The Effect of Emission Information on Housing Prices: Quasi-Experimental Evidence from the European Pollutant Release and Transfer Register

Abstract

In this paper, we study whether the release of pollutant emission information has an effect on housing prices. The event under study is the publication of the first wave of emission quantity data from the European Pollutant Release and Transfer Register in 2009. Our analysis is based on quarterly housing prices at the German postal code level for the years 2007–2011 and provides the first evidence from Europe on this research question. Estimating a differences-in-differences model and controlling for observable differences in land use, housing type distribution, tax revenues and other postal code area characteristics by means of propensity score matching, we find no significant effect of the release of emission information on the value of houses in affected postal code areas. This result survives a number of robustness checks designed to assess whether our findings are due to data aggregation issues or the actual treatment definition. This leads to the conclusion that on an aggregate level the 2009 publication of E-PRTR data did not have an immediate and noticeable effect on housing prices in Germany.

Appendix

1.1 A.1 Data on Postal Code Area Sociodemographic Characteristics

Table 15 INKAR 2010 Indicators used for matching

Sources for the INKAR database include statistical records on the German federal and state level as well as individual data collections of Federal Ministries, the Eurostat Regio database, GfK (Gesellschaft für Konsumforschung) records, official tax records and the accessibility model of the BBSR (Bundesinstitut für Bau-, Stadt- und Raumforschung). The variables in the data set are described in Table 15. The INKAR data we use is the 2010 publication, which contains information about the municipalities at the end of 2008.

The data on postal code characteristics is available at the municipal level (“Gemeinde”) and for aggregated municipalities (so-called “Gemeindeverbände”), but the unit of analysis is a postal code area as mentioned in the main text. Unfortunately, these two spatial entities do not perfectly overlap. As a result we used GIS software to match postal code areas to municipalities based on the spatial overlay of the two entities. For this purpose we combined a postal code area shape file with a shape file of 11,329 municipalities, both provided by GfK GeoMarketing GmbH as of 2012. A number of municipal reforms have led to restructuring of a few hundred municipalities between 2008 and 2012. As a result, we were unable to merge sociodemographic data to a small number of postal code areas. In total we were able to match 8194 out of our 8212 postal code areas to an INKAR unit.

The spatial overlay analysis was done based on urban area in the municipalities as the INKAR data mainly refers to sociodemographic data such as unemployment levels, tax revenues, and the prevalence of types of residential buildings. Each urban area was identified by the municipality code within which it lay. The postal code area was then allocated to the municipality with the largest share of urban area within the postal code area. Almost 300 postal code areas contained no urban areas. These were matched based on the simple spatial overlap to the municipality with the largest share of the postal code area. Once this allocation was complete based on the 2012 municipalities, we used information on changing municipalities from the German Statistical authorities (Destatis) to trace the municipalities that changed back to their 2008 identification number. For the 8194 postal code areas thus matched to the INKAR units, 4618 where allocated to a single municipality and 3576 to an aggregate (“Gemeindeverband”).

1.2 A.2 Construction of Weighted Emission Scores

In order to create postal code quantiles with respect to the severity of treatment (relative severity of emitted substances) we add up the individual emission reports for each postal code weighted by their emission thresholds. For the E-PRTR data release in 2009 this is done according to the following formula:

$$\begin{aligned} {\textit{Weighted}} \ {\textit{emission}} \ {\textit{score}}_{i}(2009)=\sum _{f_{i}=1}^{F_{i}}\left[ \sum _{p=1}^{P}\frac{{\textit{Quantity}}_{p,f_{i}}(2007)}{{\textit{Threshold}}_{p}}\right] \end{aligned}$$

(4)

Here, $f_{i}=1,\ldots ,F_{i}$ denotes the individual facilities in the respective postal code while $p=1,\ldots ,P$ denotes all substances listed in the database with $Quantity_{p,f_{i}}(2007)$ being the reported quantity of an individual substance by the respective facility for the year 2007. This emission quantity is weighted by the reporting threshold defined in the E-PRTR regulations. Overall, this measure is a good proxy for the severity of emissions recorded for a certain postal code. As the reporting thresholds are reasonable proxies for toxicity and are also available to the public as possible guidelines for interpreting the values in the database, this aggregate measure constitutes a sound basis for creating treatment quantiles. A few examples of computed Weighted Emission Scores are presented in Fig. 5.

