The PRIISM seminar series consists of research seminars of interest to an applied statistics audience, from innovative applications of applied statistics to novel statistical theory and methodology. For an archive of past events, here (2008 - 2015) and here (2015 - 2016). Links to related seminars are given here.

**Upcoming and Recent Events: Spring 2017**

Date, Time, Location | Speaker Name, Affiliation | Topic (click for more info) |
---|---|---|

2/28/2017, 9:30 - 10:30 |
Sam Pimentel (UPenn) |
Large, Sparse Optimal Matching in an Observational Study of Surgical Outcomes |

Abstract: How do health outcomes for newly-trained surgeons' patients compare with those for patients of experienced surgeons? To answer this question using data from Medicare, we introduce a new form of matching that pairs patients of 1252 new surgeons to patients of experienced surgeons, exactly balancing 176 surgical procedures and closely balancing 2.9 million finer patient categories. The new matching algorithm (which uses penalized network flows) exploits a sparse network to quickly optimize a match two orders of magnitude larger than usual in statistical matching, and allowing for extensive use of a new form of marginal balance constraint. |
||

3/1/2017, 12:30 - 1:30 |
Patrick Perry (NYU Stern) |
Scaling Latent Quantities from Text: From Black-and-White to Shades of Gray |

Abstract: Probabilistic methods for classifying texts according to the likelihood of class membership form a rich tradition in machine learning and natural language processing. For many important problems, however, class prediction is either uninteresting, because it is known, or uninformative, because it yields poor information about a latent quantity of interest. In scaling political speeches, for instance, party membership is both known and uninformative, in the sense that in systems with party discipline, what is interesting is a latent trait in the speech, such as ideological position, often at odds with party membership. Predictive tools common in machine learning, where the goal is to predict a black-or-white class--such as spam, sentiment, or authorship--are not directly designed for the measurement problem of estimating latent quantities, especially those that are not inherently unobservable through direct means.In this talk, I present a method for modeling texts not as black or white representations, but rather as explicit mixtures of perspectives. The focus shifts from predicting an unobserved discrete label to estimating the mixture proportions expressed in a text. In this "shades of gray" worldview, we are able to estimate not only the graynesses of texts but also those of the words making up a text, using likelihood-based inference. While this method is novel in its application to text, it be can situated in and compared to known approaches such as dictionary methods, topic models, and the wordscores scaling method. This new method has a fundamental linguistic and statistical foundation, and exploring this foundation exposes implicit assumptions found in previous approaches. I explore the robustness properties of the method and discuss issues of uncertainty quantification. My motivating application throughout the talk will be scaling legislative debate speeches. |
||

3/9/2017, 11:30 - 12:30 |
Ravi Shroff (NYU CUSP) |
Simple Rules for Decision-Making |

Abstract: Doctors, judges, and other experts typically rely on experience and intuition rather than statistical models when making decisions, often at the cost of significantly worse outcomes. I'll present a simple and intuitive strategy for creating statistically informed decision rules that are easy to apply, easy to understand, and perform on par with state-of-the art machine learning methods in many settings. I'll illustrate these rules with two applications to the criminal justice system: investigatory stop decisions and pretrial detention decisions. |
||

3/22/2017, 2:00 - 3:30 |
Sharif Mahmood (KSU) |
Finding common support through largest connected components and predicting counterfactuals for causal inference |

Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: Finding treatment effects in observational studies is complicated by the need to control for confounders. Common approaches for controlling include using prognostically important covariates to form groups of similar units containing both treatment and control units (e.g. statistical matching) and/or modeling responses through interpolation. Hence, treatment effects are only reliably estimated for a subpopulation under which a common support assumption holds--one in which treatment and control covariate spaces overlap. Given a distance metric measuring dissimilarity between units, we use techniques in graph theory to find common support. We construct an adjacency graph where edges are drawn between similar treated and control units. We then determine regions of common support by finding the largest connected components (LCC) of this graph. We show that LCC improves on existing methods by efficiently constructing regions that preserve clustering in the data while ensuring interpretability of the region through the distance metric. We apply our LCC method on a study of the effectiveness of right heart catheterization (RHC). To further control for confounders, we implement six matching algorithms for analyses. We find that RHC is a risky procedure for the patients and that clinical outcomes are significantly worse for patients that undergo RHC. |
||

