Colleagues tell me that a fully-specified regression model doesnt need weighting are they right?
Strictly speaking, yes they are. But the specification of such an unweighted model is difficult to achieve in practice since the model must incorporate covariates that fully account for the sample design. The covariates must first account for the variations in selection and response probabilities. One can attempt to check this by comparing weighted and unweighted estimates produced by the model. If there is no variation in the model coefficients, the sampling biases have been accounted for through the specification of the model. In the employee survey, however, one must also account for the clustering of employee observations within workplaces. This may be done through the use of multilevel modelling procedures (also known as random-coefficients models) that take account of the hierarchical nature of the data. If these two criteria are met, standard SRSWR-based inference methods can be used, enabling the user to benefit from smaller standard errors that would arise under the methods de
