Study population
The study population includes 5,239 light manufacturing workers of a global aluminum company at six U.S. facilities employed during the period from January 1, 2002 through December 31, 2007. We select workers in the light manufacturing segment of the company because this sector employs a large proportion of female employees; other sectors did not have sufficient numbers of female managers to include in this analysis. The data were obtained from sources described in previous publications [6, 10, 14–16]. Briefly, a comprehensive real-time incident management system requires recording of all first-aid and reportable injury events. This database is linked to administrative human resource, health, work-environment and socio-demographic databases for research purposes. Employees from departments in which the majority of workers were salaried or from departments with less than ten employees were not included in the analytic cohort. Further, we did not include departments for which there was no clearly defined manager. The analytic cohort consists of 4,967 hourly employees and 272 managers across 99 departments.
Identification of managers and manager gender category
An employee with a recorded job title suggestive of a leadership role (examples include “leader”, “manager”, “foreman” and “supervisor”) was considered to be a manager. Managers were defined at the midpoint of each year (July 1) and manager type was defined for all departments at this point. Departments were categorized based on whether they had female managers only, male managers only, or both female and male managers.
Outcomes
The outcome of interest is the time from the start date of an employee in a new department to his or her first acute injury. Employees who worked in more than one department during the study period may be included in the risk set more than once; however, repeated injuries within the same department are not considered as the majority of injuries in this cohort were first injuries (>70%). Injury outcomes were obtained from the incident management system. Only acute injuries, such as burns, lacerations, contusions and fractures, were considered.
Statistical methods
Cox proportional hazards models were used to model time to first acute injury. Employees were censored administratively on December 31, 2007, when they quit or were terminated, or when they changed departments, whichever occurred first. As we expect correlation among employees within the same department, we employed shared frailty models to adjust for correlation of injury risk. The shared frailty model incorporates a random effect for each department and assumes that these terms follow a gamma distribution [17–19].
Manager type (female manager only, male manager only, or both male and female managers) was defined each year for all employees and was included as a time-varying covariate. Manager type was defined for each department on July 1 of each year; employees were assigned a manager type in each year based on the department they were in on July 1. The predictor of interest was the interaction between employee gender and manager type category.
At the department level, we defined departments where the physical demands of the work was likely to be particularly high using the most common job titles within the department. If one or more of the three most common job titles within a department contained the words “metal,” “heat,” “foundry,” “welder,” “furnace,” or “kiln,” the department was defined as having jobs that required particularly high physical demand. This variable is important to include in the analyses because physical demand is related to injury risk and there are more male employees and male managers in high demand departments.
In the primary analyses, we fit three models: an unadjusted model (Model 1) adjusted for time-varying manager gender type, employee gender, and the interaction term, an adjusted model that additionally adjusted for employee race/ethnicity, employee age, employee tenure at the company, and whether the department included jobs that required high physical demand or not (Model 2), and an adjusted model that additionally included a frailty term at the department level (Model 3). All models were adjusted for fixed effects at the facility level. To address concerns about the presence of a secular trend in injuries during the study period, we further adjusted for calendar year as a time-varying covariate in all models. Employees who were ever managers during the study period were excluded from modeling.
In secondary analyses, we categorized departments as having at least one female manager or having no female managers and explored the interaction between having a female manager and the employee’s gender. We also explored the issue of discordance, i.e. male employees with female managers only and female employees with male managers only compared to all other employee-manager pairings. While only medical treatment, restricted work, and lost work time injuries are reportable to OSHA, the majority of injuries in this cohort were first aid only. Most studies of injury only include OSHA-reportable injuries (as they are likely the only ones recorded in employee databases) and do not track first-aid injuries though these types of injuries are likely very different. In secondary analyses we examined these categories of injury as outcomes separately.
To ensure the robustness of our results we conducted several sensitivity analyses. First, results from Cox models were compared to parametric survival models utilizing Gompertz and Weibull distributions to assess whether the choice of a proportional hazards model was reasonable. Second, to further examine the potential pathway of work-family conflict, we explored including having a dependent child under the age of 6 as a possible mediating variable. Third, we explored the previous year’s manager type instead of the current year’s manager gender category to address the issue of employees changing managers and/or departments in response to an injury. Finally, since the percentage of employees with each manager gender category changed dramatically during the early study period (see below), we performed a sensitivity analysis using only employee-departments starting after January 1, 2004 when manager gender categories had stabilized (Figure 1).
Analyses were performed with SAS software, Version 9.3 (SAS Institute Inc., Cary, NC) of the SAS System for Windows and Stata 12 (StataCorp, College Station, TX).
The protocol was approved, invoking the epidemiologic exemption waiving the requirement for individual consent, by the Stanford University IRB.