The mixture index of fit and minimax regression

A measure of the fit of a statistical model can be obtained by estimating the relative size of the largest fraction of the population where a distribution belonging to the model may be valid. This is the mixture index of fit that was suggested for models for contingency tables by Rudas, Clogg, Lindsay (1994) and it is extended here for models involving continuous observations. In particular, the approach is applied to regression models with normal and uniform error structures. Best fit, as measured by the mixture index of fit, is obtained with minimax estimation of the regression parameters. Therefore, whenever minimax estimation is used for these problems, the mixture index of fit provides a natural approach for measuring model fit and for variable selection. Received: September 1997     

Taking ‘Don’t Knows’ as Valid Responses: A Multiple Complete Random Imputation of Missing Data

Incomplete data is a common problem of survey research. Recent work on multiple imputation techniques has increased analysts’ awareness of the biasing effects of missing data and has also provided a convenient solution. Imputation methods replace non-response with estimates of the unobserved scores. In many instances, however, non-response to a stimulus does not result from measurement problems that inhibit accurate surveying of empirical reality, but from the inapplicability of the survey question. In such cases, existing imputation techniques replace valid non-response with counterfactual estimates of a situation in which the stimulus is applicable to all respondents. This paper suggests an alternative imputation procedure for incomplete data for which no true score exists: multiple complete random imputation, which overcomes the biasing effects of missing data and allows analysts to model respondents’ valid ‘I don’t know’ answers.     

Identification with latent variables

This is an essay on a unified approach to the identifiability problem in static models with and without hidden endogenous variables. As is well known, when some of these variables are unobserved, the prior information requirements for models when all endogenous variables are observed, are still there. In addition, extra prior information that takes the place of the means and covariances of the missing variables will have to be supplied directly or indirectly by the statistical researcher. In the paper we characterize the quality and quantity of the required information for the general linear static model and apply it when the model is i) an econometric demand and supply model with missing observations on the quantity transacted, ii) a factor analysis model with observed characteristics of the test takers and iii) a LISREL Model without fixed exogenous variables. With unknown true parameters, the exact rank conditions are seldom verifiable but we do recommend an implementable check-list that is adequate for almost all parameters.     

How to Choose a Working Model for Measuring the Statistical Evidence About a Regression Parameter

Consider using a likelihood ratio to measure the strength of statistical evidence for one hypothesis over another. Recent work has shown that when the model is correctly specified, the likelihood ratio is seldom misleading. But when the model is not, misleading evidence may be observed quite frequently. Here we consider how to choose a working regression model so that the statistical evidence is correctly represented as often as it would be under the true model. We argue that the criteria for choosing a working model should be how often it correctly represents the statistical evidence about the object of interest (regression coefficient in the true model). We see that misleading evidence about the object of interest is more likely to be observed when the working model is chosen according to other criteria (e.g., parsimony or predictive accuracy).     

Kuhn-Tucker,likelihood ratio and Wald tests for nonlinear models with inequality constraints on the parameters

This paper considers the problem of testing statistical hypothesis in nonlinear regression models with inequality constraints on the parameters. First, the Kuhn-Tucker test procedure is defined. Next, it is shown that the distribution of the Kuhn-Tucker, the likelihood ratio and the Wald test statistics converges to the same mixture of chi-square distributions under the null hypothesis. To illustrate these results two examples are considered: (1) the problem of testing that individual effects are missing in an error component model, and (2) the problem of testing equilibrium for a model of markets in disequilibrium.     

Inversion copulas from nonlinear state space models with an application to inflation forecasting

We propose the construction of copulas through the inversion of nonlinear state space models. These copulas allow for new time series models that have the same serial dependence structure as a state space model, but with an arbitrary marginal distribution, and flexible density forecasts. We examine the time series properties of the copulas, outline serial dependence measures, and estimate the models using likelihood-based methods. Copulas constructed from three example state space models are considered: a stochastic volatility model with an unobserved component, a Markov switching autoregression, and a Gaussian linear unobserved component model. We show that all three inversion copulas with flexible margins improve the fit and density forecasts of quarterly U.S. broad inflation and electricity inflation.     

