You can use a model-based multiple-imputation strategy.You can also generate models incorporating the, for example, shift pattern and generate multiple imputations assuming that model. I'm sure, R have that capacity. I would think you could use multiple plausible models as a sensitivity check.