Different Approaches for Non-Response Adjustments to Statistical Weights in the continuous NHANES (2003-04)

It is possible to use different methods to weight National Health and Nutrition Examination Survey (NHANES) data. Typically when weighting NHANES data we use a weighting class approach of predefined cells based on a limited number of variables. However, there are model based alternatives that allow for more flexibility in calculating the statistical weights. This flexibility is important because additional variables in non-response adjustments may be needed to better reduce the bias arising from the fact that non-respondents may be different from those who responded to the survey. In this paper, using NHANES data from 2003-04, we illustrate alternative methods to calculate statistical weights using both a weighting class approach and a model based approach to account for unit non-response. The WTADJUST procedure in SUDAAN version 10.0 was used to calculate weights using different approaches. Different weighting approaches were assessed using examples with the environmental subsample data and the dietary recall data from NHANES. Summary statistics derived from the different methods were compared. In addition, the distributions of the weights were compared and the influence on variance inflation was assessed by comparing the estimated design effects due to the differential weighting. We found that the comparison of the different weighting approaches varied by analysis variables with subsamples. Therefore, it is important to assess different weighting approaches for specific subsamples and specific variables when analyzing data subject to differential non-response.