Model Averaging Estimation Method by Kullback–Leibler Divergence for Multiplicative Error Model

Bayesian Model Averaging and Weighted Average Least Squares: Equivariance, Stability, and Numerical Issues

In statistical exercises where there are several candidate models, the traditional approach is to select one model using some data driven criterion and use that model for estimation, testing and other purposes, ignoring the variability of the model selection process. We discuss some problems associated with this approach. An alternative scheme is to use a model-averaged estimator, that is, a weighted average of estimators obtained under different models, as an estimator of a parameter. We show that the risk associated with a Bayesian model-averaged estimator is bounded as a function of the sample size, when parameter values are fixed. We establish conditions which ensure that a model-averaged estimator's distribution can be consistently approximated using the bootstrap. A new, data-adaptive, model averaging scheme is proposed that balances efficiency of estimation without compromising applicability of the bootstrap. This paper illustrates that certain desirable risk and resampling properties of model-averaged estimators are obtainable when parameters are fixed but unknown; this complements several studies on minimaxity and other properties of post-model-selected and model-averaged estimators, where parameters are allowed to vary.

Model uncertainty is a prominent feature in many applied settings. This is certainty true in the efficiency analysis realm where concerns over the proper distributional specification of the error components of a stochastic frontier model is, generally, still open along with which variables influence inefficiency. Given the concern over the impact that model uncertainty is likely to have on the stochastic frontier model in practice, the present research proposes two distinct model averaging estimators, one which averages over nested classes of inefficiency distributions and another that has the ability to average over distinct distributions of inefficiency. Both of these estimators are shown to produce optimal weights when the aim is to uncover conditional inefficiency at the firm level. We study the finite-sample performance of the model average estimator via Monte Carlo experiments and compare with traditional model averaging estimators based on weights constructed from model selection criteria and present a short empirical application.

In this article, we describe the estimation of linear regression models with uncertainty about the choice of the explanatory variables. We introduce the Stata commands bma and wals, which implement, respectively, the exact Bayesian model-averaging estimator and the weighted-average least-squares estimator developed by Magnus, Powell, and Pr¨ufer (2010, Journal of Econometrics 154: 139–153). Unlike standard pretest estimators that are based on some preliminary diagnostic test, these model-averaging estimators provide a coherent way of making inference on the regression parameters of interest by taking into account the uncertainty due to both the estimation and the model selection steps. Special emphasis is given to several practical issues that users are likely to face in applied work: equivariance to certain transformations of the explanatory variables, stability, accuracy, computing speed, and out-of-memory problems. Performances of our bma and wals commands are illustrated using simulated data and empirical applications from the literature on model-averaging estimation.

This report documents the Multi-Model Analysis (MMA) computer code. MMA can be used to evaluate results from alternative models of a single system using the same set of observations for all models. As long as the observations, the observation weighting, and system being represented are the same, the models can differ in nearly any way imaginable. For example, they may include different processes, different simulation software, different temporal definitions (for example, steady-state and transient models could be considered), and so on. The multiple models need to be calibrated by nonlinear regression. Calibration of the individual models needs to be completed before application of MMA. MMA can be used to rank models and calculate posterior model probabilities. These can be used to (1) determine the relative importance of the characteristics embodied in the alternative models, (2) calculate model-averaged parameter estimates and predictions, and (3) quantify the uncertainty of parameter estimates and predictions in a way that integrates the variations represented by the alternative models. There is a lack of consensus on what model analysis methods are best, so MMA provides four default methods. Two are based on Kullback-Leibler information, and use the AIC (Akaike Information Criterion) or AICc (second-order-bias-corrected AIC) model discrimination criteria. The other two default methods are the BIC (Bayesian Information Criterion) and the KIC (Kashyap Information Criterion) model discrimination criteria. Use of the KIC criterion is equivalent to using the maximum-likelihood Bayesian model averaging (MLBMA) method. AIC, AICc, and BIC can be derived from Frequentist or Bayesian arguments. The default methods based on Kullback-Leibler information have a number of theoretical advantages, including that they tend to favor more complicated models as more data become available than do the other methods, which makes sense in many situations. Many applications of MMA will be well served by the default methods provided. To use the default methods, the only required input for MMA is a list of directories where the files for the alternate models are located. Evaluation and development of model-analysis methods are active areas of research. To facilitate exploration and innovation, MMA allows the user broad discretion to define alternatives to the default procedures. For example, MMA allows the user to (a) rank models based on model criteria defined using a wide range of provided and user-defined statistics in addition to the default AIC, AICc, BIC, and KIC criteria, (b) create their own criteria using model measures available from the code, and (c) define how each model criterion is used to calculate related posterior model probabilities. The default model criteria rate models are based on model fit to observations, the number of observations and estimated parameters, and, for KIC, the Fisher information matrix. In addition, MMA allows the analysis to include an evaluation of estimated parameter values. This is accomplished by allowing the user to define unreasonable estimated parameter values or relative estimated parameter values. An example of the latter is that it may be expected that one parameter value will be less than another, as might be the case if two parameters represented the hydraulic conductivity of distinct materials such as fine and coarse sand. Models with parameter values that violate the user-defined conditions are excluded from further consideration by MMA. Ground-water models are used as examples in this report, but MMA can be used to evaluate any set of models for which the required files have been produced. MMA needs to read files from a separate directory for each alternative model considered. The needed files are produced when using the Sensitivity-Analysis or Parameter-Estimation mode of UCODE_2005, or, possibly, the equivalent capability of another program. MMA is constructed using

