| Popis: |
To reduce bias in survey estimates, most longitudinal survey organisations, nowadays, prepare and include sets of weights in public use data files for use by analysts. Aside from correcting for non-coverage, the weights are usually designed to reflect the sample design as well as to correct for non-response error by combining design weights and non-response weight adjustments together. With regard to non-response weights, many longitudinal surveys implement similar strategies (referred to as the standard weighting approach in this thesis) to create them. This approach is based upon a weighting model where: response is defined as responding at all conducted waves; all sample members whose eligibility is unknown are assumed as eligible and the model is estimated by using generic weighting variables and all sample members for which data are available on the weighting variables. However, there are several issues in longitudinal surveys that raise concerns regarding using this approach of weighting. In particular, this thesis is concerned with three challenging issues: non-monotonic response pattern which results in a large number of combinations of waves at which sample members could respond, and hence weights that result from an approach such as the one in question, which defines response as responding at all the conducted waves may not be appropriate for the analysis of data from a wave-combination that does not include all waves; unknown eligibility over time leads to including a proportion of ineligible units in the weights' calculation (if they are assumed to be eligible as in the standard approach) which may result in biased estimates unless the actual ineligible units amongst units of unknown eligibility are excluded; and the choice of the best covariates for the weighting model which may differ considerably across different subgroups of respondents in the same sample. In the standard approach only generic weighting variables are used in the weighting model, as all sample members are used in the estimation. Meanwhile, some variables, which may not be significant in predicting response for the whole sample, could be important in predicting the response in some subgroups. In this thesis, I provide three alternative approaches (each deals with one of the raised issues) for non-response weighting. I investigate each of the proposed approaches by incorporating relevant weight adjustments, as well as weights from the standard weighting approach, in a longitudinal multivariate analysis. I test the impact of weights from each alternative approach on estimates by comparing the resultant estimates with estimates resulting from the standard approach. I use data from the British Household Panel Survey (BHPS) to carry out the investigation. The findings suggest that the standard and alternative approaches, all help similarly in reducing non-response error. However, the standard approach may fail in tackling the effect of non-response in some estimates, as it does not take into account the three raised issues in the weighting of longitudinal data. In contrast, since they deal with the three issues under investigation (separately), the alternative approaches seem to handle non-response even in estimates that are not affected by the standard weighting approach. |
