|
Improving Statistical Conclusion Validity in Mediation Analysis using Bootstrap Procedures
|
| Presenter(s):
|
| MH Clark, Southern Illinois University at Carbondale, mhclark@siu.edu
|
| Steven Middleton, Southern Illinois University at Carbondale, scmidd@siu.edu
|
| Abstract:
The present study describes a bootstrap method for testing mediational relationships that has quickly gained in popularity over the past few years. The bootstrap method is a more robust test that, under certain conditions, can provide more valid results than previous methods. Because the bootstrap method uses several subsamples from the original data to measure the direct and indirect relationships between variables, sampled distributions are less skewed and small sample sizes are less likely to affect the statistical power of the test. To demonstrate its effectiveness, the bootstrap method will be compared to Barron and Kenny's (1986) traditional method for assessing mediation using a data set with a small sample size (n = 64) and non-normally distributed variables.
|
|
Comparing Analysis of Covariance (ANCOVA), Repeated-Measures Analysis of Variance (ANOVA), and Multilevel Longitudinal Design in Causal Modeling of Non-Random Clusters
|
| Presenter(s):
|
| Lihshing Wang, University of Cincinnati, leigh.wang@uc.edu
|
| Abstract:
Quasi-experimental studies involving non-random clusters of subjects are often the norm in evaluation research for drawing causal inferences. When the design involves two or more groups where pretesting is possible but randomization is not, three common statistical procedures exist for modeling the treatment effects: Analysis of Covariance (ANCOVA), which compares the posttest means using the pretest as the covariate; Repeated-Measures Analysis of Variance (RMANOVA), which estimates the interaction effect between the within-subject Time factor and between-subject Treatment factor; and Multilevel Longitudinal Design (MLD), which estimates the treatment effect after adjusting for the intraclass correlation in hierarchically clustered data with Time nested within Subject and Subject nested within Treatment. The present study reviews the theoretical framework and pragmatic utility of these three approaches, arguing for the preference of RMANOVA over ANCOVA and MLD over RMANOVA. Simulation results with known true effects are augmented to support the claim.
|
|
Spatial Regression Discontinuity: Estimating Effects of Geographically Implemented Programs and Policies
|
| Presenter(s):
|
| Christopher Moore, University of Minnesota, moor0554@umn.edu
|
| Abstract:
Estimating causal effects is an important aim in the field of program evaluation, but many programs and policies are implemented in geographically defined jurisdictions, such as school districts or states, and not by randomly assigning participants to a treatment or control group. How might evaluators estimate causal effects in the case of treatment assignment based on geographic borders? Regression discontinuity is a quasi-experimental design and statistical modeling approach that can yield causal estimates that are comparable to those derived from randomized controlled trials. Spatial regression discontinuity is a special case that recognizes geographic borders as sharp cutoff points where local effects can be estimated. This paper details how evaluators can implement spatial regression discontinuity designs that allow causal conclusions. A hierarchical spatial regression discontinuity analysis will be demonstrated in the context of a well-known study of minimum wage effects by Card and Krueger (1994).
|
| | |