Appendix B: Propensity Score Generation and Control Group Matching
Before statistical matching can be performed, a set of key demographic variables was developed to control for factors external to the Workers’ Compensation data. Because Research & Planning has access to work histories and demographics from 1992 forward, a wealth of demographic and employment data are available. Based on the available data and prior research projects (see http://doe.state.wy.us/lmi for details), the following set of variables were used in this project:
- Total inflation-adjusted earnings quintiles for 2001-2003
- Number of quarters worked from 2001-2003
- Average quarterly wage quintiles from 2001-2003
- Age group (16-19, 20-24, 25-34, 35-44, 45-54, 55-64, 65+)
- Gender
- Industry of primary employer
- Total inflation-adjusted earnings quintile in 2004
- Average quarterly wage quintile in 2004
- Injury severity for subgroups
- Worker residency status (resident or nonresident) in 2004
- Tenure with primary employer
The above indicators were used as independent variables in a binary logistic regression model (Allison, 1999) where the dependent variable was whether a worker filed a WC claim in 2004. The estimated probabilities generated from the model indicated the probability of filing a WC claim controlling simultaneously for the independent variables.
Treatment and control case matching uses the difference in estimated probabilities as a proxy for distance between cases. The cases with the closest probabilities (or closest differences) are matched first, with succeeding iterations matching cases, which are further apart. In this instance, the desired number of control cases to be matched to each treatment case was determined by trial and error. The number of controls desired was determined by optimizing the amount of treatment cases matched and the number of controls selected. Four controls per treatment using all WC claims in 2004 and two controls per treatment using the subset of more severely injured workers were chosen. For further details regarding case control matching methodology, see Parsons (2004). Parsons’ (2004) method was modified to suit the needs of Research & Planning on this project. Once the control groups (all cases and medical cases only) were created, the analysis continued as described in Results III to Results V (see pages 34-38).
"The [Quarterly Census of Employment and Wages] program derives its data from quarterly tax reports submitted to State Employment Security Agencies by over eight million employers subject to State unemployment insurance (UI) laws and from Federal agencies subject to the Unemployment Compensation for Federal Employees (UCFE) program. This includes 99.7% of all wage and salary civilian employment. These reports provide information on the number of people employed and the wages paid to the employees each quarter. The program obtains information on the location and industrial activity of each reported establishment, and assigns location and standard industrial classification codes accordingly. This establishment level information is aggregated, by industry code, to the county level, and to higher aggregate levels." (Bureau of Labor Statistics, 2008)