Sensitivity of test for overdispersion in Poisson regression

Overdispersion or extra-Poisson variation is very common for count data. This phenomenon arises when the variability of the counts greatly exceeds the mean under the Poisson assumption, resulting in substantial bias for the parameter estimates. To detect whether count data are overdispersed in the Poisson regression setting, various tests have been proposed and among them, the score tests derived by Dean (1992) are popular and easy to implement. However, such tests can be sensitive to anomalous or extreme observations. In this paper, diagnostic measures are proposed for assessing the sensitivity of Dean's score test for overdispersion in Poisson regression. Applications to the well-known fabric faults and Ames salmonella assay data sets illustrate the usefulness of the diagnostics in analyzing overdispersed count data.     

Trend Test for Overdispersed Proportions

The Cochran-Armitage test has commonly been used for a trend test in binomial proportions. The quasi-likelihood method provides a simple approach to model extra-binomial proportions. Two versions of the score and Wald tests using different parameterizations for the extra-binomial variance were investigated: one in terms of intercluster correlation, and another in terms of variance. The Monte Carlo simulation was used to evaluate the performance of the each version of the score test and the Wald test, and the Cochran-Armitage test. The simulation shows that the Cochran-Armitage test has the proper size only for the binomial sample data, and the test is no longer valid when applied to the extra-binomial data. The Wald test is more likely to exceed the nominal level than the score test under either intercluster correlation model or variance model. Both score tests performed very well even with the binomial data; the tests control the type I error and in the meantime maintain the power of detecting the dose effects. Based on the design considered in this paper, the two scores test are comparable. The score test based on the intercluster correlations model seems better controlling the Type I error but appears less powerful than that based on the variance model. An example from a developmental toxicity experiment is given.     

Significance Levels of Randomization Trend Tests in the Event of Rare Occurrences

In the statistical evaluation of data from a dose-response experiment, it is frequently of interest to test for dose-related trend: an increasing trend in response with increasing dose. The randomization trend test, a generalization of Fisher's exact test, has been recommended for animal tumorigenicity testing when the numbers of tumor occurrences are small. This paper examines the type I error of the randomization trend test, and the Cochran-Armitage and Mantel-Haenszel tests. Simulation results show that when the tumor incidence rates are less than 10%, the randomization test is conservative; the test becomes very conservative when the incidence rate is less than 5%. The Cochran-Armitage and Mantel-Haenszel tests are slightly anti-conservative (liberal) when the incidence rates are larger than 3%. Further, we propose a less conservatived method of calculating the p-value of the randomization trend test by excluding some permutations whose probabilities of occurrence are greater than the probability of the the observed outcome.     

Assessing overdispersion and dose-response in the male dominant lethal assay.

In dominant lethal studies the primary variables of interest are typically expressed as discrete counts or proportions (e.g., live implants, resorptions, percent pregnant). Simple statistical sampling models for discrete data such as binomial or Poisson generally do not fit this type of data because of extra-binomial or extra-Poisson departures from variability predicted under these simple models. Extra-variability in the fetal response may originate from parental contributions. These can lead to over- or under-dispersion seen as, e.g., extra-binomial variability in the proportion response. Utilizing a large control database, we investigated the relative impact of extra-variability from male or female contributions on the endpoints of interest. Male-related effects did not seem to contribute to overdispersion in our database; female-related effects were, however, evidenced. Various statistical methods were considered to test for significant treatment differences under these forms of sampling variability. Computer simulations were used to evaluate these methods and to determine which are most appropriate for practical use in the evaluation of dominant lethal data. Our results suggest that distribution-free statistical methods such as a nonparametric permutation test or rank-based tests for trend can be recommended for use.     

