Enhancing predictive accuracy of electoral outcomes: A comparative study of dimension reduction techniques in linear regression models

The effectiveness of Principal Component Analysis (PCA) and Factor Analysis (FA) in enhancing the predictive accuracy and interpretability of linear regression models for electoral outcomes is examined in this study, with a focus on the Croatian presidential election. Challenges such as multicollinearity and model complexity were addressed through the application of PCA and FA to socio-demographic and political variables, resulting in their transformation into a simpler, more manageable form. Multicollinearity was significantly reduced by PCA, and model stability was improved. Deeper insights into voter behavior were provided by FA through the identification of latent variables. Improvements were noted; however, issues with homoscedasticity were identified, which affected the predictive reliability of the models. The optimized linear regression model, specifically designed to address homoscedasticity (LR2), was found to demonstrate superior performance in terms of predictive accuracy and stability when compared to models that utilized PCA and FA alone. The importance of maintaining homoscedasticity in regression models is underscored by the findings, and the potential of dimension reduction techniques to improve electoral predictions is highlighted. The integration of these techniques in forecasting models is advocated for, with the aim of assisting political analysts and policymakers. It is suggested that future research could explore the combination of these methods with other advanced modeling approaches to further enhance predictive capabilities.