Fall

October 8th

3:00 PM ET

Johan Ugander
Stanford

Title: Harvesting randomness to understand computational social systems

Abstract: Modern social systems are increasingly infused with algorithmic components, designed to optimize various objectives under diverse constraints. Examples include content ranking algorithms in social media, review assignment algorithms in academic peer review, targeting strategies in social networks, and school choice mechanisms to assign students to schools. In many such systems (and in all of these examples), such algorithms are commonly randomized, motivated by fairness, strategic, or efficiency considerations. In this talk, I will argue for how such randomness can be harvested to make causal inferences not only about the effects of these systems on various outcomes, but also how the system would behave under alternative algorithmic designs, a version of off-policy evaluation. These evaluations come with unique methodological challenges, notably overlap that is often weak or nonexistent, but by developing novel methods to overcome these challenges we can gain confidence about the ways in which these systems operate and their impact on individuals and society as a whole.

November 7th

3:00 PM ET

Pei Geng
University of New Hampshire

Title:  Bias-corrected estimation in functional coefficient autoregressive models with measurement error

Abstract: The functional-coefficient autoregressive (FAR) models are flexible to fit nonlinear features in time series data with covariates. Cai et al. (2000) developed an effective local linear estimation procedure under the FAR framework. When the time series data are observed with measurement error, we first derive the asymptotic bias of the naïve local linear estimator (LLE) by ignoring the measurement error. Then, we propose a bias-corrected local linear estimation procedure for both the functional coefficients and the autoregressive error variance. Simulation studies shows superior performance of the proposed estimation method much reduced bias under various choices of FAR model settings. The asymptotic properties of the bias-corrected estimators are also established. At last, the proposed approach is applied to a cybersecurity real data example.

December 3rd

Location: Votey 209

3:00 PM ET

Peter W. MacDonald
University of Waterloo

Title:  Mesoscale two-sample testing for networks

Abstract: Networks arise naturally in many scientific fields as a representation of pairwise connections. Statistical network analysis has most often considered a single large network, but it is common in a number of applications, for example, neuroimaging, to observe multiple networks on a shared node set. When these networks are grouped by case-control status or another categorical covariate, the classical statistical question of two-sample comparison arises. In this work, we address the problem of testing for statistically significant differences in a given arbitrary subset of connections. This general framework allows an analyst to focus on a single node, a specific region of interest, or compare whole networks. Our ability to conduct mesoscale testing on a meaningful group of edges is particularly relevant for applications such as neuroimaging and distinguishes our approach from prior work, which tends to focus either on a single node or the whole network. Our approach can leverage all available network information and learn informative projections that improve testing power when a low-dimensional latent network structure is present.

Spring

February 17th

3:00 PM ET

Becky Tang
Middlebury College

Title: TBA

Abstract: TBA

March 31st

3:00 PM ET

Erika Edwards
University of Vermont

Title: TBA

Abstract: TBA

April 14th

3:00 PM ET

Jean-Gabriel Young
University of Vermont

Title: Maximum spacing estimators revisited

Abstract: TBA