Missing values in multivariate time series data, often caused by network disruptions, device or power outages, and bad weather, can pose challenges for future analysis. Common statistical methods like multiple imputation and expectation-maximization are often used to impute missing time series data. However, these methods assume that the data is missing at random and may struggle with more complex missing data mechanisms and higher missing ratios. In these cases, advanced techniques like neural networks may offer improved imputation. This study assess the effectiveness of two recurrent neural network methods LSTM and GRU, enhanced with a time decay function, named LSTM-D and GRU-D, for analyzing missing multivariate time series. Their performance is compared with three well-known statistical methods: Bootstrapped-EM, EM-ARIMA, and MICE, across different missing data scenarios. Results indicate that LSTM-D and GRU-D perform better than traditional statistical methods for two different datasets, particularly when the missing data is not random.