All Times ET
Keywords: Imputation, Hierarchical time series, Missing data
When a total is the parameter of interest in an aggregation network, calculation of higher-level aggregate totals is limited when any lower-level component of the aggregate is missing. In such cases, reporting of the higher-level aggregate requires an additional estimation step. This additional step may include: (1) estimating the whole of the higher-level aggregate independently of its aggregation components; (2) estimating the missing terminal nodes and then including these estimates in the aggregation scheme; or alternatively, (3) estimating higher-level-aggregate sums of the missing lower-level components as a partial-aggregate of the higher-level nodal total and then imputing the missing portion of the nodal total with that partial-aggregate.
A tradeoff between procedures 1 and 2 sometimes exists as higher-level-aggregate time-series are often more stable than their lower-level component series and may be modeled with less expected error, but modeling the whole of a higher-level aggregate exclusive of known aggregation components ignores valuable information about the statistic.
The partial-aggregate option is developed here: (1) for use in imputation of EIA’s International Energy Statistics when the higher-level partial-aggregate time-series is more stable than its component time-series set and some, but not all, components of the aggregate are expected to be known when the total estimate is reported; and (2), to assist analysts in examining the effects of missing data on various elements of the reporting network and identify cases where special attention may be warranted. Special software using functional and event-driven programming paradigms has been developed for visualization and process management and will be on display.