3/23/2017, 11:00 - 12:30 |
Winston Lin | Agnostic Notes on Regression Adjustments to Experimental Data: Reexamining Freedman's Critique |

Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: This talk will be mostly based on my 2013 Annals of Applied Statistics paper, which reexamines David Freedman's critique of ordinary least squares regression adjustment in randomized experiments. Random assignment is intended to create comparable treatment and control groups, reducing the need for dubious statistical models. Nevertheless, researchers often use linear regression models to adjust for random treatment-control differences in baseline characteristics. The classic rationale, which assumes the regression model is true, is that adjustment tends to reduce the variance of the estimated treatment effect. In contrast, Freedman used a randomization-based inference framework to argue that under model misspecification, OLS adjustment can lead to increased asymptotic variance, invalid estimates of variance, and small-sample bias. My paper shows that in sufficiently large samples, those problems are either minor or easily fixed. Neglected parallels between regression adjustment in experiments and regression estimators in survey sampling turn out to be very helpful for intuition. |
||

4/5/2017, 11:00 - 12:00 |
Jared Murray (CMU) |
Log-Linear Bayesian Additive Regression Trees |

Abstract: Bayesian additive regression trees (BART) have been applied to nonparametric mean regression and binary classification problems in a range of applied areas. To date BART models have been limited to models for Gaussian "data", either observed or latent, and with good reason - the Bayesian backfitting MCMC algorithm for BART is remarkably efficient in Gaussian models. But while many useful models are naturally cast in terms of observed or latent Gaussian variables, many others are not. In this talk I extend BART to a range of log-linear models including multinomial logistic regression and count regression models with zero-inflation and overdispersion. Extending to these non-Gaussian settings requires a novel prior distribution over BART's parameters. Like the original BART prior, this new prior distribution is carefully constructed and calibrated to be flexible while avoiding overfitting. With this new prior distribution and some data augmentation techniques I am able to implement an efficient generalization of the Bayesian backfitting algorithm for MCMC in log-linear (and other) BART models. I demonstrate the utility of these new methods with several examples and applications. |
||

4/19/2017, 11:00 - 12:00 |
Carlos Carvalho (UT Austin) |
Bayesian Causal Forests: Heterogeneous Treatment Effects from Observational Data |

Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: This paper develops a semi-parametric Bayesian regression model for estimating heterogeneous treatment effects from observational data. Standard nonlinear regression models, which may work quite well for prediction, can yield badly biased estimates of treatment effects when fit to data with strong confounding. Our Bayesian causal forests model avoids this problem by directly incorporating an estimate of the propensity function in the specification of the response model, implicitly inducing a covariate-dependent prior on the regression function. This new parametrization also allows treatment heterogeneity to be regularized separately from the prognostic effect of control variables, making it possible to informatively “shrink to homogeneity”, in contrast to existing Bayesian non- and semi-parametric approaches. Joint work with P. Richard Hahn and Jared Murray. |
||

4/26/2017, 11:00 - 12:00 |
Jay Verkulien (CUNY) |
Remarks on the Mean-Difference Transformation and Bland-Altman Plot |

Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: Tukey's mean-difference transformation and the Bland-Altman plot (e.g., Bland & Altman, 1986) are widely used in method comparison studies throughout the sciences, particularly in the health sciences. While intuitively appealing, easy to compute, and giving some notable advantages over simply reporting coefficients such as the concordance coefficient or intraclass correlations, they exhibit unusual behavior. In particular, one often observes systematic trends in the BA plot and they are very subject to outliers, among other issues. The purpose of this talk is to propose and study a generative model that lays out the logic of the mean-difference transformation and hence the BA plot, indicating when and why systematic trend may occur. The model provides insight into when users should expect problems with the BA plot and suggests that it should not be applied in circumstances when a more informative design such as instrumental variables is necessary. I also suggest some improvements to the graphics based on semi-parametric regression methods and discuss how putting the BA plot in a Bayesian framework could be helpful. |
||

