Paradoxes in nonparametric tests

This paper is a continuation of one (1992) in which the author studied the paradoxes that can arise when a nonparametric statistical test is used to give an ordering of k samples and the subsets of those samples. This article characterizes the projection paradoxes that can occur when using contingency tables, complete block designs, and tests of dichotomous behaviour of several samples. This is done by examining the “dictionaries” of possible orderings of each of these procedures. Specifically, it is shown that contingency tables and complete block designs, like the Kruskal-Wallis nonparametric test on k samples, minimize the number and kinds of projection paradoxes that can occur; however, using a test of dichotomous behaviour of several samples does not. An analysis is given of two procedures used to determine the ordering of a pair of samples from a set of k samples. It is shown that these two procedures may not have anything in common.     

A TEST OF HOMOGENEITY AGAINST SCALE ALTERNATIVES USING SUBSAMPLE EXTREMA1

In this paper a new class of non-parametric tests for testing homogeneity of several populations against scale alternatives is proposed. For this, independent samples of fixed sizes are drawn from each population and from these samples, all possible sub-samples of the same size are drawn and their maxima and minima are computed. Using these extreme the class of tests is obtained. Tests of this type have been offered for the two-sample slippage problem by Kochar (1978). Under certain conditions, this class of tests is shown to be consistent against ‘difference in scale’ alternatives. The test has been compared with Bhapkar's V-test (1961), Deshpande's D-test (1965), Sugiura's D_rs-test (1965) and with a classical test given by Lehmann (1959, pp. 273–275). It is shown that some members of this proposed class of tests are more efficient than the first three tests in the case of uniform, Laplace and normal distributions, when the number of populations compared is small.     

Optimizing the wald-wolfowitz runs statistic using a linkage tolerance: guidelines based on computer simulation

When performing the Wald-Wolfowitz runs test, observations from two samples are combined and ordered, and the test statistic is the number of sequences of observations from the same sample. This test statistic is equivalent to the number of links between observations from different samples, if we consider each observation to be linked to the next higher and next lower observations. While it is known that the Wald-Wolfowitz runs test is not very powerful, what would be the effect on the power of the Wald-Wolfowitz runs test if all observations within a specified Euclidean distance or “tolerance” were linked instead? This question is motivated by the simulation results of Whaley and Quade (1985), who found that for normal data, the power of the multi-dimensional runs test using a linkage tolerance compared favorably to Hotelling's T² in some instances. The results of a similar simulation procedure show that the power of the Wald-Wolfowitz runs test does indeed improve when observations are linked using a tolerance. The results also suggest that a better large sample approximation to the distribution of the test statistic needs to be found.     

A toolbox of permutation tests for structural change

The sup $LM$ test for structural change is embedded into a permutation test framework for a simple location model. The resulting conditional permutation distribution is compared to the usual (unconditional) asymptotic distribution, showing that the power of the test can be clearly improved in small samples. Furthermore, the permutation test is embedded into a general framework that encompasses tools for binary and multivariate dependent variables as well as model-based permutation testing for structural change. It is also demonstrated that the methods can not only be employed for analyzing structural changes in time series data but also for recursive partitioning of cross-section data. The procedures suggested are illustrated using both artificial data and empirical applications (number of youth homicides, employment discrimination data, carbon flux in tropical forests, stock returns, and demand for economics journals).     

On the Sampling Distributions of the Estimated Process Loss Indices with Asymmetric Tolerances

The inverse Gaussian distribution provides a flexible model for analyzing positive, right-skewed data. The generalized variable test for equality of several inverse Gaussian means with unknown and arbitrary variances has satisfactory Type-I error rate when the number of samples (k) is small (Tian, 2006). However, the Type-I error rate tends to be inflated when k goes up. In this article, we propose a parametric bootstrap (PB) approach for this problem. Simulation results show that the proposed test performs very satisfactorily regardless of the number of samples and sample sizes. This method is illustrated by an example.     

Tests for cointegration rank and choice of the alternative

This paper discusses likelihood-ratio (LR) tests on the cointegrating (CI) rank which consider any possible dimension of the CI rank under the alternative. The trace test and lambda-max test are obtained as special cases. Limit quantiles for all the tests in the class are derived. It is found that any of these tests can be used to construct an estimator of the CI rank, with no differences in asymptotic properties when the alternative is fixed. The properties of the class of tests are investigated by local asymptotic analysis, a simulation study and an empirical illustration. It is found that all the tests in the class have comparable power, which deteriorates substantially as the number of random walks increases. Tests constructed for a specific class of alternatives present minor power gains for alternatives in the class, and require the alternative to be far from the null. No test in this class is found to be asymptotically (in-)admissible. Some of the new tests in the class can also be arranged to give a constrained estimator of the CI rank, that restricts the minimum number of common trends. The power gains that these tests can obtain by constraining the minimum number of common trends appears to be limited and outweighted by the risk of inconsistency induced by the constrains. As a consequence, no value of the CI rank should be left untested, unless it can be excluded beyond any reasonable doubt.     

