Shapiro-Wilk test for normality and the detection of heteroscedasticity. These violations render the MLR coefficient estimates and significance tests unreliable. Based on scatter plot observations confirming curvilinear relationships and the linear model failures, the study concludes that non-linear approaches, such as Polynomial Regression, are required to accurately model the relationship between variables and provide statistically sound predictive performance.