Comparative Analysis of Data Correlations for Selecting Predictive Variables Using Simple Linear Regression Comparative Analysis of Data Correlations for Selecting Predictive Variables Using Simple Linear Regression

Abstract

This study aimed to compare linear relationships among variables across multiple datasets to identify highly positively correlated variable pairs suitable for predictive modeling using simple linear regression. Secondary data from five open datasets, Heart.csv, StudentMarks.csv, TestEnergyData.csv, Salary_dataset.csv, and Advertising.csv, were analyzed using Pearson’s correlation coefficient, followed by the construction of simple linear regression models to evaluate predictive performance. Key evaluation metrics included the coefficient of determination ( R2R^2R2 ), root mean squared error (RMSE), and mean absolute error (MAE). The results showed that Salary_dataset.csv exhibited the strongest linear relationship, with an R2R^2R2 of 0.9570, indicating that “Years of Experience” is a highly effective predictor of “Salary.” Similarly, StudentMarks.csv and Advertising.csv demonstrated strong correlations ( R2>0.80R^2 > 0.80R2>0.80 ), suggesting that “Study Time” and “TV Advertising” are reliable predictors of student marks and product sales, respectively. In contrast, Heart.csv and TestEnergyData.csv showed weak to moderate relationships and were deemed less suitable for prediction using simple linear regression. This research emphasizes the importance of correlation-based variable selection prior to model construction. The comparative approach adopted here offers a practical method for identifying high-potential predictor variables, supporting data-driven forecasting in fields such as education, economics, and marketing.