Pharmaceutical Statistics最新文献

The conduct of dose-finding trials can be specifically challenging in small populations, for example, in pediatric settings. Recently, research has shown that Bayesian borrowing from adult trials combined with appropriately robust prior distributions enables the conduct of pediatric dose-finding trials with very small sample sizes. However, the appropriate degree of borrowing remains a subjective choice, relying on default methods or expert opinion. This paper proposes an approach to empirically determine the degree of borrowing based on a meta-analysis of the similarity between population-specific dose-toxicity curves of other biologically similar compounds. Focusing on the pediatric use case, the approach may be applicable to any dose-finding trial with information borrowing from another population. With the ExNex and a hierarchical model, two popular statistical modeling approaches are applied. The estimated degree of similarity is then translated into the statistical model for the dose-finding algorithm using either variance inflation or robust mixture prior distributions. The performance of each combination of statistical model approaches is investigated in a simulation study. The results with mixture priors are promising for the application of the proposed methods, especially with many (20) compounds, while variance inflation models require additional fine-tuning and seem to be less robust. With fewer (3 or 7) compounds, our proposed methods are either in line with robust priors that ignore the data from other compounds or are slightly better. We further provide a case study analyzing real dose-finding data from 6 compounds with our models, demonstrating applicability in real-world situations. For clinical trials teams, the decision for or against the proposed approach might be connected to the efforts in terms of time and cost to receive the external data.

Overdispersion, a common issue in clustered multinomial data, can lead to biased standard errors and compromised statistical inference if not adequately addressed. This study describes a comprehensive procedure for constructing multiple comparisons of interest and applying multiplicity adjustments in the analysis of clustered, potentially overdispersed multinomial data. We investigate four quasi-likelihood estimators and the Dirichlet-multinomial model to account for overdispersion. Through a simulation study, we evaluate the performance of these methods under various scenarios, focusing on family-wise error rate, statistical power and coverage probability. Our findings indicate that the Afroz quasi-likelihood estimator is recommended when strict error control is required, whereas the Dirichlet-multinomial model is preferable when high statistical power is desired, albeit with a slightly increased tolerance for false positives. Additionally, we address the challenge of zero-count categories within groups, demonstrating that incorporating pseudo-observations can effectively mitigate associated estimation difficulties. Practical applications to real datasets from toxicology and flow cytometry underscore the robustness and practical utility of these methods.

There is growing interest in a hybrid control design for treatment evaluation, where a randomized controlled trial is augmented with external control data from a previous trial or a real world data source. The hybrid control design has the potential to improve efficiency but also carries the risk of introducing bias. The potential bias in a hybrid control study can be mitigated by adjusting for baseline covariates that are related to the control outcome. A key assumption for this approach is that the internal and external control outcomes are exchangeable upon conditioning on a set of measured covariates. Possible violations of the exchangeability assumption can result in bias and thus need to be addressed systematically. This article proposes a variable selection approach to addressing non-exchangeability in hybrid control studies. Under a specified outcome regression model, possible non-exchangeability can be represented as interactions between covariates and an external control indicator, some of which may be null (with zero coefficients). Null interactions support information borrowing, while non-null interactions require adjustment. Identifying non-null interactions for inclusion in the model is a variable selection problem. The adaptive lasso can be used to perform variable selection and modeling fitting, and the fitted model can be substituted into a g-computation formula. Simulation results demonstrate that, under appropriate conditions, this approach is able to improve efficiency by incorporating external control data in the absence of full exchangeability.

