The Impact of Exogenous Shocks on House Prices: the Case of the Volkswagen Emissions Scandal

Abstract

This study analyzes to what extent the announcement on 9/18/2015 of the VW diesel emissions scandal affected house prices in the vicinity of Chattanooga, TN, the location of the only VW production plant in the United States at that time. We examine the impact of the announcement with house transactions data for the 3 years from 2014 to 2016. We explore a number of alternative methods, including a permutation test, to tie down causation. Our results indicate that the brunt of the negative impact occurred 61 to 90 days after the announcement, with no statistically significant negative effects after 90 days. Although the average price discount for the study area is modest at about 3% to 4.5%, the effect tends to be significantly larger for locations closer to the VW plant. However, geographical distance has a distinctly non-linear influence on the price discount.

Appendix

We use the geographical dimension of the data to artificially create control and treated groups of observations. Apriori we expect all observations in our sample to be affected by the announcement event. Hence, any significant difference between these two groups is evidence of a treatment effect. Our proposed procedure is based on the Fisher-Pitman permutation test described, for example, in Oden and Wedel (1975) or Sprent (1998), chapter 1.6. It is an attractive alternative to the traditional t or F tests, as it does not require any distributional assumptions.

Our proposed procedure is robust to the presence of time effects or any factors with independently distributed loading coefficients. Hence, we provide an alternative to the control group approach as suggested, among others, by Hsiao et al. (2012) and Du and Zhang (2015).

As discussed in Section “Robustness Checks”, we use the distance to the VW plant to separate the observations into control and treatment group. The treatment group consists of houses near the VW plant and the control group contains houses far from the VW plant. Denote by d_i the distance of a house i to the VW plant. We define a dummy variable C_i to indicate whether the distance of observation i is larger than a chosen critical distance d_c, i.e.

$$ C_{i}=I\left( d_{i}>d_{c}\right). $$

Our proposed test statistics is then defined as:

$$ test_{m}=\frac{1}{\left( \mathbf{1}_{n\times1}-\mathbf{C}\right)^{\prime}\mathbf{T}_{m}}\left[\sum\limits_{i=1}^{n}\left( 1-C_{i}\right)T_{im}u_{i}\right]-\frac{1}{\mathbf{C}^{\prime}\mathbf{T}_{m}}\left[\sum\limits_{i=1}^{n}C_{i}T_{im}u_{i}\right], $$

where u_i are the residuals from our hedonic regression.²⁵T_m is an indicator variable for a particular month m; T_m is an n × 1 vector with elements T_im. In analogy, we use C to denote the n × 1 vector with elements C_i. Hence, ${n_{m}^{C}}=\mathbf {C}^{\prime }\mathbf {T}_{m}$ is the number observations in the control group for month m and ${n_{m}^{T}}=\left (\mathbf {1}_{n\times 1}-\mathbf {C}\right )^{\prime }\mathbf {T}_{m}$ is the number of observations in the treated group for month m.

To show the robustness of the test statistic under the null hypothesis, we can assume that the disturbances contain common factors of the following form:²⁶

$$ u_{i}=\sum\limits_{m=1}^{T}T_{im}\left( \lambda_{i}f_{m}+\varepsilon_{im}\right), $$

where λ_i is the (individual specific) loading coefficient, f_m is the time-varying common factor and ε_im is an underlying random disturbance. That is, we assume that each house has an underlying process for a price innovation (disturbance) at each time and that this is reflected in the observed transactions price (based on the date of the transaction).²⁷ Thus the test statistic becomes:

$$ \begin{array}{@{}rcl@{}} test_{m} & = & \frac{1}{{n_{m}^{T}}}\left[\sum\limits_{i=1}^{n}\left( 1-C_{i}\right)T_{im}\left( \lambda_{i}f_{m}+\varepsilon_{im}\right)\right]-\frac{1}{{n_{m}^{C}}}\left[\sum\limits_{i=1}^{n}C_{i}T_{im}\left( \lambda_{i}f_{m}+\varepsilon_{im}\right)\right]\\ & = & \frac{1}{{n_{m}^{T}}}\left[\sum\limits_{i=1}^{n}\left( 1-C_{i}\right)T_{im}\lambda_{i}f_{m}\right]-\frac{1}{{n_{m}^{C}}}\left[\sum\limits_{i=1}^{n}C_{i}T_{im}\lambda_{i}f_{m}\right]\\ & & +\frac{1}{{n_{m}^{T}}}\left[\sum\limits_{i=1}^{n}\left( 1-C_{i}\right)T_{im}\varepsilon_{im}\right]-\frac{1}{{n_{m}^{C}}}\left[\sum\limits_{i=1}^{n}C_{i}T_{im}\varepsilon_{im}\right]. \end{array} $$

Assuming that ε_im, λ_i and f_m are independently distributed with finite 2 + δ for some δ > 0 moments, we will have that for ${n_{m}^{T}}$ and ${n_{m}^{C}}$ converging to infinity (as n →∞), the test statistic will converge in probability to zero.