Valid Edgeworth Expansions of Some Estimators and Bootstrap Confidence Intervals in First‐order Autoregression (Q4939804)

For an AR1 process \(y_t=\vartheta y_{t-1}+\varepsilon_t\) where \(\varepsilon_i\) are i.i.d., estimates for the unknown \(\vartheta\), \(|\vartheta|<1\), are considered of the form \[ \hat\vartheta_{a,b}={\sum_{t=2}^n y_t y_{t-1}/( ay_1^2+\sum_{t=2}^{n-1}y_t^2 +by_n^2}),\quad a,b\geq 0. \] The least squares estimator is \(\hat\vartheta_{1,0}\), the Yule-Walker estimator is \(\hat\vartheta_{1,1}\) and the Daniels' estimator is \(\hat\vartheta_{1/2,1/2}\). In the case when \((\varepsilon_i,\varepsilon_i^2)\) satisfy Cramér's condition and \(E\varepsilon_1^{2(s+1)}<\infty\) for some \(s\geq 4\), third-order valid Edgeworth expansions are derived for the distributions of the standardized \(\hat\vartheta_{a,b}\), \[ T_1=[(n/(1-\vartheta^2)]^{1/2}(\hat\vartheta_{a,b}-\vartheta), \] and of the Studentized \(\hat\vartheta_{a,b}\), \[ T_2=[ n/(1-\hat\vartheta_{a,b})^2]^{1/2}(\hat\vartheta_{a,b}-\vartheta). \] These results are used for the construction of accelerated bias corrected bootstrap confidence intervals for \(\vartheta\).