In these regressions I use "intermediate injuries" as the dependent variable. This category, defined by the National Research Council in their 1982 report on underground coal mine safety, comprises "all fatal and permanent disability injuries as well as all injuries resulting from roof/side falls, machinery, haulage, or electrical/explosive accidents" (NRC 1982:82). The authors' intention was to identify a subset of injuries more common than fatal injuries, but less susceptible to reporting bias than disabling injuries as a whole: they "regard [the intermediate injury rate] as a compromise measure of safety that includes ample numbers of injuries for most statistical purposes and provides for reasonably good consistency between mines in the reporting of injuries" (NRC 1982:83-84). Also see Table W1b for a breakdown of injury type frequency that includes intermediate injuries.
Models with additional EIA variables (see footnote 19, p. 13)
The EIA dataset includes two variables, recoverable reserves and captive production percentage, that are unreliable prior to 1998. "Recoverable reserves" are additional coal deposits at the mine site that have yet to be mined. "Captive production percentage" refers to the portion of coal produced by the mine that is subsequently consumed by the mining operation itself. In these regressions I limit my sample to post-1997 observations but include these additional fields as covariates.
Mine fixed-effects models (see footnote 16, p. 10)
In these regressions I restrict the sample to mines that switched unionization status (i.e. unionized or de-unionized) at some point during the period of observation. There are reasons to think that this sample is unrepresentative, but I present these results for the sake of completeness.
Controller fixed-effects models (see footnote 16, p. 10)
For these regressions I restrict the sample to mines whose controllers simultaneously owned a union mine and a nonunion mine for at least one quarter. This allows me to simultaneously estimate a union effect and a fixed effect for each controller.
Propensity score matching (PSM) models (see footnote 7, p. 7)
For these regressions, I preprocess the data using a propensity score matching method so that the treated (union) mines are as similar as possible to the control (nonunion) mines. The purpose of this procedure is to reduce the dependence between the treatment variable (unionization) and the other mine characteristics I include as covariates. There are two steps to this process. First, I run a logistic regression with union status as the dependent variable. The goal here is to predict the likelihood that a mine will be unionized in a given quarter, conditional on its observable characteristics; this likelihood is referred to as its "propensity score." Second, within each quarter I identify pairs of union and nonunion observations that have very similar propensity scores. (I use the psmatch2 Stata module to perform the matches.) Mines that are not matched based on this criterion are removed from the sample.
Surface and underground models (see footnote 8, p. 7)
While I restrict the analysis in the main paper to underground coal mines, my dataset also includes all surface mines under MSHA's purview. These regressions use observations from the combined sample of surface and underground mines. Injury counts incorporate both above-ground and underground injuries, and the exposure term is total hours worked at the mine as opposed to underground hours only.
Fatal injury models, excluding "Mine Disasters" (see footnote 19, p. 13)
MSHA defines a "Mine Disaster" as an incident that results in at least five fatalities. I rerun the fatality model after excluding the four mine-years with Mine Disasters, to ensure that these few incidents are not driving my results. The four Mine Disasters occuring during the sample period are as follows: No. 5 Mine, Jim Walter Resources (9/23/01), Sago Mine
(1/2/06), Darby Mine No. 1 (5/20/06), and Crandall Canyon Mine (8/6/07). See MSHA's fact sheet on Mine Disasters for more details. NB: the Upper Big Branch disaster (4/5/10) occured after the end of my sample period.
Fatal injury models, including only fatalities from explosions/collapses (see footnote 19, p. 13)
As a robustness check, I rerun the fatality model after excluding fatalities not resulting from explosions or collapses. I include only fatalities resulting from the following: explosives,
exploding vessels under pressure, breaking agents, fall of face, fall of roof, fire, ignition or explosion of gas or dust, or inundation.
Above-ground injury models (see footnote 13, p. 8)
While I restrict the injury counts in the main paper to incidents that occur underground at underground coal mines, many underground coal mines have surface subunits where injuries and fatalities also occur. As a robustness check, I rerun all models from the main paper but alter the dependent variable injury counts so as to include all injuries occuring at a mine, regardless of where they occur.
Appleton and Baker (1984) replication (see footnote 23, p. 15)
As part of my effort to determine whether the union safety effect has emerged only recently, or has simply eluded detection in prior studies, I analyze my own data using a methodology similar to that of Appleton and Baker (1984).