Wolfgang Nagl (p), Philipp Steinbrunner, Marina Di Giacomo

Katarzyna Kopczewska

In this paper we investigate the effect of Donald Trump’s campaign for coal in his successful race for the White House in 2016. Using a spatial Durbin model we estimate the effect of coal production on the Republicans vote share in the US Presidential Election of 2016 on the county level. To avoid biased estimates we take spillover effects into account and use spatial clustering. We find a significant positive effect. The effect becomes even more pronounced when we use the vote-share difference between Mitt Romney in 2012 and Donald Trump in 2016 as the dependent variable. The positive effect of coal production on the Republican vote share are retained after allowing for non-linear effects of coal production and using coal production per worker and per working hours as main explanatory variable.

Sandy Dall'erba (p)

Wolfgang Nagl

The difference-in-difference (DID) framework is now a well-accepted method in quasi-experimental research. However, DID does not consider treatment-induced changes to a network linking treated and control units. Our instrumental variable network DID methodology controls first for the endogeneity of the network to the treatment and, second, for the direct and indirect role of the treatment on any network member. Monte Carlo simulations and an estimation of the drought impact on global wheat trade and production demonstrate the performance of our new estimator. Results show that DID disregarding the network and its changes leads to significant underestimates of overall treatment effects.

Katarzyna Kopczewska (p), Maria Kubara

Sandy Dall'erba

Akaike Information Criteria (AIC) is the most common metric to assess the quality of econometric models. It reacts to changes of variables in models and sample size. However, until now, there have been no studies to explain how much AIC responds to changes in a number of spatial neighbours indicated by spatial weights matrix W. W impacts the spatial lag – the average value in the neighbourhood, calculated with less or more observations depending on neighbourhood structure. We find that in spatial econometric models for point-pattern, AIC depends functionally on a number of k nearest neighbours (knn) selected for W and that the optimal number of k nearest neighbours knn included in W can be chosen by minimising AIC. This implies that AIC may serve as criteria for model selection and also to determine the neighbourhood structure. This opens the path to select in a non-arbitrary way the range of neighbourhood knn, which is absent in the literature. We use simulation to answer a few questions. First, we check the features of this functional relationship and try to generalise this pattern. We assess the monotonicity of functions and point of maximum. Secondly, we test how the structure of the dataset impacts this relationship by analysing the influence on AIC of proportions of different point patterns and density of points as well as the sampling from the bigger dataset. We are to detect how AIC reacts if the point pattern is a mixture of different spatial distributions overlaid on the same surface. In simulation and empirical data, we prove that AIC is a non-linear function of knn and reaches its minimum for a given knn. This is an important aspect in deriving W for efficient modelling. Solution of AIC-based selection of knn in spatial models will be especially helpful in modelling regional concentration of economic activities and local real estate / housing markets.