Abstract:

 In the process of data collection, data are often missing due to the influence of various factors such as data privacy, human negligence, process defects, etc., which can increase the complexity of the analysis, resulting in systematic bias or even erroneous statistical inference. The treatment of missing data is therefore essential. The intrinsic mechanisms of the system dictate that most real data distributions exhibit skewed distributions rather than strictly normal distributions. In this paper, for random‑missing Skew‑Normal spatial data, the effects of sample size and missing rate variation on the accuracy of six interpolation methods, namely, random interpolation, mean value interpolation, support vector machine, decision tree, random forest, and BP neural network, are explored and quantitatively evaluated using MSE and MAE. For the interpolated dataset, a Skew‑Normal Spatial Autoregressive Model is built to estimate the Bayesian parameters using the MCMC algorithm. The results of simulation studies and example analysis show that support vector machines and random forests have better fitting ability and robustness to missing spatial skewed data, and the resulting Bayesian parameter estimates are more accurate.

Key words: spatial autoregressive modeling, partial normality, missing data, machine learning