Featured White Papers
Content provided in partnership with
Government Industry
Operational readiness as a function of maintenance personnel skill level
Air Force Journal of Logistics, Fall, 2007 by Justin R. Chimka, Heather Nachtmann
Regression Modeling
Regression Model Construction
The first step of our regression modeling is to develop candidate regression models for each dependent variable. In order to find good candidate models, seven distinct regression techniques are identified and conducted as described in Figure 4. Each of these regression techniques is employed separately on two subsets of the independent variables. One subset contains percentages of each level of maintainers, number of crew chiefs, and number of total maintainers. The other subset contains the numbers of each level maintainer, number of crew chiefs, and number of total maintainers. This ensures that the percentages and numbers of each level maintainer are never included in the same model, thereby maintaining independence. Figure 5 contains resulting models from each regression technique for the MC rate dependent variable.
Regression Model Selection
The regression model construction step results in multiple candidate models for each dependent variable. The need arose to select the best model for each dependent variable by examining the linear fit of the models, the efficiency of models, and adherence to model assumptions.
The first step is to examine the linear fit of each candidate model. Any candidate model which does not result in a fit parameter (adjusted R-squared value) of 0.64 or greater was eliminated from further consideration, reducing the number of candidate models from 82 to 60. This criterion determines that no candidate model provides a good fit for flying hours and sorties. This result suggests that factors other than personnel skill level are influencing these two performance measures, and therefore flying hours and sorties are eliminated from further analysis.
The next criterion used to select the final models is model efficiency. Here, efficiency is defined as how well the model fits the data (adjusted R-squared) given the number of variable inputs needed to obtain this fit (independent variable terms). Efficient frontiers for each of the six remaining dependent variables are developed by graphing the adjusted R-squared value versus the number of variable terms for each remaining candidate model. Dominant models, or those models that lie on the efficient frontier, are identified as models that achieve better or equal adjusted R-squared values with fewer variable terms. A summary of all candidate models with fit criteria greater than 0.64 is shown in Figure 6. We have identified the most efficient models for each dependent variable, and we have reduced the number of candidate models from 60 to 18.
A summary of the efficiency analysis is given in Figure 7. An abbreviated naming scheme for the candidate models is given by regression analysis technique number and type of skill level data (P for percentage and N for number). For example, a candidate model developed for percentage of skill level data using regression 5 is Regression 5P. Figure 8 presents the efficiency analysis graph for MC rate. Here we can see that candidate models Regression 5P and Regression 7N lie on the efficient frontier as they dominate the other models.
