Estimates from a complex survey

Abstract

OBJECTIVE: To evaluate the impact of sampling design and the effect of weighting on data from the 1996 Brazilian National Survey on Demography and Health. METHODS: Secondary data analysis was performed using a sample of 1,355 interviewed women of the state of São Paulo. The sampling design of the National Survey of Household Sampling (PNAD) was used as a reference, and the municipality as primary sampling unit. The ratio estimator and Taylor's aproximation for variance were calculated using the primary sampling units and several modalities of weighting. The indicators used to evaluate precision and validity were confidence intervals, design effects (Deff) and biases. RESULTS: For the four procedures, the differences between upper and lower point estimates for prevalence were not greater than 10%. The differences on ranges of confidence intervals were less than 20%. Use of condom and hormone injection were the variables that showed design effects greater than 1.5 and biases greater than 0.20. CONCLUSIONS: According to the results, it could be said that the cluster sampling had an impact on the precision of the estimates for two out of six variables. The impact of weighting was not significant.