Inference Based on Alternative Bootstrapping Methods in Spatial Models With an Application to County Income Growth in the United States*

Abstract

This study examines aggregate county income growth across the 48 contiguous states from 1990 to 2005. To control for endogeneity, we estimate a two‐stage spatial error model and implement a number of spatial bootstrap routines to infer parameter significance. Among the results, we find that outdoor recreation and natural amenities favor positive growth in rural counties and property taxes correlate negatively with rural growth. Comparing bootstrap inference with other models, including the recent General Moment heteroskedastic‐robust spatial error estimator, we find similar conclusions suggesting bootstrapping can be effective in spatial models where asymptotic results are not well established.

Introduction: Whither Spatial Econometrics?

“…Spatial econometrics is now commonly applied in regional science research. For example, just in the December 2011 issue of the JRS , Monchuk et al (2011) examined rural income growth using a spatial error model, Rickman and Rickman (2011) examined population growth with a spatial lag approach, and Park and von Rabenau (2011) examine agglomeration spillovers with a simultaneous spatial approach.…”

Introducing bootstrap methods to investigate coefficient non-stationarity in spatial regression models

“…Alternative stationary coefficient nulls can also be considered, such conditional autoregressive nulls, in which the random part of the model is specified in terms of conditional probability distributions (Cliff and Ord, 1973). Here, advances on the bootstrap method itself may be useful, such that found in Fingleton and Legallo (2008), Lahiri (2010), Burridge and Fingleton (2010), Monchuk et al (2011), Han and Lee (2012), Herrera et al (2013). Finally moving forward to empirical studies, where the true value of this bootstrap method will be realised, the results from this study (and possible extensions) can help guide the selection of a given SVC model over an associated stationary coefficient null, that suits the properties of the real study data and the analytical questions being posed.…”

Estimating standard errors in spatial panel models with time varying spatial correlation

“…The first avenue focuses on expanding the evaluated methods to include bootstrapping and subsampling techniques. Both bootstrapping and subsampling have been explored in spatial data (less so in spatial panel data, see Warren ; Monchuk et al ) but several challenges remain. Chief among these are how to sample and whether or not the correlation structure should be maintained in the sampling process.…”