by large industrial sources with relatively similar production processes and pollution treatment technologies across facilities. Sources, however, may be geographically diverse. Deterrence effects from this dataset were examined in Gray and Deily (1996) and Deily and Gray (2007). The key data source was the EPA’s Compliance Data System (CDS), which has since been updated and incorporated into the Integrated Data for Enforcement Analysis (IDEA) system. Air compliance status outcomes, and not air emissions, are investigated.
Results from applying the simplified models for measuring specific deterrence to the steel mill dataset are presented in Table 1A and Table 1B. Table 1A presents regression specifications with key explanatory variables defined as ‘inspection lagged 1-2 years ago.’ Table 1B presents regression specifications with key explanatory variables defined as ‘inspection lagged 1-3 years ago.’ Nearly all of the models show a statistically significant specific deterrence effect. The estimated impact of lagged inspections on the compliance status dependent variable is always positive and typically strongly significant for both lag specification types. The magnitudes are also quite large. For example, in Table 1A, all three simplified models indicate that a plant with at least one inspection 1-2 years ago was 17-19 percent (.17-.19) more likely to be in compliance than a plant with no inspections 1-2 years ago. Specific deterrence effects of lagged inspections on air compliance in Table 1B are relatively similar in practice. Here, we find that plants with at least one inspection 1-3 years ago were approximately 27 percent (.27) more likely to be in compliance than plants with no inspections 1-3 years ago.
The 0.487 fixed effects estimate in Table 1B suggests that plants with at least one inspection 1-3 years ago were approximately 49 percent (.49) more likely to be in compliance than average, after controlling for changes in plants with no inspections 1-3 years ago. This latter fixed effect result is considerably larger than other estimated inspection impacts (~27 percent) and may represent an outlier. Statistical estimates can vary with the regression approach, and the presence of an anomalously large or small impact suggests that multiple regressions approaches may be useful to understand the sensitivity of empirical results to the chosen statistical technique. When results are particularly sensitive to the regression approach, conclusions should be based on the most consistent or conservative estimate. In other words, conservative, or at least average, deterrence magnitude estimates should typically be selected when estimates vary across specifications.
Most importantly, results presented in Tables 1A and 1B are reasonably similar to those found in the peer-reviewed studies Gray and Deily (1996) and Deily and Gray (2007). Since published deterrence effects for this steel dataset were derived with different models than those reported here, small adjustments are necessary to make results comparable. Gray and Deily’s (1996) key explanatory variable was an indicator measuring the existence of an inspection in two earlier years, so the results in Table 1B are most closely comparable to that study’s results. Gray and Deily (1996) found an unadjusted logit coefficient of 1.13. The equivalent unadjusted coefficient for the closely related inspection variable in columns 2 and 3 of Table 1B is approximately 1.47. Given the slight differences in variable definition across studies, deterrence effects here are