A Spatial Prior for Bayesian Vector Autoregressive Models

In this paper we develop a Bayesian prior motivated by cross-sectional spatial autoregressive models for use in time-series vector autoregressive forecasting involving spatial variables. We compare forecast accuracy of the proposed spatial prior to that from a vector autoregressive model relying on the Minnesota prior and find a significant improvement. In addition to a spatially motivated prior variance as in LeSage and Pan (1995) we develop a set of prior means based on spatial contiguity. A Theil-Goldberger estimator may be used for the proposed model making it easy to implement.     

Combining Geostatistical Methods and Hierarchical Scene Models for Analysis of Multiscale Variation in Spatial Data

Concepts from Hierarchical Analysis of Variance (ANOVA) can be combined with ideas from geostatistics to describe the multiscale structure of spatial data. Hierarchical ANOVA involves modeling spatial data as the sum of effects associated with processes acting at different spatial scales. These effects can be modeled as stationary regionalized variables, whose spatial structure can be described using the variogram. According to this model, the variogram of the spatial data is the sum of variograms and cross‐variograms of the effects. Whereas hierarchical ANOVA reveals the relationship between scale and variability, the hierarchical decomposition of the variogram relates scale with spatial structure. This analysis method can reveal otherwise undetected features of spatial data, and can guide further analysis.     

Bayesian Estimation of Limited Dependent Variable Spatial Autoregressive Models

A Gibbs sampling (Markov chain Monte Carlo) method for estimating spatial autoregressive limited dependent variable models is presented. The method can accommodate data sets containing spatial outliers and general forms of non‐constant variance. It is argued that there are several advantages to the method proposed here relative to that proposed and illustrated in McMillen (1992) for spatial probit models.     

Markovian Models of Movement Within Urban Spatial Structures

Much work has assumed that movement within urban spatial structures is an “adaptive” process. Nonetheless, mathematical models have not yet been specified and tested which formulate both how different individuals “adapt” over time in destination or route selection, and how predictions about aggregate movement can be derived from postulates about different persons. Two adaptive first-order Markov models for heterogeneous individuals are suggested by the literature. When formulated and tested, however, these models are inadequate to describe travel within urban spatial structures. This implies that the use of Markovian processes to model movement may be overrated. More confidence may be placed in other formulations such as linear learning models of route and destination choice.     

Building Spatial Choice Models from Aggregate Data

Abstract Choice model construction is usually based on information about a number of separate choice situations, for which all relevant quantities are known. This paper concerns the case where only higher level, aggregate information is available about the choice results and the prevailing conditions. We demonstrate the applicability of a generic inverse parameter estimation method in estimating a model for grocery store choice. We also propose some enhancements to standard spatial choice models and demonstrate their applicability.     

Using Spatial Filters and Exploratory Data Analysis to Enhance Regression Models of Spatial Data

Residual spatial autocorrelation is a situation frequently encountered in regression analysis of spatial data. The statistical problems arising due to this phenomenon are well‐understood. Original developments in the field of statistical analysis of spatial data were meant to detect spatial pattern, in order to assess whether corrective measures were required. An early development was the use of residual autocorrelation as an exploratory tool to improve regression analysis of spatial data. In this note, we propose the use of spatial filtering and exploratory data analysis as a way to identify omitted but potentially relevant independent variables. We use an example of blood donation patterns in Toronto, Canada, to demonstrate the proposed approach. In particular, we show how an initial filter used to rectify autocorrelation problems can be progressively replaced by substantive variables. In the present case, the variables so retrieved reveal the impact of urban form, travel habits, and demographic and socio‐economic attributes on donation rates. The approach is particularly appealing for model formulations that do not easily accommodate positive spatial autocorrelation, but should be of interest as well for the case of continuous variables in linear regression.     

Computing the Jacobian in Gaussian Spatial Autoregressive Models: An Illustrated Comparison of Available Methods

