Graduate Student Seminar
Wednesday, November 1, 2017 · 11 AM - Noon
|Session Chair||Rabab Elnaim|
|Discussant||Dr. Justin Webster|
Speaker 1: Zhou Feng
- Inverse Probability Weighting for Analyzing Survival Data in Observational Studies
- Observational studies play a big role in reviews and regulatory researches on post-marketing drug safety. Unlike in randomized clinical trials, comparator groups of observational studies are typically imbalanced on patients' characteristics and risk factors associated with outcomes, which could bias an unadjusted treatment effect estimate.
Inverse probability weighting is one of the statistical approaches that can be used to adjust for confounding and selection bias due to measured confounding. It gives each subject a weight so that in the weighted or pseudo-population, patients have balanced characteristics. Stabilized weights are often used to normalize unstabilized weights to make the weights less variable and maintain the same overall sample size of the pseudo-population as in the original sample size. A few studies have shown that for continuous and binary outcomes, stabilized and unstabilized weights lead to same results . However, it is not clear if the finding can be applied to survival data. Therefore, we focus on survival outcome and evaluate the impact of stabilized weight vs. unstabilized weights on the estimates in survival data.
Speaker 2: Joshua Hudson
- Data assimilation for the magnetohydrodynamic equations
- Data assimilation, simply stated, is a way of incorporating past measurements of an observable into a model to provide a more complete state of the system than an interpolation of the present measurement data would provide. It was first introduced as a way to predict weather phenomenon by using only sparse temperature readings. Recently, a data assimilation method using a feedback control was developed by Azouani and Titi, which has been rigorously analysed mathematically. Applying their methods to the magnetohydrodynamic equations in 2D, we obtain theoretical results and compare with numeric simulations to see how we might expect the algorithms to perform in practice.