Chi-Squared Goodness-of-Fit Tests: Cell Selection and Power

To use the Pearson chi-squared statistic to test the fit of a continuous distribution, it is necessary to partition the support of the distribution into k cells. A common practice is to partition the support into cells with equal probabilities. In that case, the power of the chi-squared test may vary substantially with the value of k. The effects of different values of k are investigated with a Monte Carlo power study of goodness-of-fit tests for distributions where location and scale parameters are estimated from the observed data. Allowing for the best choices of k, the Pearson and log-likelihood ratio chi-squared tests are shown to have similar maximum power for wide ranges of alternatives, but this can be substantially less than the power of other well-known goodness-of-fit tests.     

On diagnostic devices for proposing half-logistic and inverse half-logistic models using generalized (k) record values

In this work we consider the generalized upper (k) record values (GURV’s) and generalized lower (k) record values (GLRV’s) arising from half-logistic distribution (HLD) and inverse half-logistic distribution (IHLD). We derive some characterization results of HLD based on some moment relations of generalized upper (k) record values and those of generalized lower (k) record values and accordingly devised some diagnostic tools to identify HLD as a model to the distribution of a population. Similar characterization theorems and diagnostic tools are developed for IHLD as well. Simulation studies are conducted to validate the diagnostic tools devised for both HLD and IHLD.     

Tables of two-sided tolerance factors for a normal distribution

Given a random sample of size N from a normal distribution, we consider tolerance intervals of the form X ? ks to X + ks, where X is the sample mean and s is the sample standard deviation. The value of k is chosen so that the interval covers a given proportion P of the population with confidence γ. Exact values of k, computed from numerical integration, are given for N = 2(1)100; P = 0.75, 0.90, 0.95, 0.975, 0.99, 0.995, 0.999; and γ = 0.5, 0.75, 0.90, 0.95, 0.975, 0.99, 0.995. The exact values are compared with the values obtained from an approximation developed by Wald and Wolfowitz (1946).     

Tests for multiple upper or lower outliers in an exponential sample

SUMMARY T = \[x + ... + x ]/ Sigma x (T*= \[x + ... + x ] Sigma x ) is the max k (n- k+ 1 ) (n) i k ( 1 ) (k) i imum likelihood ratio test statistic for k upper ( lower ) outliers in an exponential sample x , ..., x . The null distributions of T for k= 1,2 were given by Fisher and by Kimber 1 n k and Stevens , while those of T*(k= 1,2) were given by Lewis and Fieller . In this paper , k the simple null distributions of T and T* are found for all possible values of k, and k k percentage points are tabulated for k= 1, 2, ..., 8. In addition , we find a way of determining k, which can reduce the masking or ' swamping ' effects .     

SAMPLING ERRORS OF GENETIC CORRELATION COEFFICIENTS CALCULATED FROM ANALYSES OF VARIANCE AND COVARIANCE

Summary. In this paper a formula is developed for estimating the sampling variance of a genetic correlation estimated from analyses of variance and covariance. The formula holds provided the heritability estimate of neither character is zero. However, the development assumes a constant number of offspring per sire, k , and the effect of varying values of k is discussed briefly. The efficiency of experiments from which genetic parameters are to be estimated has also been investigated and optimum values of k are given for various combinations of phenotypic and genetic parameters.     

Optimal linear combinations using householder transformations

In pattern classification of sampled vector valued random variables it is often essential, due to computational and accuracy considerations, to consider certain measurable transformations of the random variable. These transformations are generally of a dimension-reducing nature. In this paper we consider the class of linear dimension reducing transformations, i.e., the k × n matrices of rank k where k < n and n is the dimension of the range of the sampled vector random variable.

