Comparison of Data Imputation Techniques and their Impact

Missing and incomplete information in surveys or databases can be imputed using different statistical and soft-computing techniques. This paper comprehensively compares auto-associative neural networks (NN), neuro-fuzzy (NF) systems and the hybrid combinations the above methods with hot-deck imputation. The tests are conducted on an eight category antenatal survey and also under principal component analysis (PCA) conditions. The neural network outperforms the neuro-fuzzy system for all tests by an average of 5.8%, while the hybrid method is on average 15.9% more accurate yet 50% less computationally efficient than the NN or NF systems acting alone. The global impact assessment of the imputed data is performed by several statistical tests. It is found that although the imputed accuracy is high, the global effect of the imputed data causes the PCA inter-relationships between the dataset to become altered. The standard deviation of the imputed dataset is on average 36.7% lower than the actual dataset which may cause an incorrect interpretation of the results.

💡 Research Summary

The paper conducts a thorough comparative study of three missing‑data imputation strategies—auto‑associative neural networks (NN), neuro‑fuzzy (NF) systems, and hybrid combinations of these methods with hot‑deck imputation—using an eight‑category antenatal survey dataset. Both the raw data and a version transformed by principal component analysis (PCA) are used to assess how dimensionality reduction interacts with imputation performance. Evaluation metrics include mean absolute error, mean squared error, and overall classification accuracy. Results show that the NN consistently outperforms the NF system, achieving an average accuracy gain of 5.8 %. The hybrid approach, which augments NN or NF with hot‑deck matching, yields the highest accuracy, improving results by roughly 15.9 % relative to the standalone methods. However, this gain comes at a substantial computational cost: the hybrid models require about 50 % more processing time and memory, making them less suitable for real‑time or large‑scale applications without further optimization. Beyond point‑wise accuracy, the authors examine the global statistical impact of imputation. They find that the standard deviation of the imputed dataset is on average 36.7 % lower than that of the original data, indicating a marked compression of variability. PCA performed on the imputed data reveals altered eigenvalue structures and shifted loadings, meaning that inter‑variable relationships are distorted after imputation. Consequently, downstream analyses that rely on the original covariance structure—such as regression, clustering, or factor analysis—may produce misleading inferences if the altered variance is not accounted for. The study concludes that while advanced soft‑computing techniques can deliver high imputation accuracy, practitioners must also evaluate post‑imputation statistical properties and computational feasibility. The authors recommend integrating variance‑preserving checks, possibly through multiple‑imputation frameworks, and suggest future work to explore non‑random missingness mechanisms (MAR, MNAR) and alternative dimensionality‑reduction methods (LDA, t‑SNE) to broaden the applicability of their findings.