Uma aplicação de métodos de imputação no estudo de fatores associados ao baixo peso ao nascer

Abstract

Non-response in large databases is a common problem in most, cross-sectional and longitudinal studies. This phenomenon results in the presence of missing data which, depending on their proportion and nature, may affect the estimation of the parameters of interest in the study and lead to wrong conclusions. There are proper techniques to handle missing data, such as imputation. Proposed by Rubin (1987), imputation is a methodology based upon the substitution of plausible values for missing data, populating the database and allowing the researcher to use the traditional statistical methods for complete data. Our aim is to study different methods of imputation and to evaluate the effect of ignoring the existence of missing data in the analysis of a database of real data from the Brazilian 2006 National Demographic and Health Survey (PNDS). Moreover, a secondary objective is to evaluate factors associated with low birth weight, from this database. Low birth weight is an important indicator of health, and a risk factor for infant morbidity and mortality. The World Health Organization defines as low birth weight, less than the 2,500 kg weights. Logistic Regression models were proposed, considering maternal, gestational and newborn variables as predictors and models to both imputed and incomplete bases were compared. Our results confirm the importance of proper consideration of the problem of non-response. We hope that the study will assist researchers to be more aware of the need for adequate treatment to missing data before the phases of analysis and conclusion.