Paediatrics (RCH) - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemAn exploration of multiple imputation strategies for handling missing data in composite scores with incomplete items( 2016)Missing data are common in medical research. One area where missing data can arise is in composite scores (or scale scores) when one or more of the items that form the scale is incomplete. A method that is becoming increasingly popular for handling missing data is multiple imputation (MI). In the context of missing data in scale scores, MI can be applied at either the item level or the scale level. Various strategies have been proposed in the literature for imputing missing data at the scale level and the item level. Yet there is little comparison of these strategies in longitudinal settings and not much guidance is available about how to best implement these strategies. The challenge with using the available strategies in longitudinal studies is that one may want to impute missing data in several scales, each of which comprises a large number of items that have been measured at several waves, leading to large imputation models which may result in convergence problems. It is therefore important to evaluate the performance of these strategies in longitudinal settings to provide proper guidance for users of MI. In this thesis, I used a simulation study and a real example from the Longitudinal Study of Australian Children (LSAC) to compare the performance of the four MI strategies that are available for handling missing data in composite scores within a longitudinal setting. These strategies are: scale-level imputation using scale scores as auxiliary variables; the “standard” item-level imputation, which uses other items as auxiliary variables; item-level imputation using scale scores as auxiliary variables; and item-level imputation using principal components, derived from other items, as auxiliary variables. I also compared the effect of implementing these strategies using two MI approaches, multivariate normal imputation (MVNI) and fully conditional specification (FCS). While the literature recommends item-level imputation over scale level imputation, the research in this thesis demonstrates that when implemented using FCS, item-level imputation, with items from other scales as auxiliary variables, could produce biased parameter estimates. This research also provides support for using scales scores or principal components as auxiliary variables in item-level imputation models when the “standard” item-level imputation strategy cannot be used due to convergence problems.