Application of a new discrete distribution

In epidemiology, an infection lasting n weeks may be monitored by taking weekly serum samples. If tests on samples are independent Bernoulli trials with probability q of correctly testing positive, the apparent duration of infection ( from the first positive test to the last positive test inclusive) may be less than n weeks. This distribution of apparent length also arises when plants in a row of n each have a probability q of germinating, for example. This distribution is shown to be related to that of the number of tails obtained when tossing a coin until two heads are obtained, in a maximum of n tosses. The properties of the 'apparent length' distribution are described, and some compounded (mixed) distributions that can be derived from it are also discussed. The distribution was used to estimate the underlying distribution of the duration of infection, in a longitudinal study of infections of children. The methodology was also used to estimate the proportion of infectious episodes that were not detected. It can be similarly used to correct episode durations and rates in longitudinal studies in which episodes of any kind are detected by regular sampling.     

Nonlinear IV panel unit root testing under structural breaks in the error variance

The paper examines the behavior of a generalized version of the nonlinear IV unit root test proposed by Chang (2002) when the series’ errors exhibit nonstationary volatility. The leading case of such nonstationary volatility concerns structural breaks in the error variance. We show that the generalized test is not robust to variance changes in general, and illustrate the extent of the resulting size distortions in finite samples. More importantly, we show that pivotality is recovered when using Eicker-White heteroskedasticity-consistent standard errors. This contrasts with the case of Dickey-Fuller unit root tests, for which Eicker-White standard errors do not produce robustness and thus require computationally costly corrections such as the (wild) bootstrap or estimation of the so-called variance profile. The pivotal versions of the generalized IV tests – with or without the correct standard errors – do however have no power in $1/T$ -neighbourhoods of the null. We also study the validity of panel versions of the tests considered here.     

Spatial Clustering Tests Based on the Domination Number of a New Random Digraph Family

We use the domination number of a parametrized random digraph family called proportional-edge proximity catch digraphs (PCDs) for testing multivariate spatial point patterns. This digraph family is based on relative positions of data points from various classes. We extend the results on the distribution of the domination number of proportional-edge PCDs, and use the domination number as a statistic for testing segregation and association against complete spatial randomness. We demonstrate that the domination number of the PCD has binomial distribution when size of one class is fixed while the size of the other (whose points constitute the vertices of the digraph) tends to infinity and has asymptotic normality when sizes of both classes tend to infinity. We evaluate the finite sample performance of the test by Monte Carlo simulations and prove the consistency of the test under the alternatives. We find the optimal parameters for testing each of the segregation and association alternatives. Furthermore, the methodology discussed in this article is valid for data in higher dimensions also.     

A Generalized Distance Multi-Sample Test of Normality With Applications to Process Control

In statistical process control one typically takes periodic small samples. Statistical inferences made from these samples often assume that the samples come from normal distributions with the means and variances possibly changing over time. A multisample test of normality is proposed to test this assumption. The test statistic is the generalized distance between the standardized order statistic vector averaged across the samples and its expected value under normality. The null distribution of the statistic approaches a chi-squared distribution as the number of samples increases. A Monte Carlo study suggests that the test has desirable power properties relative to competing tests.     

Two Samples Tests for Functional Data

Data in many experiments arises as curves and therefore it is natural to use a curve as a basic unit in the analysis, which is in terms of functional data analysis (FDA). Functional curves are encountered when units are observed over time. Although the whole function curve itself is not observed, a sufficiently large number of evaluations, as is common with modern recording equipment, is assumed to be available. In this article, we consider the statistical inference for the mean functions in the two samples problem drawn from functional data sets, in which we assume that functional curves are observed, that is, we consider the test if these two groups of curves have the same mean functional curve when the two groups of curves without noise are observed. The L ²-norm based and bootstrap-based test statistics are proposed. It is shown that the proposed methodology is flexible. Simulation study and real-data examples are used to illustrate our techniques.     

C55. Quick and accurate rule of thumb for determining the critical value of the coefficient of correlation

Life table analysis techniques in epidemiology depend upon the asymptotic properties of the statistical test methods employed. In some instances, the statistical procedures indicate highly significant results which are, in reality, unjustified. The phenomenon may occur when the asymptotic methods are applied in situations where the cases of interest are few in number. This situation is illustrated by the 20 multiple myeloma deaths observed in the RERF Life Span Study cohort. A permutation test is applied to the life table data, although the test requires the false assumption that the censoring distribution is independent of the radiation dose. A simulation test is developed which does not require equal censoring, which has the same asymptotics as the usual test methods, and which is less likely to overestimate significance in small samples. It is found that both of these small-sample tests provide reasonable numerical solutions. In addition, the simulation test is recommended in general for analyzing life table data with unequal censoring. Finally, by using the small-sample tests, the frequency of death from multiple myeloma is shown to be positively associated with radiation dose (P<0.01).     

