Imputação múltipla e análise de casos completos no contexto da saúde pública: uma avaliação prática do impacto das perdas nas análises

Abstract

Researchers in the health field often deal with the problem of incomplete databases. Complete Case Analysis (CCA), which restricts the analysis to subjects with complete data, reduces the sample size and may result in biased estimates. Based on statistical grounds, the Multiple Imputation (MI) method uses all collected data and is recommended as an alternative to CCA. Data from the study Saúde em Beagá, attended by 4048 adults from two of nine health districts in the city of Belo Horizonte in 2008-2009, were used to evaluate CCA and different MI approaches in the context of logistic models with incomplete covariates. The Body Mass Index (BMI), an indicator of high public health relevance, was obtained through self-reported measures, and subsequently by direct measurements of height and weight of participants. The self-reported measures showed high percentage of missing data, which have spread to BMI based on them. However the minimum losses in direct measurements allowed BMI calculation for virtually entire sample. Given this peculiarity of the study, a hypothetical situation in which the missing data for BMI based on self-reported measures are recovered could be approached by a simple procedure. The methods of ACC and different approaches to IM were applied in a context where the BMI, with missing data, is one of the covariates in a logistic regression. The results of these methods were then compared with the results after the missing data recovery. It was found that even the more simplistic MI approach performed better than CCA since it was closer to the post-recovery results.