Characterize long-differencing data

Characterize the statistical properties of long-differencing data formed by subtracting one local time average from another, including how the choices of averaging-window length and endpoint dates affect its behavior in cross-section regressions.

Background

The paper studies time-compression procedures for non-stationary panel data, including point sampling and time averaging. Long differencing is defined as the difference between two local averages, each computed over a window of length M and centered at separate dates. Although long differencing is widely used in empirical work, the paper notes that its statistical properties are less transparent than those of the nonnegative weighting schemes analyzed formally under Assumption A.

The paper derives some variance expressions and discusses the tent-shaped weights that long differencing assigns to innovations, as well as the interaction between the low-pass filtering from local averaging and the high-pass filtering from differencing. However, it does not provide a general characterization of the resulting long-differenced data or fully establish the consequences of the tuning parameters for cross-section estimation. This unresolved issue motivates inclusion of the problem.