Reconsidering association testing methods using single-variant test statistics as alternatives to pooling tests for sequence data with rare variants

Association tests that pool minor alleles into a measure of burden at a locus have been proposed for case-control studies using sequence data containing rare variants. However, such pooling tests are not robust to the inclusion of neutral and protective variants, which can mask the association signal from risk variants. Early studies proposing pooling tests dismissed methods for locus-wide inference using nonnegative single-variant test statistics based on unrealistic comparisons. However, such methods are robust to the inclusion of neutral and protective variants and therefore may be more useful than previously appreciated. In fact, some recently proposed methods derived within different frameworks are equivalent to performing inference on weighted sums of squared single-variant score statistics. In this study, we compared two existing methods for locus-wide inference using nonnegative single-variant test statistics to two widely cited pooling tests under more realistic conditions. We established analytic results for a simple model with one rare risk and one rare neutral variant, which demonstrated that pooling tests were less powerful than even Bonferroni-corrected single-variant tests in most realistic situations. We also performed simulations using variants with realistic minor allele frequency and linkage disequilibrium spectra, disease models with multiple rare risk variants and extensive neutral variation, and varying rates of missing genotypes. In all scenarios considered, existing methods using nonnegative single-variant test statistics had power comparable to or greater than two widely cited pooling tests. Moreover, in disease models with only rare risk variants, an existing method based on the maximum single-variant Cochran-Armitage trend chi-square statistic in the locus had power comparable to or greater than another existing method closely related to some recently proposed methods. We conclude that efficient locus-wide inference using single-variant test statistics should be reconsidered as a useful framework for devising powerful association tests in sequence data with rare variants.     

Robust genomic control for association studies

Population-based case-control studies are a useful method to test for a genetic association between a trait and a marker. However, the analysis of the resulting data can be affected by population stratification or cryptic relatedness, which may inflate the variance of the usual statistics, resulting in a higher-than-nominal rate of false-positive results. One approach to preserving the nominal type I error is to apply genomic control, which adjusts the variance of the Cochran-Armitage trend test by calculating the statistic on data from null loci. This enables one to estimate any additional variance in the null distribution of statistics. When the underlying genetic model (e.g., recessive, additive, or dominant) is known, genomic control can be applied to the corresponding optimal trend tests. In practice, however, the mode of inheritance is unknown. The genotype-based chi (2) test for a general association between the trait and the marker does not depend on the underlying genetic model. Since this general association test has 2 degrees of freedom (df), the existing formulas for estimating the variance factor by use of genomic control are not directly applicable. By expressing the general association test in terms of two Cochran-Armitage trend tests, one can apply genomic control to each of the two trend tests separately, thereby adjusting the chi (2) statistic. The properties of this robust genomic control test with 2 df are examined by simulation. This genomic control-adjusted 2-df test has control of type I error and achieves reasonable power, relative to the optimal tests for each model.     

An enhanced quantile approach for assessing differential gene expressions

Summary .   Due to the small number of replicates in typical gene microarray experiments, the performance of statistical inference is often unsatisfactory without some form of information-sharing across genes. In this article, we propose an enhanced quantile rank score test (EQRS) for detecting differential expression in GeneChip studies by analyzing the quantiles of gene intensity distributions through probe-level measurements. A measure of sign correlation, δ, plays an important role in the rank score tests. By sharing information across genes, we develop a calibrated estimate of δ, which reduces the variability at small sample sizes. We compare the EQRS test with four other approaches for determining differential expression: the gene-specific quantile rank score test, the quantile rank score test assuming a common δ, a modified t -test using summarized probe-set-level intensities, and the Mack–Skillings rank test on probe-level data. The proposed EQRS is shown to be favorable for preserving false discovery rates and for being robust against outlying arrays. In addition, we demonstrate the merits of the proposed approach using a GeneChip study comparing gene expression in the livers of mice exposed to chronic intermittent hypoxia and of those exposed to intermittent room air.     

The Relationship Between Two Statistical Tests for the Comparison of Two Exponential Distributions

To compare two exponential distributions with or without censoring, two different statistics are often used; one is the F test proposed by COX (1953) and the other is based on the efficient score procedure. In this paper, the relationship between these tests is investigated and it is shown that the efficient score test is a large-sample approximation of the F test.     

