Imagine you are running a sequential experiment measuring the difference between a treatment and control condition (e.g. a pill for blood pressure via a clinical trial or testing user behavior in an Internet-based experiment). You can match similar subjects together on-the-fly (as they arrive) to achieve higher power and efficiency in the experimental results.

Imagine you are a medical practitioner treating a disease by prescribing one of two possible drugs. Which drug do you assign to patients? Is your special assignment procedure beneficial versus a naive random assignment? How much better and is the improvement statistically significant?

It is well known that cancer cells are solely dependent on glucose as their substrate for metabolism and they are not able to utilize other fuel sources such as ketones and fatty acids. It is also known that humans under heavy ketosis do not experience symptoms of hypoglycemia. In our proposal for cancer therapy, we marry these two ideas over the long term.

We develop an extension to Bayesian Additive Regression Trees (a new procedure for statistically learning non-parametric functional relationship between a set up input variables X and a response variable y). In the extension, we incorporate missing data without the need for imputation. Simulations using real data and generated models demonstrate high performance and stability over competitors.

We adapt Bayesian Additive Regression Trees (a new procedure for statistically learning non-parametric functional relationship between a set up input variables X and a response variable y). This adaptation can select "important" variables from the set of x's that affect y by employing a principled permutation-based inference procedure. We can also incorporate prior information about which variables are thought to be important before looking at the data.

We develop a tool for visualizing the model estimated by any supervised "machine" learning algorithm. We plot the variation of the fitted values across the range of a covariate of interest for all cases. These lines suggest where and to what extent heterogeneities exist between cases. We also include a visual test for model additivity in any covariate.

We present a new R package implementation of Bayesian Additive Regression Trees (a new procedure for statistically learning non-parametric functional relationship between a set up input variables X and a response variable y). The package introduces many new features for data analysis using BART such as variable selection, interaction detection, model diagnostic plots, parallelization, incorporation of missing data and the ability to save trees for future prediction.

We adapt Bayesian Additive Regression Trees (a new procedure for statistically learning non-parametric functional relationship between a set up input variables X and a response variable y). This adaptation incorporates heteroskedasticity into the model by modeling the form of heteroskedasticity as a linear model of another set of covariates (may or may not be X). In simulations, we demonstrate a reduction in overfitting and more appropriate predictive intervals than homoskedastic BART.

We develop an implementation of principal components logistic regression using a novel three split procedure. The first split trains the kernel models from a set of many kernels, the second picks the model which respects low error and correct error costs, and the third gives an honest assessment of future performance. Our implementation contains an R package and our paper applies these methods to forecasting failures to appear in court.

We conducted the first natural field experiment to explore the relationship between the "meaningfulness" of a task and worker effort, measured on three scales. We employed ~2500 workers from Amazon's Mechanical Turk, an online labor market, to label medical images. We manipulated the task to exhibit three conditions: "meaningful," "zero-context" and "shredded." We found that the meaningful treatment increased the labor supply and the shredded treatment decreased the quality of the labor.

We presented the task of predicting well-being, as measured by the "satisfaction with life scale." Using Amazon's Mechanical Turk, we created a training set of textual examples properly rated. We then used machine learning to build a high-performance model and in addition, we identify textual features that characterize well-being.

We use crowdsourcing to disambiguate 1000 words from among coarse-grained senses. Using regression, we find surprising features which drive differential WSD accuracy: (a) the number of rephrasings within a sense definition is associated with higher accuracy; (b) as word frequency increases, accuracy decreases even if the number of senses is kept constant; and (c) spending more time is associated with a decrease in accuracy.

We examine the prevalence of satisficing (mental shortcuts / cheating) on Amazon's Mechanical Turk's survey tasks. We present a question-presentation method of fading in survey questions and answers one-by-one, called "Kapcha," which we experimentally demonstrate to reduce satisficing, thereby improving the quality of survey results.

We describe the spatial organization of dendritic cells within tumor-draining lymph nodes using the software gemident. We then describe the spatial organization's association with survival outcome in cancer patients. We also characterize specific changes in number, size, maturity, and T-cell co-localization of such clusters.

We present a novel, quantitative image analysis approach incorporating 1) multi-color tissue staining, 2) high-resolution, automated whole-section imaging, 3) the use of the "gemident" image analysis software to identify cell types and locations, and 4) spatial statistical analysis. We apply our integrative approach to compare the architectural patterns of T and B cells within tumor-draining lymph nodes from breast cancer patients versus healthy lymph nodes. We found that the spatial grouping patterns of T and B cells differed between healthy and breast cancer lymph nodes, and this could be attributed to the lack of B cell localization in the extrafollicular region of the TDLNs.

We present a novel object identification algorithm developed in Java which locates objects of interest in images. Here, we apply the system to finding cells in images of immunohisto-chemically-stained lymph node tissue. The success of the method depends heavily on the use of color, the relative homogeneity of object appearance, the user's input, and the coupled statistical learning algorithm, random forests. Our system enables iterative improvements to the classification over many correction cycles, resulting in a highly accurate and user-friendly system.

Supervised learning can be used to segment regions of interest in images making use of color and morphological information. We developed a novel object identification algorithm in Java which locates phenotypes of interest in images. Our main innovation is interactive feature extraction from color images by using sums over color similarities (as measured by the Mahalanobis distance) at various radii. These features are then fed into a statistical learning algorithm to classify pixels belonging to phenotypes of interest.