In a mixture experiment, we study the behavior and properties of m mixture components, where the primary focus is on the proportions of the components that make up the mixture rather than the total amount. Mixture-amount experiments are specialized types of mixture experiments where both the proportions of the components in the mixture and the total amount of the mixture are of interest. In this paper, we consider an Order-of-Addition (OofA) mixture-amount experiment in which the response depends on both the mixture amounts of components and their order of addition. Full mixture OofA designs are constructed to maintain orthogonality between the mixture-amount model terms and the effects of the order of addition. But the number of runs in such full OofA designs increases as m increases. We employ the Threshold Accepting (TA) Algorithm to select an n-row subset from the full OofA mixture design that maximizes G-optimality while minimizing the number of experimental runs. Further, the G-efficiency criterion is used to assess how well the design supports the precise and unbiased estimation of the model parameters. These designs enable the estimation of mixture-component model parameters and the order-of-addition effects. The Fraction of Design Space (FDS) plot is used to provide a visual assessment of the prediction capabilities of a design across the entire design space.

Most phase I-II drug-combination trial designs assume that selecting the optimal dose combination based on early outcomes will also lead to maximum long-term survival benefits. However, this assumption is often violated in many clinical studies, generally due to high rates of relapse following the initial response. To address this problem, we propose the Great Wall design, a general dose optimization design for drug-combination trials. The Great Wall design employs a "divide-and-conquer" algorithm to address the issue of partial order of toxicity and uses early outcomes to eliminate dose combinations that are excessively toxic or less efficacious. It utilizes a dose randomization approach to construct a candidate set of the promising dose combinations balancing the toxicity and early efficacy outcomes. The patients assigned to the candidate set are followed to collect the survival outcomes and the final optimal dose combination is then selected to maximize the survival benefit. The simulation studies confirm the desirable operating characteristics of the Great Wall design under various clinical settings. R codes are also provided to facilitate the application. The Great Wall design is modular and practically useful in settings where investigators plan to follow patients long enough to assess survival outcomes.

Covariate adjustment aims to improve the statistical efficiency of randomized trials by incorporating information from baseline covariates. Popular methods for covariate adjustment include analysis of covariance for continuous endpoints and standardized logistic regression for binary endpoints. For survival endpoints, while some covariate adjustment methods have been developed for specific effect measures, they are not commonly used in practice for various reasons, including high demands for theoretical and methodological sophistication as well as computational skills. This article describes an augmentation approach to covariate adjustment for survival endpoints that is relatively easy to understand and widely applicable to different effect measures. This approach involves augmenting a given treatment effect estimator in a way that preserves consistency and asymptotic normality under minimal assumptions (i.e., randomization). It does not attempt to exploit other possible constraints (e.g., independent censoring, proportional hazards) on the observed data distribution. The optimal augmentation term, which minimizes the asymptotic variance of an augmented estimator, can be estimated using various statistical and machine learning methods. Simulation results demonstrate that the augmentation approach can bring substantial gains in statistical efficiency. This approach has been implemented in an R package named sleete, which is described in detail and illustrated with real data.

Duration of response (DOR) has been increasingly used as a useful measure of response to treatments in randomized clinical trials (RCT). Some estimands for DOR, such as the restricted mean DOR, although simple to use, may be sensitive to outliers and may not correctly measure treatment effects on the quantiles of DOR, such as the proportion of patients with DOR of at least 3 months. Quantile regression for survival data has been well developed. However, it is not directly applicable to DOR data in RCTs, due to the presence of non-responders for whom DOR is not defined. Although they can be treated as having zero DOR in a standard quantile regression, such an approach may not be flexible to model these subset of patients. To mitigate this issue, we propose an approach similar to the two-parts zero-inflated models, for example, for count data, so that the nonresponders are modeled as a part of the model, while DOR is modeled using quantile regression. A simulation study is conducted to examine the performance of the proposed approach. For illustration, we apply our approach to a simulated dataset of an acute myeloid leukemia trial, since the true dataset cannot be used due to confidentiality. The asymptotic properties of the proposed approach are also derived.