A Powerful and Adaptive Association Test for Rare Variants

This article focuses on conducting global testing for association between a binary trait and a set of rare variants (RVs), although its application can be much broader to other types of traits, common variants (CVs), and gene set or pathway analysis. We show that many of the existing tests have deteriorating performance in the presence of many nonassociated RVs: their power can dramatically drop as the proportion of nonassociated RVs in the group to be tested increases. We propose a class of so-called sum of powered score (SPU) tests, each of which is based on the score vector from a general regression model and hence can deal with different types of traits and adjust for covariates, e.g., principal components accounting for population stratification. The SPU tests generalize the sum test, a representative burden test based on pooling or collapsing genotypes of RVs, and a sum of squared score (SSU) test that is closely related to several other powerful variance component tests; a previous study (Basu and Pan 2011) has demonstrated good performance of one, but not both, of the Sum and SSU tests in many situations. The SPU tests are versatile in the sense that one of them is often powerful, although its identity varies with the unknown true association parameters. We propose an adaptive SPU (aSPU) test to approximate the most powerful SPU test for a given scenario, consequently maintaining high power and being highly adaptive across various scenarios. We conducted extensive simulations to show superior performance of the aSPU test over several state-of-the-art association tests in the presence of many nonassociated RVs. Finally we applied the SPU and aSPU tests to the GAW17 mini-exome sequence data to compare its practical performance with some existing tests, demonstrating their potential usefulness.     

Score test for missing at random or not under logistic missingness models

Missing data are frequently encountered in various disciplines and can be divided into three categories: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). Valid statistical approaches to missing data depend crucially on correct identification of the underlying missingness mechanism. Although the problem of testing whether this mechanism is MCAR or MAR has been extensively studied, there has been very little research on testing MAR versus MNAR. A critical challenge that is faced when dealing with this problem is the issue of model identification under MNAR. In this paper, under a logistic model for the missing probability, we develop two score tests for the problem of whether the missingness mechanism is MAR or MNAR under a parametric model and a semiparametric location model on the regression function. The implementation of the score tests circumvents the identification issue as it requires only parameter estimation under the null MAR assumption. Our simulations and analysis of human immunodeficiency virus data show that the score tests have well-controlled type I errors and desirable powers.     

ROC graphs for assessing the ability of a diagnostic marker to detect three disease classes with an umbrella ordering

Receiver operating characteristic (ROC) curves and the area under these curves are commonly used to assess the ability of a continuous diagnostic marker (e.g., DNA methylation markers) to correctly classify subjects as having a particular disease or not (e.g., cancer). These approaches, however, are not applicable to settings where the gold standard yields more than two disease states or classes. ROC surfaces and the volume under the surfaces have been proposed for settings with more than two disease classes. These approaches, however, do not allow one to assess the ability of a marker to differentiate two disease classes from a third disease class without requiring a monotone order for the three disease classes under study. That is, existing approaches do not accommodate an umbrella ordering of disease classes. This article proposes the construction of an ROC graph that is applicable for an umbrella ordering. Furthermore, this article proposes that a summary measure for this umbrella ROC graph can be used to summarize the classification accuracy, and corresponding variance estimates can be obtained using U-statistics theory or bootstrap methods. The proposed methods are illustrated using data from a study assessing the ability of a DNA methylation marker to correctly classify lung specimens into three histologic classes: squamous cell carcinoma, large cell carcinoma, and nontumor lung.     

Multivariate extensions of McNemar's test

This article considers global tests of differences between paired vectors of binomial probabilities, based on data from two dependent multivariate binary samples. Difference is defined as either an inhomogeneity in the marginal distributions or asymmetry in the joint distribution. For detecting the first type of difference, we propose a multivariate extension of McNemar's test and show that it is a generalized score test under a generalized estimating equations (GEE) approach. Univariate features such as the relationship between the Wald and score tests and the dropout of pairs with the same response carry over to the multivariate case and the test does not depend on the working correlation assumption among the components of the multivariate response. For sparse or imbalanced data, such as occurs when the number of variables is large or the proportions are close to zero, the test is best implemented using a bootstrap, and if this is computationally too complex, a permutation distribution. We apply the test to safety data for a drug, in which two doses are evaluated by comparing multiple responses by the same subjects to each one of them.     

