Weighting

RTI statisticians have done considerable research on reducing survey biases from nonresponse and coverage errors. Our researchers deploy a variety of innovative techniques from logistic modeling of unit response propensity to our unique calibration procedures utilizing generalized exponential modeling (GEM), as well as simpler techniques such as weighting-class adjustments.

Our GEM-based procedures—now available in SUDAAN® as WTADJUST—involve a generalization of logistic-model fitting that can incorporate bounds on the weighting adjustment factors to limit variance inflation. The GEM/WTADJUST calibration routine provides a unified method of weight adjustment to compensate for unit nonresponse, coverage errors, and samples that are, for whatever reason, badly balanced on characteristics believed to be correlated with key survey outcomes. Extreme design weights can be truncated before weight adjustment, and the potential impact on bias due to the truncation can be mitigated by the adjustment process. The routine can also limit the creation of extreme weights during the weight adjustment process.

Focus Areas

  • Estimation of potential survey biases
  • Reduction of nonresponse and coverage biases

Methodologies/Techniques

  • Iterative proportional fitting (or raking) adjustments
  • Poststratification
  • GEM-based calibration, which includes raking and response propensity modeling as special cases
  • Truncation and smoothing of extreme weights
  • Weighting class adjustments
  • Propensity modeling
  • Adjustments for unknown eligibility

More Information