Journal of Biopharmaceutical Statistics最新文献

Dependent samples, in which repeated measurements are made on the same subjects, eliminate potential differences among the subjects. In k-dependent samples, missing data can occur for various reasons. The Skillings-Mack test is used instead of the Friedman test for k-dependent samples with missing observations that are non-normally distributed. If a significant difference exists among groups, nonparametric multiple comparisons need to be performed. In this study, we propose an innovative approach by applying four methods to nonparametric multiple comparisons of incomplete k-dependent samples that are non-normally distributed. The four methods are two nonparametric multiple imputation methods based on machine learning (multiple imputations by chained equations utilizing classification and regression trees (MICE-CART) and random forest (MICE-RF)), one nonparametric imputation method (random hot deck imputation), and the listwise deletion method. We compare the four methods under two missing data mechanisms, four correlation coefficients, two sample sizes, and three percentages of missingness. After implementing different scenarios in a simulation study, the listwise deletion method is inferior to the other methods. MICE-CART and MICE-RF are superior to the other methods for moderate and small sample sizes with well-controlled type 1 error. The two nonparametric multiple imputation methods based on machine learning can be applied to nonparametric multiple comparisons. Therefore, we propose machine learning-based multiple imputation methods for nonparametric multiple comparisons of k-dependent samples with missing observations. The approach was also illustrated with a longitudinal dentistry clinical trial.

This article concerns p-value-based multiple testing procedures (MTPs) that can be used in a confirmatory clinical study under minimal assumptions in case the requirement for study-success is that at least k out of m primary/important hypotheses become rejected. Recently, a simple, generally valid Holm-type MTP was discussed that can be used for such a requirement for any k from one to m. It can only reject at least k (or zero) hypotheses, but this increases the power to reject k or more hypotheses compared to Holm's step-down MTP. The present article provides a simple formulation and proof of strong family-wise error rate (FWER) control for a stepwise MTP that is sharper in that for any k strictly between one and m it: (a) always rejects at least as much, and (b) can potentially reject fewer than k hypotheses. This sharper MTP too is generally valid, without any assumption about logical or stochastic relationships. It has a gatekeeping step, followed by m steps where ordered primary p-values are compared to critical constants and rejections are made in a step-down manner. These constants have the optimality property that under a natural monotonicity restriction, they cannot be increased without losing the general strong FWER control. Confidence regions like those for Holm's MTP are provided. Applications are discussed in connection with three interesting approaches proposed earlier for confirmatory studies: (a) the Superiority-Noninferiority approach; (b) Fallback tests for co-primary endpoints; and (c) Multistage gatekeeping MTPs that utilize so-called k-truncated Holm MTPs in some stages.

Crossover or treatment-switching in randomized controlled trials presents notable challenges not only in the development and approval of new drugs but also poses a complex issue in their reimbursement, especially in oncology. When the investigational treatment is superior to control, crossover from control to investigational treatment upon disease progression or for other reasons will likely cause the underestimation of treatment benefit. Rank Preserving Structural Failure Time (RPSFT) and Two-Stage Estimation (TSE) methods are commonly employed to adjust for treatment switching by estimating counterfactual survival times. However, these methods may induce informative censoring by adjusting censoring times for switchers while leaving those for non-switchers unchanged. Existing approaches such as re-censoring or inverse probability of censoring weighting (IPCW) are often used alongside RPSFT or TSE to handle informative censoring, but may result in long-term information loss or suffer from model misspecification. In this paper, Kaplan-Meier multiple imputation with bootstrap procedure (KMIB) is proposed to address the informative censoring issues in adjustment methods for treatment switching. This approach can avoid information loss and is robust to model misspecification. In the scenarios that we investigate, simulation studies show that this approach performs better than other adjustment methods when the treatment effect is small, and behave similarly under other scenarios despite different switching probability. A case study in non-small cell lung cancer (NSCLC) is also provided to demonstrate the use of this method.

