Fisher information in record data

We present some general results on Fisher information (FI) contained in upper (or lower) record values and associated record times generated from a sequence of i.i.d. continuous variables. For the record data obtained from a random sample of fixed size, we establish an interesting relationship between its FI content and the FI in the data consisting of sequential maxima. We also consider the record data from an inverse sampling plan (Samaniego and Whitaker, 1986). We apply the general results to evaluate the FI in upper as well as lower records data from the exponential distribution for both sampling plans. Further, we discuss the implication of our results to statistical inference from these record data. Received: December 2001 Acknowledgements. This research was supported by Fondo Nacional de Desarrollo Cientifico y Tecnologico (FONDECYT) grants 7990089 and 1010222 of Chile. We would like to thank the Department of Statistics at the University of Concepción for its hospitality during the stay of H. N. Nagaraja in Chile in March of 2000, when the initial work was done. We are grateful to the two referees for various comments that let to improvements in the paper.     

An information theoretic argument for the validity of the exponential model

Based on the Cramér-Rao inequality (in the multiparameter case) the lower bound of Fisher information matrix is achieved if and only if the underlying distribution is ther-parameter exponential family. This family and the lower bound of Fisher information matrix are characterized when some constraints in the form of expected values of some statistics are available. If we combine the previous results we can find the class of parametric functions and the corresponding UMVU estimators via Cramér-Rao inequality.     

A note on estimation based on record data

In data-processing standpoint, an efficient algorithm for identifying the minimum value among a set of measurements are record statistics. From a sequence of n independent identically distributed continuous random variables only about log(n) records are expected, so we expect to have little data, hence any prior information is welcome (Houchens, Record value theory and inference, Ph.D. thesis, University of California, Riverside, 1984). In this paper, non-Bayesian and Bayesian estimates are derived for the two parameters of the Exponential distribution based on record statistics with respect to the squared error and Linear-Exponential loss functions and then compared with together. The admissibility of some estimators is discussed.     

Limit theorems for the spacings of weak records

Let W(1), W(2), . . . be weak record values obtained from a sample of independent variables with common discrete distribution. In the present paper, we derive weak and strong limit theorems for the spacings W(n + m) − W(n), m ≥ 1, n → ∞.     

Shannon information in record data

In this article we present results on the Shannon information (SI) contained in upper (lower) record values and associated record times in a sequence of i.i.d continuous variables. We then establish an interesting relationship between the SI content of a random sample of fixed size, and the SI in the data consisting of sequential maxima. We also consider the information contained in the record data from an inverse sampling plan (ISP).     

Verbesserung des nicht-randomisierten exakten Testes von R.A. Fisher

Summary The so-called Exact Test of R. A. Fisher for comparing two probabilitiesp ₁ andp ₂ in a Fourfold-Table with small cell frequencies is known as a UMPU-Test. But in practice the test is used in a nonrandomized, often tabulated version. Given a certain level of significanceα it is shown: the critical region of this nonrandomized test, referred to as “Fisher 1”, can be enlarged considerably. For instance for all sample-size-sums up to 20 andα=0.01 the total number of points in the critical regions of “Fisher 1” is 552 whereas the analogous number of the new version “Fisher 2” is 788. The size of tables for “Fisher 2” can be reduced considerably because the main parts of the critical regions can be described by the aid of some Chi-square-test versions. In particular Yates’ continuity-correction turns out to be always conservative in the above mentioned region relative to “Fisher 2” whereas this is not strictly true relative to “Fisher 1”.      

