Reassessing Weights in Survey Research in Education: Implications for Multilevel Modeling

• In Event: Survey Science in Education: Methods, Analysis, and Emerging Technologies

Abstract

Sampling weights are essential in survey research using probability sampling, as they adjust for unequal selection probabilities and unit nonresponse. This is particularly important in educational surveys with complex multistage designs, such as students nested within schools. However, it remains unclear how best to incorporate weights in multilevel models applied to cross-national, cross-sectional data. We conducted a Monte Carlo simulation reflecting typical sampling and nonresponse adjustments to evaluate four strategies: (1) weights at both levels (with rescaled level-1 weights), (2) level-2 weights only, (3) level-1 weights only, and (4) no weights. We assessed bias and coverage of fixed and random effects. Results indicate that using weights at both levels (with rescaling) or only at level 2 yields the most accurate estimates.

Authors

• Umut Atasever, IEA Hamburg

• Francis H. Lim Huang, University of Missouri

• Leslie Rutkowski, Indiana University