Correlated variables in regression: Clustering and sparse estimation

We consider estimation in a high-dimensional linear model with strongly correlated variables. We propose to cluster the variables first and do subsequent sparse estimation such as the Lasso for cluster-representatives or the group Lasso based on the structure from the clusters. Regarding the first step, we present a novel and bottom-up agglomerative clustering algorithm based on canonical correlations, and we show that it finds an optimal solution and is statistically consistent. We also present some theoretical arguments that canonical correlation based clustering leads to a better-posed compatibility constant for the design matrix which ensures identifiability and an oracle inequality for the group Lasso. Furthermore, we discuss circumstances where cluster-representatives and using the Lasso as subsequent estimator leads to improved results for prediction and detection of variables. We complement the theoretical analysis with various empirical results.     

The pitfall of instrumental variables in big data: What the rule of thumb can't give you

ABSTRACT

Background: Instrumental variables (IVs) have become much easier to find in the “Big data era” which has increased the number of applications of the Two-Stage Least Squares model (TSLS). With the increased availability of IVs, the possibility that these IVs are weak has increased. Prior work has suggested a ‘rule of thumb’ that IVs with a first stage F statistic at least ten will avoid a relative bias in point estimates greater than 10%. We investigated whether or not this threshold was also an efficient guarantee of low false rejection rates of the null hypothesis test in TSLS applications with many IVs.

Objective: To test how the ‘rule of thumb’ for weak instruments performs in predicting low false rejection rates in the TSLS model when the number of IVs is large.

Method: We used a Monte Carlo approach to create 28 original data sets for different models with the number of IVs varying from 3 to 30. For each model, we generated 2000 observations for each iteration and conducted 50,000 iterations to reach convergence in rejection rates. The point estimate was set to 0, and probabilities of rejecting this hypothesis were recorded for each model as a measurement of false rejection rate. The relationship between the endogenous variable and IVs was carefully adjusted to let the F statistics for the first stage model equal ten, thus simulating the ‘rule of thumb.’

Results: We found that the false rejection rates (type I errors) increased when the number of IVs in the TSLS model increased while holding the F statistics for the first stage model equal to 10. The false rejection rate exceeds 10% when TLSL has 24 IVs and exceed 15% when TLSL has 30 IVs.

Conclusion: When more instrumental variables were applied in the model, the ‘rule of thumb’ was no longer an efficient guarantee for good performance in hypothesis testing. A more restricted margin for F statistics is recommended to replace the ‘rule of thumb,’ especially when the number of instrumental variables is large.     

Cluster Data Streams with Noisy Variables

Clustering algorithms are important methods widely used in mining data streams because of their abilities to deal with infinite data flows. Although these algorithms perform well to mining latent relationship in data streams, most of them suffer from loss of cluster purity and become unstable when the inputting data streams have too many noisy variables. In this article, we propose a clustering algorithm to cluster data streams with noisy variables. The result from simulation shows that our proposal method is better than previous studies by adding a process of variable selection as a component in clustering algorithms. The results of two experiments indicate that clustering data streams with the process of variable selection are more stable and have better purity than those without such process. Another experiment testing KDD-CUP99 dataset also shows that our algorithm can generate more stable result.     

Clustering large number of extragalactic spectra of galaxies and quasars through canopies

Abstract

Cluster analysis is the distribution of objects into different groups or more precisely the partitioning of a data set into subsets (clusters) so that the data in subsets share some common trait according to some distance measure. Unlike classification, in clustering one has to first decide the optimum number of clusters and then assign the objects into different clusters. Solution of such problems for a large number of high dimensional data points is quite complicated and most of the existing algorithms will not perform properly. In the present work a new clustering technique applicable to large data set has been used to cluster the spectra of 702248 galaxies and quasars having 1,540 points in wavelength range imposed by the instrument. The proposed technique has successfully discovered five clusters from this 702,248X1,540 data matrix.     

Clustering objects on subsets of attributes (with discussion)

Summary.  A new procedure is proposed for clustering attribute value data. When used in conjunction with conventional distance-based clustering algorithms this procedure encourages those algorithms to detect automatically subgroups of objects that preferentially cluster on subsets of the attribute variables rather than on all of them simultaneously. The relevant attribute subsets for each individual cluster can be different and partially (or completely) overlap with those of other clusters. Enhancements for increasing sensitivity for detecting especially low cardinality groups clustering on a small subset of variables are discussed. Applications in different domains, including gene expression arrays, are presented.     