A top-down approach for approximate data anonymisation

Data sharing in today's information society poses a threat to individual privacy and organisational confidentiality. k-anonymity is a widely adopted model to prevent the owner of a record being re-identified. By generalising and/or suppressing certain portions of the released dataset, it guarantees that no records can be uniquely distinguished from at least other k?1 records. A key requirement for the k-anonymity problem is to minimise the information loss resulting from data modifications. This article proposes a top-down approach to solve this problem. It first considers each record as a vertex and the similarity between two records as the edge weight to construct a complete weighted graph. Then, an edge cutting algorithm is designed to divide the complete graph into multiple trees/components. The Large Components with size bigger than 2k?1 are subsequently split to guarantee that each resulting component has the vertex number between k and 2k?1. Finally, the generalisation operation is applied on the vertices in each component (i.e. equivalence class) to make sure all the records inside have identical quasi-identifier values. We prove that the proposed approach has polynomial running time and theoretical performance guarantee O(k). The empirical experiments show that our approach results in substantial improvements over the baseline heuristic algorithms, as well as the bottom-up approach with the same approximate bound O(k). Comparing to the baseline bottom-up O(logk)-approximation algorithm, when the required k is smaller than 50, the adopted top-down strategy makes our approach achieve similar performance in terms of information loss while spending much less computing time. It demonstrates that our approach would be a best choice for the k-anonymity problem when both the data utility and runtime need to be considered, especially when k is set to certain value smaller than 50 and the record set is big enough to make the runtime have to be taken into account.     

Price asymptotics

This paper examines the pricing decisions of a seller facing an unknown demand function. It is assumed that partial information, in the form of an independent random sample of values, is available. The optimal price for the inferred demand satisfies a consistency property—as the size of the sample increases, the maximum profit and price approach the values for the case where demand is known. The main results deduced here are asymptotics for prices. Prices converge at a rate of O _p (n ^−1/3) with a limit that can be expressed as a functional of a Gaussian process. Implications for the comparison of mechanisms are discussed.      

Competitive prizes: when less scrutiny induces more effort

We consider a principal who is keen to induce his agents to work at their maximal effort levels. To this end, he samples n days at random out of the T days on which they work, and awards a prize of B dollars to the most productive agent. The principal’s policy (B, n) induces a strategic game Γ(B, n) between the agents. We show that to implement maximal effort levels weakly (or, strongly) as a strategic equilibrium (or, as dominant strategies) in Γ(B, n), at the least cost B to himself, the principal must choose a small sample size n. Thus less scrutiny by the principal induces more effort from the agents.The need for reduced scrutiny becomes more pronounced when agents have information of the history of past plays in the game. There is an inverse relation between information and optimal sample size. As agents acquire more information (about each other), the principal, so to speak, must “undo” this by reducing his information (about them) and choosing the sample size n even smaller.     

Sulla somma di due numeri aleatori egualmente distribuiti

In questo lavoro si studia il problema di ricerca della distribuzione di probabilità comune da assegnare a due numeri aleatori discreti che assumano i primin valori interi naturali in modo che la loro somma abbia moda di minima probabilità.Il problema è affrontato sia dal punto di vista teorico tramite gli strumenti della programmazione matematica, sia dal punto di vista numerico.

---

LetX andY be two random numbers with the same distribution function; in this paper we consider the problem of finding a random numberX+Y having mode with minimal probability. In particular we have considered only the case ofX andY assuming the firstn integer values, so thatp (dimensionn) is the common distribution andq (dimension 2n–1) is the distribution ofX+Y; then the problem is to minimizem=maxq ₁.In the known literature it appears that theoretical results and numerical experience have brought to various conjectures not confermed. In this paper the problem is considered from the mathematical programming point of view. Several theoretical results are obtained even if the full solution of the problem is not reached. Anyway, such results, limiting the search range of a solution, suggested extended numerical testing, also for rather large values ofn, so that non trivial conclusions can be derived.

---

---

Pervenuto il 22-1-82     

Asymptotically efficient recursive estimation for incomplete data models using the observed information

For a recursive maximum-likelihood estimator with step lengths decaying as 1/n, an adaptive matrix needs to be incorporated to obtain asymptotic efficiency. Ideally, this matrix should be chosen as the inverse Fisher information matrix, which is usually very difficult to compute for incomplete data models. In this paper we give conditions under which the observed information can be incorporated into the recursive procedure to yield an efficient estimator, and we also investigate the finite sample properties of these estimators by simulation.     

