Inférence robuste à la présence des valeurs aberrantes dans les enquêtes

Cette thèse comporte trois articles dont un est publié et deux en préparation. Le sujet central de la thèse porte sur le traitement des valeurs aberrantes représentatives dans deux aspects importants des enquêtes que sont : l’estimation des petits domaines et l’imputation en présence de non-réponse partielle. En ce qui concerne les petits domaines, les estimateurs robustes dans le cadre des modèles au niveau des unités ont été étudiés. Sinha & Rao (2009) proposent une version robuste du meilleur prédicteur linéaire sans biais empirique pour la moyenne des petits domaines. Leur estimateur robuste est de type «plugin», et à la lumière des travaux de Chambers (1986), cet estimateur peut être biaisé dans certaines situations. Chambers et al. (2014) proposent un estimateur corrigé du biais. En outre, un estimateur de l’erreur quadratique moyenne a été associé à ces estimateurs ponctuels. Sinha & Rao (2009) proposent une procédure bootstrap paramétrique pour estimer l’erreur quadratique moyenne. Des méthodes analytiques sont proposées dans Chambers et al. (2014). Cependant, leur validité théorique n’a pas été établie et leurs performances empiriques ne sont pas pleinement satisfaisantes. Ici, nous examinons deux nouvelles approches pour obtenir une version robuste du meilleur prédicteur linéaire sans biais empirique : la première est fondée sur les travaux de Chambers (1986), et la deuxième est basée sur le concept de biais conditionnel comme mesure de l’influence d’une unité de la population. Ces deux classes d’estimateurs robustes des petits domaines incluent également un terme de correction pour le biais. Cependant, ils utilisent tous les deux l’information disponible dans tous les domaines contrairement à celui de Chambers et al. (2014) qui utilise uniquement l’information disponible dans le domaine d’intérêt. Dans certaines situations, un biais non négligeable est possible pour l’estimateur de Sinha & Rao (2009), alors que les estimateurs proposés exhibent un faible biais pour un choix approprié de la fonction d’influence et de la constante de robustesse. Les simulations Monte Carlo sont effectuées, et les comparaisons sont faites entre les estimateurs proposés et ceux de Sinha & Rao (2009) et de Chambers et al. (2014). Les résultats montrent que les estimateurs de Sinha & Rao (2009) et de Chambers et al. (2014) peuvent avoir un biais important, alors que les estimateurs proposés ont une meilleure performance en termes de biais et d’erreur quadratique moyenne. En outre, nous proposons une nouvelle procédure bootstrap pour l’estimation de l’erreur quadratique moyenne des estimateurs robustes des petits domaines. Contrairement aux procédures existantes, nous montrons formellement la validité asymptotique de la méthode bootstrap proposée. Par ailleurs, la méthode proposée est semi-paramétrique, c’est-à-dire, elle n’est pas assujettie à une hypothèse sur les distributions des erreurs ou des effets aléatoires. Ainsi, elle est particulièrement attrayante et plus largement applicable. Nous examinons les performances de notre procédure bootstrap avec les simulations Monte Carlo. Les résultats montrent que notre procédure performe bien et surtout performe mieux que tous les compétiteurs étudiés. Une application de la méthode proposée est illustrée en analysant les données réelles contenant des valeurs aberrantes de Battese, Harter & Fuller (1988). S’agissant de l’imputation en présence de non-réponse partielle, certaines formes d’imputation simple ont été étudiées. L’imputation par la régression déterministe entre les classes, qui inclut l’imputation par le ratio et l’imputation par la moyenne sont souvent utilisées dans les enquêtes. Ces méthodes d’imputation peuvent conduire à des estimateurs imputés biaisés si le modèle d’imputation ou le modèle de non-réponse n’est pas correctement spécifié. Des estimateurs doublement robustes ont été développés dans les années récentes. Ces estimateurs sont sans biais si l’un au moins des modèles d’imputation ou de non-réponse est bien spécifié. Cependant, en présence des valeurs aberrantes, les estimateurs imputés doublement robustes peuvent être très instables. En utilisant le concept de biais conditionnel, nous proposons une version robuste aux valeurs aberrantes de l’estimateur doublement robuste. Les résultats des études par simulations montrent que l’estimateur proposé performe bien pour un choix approprié de la constante de robustesse.

