Gravar-mail: Correcting for non-participation bias in health surveys using record-linkage, synthetic observations and pattern mixture modelling