Empirical likelihood confidence intervals for linear regression coefficients (Q1323843)

Empirical likelihood is a nonparametric technique for constructing confidence regions. It has sampling properties similar to those of bootstrap. However, instead of putting equal probability weight \(n^{- 1}\) on each data value, empirical likelihood chooses the weights by profiling a multinomial likelihood supported on the sample. However, it is not enough to just construct confidence regions for the vector unknown regression coefficients \(\beta\). In practice, statisticians are often confronted with problems of constructing confidence intervals for a particular regression coefficient or certain linear combinations of \(\beta\). We address the above problem under the simple linear regression model. A simple linear regression model is \[ y_ i - a_ 0 + b_ 0 x_ i + \varepsilon_ i, \quad 1 \leq i \leq n,\tag{1.1} \] where all the variables appearing in (1.1) are scalars. Among them, \(x_ i\) and \(y_ i\) are the \(i \)th fixed design point and response, respectively, the \(\varepsilon_ i\)'s are independent and identically distributed random errors with mean zero and variance \(\sigma^ 2\), and \(a_ 0\) and \(b_ 0\) are the unknown intercept and slope parameters, respectively. There are two aims in this paper. First, we show how to construct empirical likelihood confidence intervals for the slope parameter \(b_ 0\) and means \(y_ 0 = a_ 0 + b_ 0 x_ 0\) for any fixed \(x_ 0\), under model (1.1). Obviously the latter case includes the intercept parameter \(a_ 0\) when one chooses \(x_ 0 = 0\). Second, we study the coverage accuracy and Bartlett correctability of empirical likelihood confidence intervals for these parameters. Analyses show that both empirical likelihood confidence intervals for \(b_ 0\) and \(y_ 0\) have coverage errors of order \(n^{-1}\), and that both confidence intervals are Bartlett correctable. Thus, simple scale adjustments can improve the coverage accuracy of those confidence intervals from order \(n^{-1}\) to order \(n^{-2}\). A simulation study is presented.