本論文的研究主軸圍繞於自我迴歸模型的估計與推論上。文獻上自我迴歸模型的估計多直接採用最小平方法, 但此估計方式卻有兩個缺點:(一)當序列具單根時,最小平方估計式的漸近分配為非正規型態,因此檢定時需透過電腦模擬得到臨界值;(二)最小平方估計式雖具一致性,但卻有嚴重的有限樣本偏誤問題。有鑑於此,我們提出一種「二階差分轉換估計式」,並證明該估計式的偏誤遠低於前述最小平方估計式,且在序列為粧定與具單根的環境下具有相同的漸近常態分配。此外,二階差分轉換估計式相當適合應用於固定效果追蹤資料模型,而據以形成的追蹤資料單根檢定在序列較短的情況下仍有不錯的檢定力。 本論文共分四章,茲分別簡單說明如下: 第1章為緒論,回顧文獻上估計與推論自我回歸模型時的問題,並說明本論文的研究目標。估計自我迴歸模型的傳統方式是直接採取最小平方法,但在序列具單根的情況下由於訊息不隨時間消逝而快速累積,使估計式的收斂速度高於序列為恒定的情況。不過,這也導致最小平方估計式的漸近分配為非標準型態,並使得進行假設檢定前必須先透過電腦模擬來獲得臨界值。其次,最小平方估計式雖具一致性,但在有限樣本下卻是偏誤的。實證上, 樣本點不多是研究者時常面臨的窘境,並使得小樣本偏誤程度格外嚴重。本章中透過對前述問題形成因素的瞭解,說明解決與改善的方法,亦即我們提出的「二階差分轉換估計式」。 第2章主要目的在於推導二階差分轉換估計式之有限樣本偏誤。我們亦推導了多階差分自我迴歸模型下二階段最小平方估計式(two stage least squares, 2SLS)與 Phillips andHan (2008)採用的一階差分轉換估計式之偏誤,以同時進行比較。本章理論與模擬結果皆顯示,一階與二階差分轉換估許式與2SLS之 $T^{−1}$ 階偏誤程度皆低於以最小平方法估計原始準模型(level model)的偏誤,其中 T 為時間序列長度。另外,一階差分轉換估計式與二階差分轉換估計式在 $T^{−1}$ 階偏誤上,分別與一階和二階差分模型下2SLS相同,但兩估計式的相對偏誤程度則因自我相關係數的大小而互有優劣。同時,我們發現估計高於二階的差分模型對小樣本偏誤並無法有更進一步的改善。最後,即使在樣本點不多的情況下,本章所推導的偏誤理論對於實際偏誤仍有良好的近似能力。 第3章主要目的在於發展二階差分轉換估計式之漸近理論。與 Phillips and Han (2008) 採用之一階差分轉換估計式相似的是,該估計式在序列為恒定與具單根的情況下收斂速度相同,並有漸近常態分配的優點。值得注意的是, 二階差分轉換估計式的漸近分配為 N(0,2),不受任何未知參數的影響。另外,當序列呈現正自我相關時,二階差分轉換估計式相較於一階差分轉換估計式具有較小的漸近變異數,進而使得據以形成的檢定統計量有較佳的對立假設偵測能力。最後, 誠如 Phillips and Han (2008) 所述,由於差分過程消除了模型中的截距項,使得此類估計方法在固定效果的動態追蹤資料模型(dynamic panel data model with fixed effect) 具相當的發展與應用價值。 本論文第4 章進一步將二階差分轉換估計式推展至固定效果的動態追蹤資料模型。文獻上估計此種模型通常利用差分來消除固定效果後,再以一般動差法 (generalized method of moments, GMM) 進行估計。然而,這樣的估計方式在序列為近單根或具單根時卻面臨了弱工具變數(weak instrument)的問題,並導致嚴重的估計偏誤。相反的,差分轉換估計式所利用的動差條件在近單根與單根的情況下仍然穩固,因此在小樣本下的估計偏誤相當輕微(甚至無偏誤)。另外,我們證明了不論序列長度(T )或橫斷面規模(n)趨近無窮大,差分轉換估計式皆有漸近常態分配的性質。與單一序列時相同的是,我們提出的二階差分轉換估計式在序列具正自我相關性時的漸近變異數較一階差分轉換估計式小;受惠於此,利用二階差分轉換估計式所建構的檢定具有較佳的檢力。值得注意的是,由於二階差分轉換估計式在單根的情況下仍有漸近常態分配的性質,我們得以直接利用該漸近理論建構追蹤資料單根檢定。電腦模擬結果發現,在小 T 大 n 的情況下,其檢力優於文獻上常用的 IPS 檢定(Im et al., 1997, 2003)。 / This thesis deals with estimation and inference in autoregressive models. Conventionally, the autoregressive models estimated by the least squares (LS) procedure may be subject to two shortcomings. First, the asymptotic distribution of the LS estimates for autoregressive coefficient is discontinuous at unity. Test statistics based on the LS estimates thus follow nonstandard distributions, and the critical values obtained need to rely on Monte Carlo techniques. Secondly, as is well known, the LS estimates of autoregressive models are biased in finite samples. This bias could be substantial and leads to serious size distortion for the test statistics built on the estimates and forecast errors. In this thesis,we consider a simple newmethod ofmoments estimator, termed the “transformed second-difference” (hereafter TSD) estimator, that is without the aforementioned problems, and has many useful applications. Notably, when applied to dynamic panel models, the associated panel unit root tests shares a great power advantage over the existing ones, for the cases with very short time span. The thesis consists of 4 chapters, which are briefly described as follows. 