Adaptively incorporating historical information into analyses of current data can improve the precision of inference without requiring additional new observation. Unfortunately, not all borrowing methods are suitable when limited historical studies are available. When a single historical study is available, the power priors control the amount of information to borrow via specification of a weight parameter that discounts the contribution of the historical data in a likelihood combined with current data. We develop a new type of conditional power prior called the historical-bias power prior using an empirical Bayes approach. It relaxes the assumption of the traditional power priors to allow for historical bias. Moreover, our new weight function controls the amount of borrowing and only borrows when historical data satisfy the borrowing criteria. This is achieved by embedding the Frequentist test-then-pool approach in the weight function. Hence, the historical-bias power prior builds a bridge between the Frequentist test-then-pool and the Bayesian power prior. In the simulation, we examine the impact of historical bias on the operating characteristics for borrowing approaches, which has not been discussed in previous literature. The results show that the historical-bias power prior yields accurate estimation and robustly powerful tests for the experimental treatment effect with good type I error control, especially when historical bias exists.

Determining the probability of success of a clinical trial using a prior distribution on the treatment effect can significantly enhance decision-making by the sponsor. In a group sequential design, the probability of success calculated at the design stage can be updated to incorporate the information disclosed by the Data Monitoring Committee (DMC), usually consisting in a simple statement that advises to continue or to stop the trial, either for efficacy or futility, following pre-specified rules defined in the protocol. We define the "probability of success post interim" as the probability of success conditioned on the assumption that the DMC recommends continuing the trial after an interim analysis. A good assessment of this probability helps mitigate the tendency of the study team to express excessive optimism or unwarranted pessimism regarding the trial's ultimate outcome after the DMC recommendation. We explore the relationship between this "probability of success post interim" and the initial probability of success, and we provide an in-depth investigation of how interim boundaries impact these probabilities. This analysis offers valuable insights that can guide the selection of boundaries for both efficacy and futility interim analyses, leading to more informed clinical trial designs.

Dose optimization is a critical stage of drug development in oncology and other disease areas. Early phase clinical trials are inherently heterogeneous due to their exploratory nature. The process of identifying an optimal dose involves careful considerations of the patient population, evaluation of therapeutic potential, and exploration of the dose-response and dose-toxicity relationships to ensure that it is safe and effective for the intended use. However, the complex mechanism of actions and uncertainties during dose optimization often introduce substantial gaps between those early phase trials and phase 3 randomized control trials. These gaps can indeed increase the chances of failure. To address these challenges, we propose a novel seamless phase I/II design, namely DOD-BART design, which utilizes machine learning technique, specifically Bayesian Additive Regression Trees (BART) to fully incorporate patient-level prognostic factors and outcomes. Our design provides a streamlined approach for dose exploration and optimization, automatically updated with emerging data to allocate patients to the most promising dose levels. DOD-BART elucidates disease relationships, analyzes and synthesizes emerging data, augments operational efficiency, and guides dose optimization for suitable population. Simulation studies demonstrate the robust performances of the DOD-BART design across a variety of realistic settings, with high probabilities of correctly identifying the optimal dose, allocating patients more to tolerable and efficacious dose levels, making less biased estimates, and efficiently utilizing patients' data.

Background: Cardiovascular diseases (CVDs) include abnormal conditions of the heart, diseased blood vessels, structural problems of the heart, and blood clots. Traditionally, CVD has been diagnosed by clinical experts, physicians, and medical specialists, which is expensive, time-consuming, and requires expert intervention. On the other hand, cost-effective digital diagnosis of CVD is now possible because of the emergence of machine learning (ML) and statistical techniques.

Method: In this research, extensive studies were carried out to classify CVD via 19 promising ML models. To evaluate the performance and rank the ML models for CVD classification, two benchmark CVD datasets are considered from well-known sources, such as Kaggle and the UCI repository. The results are analysed considering individual datasets and their combination to assess the efficiency and reliability of ML models on the basis of various performance measures, such as precision, kappa, accuracy, recall, and the F1 score. Since some of the ML models are stochastic, we repeated the simulation 50 times for each dataset using each model and applied nonparametric statistical tests to draw decisive conclusions.

Results: The nonparametric Friedman - Nemenyi hypothesis test suggests that the Extra Tree Classifier provides statistically superior accuracy and precision compared with all other models. However, the Extreme Gradient Boost (XGBoost) classifier provides statistically superior recall, kappa, and F1 scores compared with those of all the other models. Additionally, the XGBRF classifier achieves a statistically second-best rank in terms of the recall measures.