Statistical analysis of in vivo rodent micronucleus assay

The in vivo rodent micronucleus assay (MNC) is widely used as a cytogenetic assay to detect the clastogenic activity of a chemical in vivo. MNC is one of three tests in a battery recommended by the fourth International Conference on Harmonization (ICH4) of Genotoxicity Guidelines. As such it has been accepted by many regulatory authorities. However, the determination of a positive result in a genotoxicity test, including MNC, has been an issue of debate among toxicologists and biometricians. In this presentation we compare several statistical procedures that have been suggested for the analysis of MNC data and indicate which one is the most powerful. The standard protocol of MNC has at least three dose levels plus the control dose and uses at least four animals per group. For each animal, 2000 polychromatic erythrocytes (PCE) are counted. Two statistical procedures can be employed, either alone or jointly, for the analysis of the MNC dose-response curve. These are the Cochran-Armitage (C-A) trend test and the Dunnett type test. For performing Dunnett type tests, toxicologists often use negative historical control rate for the estimate of the concurrent negative control rate. Some toxicologists emphasize the reproducibility of assay results instead of the dose-response relationship for the important criterion [J. Ashby, H. Tinwell, Mutat. Res. 327 (1995) 49-55; for the rebuttal see M. Hayashi, T. Sofuni, Mutat. Res. 331 (1995) 173-174]. The following three procedures are currently employed in toxicology labs for the evaluation of MNC result. The assay response is deemed positive if it is detected by (i) the C-A trend test alone, (ii) both the C-A trend test and the Dunnett type test and (iii) either the C-A trend test or the Dunnett type test. Using Monte Carlo simulation, we first find for each procedure, sizes of tests which yield the experiment-wise type I error rate of 0.05 and show that the procedure (ii) is the most powerful against the alternatives of monotone increase. The procedure (ii) which originated from Hayashi's three-step procedure was coded in C and termed 'MNC'. The MNC software program is available in the public domain through the ftp.     

Exact and Asymptotic Power Determination For Dose‐Response Studies With Ordinal Response

In this article, we describe a conditional score test for detecting a monotone dose‐response relationship with ordinal response data. We consider three different versions of this test: asymptotic, conditional exact, and mid‐P conditional score test. Exact and asymptotic power formulae based on these tests will be studied. Asymptotic sample size formulae based on the asymptotic conditional score test will be derived. The proposed formulae are applied to a vaccination study and a developmental toxicity study for illustrative purposes. Actual significance level and exact power properties of these tests are compared in a small empirical study. The mid‐P conditional score test is observed to be the most powerful test with actual significance level close to the pre‐specified nominal level.     

On the use of historical control information for trend test in carcinogenesis

In this article a new non-model-based significance test for detecting dose-response relationship with the incorporation of historical control data is proposed. This non-model-based test is considered simpler from a regulatory perspective because it does not require validating any modeling assumptions. Moreover, our test is especially appropriate to those studies in which the intravenous doses for the investigational chemical are labeled as, e.g., low, medium and high or the dose labels do not suggest any obvious choices of dose scores. This test can be easily adopted for detecting general dose-response shape, such as an umbrella pattern. Simple adjustments will be proposed for better control of the actual Type I error. Data sets from two carcinogenesis studies will be used to illustrate our method. We also evaluate the performance of the proposed test and the famous model-based Tarone's trend test with respect to size and power.     