1. Introduction: Overview and Purpose This chapter first reviews the literature and states the purpose of this dissertation. We discuss the sources of problems in estimating autoregressive models with the conventional method. The motivation to estimate the autoregressive series with multiple-difference models, instead of the conventional level model, is provided. We then propose a new estimator, the TSD estimator, which can avoid (fully or partly) the drawbacks of the LS method, and highlight its finite-sample and asymptotic properties. 2. The Bias of 2SLSs and transformed difference estimators in Multiple-Difference AR(1) Models In this chapter, we derive approximate bias for the TSD estimator. For comparisons, the corresponding bias of the two stage least squares estimators (2SLS) in multiple-difference AR(1) models and the transformed first-difference (TFD) estimator proposed by Chowdhurry (1987) are also given as by-products. We find that: (i) All the estimators considered are much less biased than the LS ones with the level regression; (ii)The difference method can be exploited to reduce the bias only up to the order of difference 2; and (iii) The bias of the TFD and TSD estimators share the same order at $O(T^{-1})$ as that of 2SLSs. However, to the extent of bias reductions, neither the 2 considered transformed difference estimators shows a uniform dominance over the entire parameter space. Our simulation evidence lends credible supports to our bias approximation theory. 3. Gaussian Inference in AR(1) Time Series with or without a Unit Root The goal of the chapter is to develop an asymptotic theory of the TSD estimator. Similar to that of the TFD estimator shown by Phillips and Han (2008), the TSDestimator is found to have Gaussian asymptotics for all values of ρ ∈ (−1, 1] with $\sqrt{T}$ rate of convergence, where ρ is the autoregressive coefficient of interest and T is the time span. Specifically, the limit distribution of the TSD estimator is N(0,2) for all possible values of ρ. In addition, the asymptotic variance of the TSD estimator is smaller than that of the TFD estimator for the cases with ρ > 0, and the corresponding t -test thus exhibits superior power to the TFD-based one. 4. Estimation and Inference with Moment Methods for Dynamic Panels with Fixed Effects This chapter demonstrates the usefulness of the TSD estimator when applying to to dynamic panel datamodels. We find again that the TSD estimator displays a standard Gaussian limit, with a convergence rate of $\sqrt{nT}$ for all values of ρ, including unity, irrespective of how n or T approaches infinity. Particularly, the TSD estimator makes use of moment conditions that are strong for all values of ρ, and therefore can completely avoid the weak instrument problem for ρ in the vicinity of unity, and has virtually no finite sample bias. As in the time series case, the asymptotic variance of the TSD estimator is smaller than that of the TFD estimator of Han and Phillips (2009) when ρ > 0 and T > 3, and the corresponding t -ratio test is thus more capable of unveiling the true data generating process. Furthermore, the asymptotic theory can be applied directly to panel unit root test. Our simulation results reveal that the TSD-based unit root test is more powerful than the widely used IPS test (Im et al, 1997, 2003) when n is large and T is small.