Using graphical probability analysis (Bayes Nets) to evaluate a conditional DNA inclusion

This paper discusses the analysis of cases in which the inclusion or exclusion of a particular suspect, as a possible contributor to a DNA mixture, depends on the value of a variable (the number of contributors) that cannot be determined with certainty. It offers alternative ways to deal with such cases, including sensitivity analysis and object-oriented Bayesian networks, that separate uncertainty about the inclusion of the suspect from uncertainty about other variables. The paper presents a case study in which the value of DNA evidence varies radically depending on the number of contributors to a DNA mixture: if there are two contributors, the suspect is excluded; if there are three or more, the suspect is included; but the number of contributors cannot be determined with certainty. It shows how an object-oriented Bayesian network can accommodate and integrate varying perspectives on the unknown variable and how it can reduce the potential for bias by directing attention to relevant considerations and distinguishing different sources of uncertainty. It also discusses the challenge of presenting such evidence to lay audiences.

The Inflation Bias When the Central Bank Targets, the Natural Rate of Unemployment

This paper studies the proposition that an inflation bias can arise in a setup where a central banker with asymmetric preferences targets the natural unemployment rate. Preferences are asymmetric in the sense that positive unemployment deviations from the natural rate are weighted more (or less) severely than negative deviations in the central banker's loss function. The bias is proportional to the conditional variance of unemployment. The time-series predictions of the model are evaluated using data from G7 countries. Econometric estimates support the prediction that the conditional variance of unemployment and the rate of inflation are positively related.

Bayesian Estimation for Performance Measures of Two Diagnostic Tests in the Presence of Verification Bias

Sensitivity and specificity are measures that allow us to evaluate the performance of a diagnostic test. In practice, it is common to have situations where a proportion of selected individuals cannot have the real state of the disease verified, since the verification could be an invasive procedure, as occurs with biopsy. This happens, as a special case, in the diagnosis of prostate cancer, or in any other situation related to risks, that is, not practicable, nor ethical, or in situations with high cost. For this case, it is common to use diagnostic tests based only on the information of verified individuals. This procedure can lead to biased results or workup bias. In this paper, we introduce a Bayesian approach to estimate the sensitivity and the specificity for two diagnostic tests considering verified and unverified individuals, a result that generalizes the usual situation based on only one diagnostic test.

A panel data approach to economic forecasting: the bias-corrected average forecast

In this paper, we propose a novel approach to econometric forecasting of stationary and ergodic time series within a panel-data framework. Our key element is to employ the (feasible) bias-corrected average forecast. Using panel-data sequential asymptotics we show that it is potentially superior to other techniques in several contexts. In particular, it is asymptotically equivalent to the conditional expectation, i.e., has an optimal limiting mean-squared error. We also develop a zeromean test for the average bias and discuss the forecast-combination puzzle in small and large samples. Monte-Carlo simulations are conducted to evaluate the performance of the feasible bias-corrected average forecast in finite samples. An empirical exercise based upon data from a well known survey is also presented. Overall, theoretical and empirical results show promise for the feasible bias-corrected average forecast.

A panel data approach to economic forecasting: the bias-corrected average forecast

In this paper, we propose a novel approach to econometric forecasting of stationary and ergodic time series within a panel-data framework. Our key element is to employ the (feasible) bias-corrected average forecast. Using panel-data sequential asymptotics we show that it is potentially superior to other techniques in several contexts. In particular, it is asymptotically equivalent to the conditional expectation, i.e., has an optimal limiting mean-squared error. We also develop a zeromean test for the average bias and discuss the forecast-combination puzzle in small and large samples. Monte-Carlo simulations are conducted to evaluate the performance of the feasible bias-corrected average forecast in finite samples. An empirical exercise, based upon data from a well known survey is also presented. Overall, these results show promise for the feasible bias-corrected average forecast.

