Journal of Data Science

The extraordinary capabilities of large language models (LLMs) such as ChatGPT and GPT-4 are in part unleashed by aligning them with reward models that are trained on human preferences represented as rankings of responses to prompts. In this paper, we document the phenomenon of reward collapse, an empirical observation where the prevailing ranking-based approach results in an identical reward distribution for diverse prompts during the terminal phase of training. This outcome is undesirable as open-ended prompts like “write a short story about your best friend” should yield a continuous range of rewards for their completions, while specific prompts like “what is the capital city of New Zealand” should generate either high or low rewards. Our theoretical investigation reveals that reward collapse is primarily due to the insufficiency of the ranking-based objective function to incorporate prompt-related information during optimization. Then we derive closed-form expressions for the reward distribution associated with a set of utility functions in an asymptotic setting. Based on the reward distributions for different utility functions, we introduce a prompt-aware optimization scheme that provably admits a prompt-dependent reward distribution within the interpolating regime. Our experimental results suggest that our proposed prompt-aware utility functions significantly alleviate reward collapse during the training of reward models.

Physician performance is critical to caring for patients admitted to the intensive care unit (ICU), who are in life-threatening situations and require high level medical care and interventions. Evaluating physicians is crucial for ensuring a high standard of medical care and fostering continuous performance improvement. The non-randomized nature of ICU data often results in imbalance in patient covariates across physician groups, making direct comparisons of the patients’ survival probabilities for each physician misleading. In this article, we utilize the propensity weighting method to address confounding, achieve covariates balance, and assess physician effects. Due to possible model misspecification, we compare the performance of the propensity weighting methods using both parametric models and super learning methods. When the generalized propensity or the quality function is not correctly specified within the parametric propensity weighting framework, super learning-based propensity weighting methods yield more efficient estimators. We demonstrate that utilizing propensity weighting offers an effective way to assess physician performance, a topic of considerable interest to hospital administrators.

Anemia, especially among children, is a serious public health problem in Bangladesh. Apart from understanding the factors associated with anemia, it may be of interest to know the likelihood of anemia given the factors. Prediction of disease status is a key to community and health service policy making as well as forecasting for resource planning. We considered machine learning (ML) algorithms to predict the anemia status among children (under five years) using common risk factors as features. Data were extracted from a nationally representative cross-sectional survey- Bangladesh Demographic and Health Survey (BDHS) conducted in 2011. In this study, a sample of 2013 children were selected for whom data on all selected variables was available. We used several ML algorithms such as linear discriminant analysis (LDA), classification and regression trees (CART), k-nearest neighbors (k-NN), support vector machines (SVM), random forest (RF) and logistic regression (LR) to predict the childhood anemia status. A systematic evaluation of the algorithms was performed in terms of accuracy, sensitivity, specificity, and area under the curve (AUC). We found that the RF algorithm achieved the best classification accuracy of 68.53% with a sensitivity of 70.73%, specificity of 66.41% and AUC of 0.6857. On the other hand, the classical LR algorithm reached a classification accuracy of 62.75% with a sensitivity of 63.41%, specificity of 62.11% and AUC of 0.6276. Among all considered algorithms, the k-NN gave the least accuracy. We conclude that ML methods can be considered in addition to the classical regression techniques when the prediction of anemia is the primary focus.

Technological advances in software development effectively handled technical details that made life easier for data analysts, but also allowed for nonexperts in statistics and computer science to analyze data. As a result, medical research suffers from statistical errors that could be otherwise prevented such as errors in choosing a hypothesis test and assumption checking of models. Our objective is to create an automated data analysis software package that can help practitioners run non-subjective, fast, accurate and easily interpretable analyses. We used machine learning to predict the normality of a distribution as an alternative to normality tests and graphical methods to avoid their downsides. We implemented methods for detecting outliers, imputing missing values, and choosing a threshold for cutting numerical variables to correct for non-linearity before running a linear regression. We showed that data analysis can be automated. Our normality prediction algorithm outperformed the Shapiro-Wilk test in small samples with Matthews correlation coefficient of 0.5 vs. 0.16. The biggest drawback was that we did not find alternatives for statistical tests to test linear regression assumptions which are problematic in large datasets. We also applied our work to a dataset about smoking in teenagers. Because of the opensource nature of our work, these algorithms can be used in future research and projects.

Predictive automation is a pervasive and archetypical example of the digital economy. Studying how Americans evaluate predictive automation is important because it affects corporate and state governance. However, we have relevant questions unanswered. We lack comparisons across use cases using a nationally representative sample. We also have yet to determine what are the key predictors of evaluations of predictive automation. This article uses the American Trends Panel’s 2018 wave ($n=4,594$) to study whether American adults think predictive automation is fair across four use cases: helping credit decisions, assisting parole decisions, filtering job applicants based on interview videos, and assessing job candidates based on resumes. Results from lasso regressions trained with 112 predictors reveal that people’s evaluations of predictive automation align with their views about social media, technology, and politics.