Modelling the role of variables in model-based cluster analysis

In the framework of cluster analysis based on Gaussian mixture models, it is usually assumed that all the variables provide information about the clustering of the sample units. Several variable selection procedures are available in order to detect the structure of interest for the clustering when this structure is contained in a variable sub-vector. Currently, in these procedures a variable is assumed to play one of (up to) three roles: (1) informative, (2) uninformative and correlated with some informative variables, (3) uninformative and uncorrelated with any informative variable. A more general approach for modelling the role of a variable is proposed by taking into account the possibility that the variable vector provides information about more than one structure of interest for the clustering. This approach is developed by assuming that such information is given by non-overlapped and possibly correlated sub-vectors of variables; it is also assumed that the model for the variable vector is equal to a product of conditionally independent Gaussian mixture models (one for each variable sub-vector). Details about model identifiability, parameter estimation and model selection are provided. The usefulness and effectiveness of the described methodology are illustrated using simulated and real datasets.     

Clustering of Variables Around Latent Components

Abstract

Clustering of variables around latent components is investigated as a means to organize multivariate data into meaningful structures. The coverage includes (i) the case where it is desirable to lump together correlated variables no matter whether the correlation coefficient is positive or negative; (ii) the case where negative correlation shows high disagreement among variables; (iii) an extension of the clustering techniques which makes it possible to explain the clustering of variables taking account of external data. The strategy basically consists in performing a hierarchical cluster analysis, followed by a partitioning algorithm. Both algorithms aim at maximizing the same criterion which reflects the extent to which variables in each cluster are related to the latent variable associated with this cluster. Illustrations are outlined using real data sets from sensory studies.     

Penalized estimation in additive varying coefficient models using grouped regularization

Additive varying coefficient models are a natural extension of multiple linear regression models, allowing the regression coefficients to be functions of other variables. Therefore these models are more flexible to model more complex dependencies in data structures. In this paper we consider the problem of selecting in an automatic way the significant variables among a large set of variables, when the interest is on a given response variable. In recent years several grouped regularization methods have been proposed and in this paper we present these under one unified framework in this varying coefficient model context. For each of the discussed grouped regularization methods we investigate the optimization problem to be solved, possible algorithms for doing so, and the variable and estimation consistency of the methods. We investigate the finite-sample performance of these methods, in a comparative study, and illustrate them on real data examples.     

Specifying a Gaussian Markov Random Field by a Sparse Cholesky Triangle

ABSTRACT

This note discusses the approach of specifying a Gaussian Markov random field (GMRF) by the Cholesky triangle of the precision matrix. A such representation can be made extremely sparse using numerical techniques for incomplete sparse Cholesky factorization, and provide very computational efficient representation for simulating from the GMRF. However, we provide theoretical and empirical justification showing that the sparse Cholesky triangle representation is fragile when conditioning a GMRF on a subset of the variables or observed data, meaning that the computational cost increases.     

Investigating the sensitivity of Gaussian processes to the choice of their correlation function and prior specifications

A Gaussian process (GP) can be thought of as an infinite collection of random variables with the property that any subset, say of dimension n, of these variables have a multivariate normal distribution of dimension n, mean vector β and covariance matrix Σ [O'Hagan, A., 1994, Kendall's Advanced Theory of Statistics, Vol. 2B, Bayesian Inference (John Wiley & Sons, Inc.)]. The elements of the covariance matrix are routinely specified through the multiplication of a common variance by a correlation function. It is important to use a correlation function that provides a valid covariance matrix (positive definite). Further, it is well known that the smoothness of a GP is directly related to the specification of its correlation function. Also, from a Bayesian point of view, a prior distribution must be assigned to the unknowns of the model. Therefore, when using a GP to model a phenomenon, the researcher faces two challenges: the need of specifying a correlation function and a prior distribution for its parameters. In the literature there are many classes of correlation functions which provide a valid covariance structure. Also, there are many suggestions of prior distributions to be used for the parameters involved in these functions. We aim to investigate how sensitive the GPs are to the (sometimes arbitrary) choices of their correlation functions. For this, we have simulated 25 sets of data each of size 64 over the square [0, 5]×[0, 5] with a specific correlation function and fixed values of the GP's parameters. We then fit different correlation structures to these data, with different prior specifications and check the performance of the adjusted models using different model comparison criteria.     