5/3/2017, 11:00 - 12:00 |
Mireille Schnitzer (University of Montreal) |
Collaborative targeted learning using regression shrinkage |

Abstract: Causal inference practitioners are routinely presented with the challenge of wanting to adjust for large numbers of covariates despite limited sample sizes. Collaborative Targeted Maximum Likelihood Estimation (CTMLE) is a general framework for constructing doubly robust semiparametric causal estimators that data-adaptively reduce model complexity in the propensity score in order to optimize a preferred loss function. This stepwise complexity reduction is based on a loss function placed on a strategically updated model for the outcome variable, assessed through cross-validation. New work involves integrating penalized regression methods into a stepwise CTMLE procedure that may allow for a more flexible type of model selection than existing variable selection techniques. Two new algorithms are presented and assessed through simulation. The methods are then used in a pharmacoepidemiology example of the evaluation of the safety of asthma mediation during pregnancy. |
||

5/10/2017, 11:00 - 12:00 |
Mariola Moeyaert (University at Albany) |
Multilevel modeling of single-subject experimental data: Handling data and design complexities |

Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: There has been a substantial increase in the use of single-subject experimental designs (SSEDs) over the last decade of research to provide detailed examination of the effect of interventions. Whereas group comparison designs focus on the average treatment effect at one point of time, SSEDs allow researchers to investigate at the individual level the size and evolution of intervention effects. In addition, SSED studies may be more feasible than group experimental studies due to logistical and resource constraints, or due to studying a low incidence or highly fragmented population. To enhance generalizability, researchers replicate across subjects and use meta-analysis to pool effects from individuals. Our research group was one of the first to propose, develop and promote the use of multilevel models to synthesize data across subjects, allowing for estimation of the mean treatment effect, variation in effects over subjects and studies, and subject and study characteristic moderator effects (Moeyaert, Ugille, Ferron, Beretvas, & Van den Noortgate, 2013a, 2013b, 2014). Moreover, multilevel models can handle unstandardized and standardized raw data or effect sizes, linear and nonlinear time trends, treatment effects on time trends, autocorrelation and other complex covariance structures at each level. This presentation considers multiple complexities in the context of hierarchical linear modeling of SSED studies including the estimation of the variance components, which tend to be biased and imprecisely estimated. Results of a recent simulation study using Bayesian estimation techniques to deal with this issue will be discussed (Moeyaert, Rindskopf, Onghena & Van den Noortgate, 2017). |

Sign up for our mailing list to receive upcoming event reminders.

**Past Events: Fall 2016**

Date, Time, Location | Speaker Name, Affiliation | Topic (click for more info) |
---|---|---|

9/14/2016, 11:00 - 12:00 |
Stephen H. Bell (Abt Associates) |
Basing Causal Inferences about Policy Impacts on Non-Representative Samples of Sites – Risks, Consequences, and Fixes |

Speaker: Dr. Stephen Bell is an Abt Associates Fellow who holds a Ph.D. in Economics from the University of Wisconsin-Madison. He has designed and analyzed more than a dozen large-scale social experiments of policy interventions to assist disadvantaged Americans, with current work focusing on a slate of papers for IES and NSF on making findings of rigorous impact evaluations more generalizable to the nation and other inference populations. His research on methodologies for measuring social program impacts, both experimental and quasi-experimental econometric techniques, has been widely published. The work presented comes from collaborative work with Elizabeth A Stuart, Robert B. Olsen, and Larry L. Orr. Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: Randomized impact evaluations of social and educational interventions—while constituting the “gold standard” of internal validity due to the lack of selection bias between treated and untreated cases—usually lack external validity. Due to cost and convenience, or local resistance, they are almost always conducted in a set of sites that are not a probability sample of the desired inference population— the nation as a whole for social programs or a given state or school district for educational innovations. We use statistical theory and data from the Reading First evaluation to examine the risks and consequences for social experiments of non-representative site selection, asking when and to what degree policy decisions are led astray by tarnished “gold standard” evidence. We also explore possible ex ante design-based solutions to this problem and the performance of ex post methods in the literature for overcoming non-representative site selection through analytic adjustments after the fact. |
||