A partial dependence factorial analysis to deal with selection bias in observational studis

This thesis presents a creative and practical approach to dealing with the problem of selection bias. Selection bias may be the most important vexing problem in program evaluation or in any line of research that attempts to assert causality. Some of the greatest minds in economics and statistics have scrutinized the problem of selection bias, with the resulting approaches – Rubin’s Potential Outcome Approach(Rosenbaum and Rubin,1983; Rubin, 1991,2001,2004) or Heckman’s Selection model (Heckman, 1979) – being widely accepted and used as the best fixes. These solutions to the bias that arises in particular from self selection are imperfect, and many researchers, when feasible, reserve their strongest causal inference for data from experimental rather than observational studies. The innovative aspect of this thesis is to propose a data transformation that allows measuring and testing in an automatic and multivariate way the presence of selection bias. The approach involves the construction of a multi-dimensional conditional space of the X matrix in which the bias associated with the treatment assignment has been eliminated. Specifically, we propose the use of a partial dependence analysis of the X-space as a tool for investigating the dependence relationship between a set of observable pre-treatment categorical covariates X and a treatment indicator variable T, in order to obtain a measure of bias according to their dependence structure. The measure of selection bias is then expressed in terms of inertia due to the dependence between X and T that has been eliminated. Given the measure of selection bias, we propose a multivariate test of imbalance in order to check if the detected bias is significant, by using the asymptotical distribution of inertia due to T (Estadella et al. 2005) , and by preserving the multivariate nature of data. Further, we propose the use of a clustering procedure as a tool to find groups of comparable units on which estimate local causal effects, and the use of the multivariate test of imbalance as a stopping rule in choosing the best cluster solution set. The method is non parametric, it does not call for modeling the data, based on some underlying theory or assumption about the selection process, but instead it calls for using the existing variability within the data and letting the data to speak. The idea of proposing this multivariate approach to measure selection bias and test balance comes from the consideration that in applied research all aspects of multivariate balance, not represented in the univariate variable- by-variable summaries, are ignored. The first part contains an introduction to evaluation methods as part of public and private decision process and a review of the literature of evaluation methods. The attention is focused on Rubin Potential Outcome Approach, matching methods, and briefly on Heckman’s Selection Model. The second part focuses on some resulting limitations of conventional methods, with particular attention to the problem of how testing in the correct way balancing. The third part contains the original contribution proposed , a simulation study that allows to check the performance of the method for a given dependence setting and an application to a real data set. Finally, we discuss, conclude and explain our future perspectives.

Association between reported exposure to road traffic and respiratory symptoms in children: evidence of bias

BACKGROUND: Many studies showing effects of traffic-related air pollution on health rely on self-reported exposure, which may be inaccurate. We estimated the association between self-reported exposure to road traffic and respiratory symptoms in preschool children, and investigated whether the effect could have been caused by reporting bias. METHODS: In a random sample of 8700 preschool children in Leicestershire, UK, exposure to road traffic and respiratory symptoms were assessed by a postal questionnaire (response rate 80%). The association between traffic exposure and respiratory outcomes was assessed using unconditional logistic regression and conditional regression models (matching by postcode). RESULTS: Prevalence odds ratios (95% confidence intervals) for self-reported road traffic exposure, comparing the categories 'moderate' and 'dense', respectively, with 'little or no' were for current wheezing: 1.26 (1.13-1.42) and 1.30 (1.09-1.55); chronic rhinitis: 1.18 (1.05-1.31) and 1.31 (1.11-1.56); night cough: 1.17 (1.04-1.32) and 1.36 (1.14-1.62); and bronchodilator use: 1.20 (1.04-1.38) and 1.18 (0.95-1.46). Matched analysis only comparing symptomatic and asymptomatic children living at the same postcode (thus exposed to similar road traffic) showed similar ORs, suggesting that parents of children with respiratory symptoms reported more road traffic than parents of asymptomatic children. CONCLUSIONS: Our study suggests that reporting bias could explain some or even all the association between reported exposure to road traffic and disease. Over-reporting of exposure by only 10% of parents of symptomatic children would be sufficient to produce the effect sizes shown in this study. Future research should be based only on objective measurements of traffic exposure.

