具雜訊與缺失之函數型共變數的動態面板模型：平滑廣義動差估計

Abstract

本論文針對共變數有缺失觀測值與量測誤差的動態面板資料模型，提出新的動差重構估計量。此估計量將母體層級的曲面估計直接代入系統GMM動差矩陣，避免逐一個體的軌跡重構。我們透過數值模擬與實證資料，比較既有的pgmm（系統 GMM）以及naive（Nadaraya–Watson）、fpca（PACE）兩類插補式估計量與本文提出的估計量。數值模擬顯示在缺失與量測誤差並存的情境下，新的 動差重構估計量在點估計的表現優於既有估計量。實證應用方面，我們運用此方法以Booking.com資料集探討評論主題集中度與飯店平均評分之間的關聯。

This thesis proposes a new moment-reconstruction estimator for dynamic panel data models when the time-varying covariate has missing observations and measurement error. The estimator plugs population-level surface estimates directly into the system GMM moment matrices, thereby avoiding per-subject trajectory reconstruction. Through both numerical simulation and empirical data, we compare the proposed estimator against the existing pgmm (system GMM) implementation and two imputation-based estimators, naive (Nadaraya--Watson) and fpca (PACE). The simulation shows that the proposed estimator attains more accurate point estimates than the existing methods when missing observations and measurement error are present. In the real-data application, we employ the proposed method to investigate the association between review-topic concentration and average hotel ratings using the Booking.com dataset.