Estimation of the scale parameter of a log-logistic distribution

In this work the ranked set sampling technique has been applied to estimate the scale parameter $\alpha $ of a log-logistic distribution under a situation where the units in a sample can be ordered by judgement method without any error. We have evaluated the Fisher information contained in the order statistics arising from this distribution and observed that median of a random sample contains the maximum information about the parameter $\alpha $ . Accordingly we have used median ranked set sampling to estimate $\alpha $ . We have further carried out the multistage median ranked set sampling to estimate $\alpha $ with improved precision. Suppose it is not possible to rank the units in a sample according to judgement method without error but the units can be ordered based on an auxiliary variable $Z$ such that $(X, Z)$ has a Morgenstern type bivariate log-logistic distribution (MTBLLD). In such a situation we have derived the Fisher information contained in the concomitant of rth order statistic of a random sample of size $n$ from MTBLLD and identified those concomitants among others which possess largest amount of Fisher information and defined an unbalanced ranked set sampling utilizing those units in the sample and thereby proposed an estimator of $\alpha $ using the measurements made on those units in this ranked set sample.     

Inducing relations on incomplete ordinal data

Let X denote a set of n elements or stimuli which have an underlying linear ranking L based on some attribute of comparison. Incomplete ordinal data is known about L, in the form of a partial order P on X. This study considers methods which attempt to induce, or reconstruct, L based only on the ordinal information in P. Two basic methods for approaching this problem are the cardinal and sequel construction methods. Exact values are computed for the expected error of weak order approximations of L from the cardinal and sequel construction methods. Results involving interval orders and semiorders for P are also considered. Previous simulation comparisons for cardinal and sequel construction methods on interval orders were found to depend on the specific model that was used to generate random interval orders, and were not found to hold for interval orders in general. Finally, we consider the likelihood that any particular linear extension of P is the underlying L.     

Understanding the Fisher equation

It is argued that univariate long memory estimates based on ex post data tend to underestimate the persistence of ex ante variables (and, hence, that of the ex post variables themselves) because of the presence of unanticipated shocks whose short‐run volatility masks the degree of long‐range dependence in the data. Empirical estimates of long‐range dependence in the Fisher equation are shown to manifest this problem and lead to an apparent imbalance in the memory characteristics of the variables in the Fisher equation. Evidence in support of this typical underestimation is provided by results obtained with inflation forecast survey data and by direct calculation of the finite sample biases. To address the problem of bias, the paper introduces a bivariate exact Whittle (BEW) estimator that explicitly allows for the presence of short memory noise in the data. The new procedure enhances the empirical capacity to separate low‐frequency behaviour from high‐frequency fluctuations, and it produces estimates of long‐range dependence that are much less biased when there is noise contaminated data. Empirical estimates from the BEW method suggest that the three Fisher variables are integrated of the same order, with memory parameter in the range (0.75, 1). Since the integration orders are balanced, the ex ante real rate has the same degree of persistence as expected inflation, thereby furnishing evidence against the existence of a (fractional) cointegrating relation among the Fisher variables and, correspondingly, showing little support for a long‐run form of Fisher hypothesis. Copyright © 2004 John Wiley & Sons, Ltd.     

Asymptotic theory and econometric practice

The classical paradigm of asymptotic theory employed in econometrics presumes that model dimensionality, p, is fixed as sample size, n, tends to inifinity. Is this a plausible meta-model of econometric model building? To investigate this question empirically, several meta-models of cross- sectional wage equation models are estimated and it is concluded that in the wage-equation literature at least that p increases with n roughly like n ^l/4, while that hypothesis of fixed model dimensionality of the classical asymptotic paradigm is decisively rejected. The recent theoretical literature on ‘large-p’ asymptotics is then very briefly surveyed, and it is argued that a new paradigm for asymptotic theory has already emerged which explicitly permits p to grow with n. These results offer some guidance to econometric model builders in assessing the validity of standard asymptotic confidence regions and test statistics, and may eventually yield useful correction factors to conventional test procedures when p is non-negligible relative to n.     

