Graduate Student Seminar
Wednesday, April 1, 2015 · 11 AM - 12 PM
Session Chair | Wenxin Lu |
Discussant | Dr. Huang |
Speaker 1: Rowena Bastero
- Title
-
Alternative Approach to Average Treatment Effect Estimation
- Abstract
- In causal inference, the potential outcome framework posits that every unit has a pair of potential outcomes; one under the control setting and the other under the treatment setting. However, in observational studies, having both outcomes at once is impossible; one is always missing. Such types of data also have systematic differences in the covariates between the two groups, unlike randomized clinical trial data. These issues pose a problem in the estimation of the average treatment effect, which is the target estimand for inference. While propensity scores technique addresses this issue, assessment on the matches made based on propensity score continue to reflect imbalance with respect to the covariates of matched pairs (in direct matching) or matched groups (in stratification matching). With this, a modified method is proposed that guarantees a more balanced group with respect to the covariates. The proposed matching and estimation method infuses classical model-building and swapping techniques into the meta-analyses framework. The swapping procedure imputes the missing potential outcome using classical regression and the use of meta-analyses framework provides a pooled estimate for the average treatment effect. Analysis of a data with continuous response variable suggests that minimal difference in estimates is realized between the proposed method and propensity score technique. It is also concluded that better estimates are generated using the former as reflected by the smaller standard error of the estimated average treatment effect.
Speaker 2: Rabab Elnaiem
- Title
-
Unsupervised statistical learning: clustering
- Abstract
- Modern datasets are known for containing a large number of parameters, given the relatively small number of observation, which present a challenge in finding statistical approaches that meets those datasets or at least start the process of answering questions that can rise with these datasets. Unsupervised statistical learning is a set of techniques that address modern datasets needs; we will be focusing on clustering and the issue of defining a suitable of distance measure, mainly describing the two of the most popular clustering algorithm (k-means and hierarchical clustering), I will end the talk with an application of both clustering methods.