Electoral forensics: Testing the "free and fair" claim
Abstract
Democratic elections are the international norm for government legitimacy. Currently, election observers are the primary examiners of claims of democracy. Unfortunately, cost and access restrict their effectiveness. Electoral forensics complements these observers by statistically testing official results for evidence of violations of the democratic hypothesis. This research seeks to revise statistical methods to increase their applicability and improve their statistical properties vis-a-vis the data types generated in elections. It focuses on three types of data typically reported by official agencies: candidate vote count, invalidation count, and geographical location. Currently, the Benford test serves as the customary vote-count test. I improve this test by modifying it to consider the electoral division size. Similarly, present methods treat divisions as from a single population. I correct this by including a threshold (change-point) in the model. As such, the regression tests become more powerful, especially when used in conjunction with feasible generalized least squares regression. Finally, analyses tend to ignore the geographic nature of elections. I create the spatial-lag expansion model (SLEM) to compete with the popular geographically weighted regression (GWR) model. My generalized Benford test improves the Benford test. However, the power is slight for two of the versions. Allowing for two populations in the election results increases the power of the regression tests while not affecting their sizes. Finally, SLEM betters GWR in terms of speed and power; however, the distribution of its test statistic must still be estimated using simulation. Even with this advancement in the discipline, there remain a couple areas needing further treatment. First, GWR is attractive in that it allows modeling the data more closely. Further work should be done in devising an improved hypothesis-testing paradigm for it. Finally, electoral systems have unique aspects. Future research will be done in modeling electoral rules to create tests optimized for these unique conditions.
Collections
- OSU Dissertations [11222]