9/21/2016, 10:00 - 12:00 |
Vincent Dorie (NYU) |
Bayesian Inference and Stan Tutorial |

Speaker: Vince is a postdoc in NYU PRIISM program working on causal inference and nonparametrics. His recent work includes the causal inference competition at the 2016 Atlantic Causal Inference Conference and software to perform semiparametric sensitivity analyses evaluating the validity of the ignorability assumption in causal inference. Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: This two hour session is focused on getting started with Stan and how to use it in your research. Stan is an open-source Bayesian probabilistic programming environment that takes a lot of the work out of model fitting so that researchers can focus on model building and interpretation. List of topics will include: overview of Bayesian statistics, overview of Stan and MCMC, writing models in Stan, and a tutorial session where participants can write a model on their own or develop models that they have been working on independently. Stan has interfaces to numerous programming languages, but the talk will focus on R.NOTE: Please bring a laptop with RStudio and RStan installed to this session |
||

10/5/2016, 11:00 - 12:00 |
Leanna House (Virginia Tech) |
Be the Data and More: Using interactive, analytic methods to enhance learning from data for students |

Speaker: Leanna House is an Associate Professor of Statistics at Virginia Tech (VT), Blacksburg, Virginia and has been at VT since 2008. Prior to VT, she worked at Battelle Memorial Institute, Columbus, Ohio; received her Ph.D. in Statistics from Duke University, Durham, North Carolina in 2006; and subsequently served as a post-doctoral research associate for two years in the Department of Mathematical Sciences at Durham University, Durham, United Kingdom. Dr. House has authored or co-authored 25 journal papers and has been a strong statistical contributor to successful grant proposals including, "NRT-DESE: UrbComp: Data Science for Modeling, Understanding, and Advancing Urban Populations", “Usable Multiple Scale Big Data Analytics Through Interactive Visualization” , "Critical Thinking with Data Visualization", ``Examining the Taxonomic, Genetic, and Functional Diversity of Amphibian Skin Microbiota", and ``Bayesian Analysis and Visual Analytics''. Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: Datasets, no matter how big, are just tables of numbers without individuals to learn from the data, i.e., discover, process, assess, and communicate information in the data. Data visualizations are often used to present data to individuals, but most are created independently of human learning processes and lack transparency. To bridge the gap between people thinking critically about data and the utility of visualizations, we developed Bayesian Visual Analytics (BaVA) and its deterministic form, Visual to Parametric Interaction (V2PI). BaVA and V2PI transform static images of data to dynamic versions that respond to expert feedback. When applied iteratively, experts may explore data progressively in a sequence that parallels their personal sense-making processes. BaVA and V2PI have shown useful in both industry settings and the classroom. For example, we merged V2PI with motion detection software to create Be the Data. In Be the Data students physcially move in a space to communicate their expert feedback about data projected overhead. The idea is that participants have an opportuntiy to explore analytical relationships between data points by exploring relationships between themselves. This talk will focus on presenting the BaVA paradigm and its education applications. |
||

10/19/2016, 11:00 - 12:00 |
Paul De Boeck (Ohio State University) |
Why so many research hypotheses are mostly false and how to test |

