|
Abstract:
|
imputation of data with general structures (e.g., data with continuous, binary, categorical, and ordinal variables) is complicated by the fact that fully conditional specification (FCS) is often required. A key drawback of FCS is that it does not invoke an appropriate Gibbs sampling mechanism and as such convergence of the resulting Markov chain Monte Carlo procedure is not assured. Furthermore, methods that use joint modeling lack these drawbacks but have not been efficiently implemented in data of general structures. We address these issues by developing a new method that draws imputations from a latent joint multivariate normal model that underpins the generally structured data. This model is constructed using a sequence of flexible conditional linear models that enables the resulting procedure to be efficiently implemented on high dimensional datasets in practice. This method is applied to a RAND survey wherein traditional FCS procedures are shown to diverge across iterations, whereas the new method produces convergent imputations. Furthermore, the new method runs dramatically faster than FCS procedures when applied to the RAND data.
|