When estimating spatial regression models by maximum likelihood using spatial weights matrices to represent spatial processes, computing the Jacobian, ln(|I ? λW|), remains a central problem. In principle, and for smaller data sets, the use of the eigenvalues of the spatial weights matrix provides a very rapid resolution. Analytical eigenvalues are available for large regular grids. For larger problems not on regular grids, including those induced in spatial panel and dyadic (network) problems, solving the eigenproblem may not be feasible, and a number of alternatives have been proposed. This article surveys selected alternatives, and comments on their relative usefulness, covering sparse Cholesky and sparse LU factorizations, and approximations such as Monte Carlo, Chebyshev, and using lower‐order moments with interpolation. The results are presented in terms of component‐wise differences between sets of Jacobians for selected data sets. In conclusion, recommendations are made for a number of analytical settings. Al estimar modelos de regresión espacial con el método del máxima verosimilitud (máximum likelihood) y usando matrices de pesos espaciales para representar procesos espaciales, cálculo del término jacobiano (jabobian)—ln(| I ?λ W |)‐ sigue siendo un problema central. En principio, y para bases de datos más pequeñas, el uso de los valores propios (eigenvalues) de la matriz de pesos espaciales proporciona una solución muy rápida. Los eigenvalues analíticos para retículas o grillas grandes y regulares son ya conocidos. Para problemas más grandes, que no se presentan en mallas regulares ‐incluyendo aquellos que se inducen en problemas de paneles espaciales y en problemas de (redes) diádicas‐, es posible que resolver el eigenproblem no sea posible. Este artículo estudia una selección de alternativas y comenta acerca de su relativa utilidad. Se cubren las facorizaciones de tipo Cholesky disperso (sparse Cholesky) y de tipo LU dispersas (sparse LU), las aproximaciones Monte Carlo, y Chebyshev, así mismo se utiliza momentos de bajo‐orden (lower‐order) con interpolación. Los resultados se presentan en términos de diferencias de componentes entre sets de términos jacobianos para bases de datos seleccionadas. En conclusión, se hacen recomendaciones para una serie de contextos analíticos. 当采用表征空间过程的空间权重矩阵对空间回归模型进行最大似然估计时,雅可比矩阵ln(|I?λW|)的计算仍是核心问题。对于小数据集,原则上可利用空间权重矩阵的特征值提供一种快速的解决方案,对于大型规则格网数据特征值分析同样有效。但对于不规则格网大型问题,包括从空间面板和二元（网络）问题中引伸的问题,利用特征值的解决方案可能不适用,对此学术界提出了多种可选替代方案。本文选取已有的几种替代方案并评论各自的相对有效性,其中包括稀疏Cholesky分解和稀疏LU分解法,Monte Carlo和 Chebyshev近似模拟法以及低阶矩插值法。结果以所选数据集雅可比矩阵间特定组份的差异方式显示。最后,推荐了一些分析设定。     

Geographically Weighted Regression: A Method for Exploring Spatial Nonstationarity

Spatial nonstationarity is a condition in which a simple “global” model cannot explain the relationships between some sets of variables. The nature of the model must alter over space to reflect the structure within the data. In this paper, a technique is developed, termed geographically weighted regression, which attempts to capture this variation by calibrating a multiple regression model which allows different relationships to exist at different points in space. This technique is loosely based on kernel regression. The method itself is introduced and related issues such as the choice of a spatial weighting function are discussed. Following this, a series of related statistical tests are considered which can be described generally as tests for spatial nonstationarity. Using Monte Carlo methods, techniques are proposed for investigating the null hypothesis that the data may be described by a global model rather than a non-stationary one and also for testing whether individual regression coefficients are stable over geographic space. These techniques are demonstrated on a data set from the 1991 U.K. census relating car ownership rates to social class and male unemployment. The paper concludes by discussing ways in which the technique can be extended.     

Generalizing Impact Computations for the Autoregressive Spatial Interaction Model

We extend the impact decomposition proposed by LeSage and Thomas-Agnan (2015) in the spatial interaction model to a more general framework, where the sets of origins and destinations can be different, and where the relevant attributes characterizing the origins do not coincide with those of the destinations. These extensions result in three flow data configurations which we study extensively: the square, the rectangular, and the noncartesian cases. We propose numerical simplifications to compute the impacts, avoiding the inversion of a large filter matrix. These simplifications considerably reduce computation time; they can also be useful for prediction. Furthermore, we define local measures for the intra, origin, destination and network effects. Interestingly, these local measures can be aggregated at different levels of analysis. Finally, we illustrate our methodology in a case study using remittance flows all over the world.     

Prediction of Bike-sharing Trip Counts: Comparing Parametric Spatial Regression Models to a Geographically Weighted XGBoost Algorithm

Regression models are commonly applied in the analysis of transportation data. This research aims at broadening the range of methods used for this task by modeling the spatial distribution of bike-sharing trips in Cologne, Germany, applying both parametric regression models and a modified machine learning approach while incorporating measures to account for spatial autocorrelation. Independent variables included in the models consist of land use types, elements of the transport system and sociodemographic characteristics. Out of several regression models with different underlying distributions, a Tweedie generalized additive model is chosen by its values for AIC, RMSE, and sMAPE to be compared to an XGBoost model. To consider spatial relationships, spatial splines are included in the Tweedie model, while the estimations of the XGBoost model are modified using a geographically weighted regression. Both methods entail certain advantages: while XGBoost leads to far better values regarding RMSE and sMAPE and therefore to a better model fit, the Tweedie model allows an easier interpretation of the influence of the independent variables including spatial effects.     