Investigation about a screening step in model selection

In many studies a large number of variables is measured and the identification of relevant variables influencing an outcome is an important task. For variable selection several procedures are available. However, focusing on one model only neglects that there usually exist other equally appropriate models. Bayesian or frequentist model averaging approaches have been proposed to improve the development of a predictor. With a larger number of variables (say more than ten variables) the resulting class of models can be very large. For Bayesian model averaging Occam’s window is a popular approach to reduce the model space. As this approach may not eliminate any variables, a variable screening step was proposed for a frequentist model averaging procedure. Based on the results of selected models in bootstrap samples, variables are eliminated before deriving a model averaging predictor. As a simple alternative screening procedure backward elimination can be used. Through two examples and by means of simulation we investigate some properties of the screening step. In the simulation study we consider situations with fifteen and 25 variables, respectively, of which seven have an influence on the outcome. With the screening step most of the uninfluential variables will be eliminated, but also some variables with a weak effect. Variable screening leads to more applicable models without eliminating models, which are more strongly supported by the data. Furthermore, we give recommendations for important parameters of the screening step.     

Bayesian analysis of binary survey data

First a comprehensive treatment of the hierarchical-conjugate Bayesian predictive approach to binary survey data is presented, encompassing simple random, stratified, cluster, and two-stage sampling, as well as two-stage sampling within strata. For the case of two-stage sampling within strata when there is more than one variable of stratification, analysis using an unsaturated logit linear model on the prior means is proposed. This allows there to be cells containing no sampled clusters. Formulas for posterior predictive means, variances, and covariances of numbers of successes in unsampled portions of clusters are presented in terms of posterior expectations of certain functions of hyperparameters; these may be evaluated by existing methods. The technique is illustrated using a small subset of Canada Youth & AIDS Study data. A sample of students within each of various selected school boards was chosen and interviewed via questionnaire. The boards were stratified/poststratified in two dimensions, but some of the resulting cells contained no data. The additive logit linear model on the prior means produced estimates and posterior variances for boards in all cells. Data showed the additive model to be plausible.     

A probabilistic clustering method for data elements with normal distributed attributes

By considering uncertainty in the attributes common methods cannot be applicable in data clustering. In the recent years, many researches have been done by considering fuzzy concepts to interpolate the uncertainty. But when data elements attributes have probabilistic distributions, the uncertainty cannot be interpreted by fuzzy theory. In this article, a new concept for clustering of elements with predefined probabilistic distributions for their attributes has been proposed, so each observation will be as a member of a cluster with special probability. Two metaheuristic algorithms have been applied to deal with the problem. Squared Euclidean distance type has been considered to calculate the similarity of data elements to cluster centers. The sensitivity analysis shows that the proposed approach will converge to the classic approaches results when the variance of each point tends to be zero. Moreover, numerical analysis confirms that the proposed approach is efficient in clustering of probabilistic data.     

Variable Selection for Support Vector Machines

Consider using values of variables X ₁, X ₂,…, X _p to classify entities into one of two classes. Kernel-based procedures such as support vector machines (SVMs) are well suited for this task. In general, the classification accuracy of SVMs can be substantially improved if instead of all p candidate variables, a smaller subset of (say m) variables is used. A new two-step approach to variable selection for SVMs is therefore proposed: best variable subsets of size k = 1,2,…, p are first identified, and then a new data-dependent criterion is used to determine a value for m. The new approach is evaluated in a Monte Carlo simulation study, and on a sample of data sets.     

Performance of Variable Selection Methods in Regression Using Variations of the Bayesian Information Criterion

The Bayesian information criterion (BIC) is widely used for variable selection. We focus on the regression setting for which variations of the BIC have been proposed. A version that includes the Fisher Information matrix of the predictor variables performed best in one published study. In this article, we extend the evaluation, introduce a performance measure involving how closely posterior probabilities are approximated, and conclude that the version that includes the Fisher Information often favors regression models having more predictors, depending on the scale and correlation structure of the predictor matrix. In the image analysis application that we describe, we therefore prefer the standard BIC approximation because of its relative simplicity and competitive performance at approximating the true posterior probabilities.     

Bayesian covariance matrix estimation using a mixture of decomposable graphical models

We present a Bayesian approach to estimating a covariance matrix by using a prior that is a mixture over all decomposable graphs, with the probability of each graph size specified by the user and graphs of equal size assigned equal probability. Most previous approaches assume that all graphs are equally probable. We show empirically that the prior that assigns equal probability over graph sizes outperforms the prior that assigns equal probability over all graphs in more efficiently estimating the covariance matrix. The prior requires knowing the number of decomposable graphs for each graph size and we give a simulation method for estimating these counts. We also present a Markov chain Monte Carlo method for estimating the posterior distribution of the covariance matrix that is much more efficient than current methods. Both the prior and the simulation method to evaluate the prior apply generally to any decomposable graphical model.     

