Pooling states in the multinomial logit model: degrees of freedom. A correction

Received: October 4, 2000     

Adjusting the degrees of freedom for the jackknife

The rule of thumb given by Mosteller and Tukey for obtaining the degrees of freedom which are appropriate for the approximate t-statistic associated with the jackknife is discussed.Examples of possible misapplication of this rule are given.     

An approximate degrees of freedom test for heteroscedastic two-way ANOVA

Heteroscedastic two-way ANOVA are frequently encountered in real data analysis. In the literature, classical F-tests are often blindly employed although they are often biased even for moderate heteroscedasticity. To overcome this problem, several approximate tests have been proposed in the literature. These tests, however, are either too complicated to implement or do not work well in terms of size controlling. In this paper, we propose a simple and accurate approximate degrees of freedom (ADF) test. The ADF test is shown to be invariant under affine-transformations, different choices of contrast matrix for the same null hypothesis, or different labeling schemes of cell means. Moreover, it can be conducted easily using the usual F-distribution with one unknown degree of freedom estimated from the data. Simulations demonstrate that the ADF test works well in various cell sizes and parameter configurations but the classical F-tests work badly when the cell variance homogeneity assumption is violated. A real data example illustrates the methodologies.     

Computing AIC for black-box models using generalized degrees of freedom: A comparison with cross-validation

Generalized degrees of freedom (GDF), as defined by Ye (1998 Ye, J. (1998). On measuring and correcting the effects of data mining and model selection. Journal of the American Statistical Association 93(441):120–131. [Google Scholar] JASA 93:120–131), represent the sensitivity of model fits to perturbations of the data. Such GDF can be computed for any statistical model, making it possible, in principle, to derive the effective number of parameters in machine-learning approaches and thus compute information-theoretical measures of fit. We compare GDF with cross-validation and find that the latter provides a less computer-intensive and more robust alternative. For Bernoulli-distributed data, GDF estimates were unstable and inconsistently sensitive to the number of data points perturbed simultaneously. Cross-validation, in contrast, performs well also for binary data, and for very different machine-learning approaches.     

On generalized degrees of freedom with application in linear mixed models selection

The concept of degrees of freedom plays an important role in statistical modeling and is commonly used for measuring model complexity. The number of unknown parameters, which is typically used as the degrees of freedom in linear regression models, may fail to work in some modeling procedures, in particular for linear mixed effects models. In this article, we propose a new definition of generalized degrees of freedom in linear mixed effects models. It is derived from using the sum of the sensitivity of the expected fitted values with respect to their underlying true means. We explore and compare data perturbation and the residual bootstrap to empirically estimate model complexity. We also show that this empirical generalized degrees of freedom measure satisfies some desirable properties and is useful for the selection of linear mixed effects models.     

Approximating the difference of two t-variables for all degrees of freedom using truncated variables

A scaled t‐distribution is used to approximate the distribution of a linear combination of two independent t‐variables for any number of degrees of freedom, and in particular for low degrees of freedom where moments do not exist. The approximation is the method‐of‐moments solution to the analogous problem with truncated t‐variables. The approximation exists for all degrees of freedom, is very accurate for more than two degrees of freedom, and performs as well as other approximations of this form when they exist.     

On an adjustment of degrees of freedom in the minimim mean squared error ertimator

In this paper, we consider an adjustment of degrees of freedom in the minimum mean squared error (MMSE) estimator, We derive the exact MSE of the adjusted MMSE (AMMSE) estimator, and compare the MSE of the AMMSE estimator with those of the Stein-(SR), positive-part Stein-rule (PSR) and MMSE estimators by numerical evaluations. It is shown that the adjustment of degrees of freedom is effective when the noncentrality parameter is close to zero, and the MSE performance of the MMSE estimator can be improved in the wide region of the noncentrality parameter by the adjustment, ft is also shown that the AMMSE estimator can have the smaller MSE than the PSR estimator in the wide region of the noncentrality parameter     

Generalized modulus-ratio tests for analysis of factorial designs with zero degrees of freedom for error

A method of statistical analysis of single replicate and fractional factorial designs requiring no estimate of error variance is given. By comparison of the relative magnitudes of independent effect .estimates, effects corresponding to relatively large effect estimates may be asserted to be nonzero. The procedure maintains a prespecified experimentwise error rate for a general class of modulus-ratio statistics.     

The bias of the mle of box's degrees of freedom correction factor for correlation in anova

Approximate expressions for the mean and variance of the MLE of Box's 2-way ANOVA degrees of freedom factor are given for the case when the usual F test for column effects is appropriate even though there is correlation across columns. A simulation study is performed showing the approximations are good for a variety of numbers of rows and columns. The results confirm that the estimated degrees of freedom factor is badly biased when no factor at all should be used.     

Note on an extension of rational bounds for the t-tail area to arbitrary degrees of freedom

The bounds of Soms (1980a) for the tail area of the t-distribution with integral degrees of freedom are extended to arbitrary positive degrees of freedom, Comparisons are made with the bounds of Shenton and Carpenter (1965) and some numerical examples are provided.     

Maximum likelihood vs. maximum goodness of fit estimation of the three-parameter Weibull distribution

The use of statistics based on the empirical distribution function is analysed for estimation of the scale, shape, and location parameters of the three-parameter Weibull distribution. The resulting maximum goodness of fit (MGF) estimators are compared with their maximum likelihood counterparts. In addition to the Kolmogorov–Smirnov, Cramer–von Mises, and Anderson–Darling statistics, some related empirical distribution function statistics using different weight functions are considered. The results show that the MGF estimators of the scale and shape parameters are usually more efficient than the maximum likelihood estimators when the shape parameter is smaller than 2, particularly if the sample size is large.     