Quick Computation of Spatial Autoregressive Estimators

Spatial estimators usually require the manipulation of n² relations among n observations and use operations such as determinants, eigenvalues, and inverses whose operation counts grow at a rate proportional to n³. This paper provides ways to quickly compute estimates when the dependent variable follows a spatial autoregressive process, which by appropriate specification of the independent variables can subsume the case when the errors follow a spatial autoregressive process. Since only nearby observations tend to affect a given observation, most observations have no effect and hence the spatial weight matrix becomes sparse. By exploiting sparsity and rearranging computations, one can compute estimates at low cost. As a demonstration of the efficacy of these techniques, the paper provides a Monte Carlo study whereby 3,107 observation regressions require only 0.1 seconds each when using Matlab on a 200 Mhz Pentium Pro personal computer. In addition, the paper illustrates these techniques by examining voting behavior across U.S. counties in the 1980 presidential election.     

Anisotropic Variance Functions in Geographically Weighted Regression Models

Most standard methods of statistical analysis used in the social and environmental sciences are built upon the basic assumptions of independence, homogeneity, and isotropy. A notable exception to this rule is the collection of methods used in geographical analysis, which have been designed to take into account serial dependence often observed in spatial data. In addition, recent developments, in particular the method of geographically weighted regression, have provided the tools to model non‐stationary processes, and thus evidence that challenges the assumption of homogeneity. The assumption of isotropy, however, although suspect, has received considerably less attention, and there is thus a need for tools to study anisotropy in a more systematic fashion. In this paper we expand the method of geographically weighted regression in a simple yet effective way to explore the topic of anisotropy in spatial processes. We discuss two different estimation situations and exemplify the proposed technical development by means of a case study. The results suggest that anisotropy issues might be a fairly common occurrence in spatial processes and/or in the statistical modeling of spatial processes.     

A Spatial Autoregressive Poisson Gravity Model

In this article, a Poisson gravity model is introduced that incorporates spatial dependence of the explained variable without relying on restrictive distributional assumptions of the underlying data‐generating process. The model comprises a spatially filtered component—including the origin‐, destination‐, and origin‐destination‐specific variables—and a spatial residual variable that captures origin‐ and destination‐based spatial autocorrelation. We derive a two‐stage nonlinear least‐squares (NLS) estimator (2NLS) that is hetero‐scedasticity‐robust and, thus, controls for the problem of over‐ or underdispersion that often is present in the empirical analysis of discrete data or, in the case of overdispersion, if spatial autocorrelation is present. This estimator can be shown to have desirable properties for different distributional assumptions, like the observed flows or (spatially) filtered component being either Poisson or negative binomial. In our spatial autoregressive (SAR) model specification, the resulting parameter estimates can be interpreted as the implied total impact effects defined as the sum of direct and indirect spatial feedback effects. Monte Carlo results indicate marginal finite sample biases in the mean and standard deviation of the parameter estimates and convergence to the true parameter values as the sample size increases. In addition, this article illustrates the model by analyzing patent citation flows data across European regions. En el presente artículo, se introduce un modelo de gravedad Poisson, que incorpora la dependencia espacial de la variable explicada, sin apoyarse en presunciones de distribución restrictivas del proceso subyacente de generación de datos. El modelo comprende de un componente espacialmente filtrado, que incluye las variables de origen, destino y origen‐destino específico; y una variable espacial residual que captura la auto‐correlación espacial basada en el origen y destino. Se deriva del calculador (2NLS) de dos etapas no lineales de mínimos cuadrados (NLS), el cual es robusto en heterocedasticidad, y por ello controla el problema de sobre‐dispersión o baja‐dispersión (over and under dispersion), que a menudo se presenta en el análisis empírico de datos discretos; o, en el caso de de sobre‐dispersión, cuando se presenta la auto correlación espacial. Este calculador puede demostrar tener propiedades deseables para diferentes supuestos distribucionales, como los flujos observados un componente (espacialmente) filtrado, ya sea Poisson o binomial negativo. En nuestra especificación de modelo espacial auto regresivo (SAR), las estimaciones de los parámetros resultantes se pueden interpretar como los efectos de impacto total implícitos, definidos como la suma de efectos espaciales, directos o indirectos, de retroalimentación (feedback). Los resultados Monte Carlo indican sesgos marginales de muestras finitas en la media y la desviación estándar de los parámetros estimados, y la convergencia de los valores de los parámetros reales, a medida que aumenta el tamaño de muestra. Este artículo ilustra el modelo mediante el análisis de flujos de datos de citas de patentes, a través de las regiones europeas. 本文提出了一种蕴含空间依赖的泊松引力模型,该模型中解释变量无需依赖潜在数据生成过程的限制性分布假设。该模型由包含起点、终点、起点‐终点特定变量的空间滤波组分和空间残差变量组成,能捕捉到基于起点和终点的空间自相关。我们推导出一个二阶非线性最小二乘（NLS）估计（2NLS）,它对异方差具有鲁棒性,从而可控制对于离散或过离散数据经验性分析中经常出现的过离散和低离散问题。如果空间自相关存在,过离散数据分析就是一个例子。对于不同的分布假设,如或泊松分布或是负二项式分布的观测流或（空间）滤波组分,该估计量显示出令人满意的性能。在本文的空间自回归（SAR）模型设定中,参数估计结果可解释为隐含的全局影响效应,并可被定义为直接和间接的空间反馈效应之和。蒙特卡罗结果给出了参数估计中均值、标准差的临界有限样本偏差,且随样本量增大收敛于真正参数值。此外,本文基于欧洲地区专利引用的流数据进行了模型验证。     