Estimating systematic continuous-time trends in recidivism using a non-Gaussian panel data model

We model panel data of crime careers of juveniles from a Dutch Judicial Juvenile Institution. The data are decomposed into a systematic and an individual-specific component, of which the systematic component reflects the general time-varying conditions including the criminological climate. Within a model-based analysis, we treat (1) shared effects of each group with the same systematic conditions, (2) strongly non-Gaussian features of the individual time series, (3) unobserved common systematic conditions, (4) changing recidivism probabilities in continuous time and (5) missing observations. We adopt a non-Gaussian multivariate state-space model that deals with all these issues simultaneously. The parameters of the model are estimated by Monte Carlo maximum likelihood methods. This paper illustrates the methods empirically. We compare continuous time trends and standard discrete-time stochastic trend specifications. We find interesting common time variation in the recidivism behaviour of the juveniles during a period of 13 years, while taking account of significant heterogeneity determined by personality characteristics and initial crime records.     

Nonlinear dynamics of interest rate and inflation

According to several empirical studies US inflation and nominal interest rates as well as the real interest rate can be described as unit root processes. These results imply that nominal interest rates and expected inflation do not move one‐for‐one in the long run, which is incongruent with theoretical models. In this paper we introduce a new nonlinear bivariate mixture autoregressive model that seems to fit quarterly US data (1953 : II–2004 : IV) reasonably well. It is found that the three‐month Treasury bill rate and inflation share a common nonlinear component that explains a large part of their persistence. The real interest rate is devoid of this component, indicating one‐for‐one movement of the nominal interest rate and inflation in the long run and, hence, stationarity of the real interest rate. Copyright © 2006 John Wiley & Sons, Ltd.     

Estimation in Two-Stage Models with Heteroscedasticity

A surprising number of important problems can be cast in the framework of estimating a mean and variance using data arising from a two-stage structure. The first stage is a random sampling of "units" with some quantity of interest associated with the unit. The second stage produces an estimate of that quantity and usually, but not always, an estimated standard error, which may change considerably across units. Heteroscedasticity in the estimates over different units can arise for a number of reasons, including variation associated with the unit and changing sampling effort over units. This paper presents a broad discussion of the problem of making inferences for the population mean and variance associated with the unobserved true values at the first stage of sampling. A careful discussion of the causes of heteroscedasticity is given, followed by an examination of ways in which inferences can be carried out in a manner that is robust to the nature of the within unit heteroscedasticity. Among the conclusions are that under any type of heteroscedasticity, an unbiased estimate of the mean and the variance of the estimated mean can be obtained by using the estimates as if they were true unobserved values from the first stage. The issue of using the mean versus a weighted average which tries to account for the heteroscedasticity is also discussed. An unbiased estimate of the population variance is given and the variance of this estimate and its covariance with the estimated mean is provided under various types of heteroscedasticity. The two-stage setting arises in many contexts including the one-way random effects models with replication, meta-analysis, multi-stage sampling from finite populations and random coefficients models. We will motivate and illustrate the problem with data arising from these various contexts with the goal of providing a unified framework for addressing such problems.     

A semiparametric random effects approach to unobserved heterogeneity in count data models of organizational founding

The paper discusses a semiparametric random-effects approach to the problem of unobserved population heterogeneity in organizational research based on models for pooled cross-sectional time series count data. The analytical value of this approach rests in its ability to produce estimates of the structural parameters that do not depend on any specific assumption about the distribution of the heterogeneity components in the population. The practical value of the method proposed is illustrated in an empirical application to processes of organizational founding, and to the relation between density dependence and unobserved heterogeneity in spatially distributed organizational populations. The empirical evidence produced suggests that future studies of organizational founding at the population level will have to account for variation in observed as well as unmeasured (or unobservable) variables.     