Exact estimation of multiple directed acyclic graphs

This paper considers structure learning for multiple related directed acyclic graph (DAG) models. Building on recent developments in exact estimation of DAGs using integer linear programming (ILP), we present an ILP approach for joint estimation over multiple DAGs. Unlike previous work, we do not require that the vertices in each DAG share a common ordering. Furthermore, we allow for (potentially unknown) dependency structure between the DAGs. Results are presented on both simulated data and fMRI data obtained from multiple subjects.     

Data skeletons: simultaneous estimation of multiple quantiles for massive streaming datasets with applications to density estimation

We consider the problem of density estimation when the data is in the form of a continuous stream with no fixed length. In this setting, implementations of the usual methods of density estimation such as kernel density estimation are problematic. We propose a method of density estimation for massive datasets that is based upon taking the derivative of a smooth curve that has been fit through a set of quantile estimates. To achieve this, a low-storage, single-pass, sequential method is proposed for simultaneous estimation of multiple quantiles for massive datasets that form the basis of this method of density estimation. For comparison, we also consider a sequential kernel density estimator. The proposed methods are shown through simulation study to perform well and to have several distinct advantages over existing methods.     

Estimation of multiple gamma scale-parameters: bayes estimation subject to uniform domination

Simultaneous estimation of p gamma scale-parameters is considered under squared-error loss. The problem of minimizing, subject to uniform risk domination, the Bayes risk (or more generally the posterior expected loss) against certain conjugate or mixtures of conjugate priors is considered. Rather surprisingly, it is shown that the minimization can be done conditionally, thus avoiding variational arguments. Relative savings loss (and a posterior version thereof) are found, and it is found that in the most favorable situations, Bayesian robustness can be achieved without sacrificing substantial subjective Bayesian gains.     

Nonparametric estimation of a multivariate multiple regression function

Let (X₁, X₂, Y₁, Y₂) be a four dimensional random variable having the joint probability density function f(x₁, x₂, y₁, y₂). In this paper we consider the problem of estimating the regression function \({{E[(_{Y_2 }^{Y_1 } )} \mathord{\left/ {\vphantom {{E[(_{Y_2 }^{Y_1 } )} {_{X_2 = X_2 }^{X_1 = X_1 } }}} \right. \kern-0em} {_{X_2 = X_2 }^{X_1 = X_1 } }}]\) on the basis of a random sample of size n. We have proved that under certain regularity conditions the kernel estimate of this regression function is uniformly strongly consistent. We have also shown that under certain conditions the estimate is asymptotically normally distributed.     

Interval estimation for conformance proportions of multiple quality characteristics

A conformance proportion is an important and useful index to assess industrial quality improvement. Statistical confidence limits for a conformance proportion are usually required not only to perform statistical significance tests, but also to provide useful information for determining practical significance. In this article, we propose approaches for constructing statistical confidence limits for a conformance proportion of multiple quality characteristics. Under the assumption that the variables of interest are distributed with a multivariate normal distribution, we develop an approach based on the concept of a fiducial generalized pivotal quantity (FGPQ). Without any distribution assumption on the variables, we apply some confidence interval construction methods for the conformance proportion by treating it as the probability of a success in a binomial distribution. The performance of the proposed methods is evaluated through detailed simulation studies. The results reveal that the simulated coverage probability (cp) for the FGPQ-based method is generally larger than the claimed value. On the other hand, one of the binomial distribution-based methods, that is, the standard method suggested in classical textbooks, appears to have smaller simulated cps than the nominal level. Two alternatives to the standard method are found to maintain their simulated cps sufficiently close to the claimed level, and hence their performances are judged to be satisfactory. In addition, three examples are given to illustrate the application of the proposed methods.     

Isotonic estimation of the intensity of a nonhomogeneous Poisson process: The multiple realization setup

In the 1950s Brunk and Van Eeden each obtained maximum-likelihood estimators of a finite product of probability density functions under partial or complete ordering of their parameters. Their results play an important role in the general theory of inference under order restrictions and lead to an isotonic estimator of the intensity of a nonhomogeneous Poisson process. Here an elementary derivation of the maximum likelihood estimator (m.l.e.) for the intensity of a nonhomogeneous Poisson process is given when several (possibly censored) realizations are available. Boswell obtained the m.l.e. based on a single realization as well as a conditional m.l.e. under the same conditions. An example is given to show that in the multirealization setup a conditional m.l.e. may not exist; the proofs are, we believe, new and elementary. An illustrative application is given.     

A bivariate F distribution with marginals on arbitrary numerator and denominator degrees of freedom, and related bivariate beta and t distributions

The classical bivariate F distribution arises from ratios of chi-squared random variables with common denominators. A consequent disadvantage is that its univariate F marginal distributions have one degree of freedom parameter in common. In this paper, we add a further independent chi-squared random variable to the denominator of one of the ratios and explore the extended bivariate F distribution, with marginals on arbitrary degrees of freedom, that results. Transformations linking F, beta and skew t distributions are then applied componentwise to produce bivariate beta and skew t distributions which also afford marginal (beta and skew t) distributions with arbitrary parameter values. We explore a variety of properties of these distributions and give an example of a potential application of the bivariate beta distribution in Bayesian analysis.