specific random effects, and (3) logit regression with plant-specific conditional random effects. The discrete nature of the 0/1 compliance status metric suggests non-linear logit regressions are more appropriate than linear regression models. All methods are discussed in Section 2 and the technical appendix of the Task 1 white paper. All are easily implemented (pre-programmed) with modern statistical software. All models include year-specific indicator variables and period-specific indicator variables.
Key Simplification and Justification 1: Detailed variables representing plant and community characteristics assembled from non-EPA datasets are omitted. State indicator variables, time indicator variables, and panel data statistical techniques account for these omitted factors in our simplified models. State indicator variables capture community and regulatory differences across states. Year indicator variables capture common technological change, sector maturation, and economic fluctuations over time. Panel data statistical techniques (fixed effects, random effects, conditional random effects) capture systematic plant characteristics like age, capacity, industrial sub-category, and profitability. The key assumption underlying this simplification is that technical change is relatively modest, regulations are fairly static, and managerial attitudes are not evolving rapidly for most facilities over the sample period.
Key Simplification and Justification 2: Concerns about reverse causality are substantially less significant for the measurement of general deterrence than for the measurement of specific deterrence. In short, monitoring and enforcement targeting at any given plant has less to do with emissions or non-compliance at other facilities than emissions or non-compliance at the plant in question. Lags do not typically need to be as far in the past and reverse causality is not typically crucial. Therefore, only minor additional simplifications are required for measuring general deterrence with cost effective methods. These simplifications include ignoring statistical techniques designed to improve the statistical precision of the estimation. The magnitudes of the deterrence estimates are unaffected by these more minor statistical considerations.
5. Benchmarking the Simplified Methods
In this section, we first benchmark the simplified models for measuring specific deterrence presented in Section 3. We then benchmark the simplified models for measuring general deterrence in Section 4. In each case, we use a dataset that has been extensively analyzed in the existing peer-reviewed literature. The goal is to evaluate the simplified deterrence models and compare key results to those published using the same datasets with more expensively implemented statistical methods.
Benchmarking Specific Deterrence Methods
Our benchmark dataset for the specific deterrence measurement methods is a steel industry dataset of 41 mills for the period 1980-1989. The steel industry is characterized