In many scientific studies, the underlying data-generating process is unknown and multiple statistical models are considered to describe it. For example, in a factorial experiment we might consider models involving just main effects, as well as those that include interactions. Model-averaging is a commonly-used statistical technique to allow for model uncertainty in parameter estimation. In the frequentist setting, the model-averaged estimate of a parameter is a weighted mean of the estimates from the individual models, with the weights typically being based on an information criterion, cross-validation, or bootstrapping. One approach to building a model-averaged confidence interval is to use a Wald interval, based on the model-averaged estimate and its standard error. This has been the default method in many application areas, particularly those in the life sciences. The MA-Wald interval, however, assumes that the studentized model-averaged estimate has a normal distribution, which can be far from true in practice due to the random, data-driven model weights. Recently, the model-averaged tail area Wald interval (MATA-Wald) has been proposed as an alternative to the MA-Wald interval, which only assumes that the studentized estimate from each model has a N(0, 1) or t-distribution, when that model is true. This alternative to the MA-Wald interval has been shown to have better coverage in simulation studies. However, when we have a response variable that is skewed, even these relaxed assumptions may not be valid, and use of these intervals might therefore result in poor coverage. We propose a new interval (MATA-SBoot) which uses a parametric bootstrap approach to estimate the distribution of the studentized estimate for each model, when that model is true. This method only requires that the studentized estimate from each model is approximately pivotal, an assumption that will often be true in practice, even for skewed data. We illustrate use of this new interval in the analysis of a three-factor marine global change experiment in which the response variable is assumed to have a lognormal distribution. We also perform a simulation study, based on the example, to compare the lower and upper error rates of this interval with those for existing methods. The results suggest that the MATA-SBoot interval can provide better error rates than existing intervals when we have skewed data, particularly for the upper error rate when the sample size is small.

Label distribution learning (LDL) is the state-of-the-art approach to dealing with a number of real-world applications, such as chronological age estimation from a face image, where there is an inherent similarity among adjacent age labels. LDL takes into account the semantic similarity by assigning a label distribution to each instance. The well-known Kullback-Leibler (KL) divergence is the widely used loss function for the LDL framework. However, the KL divergence does not fully and effectively capture the semantic similarity among age labels, thus leading to suboptimal performance. In this article, we propose a novel loss function based on the optimal transport theory for the LDL-based age estimation. A ground metric function plays an important role in the optimal transport formulation. It should be carefully determined based on the underlying geometric structure of the label space of the application in-hand. The label space in the age estimation problem has a specific geometric structure, that is, closer ages have more inherent semantic relationships. Inspired by this, we devise a novel ground metric function, which enables the loss function to increase the influence of highly correlated ages; thus exploiting the semantic similarity among ages more effectively than the existing loss functions. We then use the proposed loss function, namely, γ -Wasserstein loss, for training a deep neural network (DNN). This leads to a notoriously computationally expensive and nonconvex optimization problem. Following the standard methodology, we formulate the optimization function as a convex problem and then use an efficient iterative algorithm to update the parameters of the DNN. Extensive experiments in age estimation on different benchmark datasets validate the effectiveness of the proposed method, which consistently outperforms state-of-the-art approaches.

Applied researchers often confront two issues when using the fixed effect-two-stage least squares (FE-2SLS) estimator for panel data models. One is that it may lose its consistency due to too many instruments. The other is that the gain of using FE-2SLS may not exceed its loss when the endogeneity is weak. In this paper, an L 2 Boosting regularization procedure for panel data models is proposed to tackle the many instruments issue. We then construct a Stein-like model-averaging estimator to take advantage of FE and FE-2SLS-Boosting estimators. Finite sample properties are examined in Monte Carlo and an empirical application is presented.