Non-inferiority of new procedure to standard procedure in stratified matched-pair design

We consider the statistical testing for non-inferiority of a new treatment compared with the standard one under matched-pair setting in a stratified study or in several trials. A non-inferiority test based on the efficient scores and a Mantel-Haenszel (M-H) like procedure with restricted maximum likelihood estimators (RMLEs) of nuisance parameters and their corresponding sample size formulae are presented. We evaluate the above tests and the M-H type Wald test in level and power. The stratified score test is conservative and provides the best power. The M-H like procedure with RMLEs gives an accurate level. However, the Wald test is anti-conservative and we suggest caution when it is used. The unstratified score test is not biased but it is less powerful than the stratified score test when base-line probabilities related to strata are not the same. This investigation shows that the stratified score test possesses optimum statistical properties in testing non-inferiority. A common difference between two proportions across strata is the basic assumption of the stratified tests, we present appropriate tests to validate the assumption and related remarks.     

Estimation and test for linkage between markers: a comparison of lod score and χ 2 test in a linkage study of maritime pine (Pinus pinaster Ait.)

The first step in the construction of a linkage map involves the estimation and test for linkage between all possible pairs of markers. The lod score method is used in many linkage studies for the latter purpose. In contrast with classical statistical tests, this method does not rely on the choice of a first-type error level. We thus provide a comparison between the lod score and a ² test on linkage data from a gymnosperm, the maritime pine. The lod score appears to be a very conservative test with the usual thresholds. Its severity depends on the type of data used.     

The Power of the Mantel-Haenszel Test when Adjustment is by the True or Estimated Propensity Score

The power of the Mantel-Haenszel test for no treatment effect in the case of binary exposure and response variates was examined through simulation studies when subclasses were formed on the basis of the true and estimated propensity scores and by direct stratification on two continuous covariates. The power of these tests was also compared to the score test in a misspecified logistic regression model. In general adjustment by the true propensity score was most likely to reject a false null hypothesis, the score test was more likely to reject a false null hypothesis than the Mantel-Haenszel test when adjustment is by the estimated propensity score or subclassification on the covariates. There was litte difference in the observed powers of the Mantel-Haenszel tests between adjustment by the estimated propensity score and subclassification on the covariates.     

A comparison of likelihood ratio tests and Rao's score test for three separable covariance matrix structures

The problem of testing the separability of a covariance matrix against an unstructured variance‐covariance matrix is studied in the context of multivariate repeated measures data using Rao's score test (RST). The RST statistic is developed with the first component of the separable structure as a first‐order autoregressive (AR(1)) correlation matrix or an unstructured (UN) covariance matrix under the assumption of multivariate normality. It is shown that the distribution of the RST statistic under the null hypothesis of any separability does not depend on the true values of the mean or the unstructured components of the separable structure. A significant advantage of the RST is that it can be performed for small samples, even smaller than the dimension of the data, where the likelihood ratio test (LRT) cannot be used, and it outperforms the standard LRT in a number of contexts. Monte Carlo simulations are then used to study the comparative behavior of the null distribution of the RST statistic, as well as that of the LRT statistic, in terms of sample size considerations, and for the estimation of the empirical percentiles. Our findings are compared with existing results where the first component of the separable structure is a compound symmetry (CS) correlation matrix. It is also shown by simulations that the empirical null distribution of the RST statistic converges faster than the empirical null distribution of the LRT statistic to the limiting χ² distribution. The tests are implemented on a real dataset from medical studies.     

A fast, unbiased and exact allelic test for case-control association studies

Association studies are traditionally performed in the case-control framework. As a first step in the analysis process, comparing allele frequencies using the Pearson's chi-square statistic is often invoked. However such an approach assumes the independence of alleles under the hypothesis of no association, which may not always be the case. Consequently this method introduces a bias that deviates the expected type I error-rate. In this article we first propose an unbiased and exact test as an alternative to the biased allelic test. Available data require to perform thousands of such tests so we focused on its fast execution. Since the biased allelic test is still widely used in the community, we illustrate its pitfalls in the context of genome-wide association studies and particularly in the case of low-level tests. Finally, we compare the unbiased and exact test with the Cochran-Armitage test for trend and show it perfoms similarly in terms of power. The fast, unbiased and exact allelic test code is available in R, C++ and Perl at: http://stat.genopole.cnrs.fr/software/fueatest.