Adaptive model selection using empirical complexities

Given $n$ independent replicates of a jointly distributed pair $(X,Y)\in {\cal R}^d \times {\cal R}$, we wish to select from a fixed sequence of model classes ${\cal F}_1, {\cal F}_2, \ldots$ a deterministic prediction rule $f: {\cal R}^d \to {\cal R}$ whose risk is small. We investigate the possibility of empirically assessingthe {\em complexity} of each model class, that is, the actual difficulty of the estimation problem within each class. The estimated complexities are in turn used to define an adaptive model selection procedure, which is based on complexity penalized empirical risk.The available data are divided into two parts. The first is used to form an empirical cover of each model class, and the second is used to select a candidate rule from each cover based on empirical risk. The covering radii are determined empirically to optimize a tight upper bound on the estimation error. An estimate is chosen from the list of candidates in order to minimize the sum of class complexity and empirical risk. A distinguishing feature of the approach is that the complexity of each model class is assessed empirically, based on the size of its empirical cover.Finite sample performance bounds are established for the estimates, and these bounds are applied to several non-parametric estimation problems. The estimates are shown to achieve a favorable tradeoff between approximation and estimation error, and to perform as well as if the distribution-dependent complexities of the model classes were known beforehand. In addition, it is shown that the estimate can be consistent,and even possess near optimal rates of convergence, when each model class has an infinite VC or pseudo dimension.For regression estimation with squared loss we modify our estimate to achieve a faster rate of convergence.

A new light on Minkowski's? $(x)$ function

The well--known Minkowski's? $(x)$ function is presented as the asymptotic distribution function of an enumeration of the rationals in (0,1] based on their continued fraction representation. Besides, the singularity of ?$(x)$ is clearly proved in two ways: by exhibiting a set of measure one in which ?ï$(x)$ = 0; and again by actually finding a set of measure one which is mapped onto a set of measure zero and viceversa. These sets are described by means of metrical properties of different systems for real number representation.

The colonial origins of Civil War

The dominant hypothesis in the literature that studies conflict is that poverty is the main cause of civil wars. We instead analyze the effect of institutions on civil war, controlling for income per capita. In our set up, institutions are endogenous and colonial origins affect civil wars through their legacy on institutions. Our results indicate that institutions, proxied by the protection of property rights, rule of law and the efficiency of the legal system, are a fundamental cause of civil war. In particular, an improvement in institutions from the median value in the sample to the 75th percentile is associated with a 38 percentage points reduction in the incidence of civil wars. Moreover, once institutions are included as explaining civil wars, income does not have any effect on civil war, either directly or indirectly.

Measuring subcompositional incoherence

Subcompositional coherence is a fundamental property of Aitchison s approach to compositional data analysis, and is the principal justification for using ratios of components. We maintain, however, that lack of subcompositional coherence, that is incoherence, can be measured in an attempt to evaluate whether any given technique is close enough, for all practical purposes, to being subcompositionally coherent. This opens up the field to alternative methods, which might be better suited to cope with problems such as data zeros and outliers, while being only slightly incoherent. The measure that we propose is based on the distance measure between components. We show that the two-part subcompositions, which appear to be the most sensitive to subcompositional incoherence, can be used to establish a distance matrix which can be directly compared with the pairwise distances in the full composition. The closeness of these two matrices can be quantified using a stress measure that is common in multidimensional scaling, providing a measure of subcompositional incoherence. The approach is illustrated using power-transformed correspondence analysis, which has already been shown to converge to log-ratio analysis as the power transform tends to zero.

Crecimiento de las regiones españolas: estructura sectorial, dinámica regional y distribución de rentas.

El análisis de las regiones españolas en el período 1980-1995 indicaque la composición sectorial explica la mayor parte de la evolución del empleo y de las diferencias en productividad, salarios medios y participación de las rentas del trabajo. Para el VAB el componenteregional es más importante que el sectorial, aunque éste no esdespreciable. Nuestro análisis permite identificar a lo largo del tiempo aquellas regiones que han crecido más (menos) que lo esperado dada su composición sectorial. Identificamos una clara relación inversa entre la participación de las rentas del trabajo en el producto y el componente puramente regional del crecimiento del empleo. Sin embargo no observamos relación entre la tasa de paro y la distribución del producto. Ello sugiere que los salarios son poco elásticos a las condiciones del mercado de trabajo, pero el crecimiento del empleo sí lo es a la evolución de las rentas del capital de la región.