On the Behaviour of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models

In linear mixed models, model selection frequently includes the selection of random effects. Two versions of the Akaike information criterion (AIC) have been used, based either on the marginal or on the conditional distribution. We show that the marginal AIC is no longer an asymptotically unbiased estimator of the Akaike information, and in fact favours smaller models without random effects. For the conditional AIC, we show that ignoring estimation uncertainty in the random effects covariance matrix, as is common practice, induces a bias that leads to the selection of any random effect not predicted to be exactly zero. We derive an analytic representation of a corrected version of the conditional AIC, which avoids the high computational cost and imprecision of available numerical approximations. An implementation in an R package is provided. All theoretical results are illustrated in simulation studies, and their impact in practice is investigated in an analysis of childhood malnutrition in Zambia.

REMOVING TECHNICAL VARIABILITY IN RNA-SEQ DATA USING CONDITIONAL QUANTILE NORMALIZATION

The ability to measure gene expression on a genome-wide scale is one of the most promising accomplishments in molecular biology. Microarrays, the technology that first permitted this, were riddled with problems due to unwanted sources of variability. Many of these problems are now mitigated, after a decade’s worth of statistical methodology development. The recently developed RNA sequencing (RNA-seq) technology has generated much excitement in part due to claims of reduced variability in comparison to microarrays. However, we show RNA-seq data demonstrates unwanted and obscuring variability similar to what was first observed in microarrays. In particular, we find GC-content has a strong sample specific effect on gene expression measurements that, if left uncorrected, leads to false positives in downstream results. We also report on commonly observed data distortions that demonstrate the need for data normalization. Here we describe statistical methodology that improves precision by 42% without loss of accuracy. Our resulting conditional quantile normalization (CQN) algorithm combines robust generalized regression to remove systematic bias introduced by deterministic features such as GC-content, and quantile normalization to correct for global distortions.

Effects of stereotypic behaviour and chronic mild stress on judgement bias in laboratory mice

Cognitive processes are influenced by underlying affective states, and tests of cognitive bias have recently been developed to assess the valence of affective states in animals. These tests are based on the fact that individuals in a negative affective state interpret ambiguous stimuli more pessimistically than individuals in a more positive state. Using two strains of mice we explored whether unpredictable chronic mild stress (UCMS) can induce a negative judgement bias and whether variation in the expression of stereotypic behaviour is associated with variation in judgement bias. Sixteen female CD-1 and 16 female C57BL/6 mice were trained on a tactile conditional discrimination test with grade of sandpaper as a cue for differential food rewards. Once they had learned the discrimination, half of the mice were subjected to UCMS for three weeks to induce a negative affective state. Although UCMS induced a reduced preference for the higher value reward in the judgement bias test, it did not affect saccharine preference or hypothalamic–pituitary–adrenal (HPA) activity. However, UCMS affected responses to ambiguous (intermediate) cues in the judgement bias test. While control mice showed a graded response to ambiguous cues, UCMS mice of both strains did not discriminate between ambiguous cues and tended to show shorter latencies to the ambiguous cues and the negative reference cue. UCMS also increased bar-mouthing in CD-1, but not in C57BL/6 mice. Furthermore, mice with higher levels of stereotypic behaviour made more optimistic choices in the judgement bias test. However, no such relationship was found for stereotypic bar-mouthing, highlighting the importance of investigating different types of stereotypic behaviour separately.