A note on reliability properties of k-record statistics

LetY _k,n denote the nth (upper) k-record value of an infinite sequence of independent, identically distributed random variables with common continuous distribution function F. We show that if the nth k-record valueY _k,n has an increasing failure rate (IFR), thenY _l,n (l<k) andY _k+1,n+1(n≤k+1) also have IFR distributions. On the other hand, ifY _k,n has a decreasing failure rate (DFR), thenY _l,n (1>k) has also a DFR distribution. We also present some results concerning log convexity and log concavity ofY _k,n .     

Developing a Reputation for Reticence

A sender who has disclosable information with probability less than one may partially conceal bad news by choosing to withhold information and pooling with uninformed types. The success of this strategy depends on receivers' beliefs about the probability that the sender has disclosable news. In a dynamic context, informed senders try to cultivate a reputation for reticence either by concealing good news along with the bad, or by concealing some good news and disclosing some bad news. A reputation for reticence is valuable because it makes receivers less skeptical of past or future nondisclosures. The model provides insight into the choice by firms such as Google not to disclose quarterly earnings guidance to analysts, as well as Tony Blair's reticence over his son's vaccine record during the measles–mumps–rubella scare in the United Kingdom.     

Choice of optimal initial designs in sequential experiments

Combined-optimal designs (Li and Lin, 2003) are obviously the best choices for the initial designs if we partition the experiment into two parts with equal size to obtain some information about the process, especially for the case not considering the blocking factor. In this paper, the definition of combined-optimal design is extended to the case when blocking factor is significant, and this new class of designs is called blocked combined-optimal designs. Some general results are obtained which relate 2^k−p_III initial designs with their complementary designs when , where n=2^k−p. By applying these results, we are able to characterize 2^k−p_III combined-optimal designs or blocked combined-optimal designs in terms of their complementary designs. It is also proved that both 2^k−p_III combined-optimal and blocked combined-optimal designs are not minimum aberration designs when and n−1−k > 2. And some combined-optimal and blocked combined-optimal designs with 16 and 32 runs are constructed for illustration. 2000 Mathematics Subject Classifications: 62K15, 62K05     

Slow convergence of the number of near-maxima for Burr XII distributions

A near-maximum is an observation which falls within a distance a of the maximum observation in an independent and identically distributed sample of size n. Subject to some conditions on the tail thickness of the population distribution, the number K _n (a) of near-maxima is known to converge in probability to one or infinity, or in distribution to a shifted geometric law. In this paper we show that for all Burr XII distributions K _n (a) converges almost surely to unity, but this convergence property may not become clear under certain cases even for very large n. We explore the reason of such slow convergence by studying a distributional continuity between Burr XII and Weibull distributions. We have also given a theoretical explanation of slow convergence of K _n (a) for the Burr XII distributions by showing that the rate of convergence in terms of P{K _n (a) > 1} tending to zero changes very little with the sample size n. Illustrations of the limiting behaviour K _n (a) for the Burr XII and the Weibull distributions are given by simulations and real data. The study also raises an important issue that although the Burr XII provides overall better fit to a given data set than the Weibull distribution, cautions should be taken for the extrapolation of the upper tail behaviour in the case of slow convergence.      

Optimal Design Approach to GMM Estimation of Parameters Based on Empirical Transforms

Parameter estimation based on the generalized method of moments (GMM) is proposed. The proposed method employs a distance between an empirical and the corresponding theoretical transform. Estimation by the empirical characteristic function (CF) is a typical example, but alternative empirical transforms are also employed, such as the empirical Laplace transform when dealing with non‐negative random variables. D‐optimal designs are discussed, whereby the arguments of the empirical transform are chosen by maximizing the determinant of the asymptotic Fisher information matrix for the resulting estimators. The methods are applied to some parametric models for which classical inference is complicated.