Speaker: Paul De Boeck is a professor of quantitative psychology at the Ohio State University. Before moving to OSU in 2012 he was a professor of psychological methods at the University of Amsterdam (Netherlands) and a professor of psychological assessment at the KULeuven (Belgium). He was president of the Psychometric Society in 2008 and he is the founding editor of the Applied Research and Case Studies section of Psychometrika. His research interests are generalized linear mixed models and explanatory item response theory, and applications of these approaches in the domains of individual differences in cognition, emotion, and psychopathology. More recently he tries to get his work published on the credibility crisis in psychology and feasible but perhaps uncommon methods that may be useful as a response to the crisis. Location: Kimball Hall, 246 Greene Street, 3rd floorAbstract: From a recent Science article with a large number of replications of psychological studies the base rate of the null hypothesis of no effect can be estimated. It turns out to be extremely high, which implies that many research hypotheses are false. As I will explain they are perhaps not fully false but mostly false. A possible explanation for why unlikely hypotheses tend to be selected for empirical studies can be found in expected utility theory. It can be shown that for low to moderately high power rates, the expected utility of studies increases with the probability of the null hypothesis being true. A high probability of the null hypothesis being true can be understood as reflecting a contextual variation of effects that are in general not much different from zero. Increasing the power of studies has become a popular remedy to counter the replicability crisis but this strategy is highly misleading if effects vary. Meta-analysis is considered another remedy but it is a suboptimal and labor-intensive approach and it is only long-term method. Two more feasible methods will be discussed to deal with contextual variation. |
||

11/2/2016, 11:00 - 12:00 |
Catherine Calder (Ohio State University) |
Latent Space Models for Affiliation Networks |

Speaker: Catherine (“Kate”) Calder is professor of statistics at The Ohio State University, where she has served on the faculty since 2003. Her research interests include spatial statistics, Bayesian modeling and computation, and network analysis, with application to problems in the social, environmental, and health sciences. Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: An affiliation network is a particular type of two-mode social network that consists of a set of `actors' and a set of `events' where ties indicate an actor's participation in an event. Methods for the analysis of affiliation networks are particularly useful for studying patterns of segregation and integration in social structures characterized by both people and potentially shared activities (e.g., parties, corporate board memberships, church attendance, etc.) One way to analyze affiliation networks is to consider one-mode network matrices that are derived from an affiliation network, but this approach may lead to the loss of important structural features of the data. The most comprehensive approach is to study both actors and events simultaneously. Statistical methods for studying affiliation networks, however, are less well developed than methods for studying one-mode, or actor-actor, networks. In this talk, I will describe a bilinear generalized mixed-effects model, which contains interacting random effects representing common activity pattern profiles and shared patterns of participation in these profiles. I will demonstrate how the proposed model is able to capture forth-order dependence, a common feature of affiliation networks, and describe a Markov chain Monte Carlo algorithm for Bayesian inference. I then will use the latent space interpretation of model components to explore patterns in extracurricular activity membership of students in a racially-diverse high school in a Midwestern metropolitan area. Using techniques from spatial point pattern analysis, I will show how our model can provide insight into patterns of racial segregation in the voluntary extracurricular activity participation profiles of adolescents. This talk is based on joint work with Yanan Jia and Chris Browning. |
||

12/7/2016, 11:00 - 12:00 |
Kathryn Vasilaky (Columbia University) |
Generalized Ridge Regression Using an Iterative Solution |

Speaker: Kathryn is a postdoc at Columbia University's Earth Institute. Her PhD is in applied economics with interests in development economics, and applied statistics. Location: Kimball Hall, 246 Greene Street, 3rd Floor conference roomAbstract: An iterative method is introduced for solving noisy, ill-conditioned inverse problems, where the standard ridge regression is just the first iteration of the iterative method to be presented. In addition to the regularization parameter, lambda, we introduce an iteration parameter k, which generalizes the ridge regression. The derived noise damping filter is a generalization of the standard ridge regression filter (also known as Tikhonov). Application of the generalized solution performs better than the pseudo-inverse (the default solution to OLS in most statistical packages), and better than standard ridge regression (L-2 regularization), when the covariate matrix or design matrix is ill-conditioned, or highly collinear. A few examples are presented using both simulated and real data. |

Sign up for our mailing list to receive upcoming event reminders.

**Links to Related Seminars:**