Imputación de datos perdidos en series de mediciones de contaminantes atmosféricos insumo para la vigilancia en salud ambiental

During the operation of air quality monitoring stations, amounts of missing data can be found in the pollutant's concentrations time series, which can reach up to 10% of the expected records. To explore the association between atmospheric pollutants and variations in the frequency of occurrence of health events, it is convenient to have complete atmospheric pollutants data, in order to avoid introducing biases into the results. A useful strategy for obtaining complete time series is the application of missing data imputation methods. In this work, we compare the performance of two imputation techniques in particulate matter data: Multiple Imputation and Neural Networks. Data from two stations of the Air Quality Monitoring Network of the Aburrá Valley, Antioquia were used. In general, the applied techniques presented a good fit, where the results observed when applying each method in situations of presence or absence of meteorological data (or predictor variables) were differential. In situations where there is no access to predictor variables and there are percentages of missing data greater than 25%, the Neural Networks technique showed a better performance. The presence of missing values in the predictor variables, used in the Multiple Imputation method, limited the technique's performance.