Estimating conditional volatility with neural networks

It is well known that one of the obstacles to effective forecasting of exchange rates is heteroscedasticity (non-stationary conditional variance). The autoregressive conditional heteroscedastic (ARCH) model and its variants have been used to estimate a time dependent variance for many financial time series. However, such models are essentially linear in form and we can ask whether a non-linear model for variance can improve results just as non-linear models (such as neural networks) for the mean have done. In this paper we consider two neural network models for variance estimation. Mixture Density Networks (Bishop 1994, Nix and Weigend 1994) combine a Multi-Layer Perceptron (MLP) and a mixture model to estimate the conditional data density. They are trained using a maximum likelihood approach. However, it is known that maximum likelihood estimates are biased and lead to a systematic under-estimate of variance. More recently, a Bayesian approach to parameter estimation has been developed (Bishop and Qazaz 1996) that shows promise in removing the maximum likelihood bias. However, up to now, this model has not been used for time series prediction. Here we compare these algorithms with two other models to provide benchmark results: a linear model (from the ARIMA family), and a conventional neural network trained with a sum-of-squares error function (which estimates the conditional mean of the time series with a constant variance noise model). This comparison is carried out on daily exchange rate data for five currencies.

The conditional performance of euro bond funds: evidence from Portugal during the debt crisis

This paper evaluates the performance of a survivorship bias-free data set of Portuguese funds investing in Euro-denominated bonds by using conditional models that consider the public information available to investors when the returns are generated. We find that bond funds underperform the market significantly and by an economically relevant magnitude. This underperformance cannot be explained by the expenses they charge. Our findings support the use of conditional performance evaluation models, since we find strong evidence of both time-varying risk and performance, dependent on the slope of the term structure and the inverse relative wealth variables. We also show that survivorship bias has a significant impact on performance estimates. Furthermore, during the European debt crisis, bond fund managers performed significantly better than in non-crisis periods and were able to achieve neutral performance. This improved performance throughout the crisis seems to be related to changes in funds’ investment styles.

Hubble parameter reconstruction from a principal component analysis: minimizing the bias

Aims. A model-independent reconstruction of the cosmic expansion rate is essential to a robust analysis of cosmological observations. Our goal is to demonstrate that current data are able to provide reasonable constraints on the behavior of the Hubble parameter with redshift, independently of any cosmological model or underlying gravity theory. Methods. Using type Ia supernova data, we show that it is possible to analytically calculate the Fisher matrix components in a Hubble parameter analysis without assumptions about the energy content of the Universe. We used a principal component analysis to reconstruct the Hubble parameter as a linear combination of the Fisher matrix eigenvectors (principal components). To suppress the bias introduced by the high redshift behavior of the components, we considered the value of the Hubble parameter at high redshift as a free parameter. We first tested our procedure using a mock sample of type Ia supernova observations, we then applied it to the real data compiled by the Sloan Digital Sky Survey (SDSS) group. Results. In the mock sample analysis, we demonstrate that it is possible to drastically suppress the bias introduced by the high redshift behavior of the principal components. Applying our procedure to the real data, we show that it allows us to determine the behavior of the Hubble parameter with reasonable uncertainty, without introducing any ad-hoc parameterizations. Beyond that, our reconstruction agrees with completely independent measurements of the Hubble parameter obtained from red-envelope galaxies.

Experimental studies of di-jet survival and surface emission bias in Au plus Au collisions via angular correlations with respect to back-to-back leading hadrons

We report first results from an analysis based on a new multi-hadron correlation technique, exploring jet-medium interactions and di-jet surface emission bias at the BNL Relativistic Heavy Ion Collider (RHIC). Pairs of back-to-back high-transverse-momentum hadrons are used for triggers to study associated hadron distributions. In contrast with two-and three-particle correlations with a single trigger with similar kinematic selections, the associated hadron distribution of both trigger sides reveals no modification in either relative pseudorapidity Delta eta or relative azimuthal angle Delta phi from d + Au to central Au + Au collisions. We determine associated hadron yields and spectra as well as production rates for such correlated back-to-back triggers to gain additional insights on medium properties.