Normal approximation and smoothness for sums of means of lattice-valued random variables (Q373527)

Let \( \hat{\theta}\) be a given statistic for which a central limit theorem applies. To obtain estimates for the difference between the exact distribution of \(T=(\hat{\theta} - E[ \hat{\theta}])/\sqrt{\mathrm{Var} ( \hat{\theta})}\) and the standard normal distribution \(\Phi (x)\), several expansions for the distribution of \(T\) are available. An important class of these expansions is given by Edgeworth expansions, which are expansions of the form \[ P (T \leq x) = \Phi (x) + \sum_{j=0}^{r} \frac{p_j(x)}{n^{j/2}} \, \phi (x) + O(n^{-(r+1)/2}), \qquad r \geq 0, \] where \(p_0 (x) \equiv 0\), \(\phi (x)\) is the derivative of \(\Phi (x)\), and for \(j \geq 1\), \(p_j(x)\) are polynomials whose coefficients depend on the cumulants of \(\hat{\theta} - E[ \hat{\theta}]\). For examples, see [\textit{P. Hall}, The bootstrap and Edgeworth expansion. Springer Series in Statistics. New York etc.: Springer-Verlag. (1992; Zbl 0744.62026)], and [\textit{X. H. Zhou, C. M. Li} and \textit{Z. Yang}, ``Improving interval estimation of binomial proportions'', Phil. Trans. Roy. Soc. Ser. A 366, 2405--2418 (2001)]. In this article the authors investigate the first order Edgeworth expansions of sums of independent means of independent lattice-valued random variables. Sums or differences of binomial proportions are special cases of the problem under investigation. Let \(\{X_{j1}, X_{j2}, \dots , X_{jn_j}\}\), \(j=1,2,\dots , k\), \(k\geq 2\), be \(k\) independent samples of independent lattice-valued random variables, with \(E[|X_{j1}|^3]< +\infty\). Put \(\mu_j = E(X_{j1})\), \(\sigma_j^2 = \mathrm{Var} (X_{j1})\), \(\overline{X}_j = n_j^{-1} \sum_i X_{ji}\), and \(S = \sum_{j=1}^k \overline{X}_j \). Under these assumptions one would expect \(S\) to have a first order Edgeworth expansion of the form \[ P\left(\frac{S - E(S)}{\sqrt{\mathrm{Var} (S)}}\leq x \right) = \Phi (x) + \frac{\beta (1-x^2) \phi (x) }{6 \sqrt{n}} + \frac{d_n(x) \phi (x)}{\sqrt{n}} + o(n^{-1/2}), \] where \(n = n_1 + \dots + n_k\), \[ \beta = \beta (n) = \frac{\sqrt{n} \, E[(S - E(S))^3]}{{(\mathrm{Var} (S))^{3/2}}}, \] and \(d_n\) is a discontinuous term in general needed when dealing with lattice distributions see [\textit{C.-G. Esseen}, Acta Math. 77, 1--125 (1945; Zbl 0060.28705)]. The terms \(d_n\) are often referred to as continuity corrections. The authors investigate the distribution of S, and describe a methodology and conditions under which continuity corrections are not needed for this multi-sample problem. Specifically, suppose that the sample sizes \(n_1, n_2, \dots , n_k\) are changing in such a way that the correspondent sequence of values of \(n\) is strictly increasing, and that \[ \min_{1 \leq j \leq k} \liminf_{n\rightarrow +\infty} \frac{n_j}{n} > 0. \] Let \(e_j\) denote the span of the distribution of \(X_{j1}\), and for every \(1 \leq j_1 < j_2 \leq k\) put \(\rho_{j_1j_2} = (e_{j_2}n_{j_1})/ (e_{j_1}n_{j_2})\). The authors prove that if for at least one of the \(\rho_{j_1j_2}\), \[ \lim_{n\rightarrow +\infty} \sqrt{n} \, | \sin (l \rho_{j_1j_2} \pi)| = +\infty \qquad \text{ for every positive integer \(l\), } \] then \[ P\left(\frac{S - E(S)}{\sqrt{\mathrm{Var} (S)}}\leq x \right) = \Phi (x) + \frac{\beta (1-x^2) \phi (x) }{6 \sqrt{n}} + o(n^{-1/2}) \] holds uniformly in \(x\). The authors also give conditions under which a continuity correction \(d_n\) is needed, and for the case \(k=2\), \(d_n\) is derived. Extensions to problems where distributions are estimated using the bootstrap are also given.