Missing Values Treatment in Agronomy Dataset Using PCA-Based Multiple Imputation (Bootstrap Versus Bayesian)

Abstract

Missing values are prevalent in agronomy datasets and need consideration to ensure the applicability of statistical methods and avoid bias in treating them. Previous studies indicate that multiple imputation is more effective than single imputation, with Principal Component Analysis (PCA)-based methods effectively handling multicollinearity in multivariate data. However, such approaches are rarely applied to agronomy data, hence there is a need to assess their performance to add knowledge in the area. This study evaluates the performance of two PCA-based multiple imputation approaches on missing multivariate agronomy data: multiple imputation using regularised PCA through bootstrap procedure (BootMI-REM-PCA) and multiple imputation using regularised PCA through Bayesian procedure (BayesMI-REM-PCA). The data were obtained from the Department of Agriculture Sarawak. A simulation study was conducted using 500 simulated datasets at 5%, 10%, and 20% missingness. Results showed comparable performance between BootMI-REM-PCA and BayesMI-REM-PCA at 5% missingness, with equal coefficient of determination (R²) values of 0.998, while BootMI-REMPCA exhibited slightly lower root mean squared error (RMSE) of 1.527 and mean absolute error (MAE) of 0.160. However, BayesMI-REM-PCA outperformed at higher missing rates, achieving the lowest RMSE (2.238 at 10% and 3.051 at 20%) and MAE (0.315 at 10% and 0.601 at 20%), along with the highest R² values of 0.996 and 0.993, respectively. While imputation accuracy declines as missing data increases, BayesMI-REM-PCA preserves the characteristics of real data. The findings are expected to help agricultural scientists and researchers prepare high-quality data for accurate analysis.