We introduce and study a novel model-selection strategy for Bayesian learning, based on optimal transport, along with its associated predictive posterior law: the Wasserstein population barycenter of the posterior law over models. We first show how this estimator, termed Bayesian Wasserstein barycenter (BWB), arises naturally in a general, parameter-free Bayesian model-selection framework, when the considered Bayesian risk is the Wasserstein distance. Examples are given, illustrating how the BWB extends some classic parametric and non-parametric selection strategies. Furthermore, we also provide explicit conditions granting the existence and statistical consistency of the BWB, and discuss some of its general and specific properties, providing insights into its advantages compared to usual choices, such as the model average estimator. Finally, we illustrate how this estimator can be computed using the stochastic gradient descent (SGD) algorithm in Wasserstein space introduced in a companion paper, and provide a numerical example for experimental validation of the proposed method.

The dynamic average consensus problem for a group of agents is considered. Each agent in the group is supposed to estimate the average of inputs applied to all agents and the estimation should be done in a distributed way. By reinterpreting the proportional integral type estimator, a new structure for the average estimator which can embed the internal model of inputs is proposed and conditions which result in the zero estimation error in the steady state are derived. We present constructive design procedures for the cases of constant inputs and time-varying inputs employing the root locus for the former and LQR-based design for the latter. The theory is validated through numerical simulations.

Knowledge distillation (KD), transferring knowledge from a cumbersome teacher model to a lightweight student model, has been investigated to design efficient neural architectures. Generally, the objective function of KD is the Kullback-Leibler (KL) divergence loss between the softened probability distributions of the teacher model and the student model with the temperature scaling hyperparameter τ. Despite its widespread use, few studies have discussed how such softening influences generalization. Here, we theoretically show that the KL divergence loss focuses on the logit matching when τ increases and the label matching when τ goes to 0 and empirically show that the logit matching is positively correlated to performance improvement in general. From this observation, we consider an intuitive KD loss function, the mean squared error (MSE) between the logit vectors, so that the student model can directly learn the logit of the teacher model. The MSE loss outperforms the KL divergence loss, explained by the penultimate layer representations difference between the two losses. Furthermore, we show that sequential distillation can improve performance and that KD, using the KL divergence loss with small τ particularly, mitigates the label noise. The code to reproduce the experiments is publicly available online at https://github.com/jhoon-oh/kd_data/.

Model averaging procedure for partially linear single-index models

In this paper, we study optimal model averaging estimators of regression coefficients in a multinomial logit model, which is commonly used in many scientific fields. A Kullback–Leibler (KL) loss-based weight choice criterion is developed to determine averaging weights. Under some regularity conditions, we prove that the resulting model averaging estimators are asymptotically optimal. When the true model is one of the candidate models, the averaged estimators are consistent. Simulation studies suggest the superiority of the proposed method over commonly used model selection criterions, model averaging methods, as well as some other related methods in terms of the KL loss and mean squared forecast error. Finally, the website phishing data is used to illustrate the proposed method.

In supervised classification problems using Deep Neural Networks, the loss function is typically based on the Kullback–Leibler divergence. However, alternative entropic divergence formulations, such as the Jensen–Shannon Divergence (JSD), have recently garnered attention for their unique properties. In this study, we delve deeper into the interpretation of the JSD and its generalized form, the Jensen–Tsallis Divergence (JTD), as alternative loss functions for supervised classification. When provided with one-hot encoded distributions for the true label probabilities, we demonstrate that these novel divergences impose an intrinsic output confidence regularization that prevents overfitting. Additionally, the q non-extensive parameter of the JTD directly influences the structure of the regularizer, offering increased flexibility in the formulation of the loss function. Through experiments conducted on artificially imbalanced versions of MNIST, Fashion-MNIST, SVHN and CIFAR-10 we showcase how JTD outperforms JSD and other traditional loss functions in terms of generalization performance, especially for highly imbalanced datasets.

The latent variable prior of the variational autoencoder (VAE) often utilizes a standard Gaussian distribution because of the convenience in calculation, but has an underfitting problem. This paper proposes a variational autoencoder with optimizing Gaussian mixture model priors. This method utilizes a Gaussian mixture model to construct prior distribution, and utilizes the Kullback-Leibler (KL) distance between posterior and prior distribution to implement an iterative optimization of the prior distribution based on the data. The greedy algorithm is used to solve the KL distance for defining the approximate variational lower bound solution of the loss function, and for realizing the VAE with optimizing Gaussian mixture model priors. Compared with the standard VAE method, the proposed method obtains state-of-the-art results on MNIST, Omniglot, and Frey Face datasets, which shows that the VAE with optimizing Gaussian mixture model priors can learn a better model.