On page 323, shortly after equation (22), you mention that the estimates of both total variance and its related covariance term are set equal to zero if the calculation would otherwise yield a negative variance estimate.
I am not sure of the extent to which this is a sensible question, but have you had a chance to explore the sensitivity of your results to this practical detail, i.e. does there exist some alternate weighting scheme (or schemes) that leaves you with a nonzero (positive) variance estimate? More generally, does something vaguely similar to the Newey-West standard error correction (that I am used to seeing in a time series context) exist for this type of estimation problem?
First, thank you for presenting!
On page 323, shortly after equation (22), you mention that the estimates of both total variance and its related covariance term are set equal to zero if the calculation would otherwise yield a negative variance estimate.
I am not sure of the extent to which this is a sensible question, but have you had a chance to explore the sensitivity of your results to this practical detail, i.e. does there exist some alternate weighting scheme (or schemes) that leaves you with a nonzero (positive) variance estimate? More generally, does something vaguely similar to the Newey-West standard error correction (that I am used to seeing in a time series context) exist for this type of estimation problem?