Subsampling in Longitudinal Models
Subsampling in Longitudinal Models
复制标题
DOI：
10.1007/s11009-023-10015-4
复制
发表时间：
2023-02
影响因子：
0.9
通讯作者：
Ziyang Wang;Haiying Wang;N. Ravishanker
中科院分区：
文献类型：
--
作者：
Ziyang Wang;Haiying Wang;N. Ravishanker
对于大规模数据，通常使用子抽样方法来近似全数据参数估计。理想的二次抽样方法从完整数据中挑选一小部分信息观测值，并使用更少的计算能力产生准确的近似完整数据估计。现有的二次抽样方法的研究主要集中在独立响应上。本文讨论了纵向数据的子抽样方法，其中在一个块内的观察是相关的，并开发最佳子抽样方法来近似模型参数的全数据极大似然估计。本文首先在一般抽样概率下建立了子样估计量的条件渐近分布，然后导出了使子样估计量的渐近均方误差最小的最优抽样方法。为了评估所提出的方法的有限样本性能，我们提供了基于模拟数据的数值实验结果。
For large scale data, subsampling methods are often used to approximate the full-data parameter estimates. An ideal subsampling method picks a small proportion of informative observations from the full data and produces an accurate approximate to the full-data estimate using much less computing power. Existing studies on subsampling methods focus on independent responses. This paper discusses subsampling methods for longitudinal data where observations within a block are correlated, and develops optimal subsampling methods to approximate the full-data maximum likelihood estimators of the model parameters. We first establish the conditional asymptotic distribution of the subsample estimator with general subsampling probabilities, and then derive theoptimalsubsampling method that minimizes the asymptotic mean squared error of the subsample estimator. To evaluate the finite sample performance of the proposed method, we provide results based on numerical experiments with simulated data.