DEMAND FOR MEDICAL CARE BY THE ELDERLY: A FINITE MIXTURE APPROACH

In this article we develop a finite mixture negative binomial count model that accommodates unobserved heterogeneity in an intuitive and analytically tractable manner. This model, the standard negative binomial model, and its hurdle extension are estimated for six measures of medical care demand by the elderly using a sample from the 1987 National Medical Expenditure Survey. The finite mixture model is preferred overall by statistical model selection criteria. Two points of support adequately describe the distribution of the unobserved heterogeneity, suggesting two latent populations, the ‘healthy’ and the ‘ill’ whose fitted distributions differ substantially from each other. © 1997 John Wiley & Sons, Ltd.     

The error structure of time series cross-section hedonic models with sporadic event timing and serial correlation

When estimating hedonic models of housing prices, the use of time series cross-section repeat sales data can provide improvements in estimator efficiency and correct for unobserved characteristics. However, in cases where serial correlation is present, the irregular timing of sales should also be considered. In this paper we develop a model that uses information on the timing of events to account for the sporadic occurrence of events. The model presumes that the serial correlation process can be decomposed into a time-independent (event-wise) component and a time-dependent (time-wise) component. Empirical tests cannot reject the presence of sporadic correlation patterns, while simulations show that the failure to account for sporadic correlation leads to significant losses in efficiency, and that the losses from ignoring sporadic correlation when it exists are larger than losses when sporadic correlation is falsely assumed. Copyright © 1999 John Wiley & Sons, Ltd.     

Lock-in and unobserved preferences in server operating systems: A case of Linux vs. Windows

This paper investigates to what extent the persistence of Microsoft Windows in the market for server operating systems is due to lock-in or unobserved preferences. While the hypothesis of lock-in plays an important role in the antitrust policy debate for the operating systems market, it has not been extensively documented empirically. To account for unobserved preferences, we use a panel data identification approach based on time-variant group fixed effects, and estimate the dynamic discrete choice panel data model developed by Arellano and Carrasco (2003). Using detailed establishment-level data, we find that once we account for unobserved preferences, the estimated magnitudes of lock-in are considerably smaller than those from the conventional approaches, suggesting that unobserved preferences play a major role in the persistence of Windows. Further robustness checks are consistent with our findings.     

Empirical evidence of income dynamics across EU regions

This paper analyses the distribution of purchasing power standardized per capita income across EU‐12 regions between 1977 and 1996. Dispersion of incomes between regions is measured taking into account their population sizes. The cross‐sectional distributions are initially described by weighted kernel density estimates, revealing a multimodal structure of the distributions, less evident over the period. This evidence is supported by a bootstrap test. To detect homogeneous groups of regions, the empirical distributions are approximated by a finite mixture of normal densities. The components of the mixture represent clusters of poor/rich regions, while the mixing proportions the allocation over the poor and the rich components. The number of components is assessed by a bootstrap LR test, and the goodness of fit by a kernel density‐based test. Income mobility is modelled by the stochastic kernel, the continuous counterpart of the transition probability matrix. The main implication is a very slow process of catching up of the poorest regions with the richer ones and a process of shifting away of a small group of very rich regions. This evidence is reflected in the shape of the ergodic distribution, which is well fitted by a two‐component mixture model. Copyright © 2006 John Wiley & Sons, Ltd.     

Statistical Implementations of Agent-Based Demographic Models

A variety of demographic statistical models exist for studying population dynamics when individuals can be tracked over time. In cases where data are missing due to imperfect detection of individuals, the associated measurement error can be accommodated under certain study designs (e.g. those that involve multiple surveys or replication). However, the interaction of the measurement error and the underlying dynamic process can complicate the implementation of statistical agent-based models (ABMs) for population demography. In a Bayesian setting, traditional computational algorithms for fitting hierarchical demographic models can be prohibitively cumbersome to construct. Thus, we discuss a variety of approaches for fitting statistical ABMs to data and demonstrate how to use multi-stage recursive Bayesian computing and statistical emulators to fit models in such a way that alleviates the need to have analytical knowledge of the ABM likelihood. Using two examples, a demographic model for survival and a compartment model for COVID-19, we illustrate statistical procedures for implementing ABMs. The approaches we describe are intuitive and accessible for practitioners and can be parallelised easily for additional computational efficiency.     