In many acute myeloid leukemia (AML) studies, event-free survival (EFS) has been accepted as a primary efficacy endpoint. In those studies, the patients who do not achieve complete remission (CR) in the induction period are regarded as induction treatment failure (ITF). The recent FDA guidance on AML (2022) has clearly specified ITF as the event at Day 1 of randomization, considering the variability of length of individual induction treatment periods among studies. Xu et al. (2021) suggested decomposing the log-rank test statistic into the ITF portion, and the Non-ITF portion that is defined as the patients who achieved CR, and assumed proportional hazards for the Non-ITF portion. However, especially in the newly diagnosed AML study, there is some indication of the cured patients who achieve CR during the induction period. As a result, Non-zero ITF rates and cured patients invalidate the proportional hazards assumption and therefore, the conventional power calculation based on the number of events may be problematic in this setting. Our research follows the same decomposition of the log-rank test statistic as Xu et al. (2021) and suggests a new sample size calculation method accounting for the presence of both ITF and cured patients. The result shows that the analytically calculated power of log-rank test based on our proposal was very similar to the empirical power based on simulations in various finite sample settings and was also useful to protect from overestimation and underestimation of the required sample size in the presence of cure fraction.

There is a renewed interest in defining the target of estimation when designing randomized trials. Motivated by design work in trials of HIV-1 curative interventions, we compare the Wilcoxon-Mann-Whitney (WMW) estimand to a difference in medians or means in a two-arm study. First, we define each estimand along with an appropriate estimator. Then, we highlight relevant asymptotic relative efficiency (ARE) results for the estimators under normal distributions (ARE: WMW/mean = $3/\pi$ , median/mean = $2/\pi$ , median/WMW = $2/3$ ), as well as normal mixtures. Measurement of outcomes related to HIV-1 cure involve laboratory assays with lower limits of quantification giving rise to left-censored data. In our simulation study, we compare the estimators in the presence of left-censored observations and at small sample sizes, illustrating that under a censored normal mixture distribution the WMW approach is unbiased, powerful, and has confidence intervals with nominal coverage. We apply our findings to a randomized trial designed to reduce HIV-1 reservoirs. We further expose several extensions of the WMW approach that allows for assessment of interactions between subgroups in a trial, adjustment for covariates, and general ranking methods for clinical outcomes in other disease areas. We end with a discussion summarizing the merits of a WMW based intervention effect estimate versus an estimate summarized on the scale the intervention was originally measured such as the difference in medians or means.

Flexible machine learning algorithms are increasingly utilized in real-world data analyses. When integrated within double robust methods, such as the Targeted Maximum Likelihood Estimator (TMLE), complex estimators can result in significant undercoverage-an issue that is even more pronounced in singly robust methods. The Double Cross-Fitting (DCF) procedure complements these methods by enabling the use of diverse machine learning estimators, yet optimal guidelines for the number of data splits and repetitions remain unclear. This study aims to explore the effects of varying the number of splits and repetitions in DCF on TMLE estimators through statistical simulations and a data analysis. We discuss two generalizations of DCF beyond the conventional three splits and apply a range of splits to fit the TMLE estimator, incorporating a super learner without transforming covariates. The statistical properties of these configurations are compared across two sample sizes (3000 and 5000) and two DCF generalizations (equal splits and full data use). Additionally, we conduct a real-world analysis using data from the National Health and Nutrition Examination Survey (NHANES) 2017-18 cycle to illustrate the practical implications of varying DCF splits, focusing on the association between obesity and the risk of developing diabetes. Our simulation study reveals that five splits in DCF yield satisfactory bias, variance, and coverage across scenarios. In the real-world application, the DCF TMLE method showed consistent risk difference estimates over a range of splits, though standard errors increased with more splits in one generalization, suggesting potential drawbacks to excessive splitting. This research underscores the importance of judicious selection of the number of splits and repetitions in DCF TMLE methods to achieve a balance between computational efficiency and accurate statistical inference. Optimal performance seems attainable with three to five splits. Among the generalizations considered, using full data for nuisance estimation offered more consistent variance estimation and is preferable for applied use. Additionally, increasing the repetitions beyond 25 did not enhance performance, providing crucial guidance for researchers employing complex machine learning algorithms in causal studies and advocating for cautious split management in DCF procedures.