Dennis L. Sun

Exact post-selection inference, with application to the lasso

We develop a general approach to valid inference after model selection. In a nutshell, our approach produces post-selection inferences with the same frequency guarantees as those given by data splitting but are more powerful. At the core of our framework is a result that characterizes the distribution of a post-selection estimator conditioned on the selection event. We specialize the approach to model selection by the lasso to form valid condence intervals for the selected coecients and test whether all relevant variables have been included in the model.

Alternating direction method of multipliers for non-negative matrix factorization with the beta-divergence

Non-negative matrix factorization (NMF) is a popular method for learning interpretable features from non-negative data, such as counts or magnitudes. Different cost functions are used with NMF in different applications. We develop an algorithm, based on the alternating direction method of multipliers, that tackles NMF problems whose cost function is a beta-divergence, a broad class of divergence functions. We derive simple, closed-form updates for the most commonly used beta-divergences. We demonstrate experimentally that this algorithm has faster convergence and yields superior results to state-of-the-art algorithms for this problem.

Universal speech models for speaker independent single channel source separation

Supervised and semi-supervised source separation algorithms based on non-negative matrix factorization have been shown to be quite effective. However, they require isolated training examples of one or more sources, which is often difficult to obtain. This limits the practical applicability of these algorithms. We examine the problem of efficiently utilizing general training data in the absence of specific training examples. Specifically, we propose a method to learn a universal speech model from a general corpus of speech and show how to use this model to separate speech from other sound sources. This model is used in lieu of a speech model trained on speaker-dependent training examples, and thus circumvents the aforementioned problem. Our experimental results show that our method achieves nearly the same performance as when speaker-dependent training examples are used. Furthermore, we show that our method improves performance when training data of the non-speech source is available.

Non-negative matrix completion for bandwidth extension: A convex optimization approach

Bandwidth extension is the problem of recovering missing bandwidth in audio signals that have been band-passed, typically for compression purposes. One approach that has been shown to be successful for bandwidth extension is non-negative matrix factorization (NMF). The disadvantage of NMF is that it is non-convex and intractable to solve in general. However, in bandwidth extension, only the reconstruction is needed and not the explicit factors. We formulate bandwidth extension as a convex optimization problem, propose a simple algorithm, and demonstrate the effectiveness of this approach on practical examples.

Neighborhood Variation in the Price of Cheap Tobacco Products in California: Results From Healthy Stores for a Healthy Community

Background Retail marketing surveillance research highlights concerns about lower priced cigarettes in neighborhoods with a higher proportion of racial/ethnic minorities but focuses almost exclusively on premium brands. To remedy this gap in the literature, the current study examines neighborhood variation in prices for the cheapest cigarettes and a popular brand of cigarillos in a large statewide sample of licensed tobacco retailers in a low-tax state. Methods All 61 local health departments in California trained data collectors to conduct observations in a census of eligible licensed tobacco retailers in randomly selected zip codes (n = 7393 stores, completion rate=91%). Data were collected in 2013, when California had a low and stagnant tobacco tax. Two prices were requested: the cheapest cigarette pack regardless of brand and a single, flavored Swisher Sweets cigarillo. Multilevel models (stores clustered in tracts) examined prices (before sales tax) as a function of neighborhood race/ethnicity and proportion of school-age youth (aged 5-17). Models adjusted for store type and median household income. Results Approximately 84% of stores sold cigarettes for less than

Inequalities in tobacco outlet density by race, ethnicity and socioeconomic status, 2012, USA: results from the ASPiRE Study

5 and a Swisher Sweets cigarillo was available for less than

Peer Assessment Enhances Student Learning: The Results of a Matched Randomized Crossover Experiment in a College Statistics Class.

1 in 74% of stores that sold the brand. The cheapest cigarettes cost even less in neighborhoods with a higher proportion of school-age residents and Asian/Pacific Islanders. Conclusions Neighborhood disparities in the price of the cheapest combustible tobacco products are a public health threat. Policy changes that make all tobacco products, especially combustible products, less available and more costly may reduce disparities in their use and protect public health. Implications Much of what is known about neighborhood variation in the price of combustible tobacco products focuses on premium brand cigarettes. The current study extends this literature in two ways, by studying prices for the cheapest cigarette pack regardless of brand and a popular brand of flavored cigarillos and by reporting data from the largest statewide sample of licensed tobacco retailers. Significantly lower prices in neighborhoods with a higher proportion of youth and of racial/ethnic groups with higher smoking prevalence are a cause of concern. The study results underscore the need for policies that reduce availability and increase price of combustible tobacco products, particularly in states with low, stagnant tobacco taxes.

Optimal Inference After Model Selection

Background Evidence of racial/ethnic inequalities in tobacco outlet density is limited by: (1) reliance on studies from single counties or states, (2) limited attention to spatial dependence, and (3) an unclear theory-based relationship between neighbourhood composition and tobacco outlet density. Methods In 97 counties from the contiguous USA, we calculated the 2012 density of likely tobacco outlets (N=90 407), defined as tobacco outlets per 1000 population in census tracts (n=17 667). We used 2 spatial regression techniques, (1) a spatial errors approach in GeoDa software and (2) fitting a covariance function to the errors using a distance matrix of all tract centroids. We examined density as a function of race, ethnicity, income and 2 indicators identified from city planning literature to indicate neighbourhood stability (vacant housing, renter-occupied housing). Results The average density was 1.3 tobacco outlets per 1000 persons. Both spatial regression approaches yielded similar results. In unadjusted models, tobacco outlet density was positively associated with the proportion of black residents and negatively associated with the proportion of Asian residents, white residents and median household income. There was no association with the proportion of Hispanic residents. Indicators of neighbourhood stability explained the disproportionate density associated with black residential composition, but inequalities by income persisted in multivariable models. Conclusions Data from a large sample of US counties and results from 2 techniques to address spatial dependence strengthen evidence of inequalities in tobacco outlet density by race and income. Further research is needed to understand the underlying mechanisms in order to strengthen interventions.

Exact post-selection inference with the lasso

Feedback has a powerful influence on learning, but it is also expensive to provide. In large classes it may even be impossible for instructors to provide individualized feedback. Peer assessment is one way to provide personalized feedback that scales to large classes. Besides these obvious logistical benefits, it has been conjectured that students also learn from the practice of peer assessment. However, this has never been conclusively demonstrated. Using an online educational platform that we developed, we conducted an in-class matched-set, randomized crossover experiment with high power to detect small effects. We establish that peer assessment causes a small but significant gain in student achievement. Our study also demonstrates the potential of web-based platforms to facilitate the design of high-quality experiments to identify small effects that were previously not detectable.