Genetic Algorithm in the Wavelet Domain for Large p Small n Regression

Many areas of statistical modeling are plagued by the “curse of dimensionality,” in which there are more variables than observations. This is especially true when developing functional regression models where the independent dataset is some type of spectral decomposition, such as data from near-infrared spectroscopy. While we could develop a very complex model by simply taking enough samples (such that n > p), this could prove impossible or prohibitively expensive. In addition, a regression model developed like this could turn out to be highly inefficient, as spectral data usually exhibit high multicollinearity. In this article, we propose a two-part algorithm for selecting an effective and efficient functional regression model. Our algorithm begins by evaluating a subset of discrete wavelet transformations, allowing for variation in both wavelet and filter number. Next, we perform an intermediate processing step to remove variables with low correlation to the response data. Finally, we use the genetic algorithm to perform a stochastic search through the subset regression model space, driven by an information-theoretic objective function. We allow our algorithm to develop the regression model for each response variable independently, so as to optimally model each variable. We demonstrate our method on the familiar biscuit dough dataset, which has been used in a similar context by several researchers. Our results demonstrate both the flexibility and the power of our algorithm. For each response variable, a different subset model is selected, and different wavelet transformations are used. The models developed by our algorithm show an improvement, as measured by lower mean error, over results in the published literature.     

大维变量选择、混频因子模型与新冠肺炎疫情冲击下的GDP现时预测

新冠肺炎疫情不仅对我国宏观经济造成了巨大冲击,也为准确预测我国宏观经济未来走势带来挑战。本文从新冠肺炎疫情冲击出发,将模型置信集检验与U-MIDAS模型组合,设计了一种在混频情形下利用预测变量的异质性波动从大维数据集中选取对GDP具有稳定预测效果变量的方法。通过利用选取出的稳定性变量构建多种形式的混频目标因子模型并与其他类型的混频因子模型对比,全面评估了不同模型在疫情前后对GDP进行高频现时预测的效果。研究发现,在疫情冲击前的平稳时期,利用覆盖范围较广的变量构建双因子MIDAS模型预测效果最优;利用稳定性变量构建的单因子U-MIDAS模型同样具有良好的预测效果。当经济从冲击中持续恢复时,利用部分稳定性变量构建的双因子U-MIDAS模型在捕捉到GDP的核心变化后率先对其连续做出准确的现时预测。经济稳定时,对预测变量设定较长的滞后阶数会提升预测效果;在冲击后的恢复期中则应减少滞后阶数,避免变量在冲击中出现的异常值对预测产生负面影响。本文也为当经济受到巨大外生冲击或处于冲击后的恢复期时其他宏观经济指标的预测提供了有价值的参考。     

Variable selection in gamma regression models via artificial bee colony algorithm

Variable selection is an important task in regression analysis. Performance of the statistical model highly depends on the determination of the subset of predictors. There are several methods to select most relevant variables to construct a good model. However in practice, the dependent variable may have positive continuous values and not normally distributed. In such situations, gamma distribution is more suitable than normal for building a regression model. This paper introduces an heuristic approach to perform variable selection using artificial bee colony optimization for gamma regression models. We evaluated the proposed method against with classical selection methods such as backward and stepwise. Both simulation studies and real data set examples proved the accuracy of our selection procedure.     

Sparse Bayesian variable selection in multinomial probit regression model with application to high-dimensional data classification

Here we consider a multinomial probit regression model where the number of variables substantially exceeds the sample size and only a subset of the available variables is associated with the response. Thus selecting a small number of relevant variables for classification has received a great deal of attention. Generally when the number of variables is substantial, sparsity-enforcing priors for the regression coefficients are called for on grounds of predictive generalization and computational ease. In this paper, we propose a sparse Bayesian variable selection method in multinomial probit regression model for multi-class classification. The performance of our proposed method is demonstrated with one simulated data and three well-known gene expression profiling data: breast cancer data, leukemia data, and small round blue-cell tumors. The results show that compared with other methods, our method is able to select the relevant variables and can obtain competitive classification accuracy with a small subset of relevant genes.