Online Appendix I: Additional Robustness Checks
Robustness check #1: System GMM results using subsets of instruments
Table A1 shows that our main system GMM results on the effect of RC are robust to the use of subsets of instruments. Several specifications yield positive and statistically significant estimates that are similar in magnitude to our main results (Table A1, columns 1, 3, 4, 6, 7, 9). In no specifications, do we find that RC reduces plants’ pollution (Table A1).
Robustness check #2: Redefinition of instruments
We re-construct our BLP-type instruments) using plants owned by the same firm located in different states to minimize potential spillovers across nearby plants. This strategy builds on findings in Shimshack and Ward (2005) and Gray and Shadbegian (2007) that inspections at a plant can result in general deterrence at other plants in that state; but this general deterrence is limited within the state. This specification reduces our number of observations to plants owned by firms that operate in multiple states, but did not qualitatively change our parameter estimates or our over-identification test statistics.
Robustness check # 3: Alternative dependent variables
Our results that RC raises pollution is robust to the alternative measure of air pollution i.e., pounds of air pollution. We find that RC raises pounds of air pollution by 6.1% in the multi-plant sample and by 9.8% in the all plant sample (Table A2, column 3-4). These estimates are about one third to two thirds the size of estimates for toxicity-weighted air pollution.
Online Appendix II: Propensity Score Matching (PSM)
We usepropensity score matching (PSM) as an alternative method to estimate the effect of RC. The PSM method, unlike the System GMM method, does not rely on the IV exclusion restrictions. The PSM estimates the average treatment effects on the treated. We apply several matching procedures as a check for robustness, i.e., kernel matching (Table A3, column 1), 5-nearest neighbors matching (Table A3, column 2 and 3), 1-nearest neighbor matching (column 4 and 5). Our preferred estimates from the nearest neighbor matching procedure are matches with common support restrictions (Smith and Todd, 2005).
The PSM estimates, particularly, those from kernel matching (0.147) and the 5-nearest neighbor matching with common support (0.169) are fairly similar in magnitude to our GMM estimates (0.159 to 0.204). The PSM estimate from the 1-nearest neighbor matching with common support (0.068) is about one third to almost one half the size of the GMM estimates. The PSM estimates, however, are not statistically significant at conventional levels.
Notes: Kernel matching uses all control observations in matching against the treated observations, but weights the observations by the closeness of the match using an Epanechnikov kernel. For the nearest neighbor procedures with the common support restrictions, we exclude the treated observations whose estimated likelihood of treatment is greater than the highest estimated likelihood of the untreated observations. When matching is done with replacement, a control case can be used more than once, allowing for more precise matching. We use a probit specification to generate our propensity score estimates.
Online Appendix III: Impact of RC on plant-level pollution intensity
We examine the impact of RC on plant-level pollution intensity, defined as the ratio of toxicity-weighted air pollution to their number of employees. Lower pollution intensity would imply a more favorable trade-off between production and pollution (Cole et al., 2005); i.e., the same production would result in lower pollution.
We find that RC raises pollution intensity by 15.1% in the multi-plant sample (Table A4, column 3). We also find that RC raises pollution intensity by 12.6% in the all plant sample (though we note that the instruments for the all plant sample are imperfect) (Table A4 column 4). We also use the alternative denominator of plant-level employee squared. We continue to find positive coefficients in both the multi-plant and all-plant samples, though these estimates are not statistically significant. We conclude that RC fails to reduce pollution intensity.
The alternative normalization strategy addresses one potential estimation issue from using the number of employee as the denominator. This estimation issue arises if plants respond to RC by choosing a production process that is less labor intensive, but that does not raise pollution per unit of production. Should larger plants increase their output at a faster rate than their use of labor, our denominator for large plants may be too small, resulting in too large a measure of pollution intensity. Given that RC participants typically have larger plants, this mismeasurement of pollution intensity could bias our estimates on the impact of RC. Nevertheless, when we address this potential source of bias using employee squared as a denominator, our conclusions that RC does not reduce pollution intensity continues to hold.
In defining pollution intensity, we use plant-level employee from Dun & Bradstreet to proxy for preferred denominator of plant-level output. This approach is used in other studies, given the absence of publicly available plant-level data (Holladay, 2010). We provide evidence that plant-level employee can serve as a proxy for plant-level output. In the absence of publicly available plant-level data, we examine the relationship between the number of employees and output, measured as value-added and value of shipments, using data at the SIC-4 level from the NBER-CES Manufacturing Industry Database between 1988 and 2001 (Becker and Gray, 2009).
We restrict our analysis to industries in the chemical manufacturing sub-sector (SIC-28). We regress measures of output (value added or value of shipments, normalized to 1987 prices) on the number of employees and SIC-4 and year dummies. We find a strong positive relationship between the number of employees and both measures of output. We find positive and statistically significant coefficients in regressions with either measure of output (Table A5, column 1 and 4). When we add the variable employee squared as an additional explanatory variable, we continue to find strong positive relationships between the number of employees and both measures of output (column 2 and 5). In the regression with the number of employees squared, SIC-4 and year dummies, we find a strong positive relationship between the number of employees squared and both measures of output (column 3 and 6).
Online Appendix IV: Factors associated with RC participation
References for the Online Appendix
Becker, Randy A. and Wayne B. Gray. 2009. “NBER-CES Manufacturing Industry Database:
Updated Data 1958-2005 - SIC and NAICS versions” Retrieved from
Gray, Wayne B. and Ronald J. Shadbegian. 2007. “The Environmental Performance of Polluting
Plants: A Spatial Analysis.” Journal of Regional Science 47(1): 63-84.
Holladay, J. Scott. 2010. “Abate or Abscond: Polluters’ Response to Environmental
Regulation.” presented at the Georgia Institute of Technology economics seminar, October 2010.
Shimshack, Jay and Michael Ward. 2005. “Regulator Reputation, Enforcement, and
Environmental Compliance.” Journal of Environmental Economics and Management 50(3): 519-540
Smith, Jeffrey A. and Petra E. Todd. 2005. Does Matching Overcome LaLonde’s Critique of
Nonexperimental Estimators? Journal of Econometrics 125 (1-2) 305-353.
1