Search

Transportability analysis is a causal inference framework used to evaluate the external validity of randomized clinical trials (RCTs) or observational studies. Most existing transportability analysis methods require individual patient-level data (IPD) for both the source and the target population, narrowing its applicability when only target aggregate-level data (AgD) is available. Besides, accounting for censoring is essential to reduce bias in longitudinal data, yet AgD-based transportability methods in the presence of censoring remain underexplored. Here, we propose a two-stage weighting framework named "Target Aggregate Data Adjustment" (TADA) to address the mentioned challenges simultaneously. TADA is designed as a two-stage weighting scheme to simultaneously adjust for both censoring bias and distributional imbalances of effect modifiers (EM), where the final weights are the product of the inverse probability of censoring weights and participation weights derived using the method of moments. We have conducted an extensive simulation study to evaluate TADA's performance. Our results indicate that TADA can effectively control the bias resulting from censoring within a non-extreme range suitable for most practical scenarios, and enhance the application and clinical interpretability of transportability analyses in settings with limited data availability.

The ICH E9(R1) addendum provides guidelines on accounting for intercurrent events in clinical trials using the estimands framework. However, there has been limited attention on the estimands framework for meta-analysis. Using treatment switching, a well-known intercurrent event that occurs frequently in oncology, we conducted a simulation study to explore the bias introduced by pooling together estimates targeting different estimands in a meta-analysis of randomized clinical trials (RCTs) that allowed for treatment switching. We simulated overall survival data of a collection of RCTs that allowed patients in the control group to switch to the intervention treatment after disease progression under fixed-effects and random-effects models. For each RCT, we calculated effect estimates for a treatment policy estimand that ignored treatment switching, and a hypothetical estimand that accounted for treatment switching by censoring switchers at the time of switching. Then, we performed random-effects and fixed-effects meta-analyses to pool together RCT effect estimates while varying the proportions of treatment policy and hypothetical effect estimates. We compared the results of meta-analyses that pooled different types of effect estimates with those that pooled only treatment policy or hypothetical estimates. We found that pooling estimates targeting different estimands results in pooled estimators that reflect neither the treatment policy estimand nor the hypothetical estimand. This finding shows that pooling estimates of varying target estimands can generate misleading results, even under a random-effects model. Adopting the estimands framework for meta-analysis may improve alignment between meta-analytic results and the clinical research question of interest.

The SARS-CoV-2 pandemic has so far unfolded diversely across the fifty United States of America, reflected both in different time progressions of infection "waves" and in magnitudes of local infection rates. Despite a marked diversity of presentations, most U.S. states experienced their single greatest surge in daily new cases during the transition from Fall 2020 to Winter 2021. Popular media also cite additional similarities between states -- often despite disparities in governmental policies, reported mask-wearing compliance rates, and vaccination percentages. Here, we identify a set of robust, low-dimensional clusters that 1) summarize the timings and relative heights of four historical COVID-19 "wave opportunities" accessible to all 50 U.S. states, 2) correlate with geographical and intervention patterns associated with those groups of states they encompass, and 3) predict aspects of the "fifth wave" of new infections in the late Summer of 2021. In particular, we argue that clustering elucidates a negative relationship between vaccination rates and subsequent case-load variabilities within state groups. We advance the hypothesis that vaccination acts as a ``seat belt," in effect constraining the likely range of new-case upticks, even in the context of the Summer 2021, variant-driven surge., Comment: 22 pages, 4 figures; submitted to PLOS ONE

Deep Neural Network (DNN) models have continuously been growing in size in order to improve the accuracy and quality of the models. Moreover, for training of large DNN models, the use of heterogeneous GPUs is inevitable due to the short release cycle of new GPU architectures. In this paper, we investigate how to enable training of large DNN models on a heterogeneous GPU cluster that possibly includes whimpy GPUs that, as a standalone, could not be used for training. We present a DNN training system, HetPipe (Heterogeneous Pipeline), that integrates pipelined model parallelism (PMP) with data parallelism (DP). In HetPipe, a group of multiple GPUs, called a virtual worker, processes minibatches in a pipelined manner, and multiple such virtual workers employ data parallelism for higher performance. We also propose a novel parameter synchronization model, which we refer to as Wave Synchronous Parallel (WSP) to accommodate both PMP and DP for virtual workers, and provide convergence proof of WSP. Our experimental results on a given heterogeneous setting show that with HetPipe, DNN models converge up to 49% faster compared to the state-of-the-art DP technique.

The Convolutional Neural Network (CNN) model, often used for image classification, requires significant training time to obtain high accuracy. To this end, distributed training is performed with the parameter server (PS) architecture using multiple servers. Unfortunately, scalability has been found to be poor in existing architectures. We find that the PS network is the bottleneck as it communicates a large number of gradients and parameters with the many workers. This is because synchronization with the many workers has to occur at every step of training. Depending on the model, communication can be in the several hundred MBs per synchronization. In this paper, we propose a scheme to reduce network traffic through layer placement that considers the resources that each layer uses. Through analysis of the characteristics of CNN, we find that placement of layers can be done in an effective manner. We then incorporate this observation within the TensorFlow framework such that layers can be automatically placed for more efficient training. Our evaluation making use of this placement scheme show that training time can be significantly reduced without loss of accuracy for many CNN models.