In this connection, we use certain results (Decell and Quirein, 1973), that guarantee, relative to various class separability criteria, the existence of an extremal transformation. These results also guarantee that the extremal transformation can be expressed in the form (I_k∣ Z)U where I_k is the k × k identity matrix and U is an orthogonal n × n matrix. These results actually limit the search for the extremal linear transformation to a search over the obviously smaller class of k × n matrices of the form (I_k ∣Z)U. In this paper these results are refined in the sense that any extremal transformation can be expressed in the form (I_K∣Z)H_p … H₁ where p ≤ min{k, n?k} and H_i is a Householder transformation i=l,…, p, The latter result allows one to construct a sequence of transformations (L_K∣ Z)H₁, (I_K Z)H₂H₁ … such that the values of the class separability criterion evaluated at this sequence is a bounded, monotone sequence of real numbers. The construction of the i-th element of the sequence of transformations requires the solution of an n-dimensional optimization problem. The solution, for various class separability criteria, of the optimization problem will be the subject of later papers. We have conjectured (with supporting theorems and empirical results) that, since the bounded monotone sequence of real class separability values converges to its least upper bound, this least upper bound is an extremal value of the class separability criterion.

Several open questions are stated and the practical implications of the results are discussed.     

Extended and updated tables for the Friedman rank test

The Friedman's test is used for assessing the independence of repeated experiments resulting in ranks, summarized as a table of integer entries ranging from 1 to k, with k columns and N rows. For its practical use, the hypothesis testing can be derived either from published tables with exact values for small k and N, or using an asymptotic analytical approximation valid for large N or large k. The quality of the approximation, measured as the relative difference of the true critical values with respect those arising from the asymptotic approximation is simply not known. The literature review shows cases where the wrong conclusion could have been drawn using it, although it may not be the only cause of opposite decisions. By Monte Carlo simulation we conclude that published tables do not cover a large enough set of (k, N) values to assure adequate accuracy. Our proposal is to systematically extend existing tables for k and N values, so that using the analytical approximation for values outside it will have less than a prescribed relative error. For illustration purposes some of the tables have been included in the paper, but the complete set is presented as a source code valid for Octave/Matlab/Scilab etc., and amenable to be ported to other programming languages.     

On the Use of Adaptive Nearest Neighbors for Missing Value Imputation

A popular nonparametric treatment of missing value imputation uses methods based on k-nearest neighbors, where the number k of nearest neighbors is fixed without any consideration of the local features of missing values. This article proposes an alternative imputation method based on adaptive nearest neighbors, which takes into account the local features of the data. The proposed method adapts the number of neighbors in imputing the missing values according to the location of the missing values. Efficiency evaluation is then gauged through simulation studies using both simulated and real data. It is shown that the proposed method has distinct advantages over the imputation method based on k-nearest neighbors.     

Biased discriminant analysis: Evaluation of the optimum probability of misclassification

This article extends the work of DiPillo (1976) on the Biased Minimum x² Rule. The optimum value of k (the biasing factor) Is determined and the true probability of misclassification is found. The proportion improvements reported in the 1976 paper are shown to be conservative. Some suggestions for algorithms to determine the optimal value of k are presented.     

Normal quantile estimation based on minimizing the error in predicted distribution functions

Estimators of the form [Xbar] + kS for estimating the p quantile of a normal distribution are studied when k is chosen to either minimize the mean square error in the predicted distribution function or to make the predicted distribution function unbiased for p. Here, [Xbar] and S are the usual sample mean and standard deviation, respectively, and the predicted distribution function is the true (normal) distribution function evaluated at the estimated quantile.

These k values are presented for various sample sizes and values of p, and application to warranty determination is discussed.     

A new look at rank tests in ordered-2×k contingency tables

We consider testing for association in contingency tables with 2 rows and k columns, where the columns represent ordered categories. If the rows are treatments and the columns are outcomes, this may be treated as a two-sample problem with all the outcomes tied at one of only k values. Then rank tests may be applied even without knowing the values. Some special considerations apply, however, and the most usual rank tests may not be the best ones. We use a graphical technique to compare the properties of various rank tests.     

Multiple decision rules for comparing several populations with a fixed known standard

Independent observations are available from k univariate distributions indexed by a real parameter θ. It is desired to select that distribution with the largest parameter value unless this value is smaller than some fixed standard θ₀ in which case no distribution is to be selected. Various single-stage procedures for this (k+l)-decision problem are discussed, using indifference zone, decision theoretic, Bayesian, and subset selection approaches.     