Measuring underlying economic activity

Recently, interest in the methodology of constructing coincident economic indicators has been revived by the work of Stock and Watson (1989b). They adopt the framework of the state space form and Kalman filter in which to construct an optimal estimate of an unobserved component. This is interpreted as corresponding to underlying economic activity derived from a set of observed indicator variables. In this paper we apply the Stock and Watson approach to the UK where the observed indicator variables are those that make up the Central Statistical Office (CSO) coincident indicator. The time series properties of the indicator variables are examined and three of the five variables are first difference stationary and are cointegrated, the remaining two are stationary in levels. We then construct two alternative measures of economic activity, each of which deals with the different orders of stationarity of the variables. The first uses the levels of the observed component variables that allows for the cointegrating relationship. The second imposes stationarity on the I(1) variables before the estimation by taking first differences. The levels index is viewed as the preferred specification as it allows for the long-run relationships between the variables and has a superior statistical fit. ©1996 John Wiley & Sons, Ltd.     

Likelihood inference in some finite mixture models

Parametric mixture models are commonly used in applied work, especially empirical economics, where these models are often employed to learn for example about the proportions of various types in a given population. This paper examines the inference question on the proportions (mixing probability) in a simple mixture model in the presence of nuisance parameters when sample size is large. It is well known that likelihood inference in mixture models is complicated due to (1) lack of point identification, and (2) parameters (for example, mixing probabilities) whose true value may lie on the boundary of the parameter space. These issues cause the profiled likelihood ratio (PLR) statistic to admit asymptotic limits that differ discontinuously depending on how the true density of the data approaches the regions of singularities where there is lack of point identification. This lack of uniformity in the asymptotic distribution suggests that confidence intervals based on pointwise asymptotic approximations might lead to faulty inferences. This paper examines this problem in details in a finite mixture model and provides possible fixes based on the parametric bootstrap. We examine the performance of this parametric bootstrap in Monte Carlo experiments and apply it to data from Beauty Contest experiments. We also examine small sample inferences and projection methods.     

A Predictive Estimator of the Mean with Missing Data

One of the most difficult problems confronting investigators who analyze data from surveys is how treat missing data. Many statistical procedures can not be used immediately if any values are missing. This paper considers the problem of estimating the population mean using auxiliary information when some observations on the sample are missing and the population mean of the auxiliary variable is not available. We use tools of classical statistical estimation theory to find a suitable estimator. We study the model and design properties of the proposed estimator. We also report the results of a broad-based simulation study of the efficiency of the estimator, which reveals very promising results.     

Fractional imputation for incomplete two-way contingency tables

Imputation procedures such as fully efficient fractional imputation (FEFI) or multiple imputation (MI) create multiple versions of the missing observations, thereby reflecting uncertainty about their true values. Multiple imputation generates a finite set of imputations through a posterior predictive distribution. Fractional imputation assigns weights to the observed data. The focus of this article is the development of FEFI for partially classified two-way contingency tables. Point estimators and variances of FEFI estimators of population proportions are derived. Simulation results, when data are missing completely at random or missing at random, show that FEFI is comparable in performance to maximum likelihood estimation and multiple imputation and superior to simple stochastic imputation and complete case anlaysis. Methods are illustrated with four data sets.     

Time and causality: A Monte Carlo assessment of the timing-of-events approach

We present new Monte Carlo evidence regarding the feasibility of separating causality from selection within non-experimental duration data, by means of the non-parametric maximum likelihood estimator (NPMLE). Key findings are: (i) the NPMLE is extremely reliable, and it accurately separates the causal effects of treatment and duration dependence from sorting effects, almost regardless of the true unobserved heterogeneity distribution; (ii) the NPMLE is normally distributed, and standard errors can be computed directly from the optimally selected model; and (iii) unjustified restrictions on the heterogeneity distribution, e.g., in terms of a pre-specified number of support points, may cause substantial bias.