Hypothesis testing for Markov chain Monte Carlo

Testing between hypotheses, when independent sampling is possible, is a well developed subject. In this paper, we propose hypothesis tests that are applicable when the samples are obtained using Markov chain Monte Carlo. These tests are useful when one is interested in deciding whether the expected value of a certain quantity is above or below a given threshold. We show non-asymptotic error bounds and bounds on the expected number of samples for three types of tests, a fixed sample size test, a sequential test with indifference region, and a sequential test without indifference region. Our tests can lead to significant savings in sample size. We illustrate our results on an example of Bayesian parameter inference involving an ODE model of a biochemical pathway.     

EXACT TWO-SAMPLE MEDIAN TESTS WHEN ONE "SAMPLE" VALUE POSSIBLY FROM A DIFFERENT POPULATION

Random samples are assumed for the univariate two-sample problem. Sometimes this assumption may be violated in that an observation in one “sample”, of size m, is from a population different from that yielding the remaining m—1 observations (which are a random sample). Then, the interest is in whether this random sample of size m—1 is from the same population as the other random sample. If such a violation occurs and can be recognized, and also the non-conforming observation can be identified (without imposing conditional effects), then that observation could be removed and a two-sample test applied to the remaining samples. Unfortunately, satisfactory procedures for such a removal do not seem to exist. An alternative approach is to use two-sample tests whose significance levels remain the same when a non-conforming observation occurs, and is removed, as for the case where the samples were both truly random. The equal-tail median test is shown to have this property when the two “samples” are of the same size (and ties do not occur).     

The effect of non-normal disturbances and conditional heteroskedasticity on multiple cointegration tests

!t is well-known that Johansen's multiple cointegration tests' results and those of Johansen and Juselius' tests for restricrions on cointegrating vectors and their weights have far-reaching implications for economic modelling and analysis. Therefore, it is important to ensure that the tests have desirable finite sample properties. Although the statistics are derived under Gaussian distribution,the asympotic results are derived under a much wider class of distributions. Using simulation, this paper investigates the effect of non-normal disturbances on these tests in finite samples. Further, ARCH/GARCH type conditional heteroskedasticity is present in many economic and financial time series. This paper examines the finite properties of the tests when the error term follows ARCH/GARCH type processes. From the evidence, it appears that researchers should not be overly concerned by the possibility of small departures from non-normality when using Johansen's suggested techniques even in finite samples. ARCH and GARCH effects may be more problematic, however. In particular it becomes more important ro test whether the restriction implicit in the integrated (or near-integrated) ARCH-type Drocess actually holds in time series for the application of the cointegraiion rank tests and the test for restrictions on cointegrating weights. The tests for restrictions on cointegrating vectors apper to be robust for non-normal errors and for all ARCH and GARCH type processes considered.     

A pooled Bayes test of independence for sparse contingency tables from small areas

We study the association between bone mineral density (BMD) and body mass index (BMI) when contingency tables are constructed from the several U.S. counties, where BMD has three levels (normal, osteopenia and osteoporosis) and BMI has four levels (underweight, normal, overweight and obese). We use the Bayes factor (posterior odds divided by prior odds or equivalently the ratio of the marginal likelihoods) to construct the new test. Like the chi-squared test and Fisher's exact test, we have a direct Bayes test which is a standard test using data from each county. In our main contribution, for each county techniques of small area estimation are used to borrow strength across counties and a pooled test of independence of BMD and BMI is obtained using a hierarchical Bayesian model. Our pooled Bayes test is computed by performing a Monte Carlo integration using random samples rather than Gibbs samples. We have seen important differences among the pooled Bayes test, direct Bayes test and the Cressie-Read test that allows for some degree of sparseness, when the degree of evidence against independence is studied. As expected, we also found that the direct Bayes test is sensitive to the prior specifications but the pooled Bayes test is not so sensitive. Moreover, the pooled Bayes test has competitive power properties, and it is superior when the cell counts are small to moderate.     

A Cautionary Note on the Use of the Grønnesby and Borgan Goodness-of-Fit Test for the Cox Proportional Hazards Model

Grønnesby and Borgan (1996, Lifetime Data Analysis 2, 315–328) propose an omnibus goodness-of-fit test for the Cox proportional hazards model. The test is based on grouping the subjects by their estimated risk score and comparing the number of observed and a model based estimated number of expected events within each group. We show, using extensive simulations, that even for moderate sample sizes the choice of number of groups is critical for the test to attain the specified size. In light of these results we suggest a grouping strategy under which the test attains the correct size even for small samples. The power of the test statistic seems to be acceptable when compared to other goodness-of-fit tests.     

Some distribution-free rane-like statistics having the mann-whitney-wilcoxon null distribution

In this note we suggest a class of two-sample test statistics iich have, as their null distribution,the Mann-Whitney-Wilcoxon

ill distribution. An interesting property of these statistics is lat many are not rank statistics; that is, they cannot be coumplited from, the ranks of the original observations. However, they %e still distribution-free when the two populations are identi-il. This class contains the Mann-Whitney-Wilcoxon test for the niality of location parameters of two distributions and a two-aiaple test for equality of spreads of two distributions recently ivestigated by Fligner and Killeen (1976)