In the last years it has been consolidated the use of artificial neural nets as a complement to statistical methods. However, it has not been deeply studied neither how the presence of missing data affects artificial neuronal nets nor the establishment of the best strategies to treat missing data in the stage of statistical analysis. In our work we investigate the effectiveness of different techniques to face missing data in univariant descriptive analysis and in the generation of classification models, including multilayer perceptron and radial basis function neural nets. Our results suggest that, in general, artificial neural nets are more effective in decreasing the imputation error than other broadly used analysis techniques.
|Publication status||Published - 1 Aug 2000|