A scaled difference chi-square test statistic for moment structure analysis

A family of scaling corrections aimed to improve the chi-square approximation of goodness-of-fit test statistics in small samples, large models, and nonnormal data was proposed in Satorra and Bentler (1994). For structural equations models, Satorra-Bentler's (SB) scaling corrections are available in standard computer software. Often, however, the interest is not on the overall fit of a model, but on a test of the restrictions that a null model say ${\cal M}_0$ implies on a less restricted one ${\cal M}_1$. If $T_0$ and $T_1$ denote the goodness-of-fit test statistics associated to ${\cal M}_0$ and ${\cal M}_1$, respectively, then typically the difference $T_d = T_0 - T_1$ is used as a chi-square test statistic with degrees of freedom equal to the difference on the number of independent parameters estimated under the models ${\cal M}_0$ and ${\cal M}_1$. As in the case of the goodness-of-fit test, it is of interest to scale the statistic $T_d$ in order to improve its chi-square approximation in realistic, i.e., nonasymptotic and nonnormal, applications. In a recent paper, Satorra (1999) shows that the difference between two Satorra-Bentler scaled test statistics for overall model fit does not yield the correct SB scaled difference test statistic. Satorra developed an expression that permits scaling the difference test statistic, but his formula has some practical limitations, since it requires heavy computations that are notavailable in standard computer software. The purpose of the present paper is to provide an easy way to compute the scaled difference chi-square statistic from the scaled goodness-of-fit test statistics of models ${\cal M}_0$ and ${\cal M}_1$. A Monte Carlo study is provided to illustrate the performance of the competing statistics.

Analysis of matched matrices

We consider the joint visualization of two matrices which have common rowsand columns, for example multivariate data observed at two time pointsor split accord-ing to a dichotomous variable. Methods of interest includeprincipal components analysis for interval-scaled data, or correspondenceanalysis for frequency data or ratio-scaled variables on commensuratescales. A simple result in matrix algebra shows that by setting up thematrices in a particular block format, matrix sum and difference componentscan be visualized. The case when we have more than two matrices is alsodiscussed and the methodology is applied to data from the InternationalSocial Survey Program.

Measuring asymmetries in brand associations using correspondence analysis

Correspondence analysis is introduced in the brand associationliterature as an alternative tool to measure dominance, for theparticular case of free choice data. The method is also used to analysedifferences, or asymmetries, between brand-attribute associations whereattributes are associated with evoked brands, and brand-attributeassociations where brands are associated with the attributes. Anapplication to a sample of deodorants is used to illustrate the proposedmethodology.

An inequality for uniform deviations of sample averages from their means

We derive a new inequality for uniform deviations of averages from their means. The inequality is a common generalization of previous results of Vapnik and Chervonenkis (1974) and Pollard (1986). Usingthe new inequality we obtain tight bounds for empirical loss minimization learning.

Small-area estimation with spatial similarity

A class of composite estimators of small area quantities that exploit spatial (distancerelated)similarity is derived. It is based on a distribution-free model for the areas, but theestimators are aimed to have optimal design-based properties. Composition is applied alsoto estimate some of the global parameters on which the small area estimators depend.It is shown that the commonly adopted assumption of random effects is not necessaryfor exploiting the similarity of the districts (borrowing strength across the districts). Themethods are applied in the estimation of the mean household sizes and the proportions ofsingle-member households in the counties (comarcas) of Catalonia. The simplest version ofthe estimators is more efficient than the established alternatives, even though the extentof spatial similarity is quite modest.

Fighting against Malaria: Prevent wars while waiting for the "miraculous" vaccine

The World Health Organization estimates that 300 million clinical cases of malaria occur annually and observed that during the 80's and part of the 90's its incidence increased. In this paper we explore the influence of refugees from civil wars on the incidence of malaria in the refugee-receiving countries. Using civil wars as an instrumental variable we show that for each 1,000 refugees there are between 2,000 and 2,700 cases of malaria in the refugee receiving country. On average 13% of the cases of malaria reported by the WHO are caused by forced migration as a consequence of civil wars.

Estimating parliamentary composition through electoral polls

Any electoral system has an electoral formula that converts voteproportions into parliamentary seats. Pre-electoral polls usually focuson estimating vote proportions and then applying the electoral formulato give a forecast of the parliament's composition. We here describe theproblems arising from this approach: there is always a bias in theforecast. We study the origin of the bias and some methods to evaluateand to reduce it. We propose some rules to compute the sample sizerequired for a given forecast accuracy. We show by Monte Carlo simulationthe performance of the proposed methods using data from Spanish electionsin last years. We also propose graphical methods to visualize how electoralformulae and parliamentary forecasts work (or fail).

University admission marks in Catalonia: Some highlights from the empirical research

The results of the examinations taken by graduated high school studentswho want to enrol at a Catalan university are here studied. To do so,the authors address several issues related to the equity of the system:reliability of grading, difficulty and discrimination power of the exams.The general emphasis is put upon the concurrent research and empiricalevidence about the properties of the examination items and scores. Aftera discussion about the limitations of the exams' format and appropriatenessof the instruments used in the study, the article concludes with somesuggestions to improve such examinations.

Inference with the lognormal distribution

Several estimators of the expectation, median and mode of the lognormal distribution are derived. They aim to be approximately unbiased, efficient, or have a minimax property in the class of estimators we introduce. The small-sample properties of these estimators are assessed by simulations and, when possible, analytically. Some of these estimators of the expectation are far more efficient than the maximum likelihood or the minimum-variance unbiased estimator, even for substantial samplesizes.

On the short-time behavior of the implied volatility for jump-diffusion models with stochastic volatility

In this paper we use Malliavin calculus techniques to obtain an expression for the short-time behavior of the at-the-money implied volatility skew for a generalization of the Bates model, where the volatility does not need to be neither a difussion, nor a Markov process as the examples in section 7 show. This expression depends on the derivative of the volatility in the sense of Malliavin calculus.