Fig. 5

Examples for the construction of weighted emission scores

1.3 A.3 Logit Estimations for Propensity Score Matching

See Tables 16 and 17.

Table 16 Logit estimates for matching, part I

Table 17 Logit estimates, part II

1.4 A.4 Development of Housing Prices in Germany

See Fig. 6.

Fig. 6

Price trends (House Price Index), unmatched sample

1.5 A.5 Testing the Common Trend Assumption

The common trend assumption is tested by estimating the following regression setup:

$$\begin{aligned} Y_{it}= & {} \beta ^{0}+\beta ^{1}(1-Post_{t})\cdot t+\beta ^{2}Post_{t}\cdot t+\beta ^{3}(1-Post_{t})\cdot Treated{}_{i}\cdot t\nonumber \\&+\,\beta ^{4}Post_{t}\cdot Treated_{i}\cdot t+\beta ^{5}Post_{t}\cdot Treated_{i}+\beta ^{6}Post_{t}+\beta _{i}^{7}+\varepsilon _{it} \end{aligned}$$

(5)

The main coefficient of interest is the one for a possible trend divergence prior to treatment for the treatment group ($\beta ^{3}$, “Pre-Trend*Treatment” in Table 18). Given the graphical representation of the trends in housing prices, a linear trend model is the preferable choice for our analysis. A sharp turn in the development of German housing prices can be seen to coincide roughly with the publication of the first E-PRTR wave and is not restricted to either of the groups but a universal feature of housing prices in Germany, which can be explained by the notion that there has been a surge in housing prices after the recent financial crisis due to other investment options becoming less attractive. Most importantly, we observe in Table 18 that the common trend assumption holds for both parts of Germany after the matching procedure has been completed. Any pre-treatment trend differences that may have prevailed in the unmatched sample between treatment and control group are eliminated by nearest neighbor matching. It is also worth noting that in the matched sample there is no significant trend differential after 2009Q2, which already indicates that the treatment (consisting of at least one E-PRTR report within a postal-code) may have had little effect on the trend in housing prices.

Table 18 Common trend regressions

1.6 A.6 Histograms of Propensity Scores

See Figs. 7, 8, 9, 10.

Fig. 7

Propensity scores, Western Germany, no buffer

Fig. 8

Propensity scores, Western Germany, 500 m buffer

Fig. 9

Propensity scores, Eastern Germany, no buffer

Fig. 10

Propensity scores, Eastern Germany, 500 m buffer

1.7 A.7 Histograms of Propensity Scores (urban areas only)

See Fig. 11, 12.

Fig. 11

Histograms of propensity scores, urban areas, Western Germany

Fig. 12

Histograms of propensity scores, urban areas, Eastern Germany

1.8 A.8 Geographical Distribution of Treatment and Control Group (matched sample)

See Fig. 13.

Fig. 13

Treatment and control groups with NN matching

1.9 A.9 Geographical Distribution Treatment and Control Group (matched sample, urban areas only)

See Fig. 14.

Fig. 14

Map of treated areas and matched controls, urban areas

1.10 A.10 Mean Characteristics and Bias Comparison for Treatment and Control Group (urban areas only)

See Tables 19, 20.

Table 19 Urban areas only: Treatment and control group before and after matching (A), Western Germany

Table 20 Urban areas only: Treatment and control group before and after matching (A), Eastern Germany

1.11 A.11 Summary of Mean Comparisons and Regression Results

See Table 21.

Table 21 ATET versus regression coefficients (full and matched sample)