Least significant spacing for ‘one versus the rest’ normal populations

Consider sample means from k(≥2) normal populations where the variances and sample sizes are equal. The problem is to find the ‘least significant difference’ or ‘spacing’ (LSS) between the two largest means, so that if an observed spacing is larger we have confidence 1 - α that the population with largest sample mean also has the largest population mean.

When the variance is known it is shown that the maximum LSS occurs when k = 2, provided a < .2723. In other words, for any value of k we may use the usual (one-tailed) least significant difference to demonstrate that one population has a population mean greater than (or equal to) the rest.

When the variance is estimated bounds are obtained for the confidence which indicate that this last result is approximately correct.     

Confidence intervals for the threshold parameter

Confidence intervals for the threshold parameter (guarantee-life ) are considered. The first k failure-times from a sample of size n are observed. Under the assumption that as n →∞ the first failure-time is attracted to the Weibull distribution, confidence intervals based on the observed range are constructed. It is shown that as k(k ≥ 2) increases the expected length of the confidence interval is substantially reduced. However, when k = 10 (or 20 in some cases) the expected length is near its minimum value.     

Large-sample quantile estimation in pareto laws

A large-sample method of estimation for the parameters of Pareto laws is investigatedo The estimates are derived by using a small subset of k sample quantiles out of the original observations. The optimum spacing of the k quantiles is also examined. A Monte Carlo study compares this method with the method of moments and that of maximum likelihood for a selected set of parameter values and sample sizes.     

Partitioning a distribution

Suppose it is desired to partition a distribution into k groups (classes) using squared error or absolute error as the mea¬sure of information retained. An algorithm to obtain the optimal boundaries (or class probabilities) is giTen. For the case of squared error optimal class probabilities were obtained for k = 2 to 15 for beta (for various values of the parameters), chi-square (12 d.f.) exponential, normals and uniform distributions. Results obtained are compared and analysed in light of existing papers, Special attention is given to the case k =5, corresponding to the assignment of the letter grades A, B, C, D9 P in courses, and to the case k = 9 corresponding to stanines.     

A Simulation Study of Some Biasing Parameters for the Ridge Type Estimation of Poisson Regression

This article proposes several estimators for estimating the ridge parameter k based on Poisson ridge regression (RR) model. These estimators have been evaluated by means of Monte Carlo simulations. As performance criteria, we have calculated the mean squared error (MSE), the mean value, and the standard deviation of k. The first criterion is commonly used, while the other two have never been used when analyzing Poisson RR. However, these performance criteria are very informative because, if several estimators have an equal estimated MSE, then those with low average value and standard deviation of k should be preferred. Based on the simulated results, we may recommend some biasing parameters that may be useful for the practitioners in the field of health, social, and physical sciences.     

Consistency of the GMLE with Mixed Case Interval-Censored Data

In this paper we consider an interval censorship model in which the endpoints of the censoring intervals are determined by a two stage experiment. In the first stage the value k of a random integer is selected; in the second stage the endpoints are determined by a case k interval censorship model. We prove the strong consistency in the L ₁( μ )-topology of the non-parametric maximum likelihood estimate of the underlying survival function for a measure μ which is derived from the distributions of the endpoints. This consistency result yields strong consistency for the topologies of weak convergence, pointwise convergence and uniform convergence under additional assumptions. These results improve and generalize existing ones in the literature.     

Simultaneous prediction intervals for regression models with an intercept

Consider the usual linear regression model y = x’β+?, relating a response y to a vector of predictors x. Suppose that n observations on y together with the corresponding values of x are available , and it is desired to construct simultaneous prediction intervals for k future values of y at values of x which can not be ascertained beforehand. In most applications the regression model contains an intercept. This paper presents two sets of prediction intervals appropriate to this case. The proposed intervals are compared with those of Carlstein (1986), and the improvements are illustrated in the case of simple linear regression.