Discussion of 'Statistical inference for streamed longitudinal data'.

This article explores the trade-off between computational efficiency and estimation efficiency in the statistical analysis of correlated streaming data. The authors compare the use of subsampling and ignoring correlation in the analysis of massive data and find that subsampling can be more effective for correlated data. They provide numerical illustrations to demonstrate the effectiveness of subsampling with correlated data streams. The article also discusses the implementation of ordinary least squares (OLS) and weighted least squares (WLS) estimators for data analysis, showing that there is some efficiency loss with the OLS estimator compared to the WLS estimator. The use of subsampling methods, such as uniform subsampling, can achieve good estimation results with a small fraction of the full data. The text also discusses the bias-variance trade-off in weighting methods for estimation and the convergence properties of a statistical estimator in the context of batch learning. The authors conclude by mentioning that the efficiency of their proposed methods might not be fully demonstrated in previous studies and suggest that adjusting the bias for inference is an open question. The work was supported by the National Science Foundation, and supplementary material with detailed derivations of equations is provided by the authors. [Extracted from the article]