Scholarlink

<p>Missing data in climatological time series compromises statistical analysis, a critical challenge in high-variability tropical regions. This study has two linked objectives. First, it aims to identify the optimal multiple imputation method for long-term tropical climate data. Second, it applies Granger causality analysis to the optimally imputed series to examine predictive interdependencies among temperature, humidity, rainfall, and sunshine duration. The main contribution is an integrated framework that combines methodological validation with climatological inference. This study utilized a 15-year (2010–2024) daily time-series dataset (<i>N</i> = 21,872 observations) for temperature, humidity, rainfall, and sunshine duration from Banda Aceh, Indonesia. We systematically evaluated the performance of three MI.&#xa0;Expectation-Maximization (EM), Fully Conditional Specification (FCS), and Predictive Mean Matching (PMM) by simulating data loss (10%, 20%, 30%) and assessing imputation accuracy via RMSE, MAE, MSE, and WAPE. The superior method was then used to create a complete dataset for Granger causality analysis, following stationarity testing (ADF, Box-Cox) and AIC-based optimal lag selection. PMM was unequivocally the superior imputation method, consistently yielding the lowest error metrics across all tested levels of missingness. The findings show that PMM was the optimal imputation method because it preserved the original non-normal distribution by using observed donor values. In contrast, EM and FCS rely on distributional assumptions that may underestimate variance and distort extreme values in highly skewed tropical climate data. Using the completed dataset, the subsequent Granger causality analysis (optimal lag = 14) identified significant (<i>p</i> &lt; 0.05) bidirectional causal relationships among temperature, humidity, rainfall, and sunshine duration. A significant unidirectional causal relationship between sunshine duration and rainfall was also confirmed. These findings provide a reliable data framework for refining predictive weather models and supporting evidence-based climate action (SDG 13).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing multiple imputation methods for causal analysis of long-term tropical climate data

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Novi Reandy Sasmita,
  • Zatul Aklya,
  • Latifah Rahayu,
  • Feby Apriliansyah,
  • Fathin Nafisa

摘要

Missing data in climatological time series compromises statistical analysis, a critical challenge in high-variability tropical regions. This study has two linked objectives. First, it aims to identify the optimal multiple imputation method for long-term tropical climate data. Second, it applies Granger causality analysis to the optimally imputed series to examine predictive interdependencies among temperature, humidity, rainfall, and sunshine duration. The main contribution is an integrated framework that combines methodological validation with climatological inference. This study utilized a 15-year (2010–2024) daily time-series dataset (N = 21,872 observations) for temperature, humidity, rainfall, and sunshine duration from Banda Aceh, Indonesia. We systematically evaluated the performance of three MI. Expectation-Maximization (EM), Fully Conditional Specification (FCS), and Predictive Mean Matching (PMM) by simulating data loss (10%, 20%, 30%) and assessing imputation accuracy via RMSE, MAE, MSE, and WAPE. The superior method was then used to create a complete dataset for Granger causality analysis, following stationarity testing (ADF, Box-Cox) and AIC-based optimal lag selection. PMM was unequivocally the superior imputation method, consistently yielding the lowest error metrics across all tested levels of missingness. The findings show that PMM was the optimal imputation method because it preserved the original non-normal distribution by using observed donor values. In contrast, EM and FCS rely on distributional assumptions that may underestimate variance and distort extreme values in highly skewed tropical climate data. Using the completed dataset, the subsequent Granger causality analysis (optimal lag = 14) identified significant (p < 0.05) bidirectional causal relationships among temperature, humidity, rainfall, and sunshine duration. A significant unidirectional causal relationship between sunshine duration and rainfall was also confirmed. These findings provide a reliable data framework for refining predictive weather models and supporting evidence-based climate action (SDG 13).