Search

Let $F=F_N$ be the distribution of a finite real population of size $N$. Let $\widehat{F}=F_N$ be the empirical distribution of a sample of size $n$ drawn from the population without replacement. We prove the following remarkable {\it inversion principle} for obtaining unbiased estimates. Let $ T \left(F_N\right)$ be any product of the moments or cumulants of $F_N$. Let $T_{n, N} \left( F_N \right) = E T \left( F_n \right)$. Then $E T_{N, n} \left( F_n \right) = T \left( F_N \right)$. We also obtain an explicit expression for $T_{n, N} \left(F_N\right)$ for all $ T \left( F_N \right)$ of order up to 6. We also prove the following related result. If $F_n$ and $F_N$ are the sample and population distributions, the only functionals for which $E T \left( F_n \right) = \lambda_{n, N} T \left( F_N \right)$ are noncentral moments, and generalized second and third order central moments. For these three cases the eigenvalues are $\lambda_{n, N}=1$, $\left( 1 - n^{-1} \right) \left( 1 - N^{-1} \right)^{-1}$, and $\left( 1 - n^{-1} \right) \left( 1 - 2n^{-1} \right) \left( 1 - N^{-1} \right)^{-1} \left( 1 - 2N^{-1} \right)^{-1}$ respectively.

We give the cumulative distribution function of $M_n$, the maximum of a sequence of $n$ observations from an ARMA(1, 1) process. Solutions are first given in terms of repeated integrals and then for the case, where the underlying random variables are absolutely continuous. The distribution of $M_n$ is then given as a weighted sum of the $n$th powers of the eigenvalues of a non-symmetric Fredholm kernel. The weights are given in terms of the left and right eigenfunctions of the kernel. These results are large deviations expansions for estimates, since the maximum need not be standardized to have a limit. In fact, such a limit need not exist., Comment: arXiv admin note: text overlap with arXiv:1001.5265

The chain rule for derivatives of a function of a function is extended to a function of a statistical functional, and applied to obtain approximations to the cumulants, distribution and quantiles of functions of sample moments, and so to obtain third order confidence intervals and estimates of reduced bias for functions of moments. As an example we give the distribution of the standardized skewness for a normal sample to magnitude $O(n^{-2})$, where $n$ is the sample size.

We give expansions for the distribution, density, and quantiles of an estimate, building on results of Cornish, Fisher, Hill, Davis and the authors. The estimate is assumed to be non-lattice with the standard expansions for its cumulants. By expanding about a skew variable with matched skewness, one can drastically reduce the number of terms needed for a given level of accuracy. The building blocks generalize the Hermite polynomials. We demonstrate with expansions about the gamma.

A great deal of inference in statistics is based on making the approximation that a statistic is normally distributed. The error in doing so is generally $O(n^{-1/2})$ and can be very considerable when the distribution is heavily biased or skew. This note shows how one may reduce this error to $O(n^{-(j+1)/2})$, where $j$ is a given integer. The case considered is when the statistic is the mean of the sample values from a continuous one-parameter distribution, after the sample has undergone an initial transformation.

We consider the problem of estimating an arbitrary smooth functional of $k \geq 1 $ distribution functions (d.f.s.) in terms of random samples from them. The natural estimate replaces the d.f.s by their empirical d.f.s. Its bias is generally $\sim n^{-1}$, where $n$ is the minimum sample size, with a {\it $p$th order} iterative estimate of bias $ \sim n^{-p}$ for any $p$. For $p \leq 4$, we give an explicit estimate in terms of the first $2p - 2$ von Mises derivatives of the functional evaluated at the empirical d.f.s. These may be used to obtain {\it unbiased} estimates, where these exist and are of known form in terms of the sample sizes; our form for such unbiased estimates is much simpler than that obtained using polykays and tables of the symmetric functions. Examples include functions of a mean vector (such as the ratio of two means and the inverse of a mean), standard deviation, correlation, return times and exceedances. These $p$th order estimates require only $\sim n $ calculations. This is in sharp contrast with computationally intensive bias reduction methods such as the $p$th order bootstrap and jackknife, which require $\sim n^p $ calculations.

This paper extends Edgeworth-Cornish-Fisher expansions for the distribution and quantiles of nonparametric estimates in two ways. Firstly it allows observations to have different distributions. Secondly it allows the observations to be weighted in a predetermined way. The use of weighted estimates has a long history including applications to regression, rank statistics and Bayes theory. However, asymptotic results have generally been only first order (the CLT and weak convergence). We give third order asymptotics for the distribution and percentiles of any smooth functional of a weighted empirical distribution, thus allowing a considerable increase in accuracy over earlier CLT results. Consider independent non-identically distributed ({\it non-iid}) observations $X_{1n}, ..., X_{nn}$ in $R^s$. Let $\hat{F}(x)$ be their {\it weighted empirical distribution} with weights $w_{1n}, ..., w_{nn}$. We obtain cumulant expansions and hence Edgeworth-Cornish-Fisher expansions for $T(\hat{F})$ for any smooth functional $T(\cdot)$ by extending the concepts of von Mises derivatives to signed measures of total measure 1. As an example we give the cumulant coefficients needed for Edgeworth-Cornish-Fisher expansions to $O(n^{-3/2})$ for the sample variance when observations are non-iid.

We give the distribution function of $M_n$, the maximum of a sequence of $n$ observations from an autoregressive process of order 2. Solutions are first given in terms of repeated integrals and then for the case, where the underlying random variables are absolutely continuous. When the correlations are positive, P(M_n \leq x) =a_{n,x}, where a_{n,x}= \sum_{j=1}^\infty \beta_{jx} \nu_{jx}^{n} = O (\nu_{1x}^{n}), where $\{\nu_{jx}\}$ are the eigenvalues of a non-symmetric Fredholm kernel, and $\nu_{1x}$ is the eigenvalue of maximum magnitude. The weights $\beta_{jx}$ depend on the $j$th left and right eigenfunctions of the kernel. These results are large deviations expansions for estimates, since the maximum need not be standardized to have a limit. In fact such a limit need not exist., Comment: 8 pages This version removes an inappropriate note