You are here:
A Bayesian network analysis of the Federal Employee Viewpoint Survey (FEVS) for the U.S. Environmental Protection Agency
Citation:
Carriger, John F., C. Acheson, AND R. Herrmann. A Bayesian network analysis of the Federal Employee Viewpoint Survey (FEVS) for the U.S. Environmental Protection Agency. 2019 Bayesialab Conference, Durham, North Carolina, October 10 - 11, 2019.
Impact/Purpose:
Present exploratory analysis of Bayesian networks for survey data with a case study using the Federal Employee Viewpoint Survey data from 2018. This will be presented at an applied Bayesian network-focused conference with colleagues from around the world with a diversity of backgrounds and applications of Bayesian networks. Bayesian networks can be insightful tools for generating insights into survey data.
Description:
Bayesian networks are useful for generating insights from survey data on workforce satisfaction and beyond. Employee viewpoint survey interpretations may be supported by data-driven probabilistic graphical support tools. The capabilities for a Bayesian network survey analysis is demonstrated and explored through an initial analysis of the 2018 Federal Employee Viewpoint Survey (FEVS) response data from personnel at the U.S. Environmental Protection Agency (EPA). The FEVS is a voluntarily-taken survey that has been administered annually to federal employees across the U.S. since 2002. A focused analysis of EPA data was conducted to examine the insights from applying Bayesian networks. First, EPA data were isolated from the rest of the federal employee responses. Three partitions of the EPA survey response data were further made for separate analyses: all data from the EPA, data from only the Office of Research and Development personnel, and data from all personnel except from the Office of Research and Development. Core survey questions were used for this analysis that comprised questions related to viewpoints on workplace experiences, supervision, and employee satisfaction. Demographics and work/life balance questions were not included in this analysis. Each of the three partitions of responses was separately analyzed with Bayesian networks and then compared. An exploratory analysis was first conducted to examine the importance of each variable from contribution to the joint probability of a tree-based network. Node force statistics provided quantitative measures for the centrality of the response question in the model and the visual relationships and arc force measures were used to examine associations. Next, supervised learning was conducted to examine the relationships of the core questions on responses to a target question. The resulting model was used with dynamic profile and target optimization tree methods to develop a priority order and pathway proposals for maximizing a positive response to the target question. Additional approaches for generating insights with the survey data, including clustering of survey questions, were also examined but not fully implemented in this exploratory analysis. Advances in Bayesian network methods for handling large and complex data sets from surveys can allow for clear insights from multivariate survey data and a clarification of potential pathways for optimization under uncertainty.