A statistical methodology named "capture recapture", a Kaplan-Meier Summary Statistic, and an urn model framework are presented to describe the elicitation, then estimate both the number of interviews and the total number of items ("codes") that will be elicited during patient interviews, and present a summary graphical statistic that "saturation" has occurred. This methodology is developed to address a gap in the FDA 2009 PRO and 2012 PFDD guidance for determining the number of interviews (sample size). This estimate of the number of interviews (sample size) uses a two-step procedure. The estimate of the total number of items is then used to estimate the number of interviews to elicit all items. A framework called an urn model is a framework for describing the elicitation and demonstrate the algorithm for declaring saturation "first interview with zero new codes". A caveat emptor is that due to independence assumptions, the urn model is not used as a method for estimating probabilities. The URN model provides a framework to demonstrate that an algorithm such as "first interview with zero new codes" may establish that all codes have been elicited. The limitations of the Urn model, capture recapture, and Kaplan-Meier are summarized. The statistical methods and the estimates supplement but do not replace expert judgement and declaration of "saturation." A graphical summary statistic is presented to summarize "saturation," after expert declaration for two algorithms. An example of a capture-recapture estimate, using simulated data is provided. The example suggests that the estimate of total number of codes may be accurate when prepared as early as the second interview. A second simulation is presented with an URN model, under a strong assumption of independence that an algorithm such as 'first interview with zero new codes" may fail to identify all codes. Potential errors in declaration of saturation are presented. Recommendations are presented for additional research and the use of the algorithm "first interview with zero new codes."

We propose a Bayesian optimal phase 2 design for jointly monitoring efficacy and toxicity, referred to as BOP2-TE, to improve the operating characteristics of the BOP2 design proposed by Zhou. BOP2-TE utilizes a Dirichlet-multinomial model to jointly model the distribution of toxicity and efficacy endpoints, making go/no-go decisions based on the posterior probability of toxicity and futility. In comparison to the original BOP2 and other existing designs, BOP2-TE offers the advantage of providing rigorous type I error control in cases where the treatment is toxic and futile, effective but toxic, or safe but futile, while optimizing power when the treatment is effective and safe. As a result, BOP2-TE enhances trial safety and efficacy. We also explore the incorporation of BOP2-TE into multiple-dose randomized trials for dose optimization, and consider a seamless design that integrates phase I dose finding with phase II randomized dose optimization. BOP2-TE is user-friendly, as its decision boundary can be determined prior to the trial's onset. Simulations demonstrate that BOP2-TE possesses desirable operating characteristics. We have developed a user-friendly web application as part of the BOP2 app, which is freely available at https://www.trialdesign.org.

Symptoms post-SARS-CoV-2 infection may persist for months and cause significant impairment and impact to quality of life. Acute symptoms of SARS-CoV-2 infection are well studied, yet data on clusters of symptoms over time, or post-acute sequelae of SARS-CoV-2 infection (PASC), are limited. We aim to characterize PASC phenotypes by identifying symptom clusters over a six-month period following infection in individuals vaccinated (boosted and not) and those unvaccinated. Subjects with ≥1 self-reported symptom and positive RT-PCR for SARS-CoV-2 at CVS Health US test sites were recruited between January and April 2022. Patient-reported outcomes symptoms, health-related quality of life (HRQoL), work productivity and activity impairment (WPAI) were captured at 1 month, 3 months, and 6 months post-acute infection. Phenotypes of PASC were determined based on subject matter knowledge and balanced consideration of statistical criteria (lower AIC, lower BIC, and adequate entropy) and interpretability. Generalized estimation equation approach was used to investigate relationship between QoL, WPAI and number of symptoms and identified phenotypes, and relationship between phenotypes and vaccination status as well. LCA identified three phenotypes that are primarily differentiated by number of symptoms. These three phenotypes remained consistent across time periods. Subjects with more symptoms were associated with lower HRQoL, and worse WPAI scores. Vaccinated individuals were more likely to be in the low symptom burden latent classes at all time points compared to unvaccinated individuals.