Beyond Moran's I: Testing for Spatial Dependence Based on the Spatial Autoregressive Model

The statistic known as Moran's I is widely used to test for the presence of spatial dependence in observations taken on a lattice. Under the null hypothesis that the data are independent and identically distributed normal random variates, the distribution of Moran's I is known, and hypothesis tests based on this statistic have been shown in the literature to have various optimality properties. Given its simplicity, Moran's I is also frequently used outside of the formal hypothesis-testing setting in exploratory analyses of spatially referenced data; however, its limitations are not very well understood. To illustrate these limitations, we show that, for data generated according to the spatial autoregressive (SAR) model, Moran's I is only a good estimator of the SAR model's spatial-dependence parameter when the parameter is close to 0. In this research, we develop an alternative closed-form measure of spatial autocorrelation, which we call APLE , because it is an approximate profile-likelihood estimator (APLE) of the SAR model's spatial-dependence parameter. We show that APLE can be used as a test statistic for, and an estimator of, the strength of spatial autocorrelation. We include both theoretical and simulation-based motivations (including comparison with the maximum-likelihood estimator), for using APLE as an estimator. In conjunction, we propose the APLE scatterplot, an exploratory graphical tool that is analogous to the Moran scatterplot, and we demonstrate that the APLE scatterplot is a better visual tool for assessing the strength of spatial autocorrelation in the data than the Moran scatterplot. In addition, Monte Carlo tests based on both APLE and Moran's I are introduced and compared. Finally, we include an analysis of the well-known Mercer and Hall wheat-yield data to illustrate the difference between APLE and Moran's I when they are used in exploratory spatial data analysis.     

Geographically Weighted Cox Regression for Prostate Cancer Survival Data in Louisiana

The Cox proportional hazard model is one of the most popular tools in analyzing time-to-event data in public health studies. When outcomes observed in clinical data from different regions yield a varying pattern correlated with location, it is often of great interest to investigate spatially varying effects of covariates. In this paper, we propose a geographically weighted Cox regression model for sparse spatial survival data. In addition, a stochastic neighborhood weighting scheme is introduced at the county level. Theoretical properties of the proposed geographically weighted estimators are examined in detail. A model selection scheme based on the Takeuchi’s model robust information criteria is discussed. Extensive simulation studies are carried out to examine the empirical performance of the proposed methods. We further apply the proposed methodology to analyze real data on prostate cancer from the Surveillance, Epidemiology, and End Results cancer registry for the state of Louisiana.     

Aggregation and Ecological Effects in Geographically Based Data

Statistics calculated using the means of geographic areas can differ substantially from the corresponding statistics based on data from individuals. Analysts who base their conclusions about individual-level relationships on area-level analyses run the risk of committing the ecological fallacy. Statistical models are proposed that capture the essential features of the structure of a population composed of geographically defined groups and can encompass grouping processes and contextual effects. These models are used to show how small effects in the analysis of individual-level data can be magnified substantially when the corresponding analysis based on aggregated data is carried out. Thus the source of aggregation effects is exposed. While aggregation effects have been studied by many authors, no general approach has been offered to the problem of adjusting an area-level analysis so as to correct for aggregation effects and hence remove, or at least reduce, the bias that leads to the ecological fallacy. The statistical models proposed are used to provide an approach to this problem. Data from the 1991 U.K. Census of Housing and Population are used to illustrate the size of the aggregation effects and the extent to which the proposed adjustments succeed in their objective.