OBJECTIVE: In the context of clinical research, there is an increasing need for new study designs that help to incorporate already available data. With the help of historical controls, the existing information can be utilized to support the new study design, but of course, inclusion also carries the risk of bias in the study results.
METHODS: To combine historical and randomized controls we investigate the Fill-it-up-design, which in the first step checks the comparability of the historical and randomized controls performing an equivalence pre-test. If equivalence is confirmed, the historical control data will be included in the new RCT. If equivalence cannot be confirmed, the historical controls will not be considered at all and the randomization of the original study will be extended. We are investigating the performance of this study design in terms of type I error rate and power.
RESULTS: We demonstrate how many patients need to be recruited in each of the two steps in the Fill-it-up-design and show that the family wise error rate of the design is kept at 5 % . The maximum sample size of the Fill-it-up-design is larger than that of the single-stage design without historical controls and increases as the heterogeneity between the historical controls and the concurrent controls increases.
CONCLUSIONS: The two-stage Fill-it-up-design represents a frequentist method for including historical control data for various study designs. As the maximum sample size of the design is larger, a robust prior belief is essential for its use. The design should therefore be seen as a way out in exceptional situations where a hybrid design is considered necessary.

The geometric median, which is applicable to high-dimensional data, can be viewed as a generalization of the univariate median used in 1-dimensional data. It can be used as a robust estimator for identifying the location of multi-dimensional data and has a wide range of applications in real-world scenarios. This paper explores the problem of high-dimensional multivariate analysis of variance (MANOVA) using the geometric median. A maximum-type statistic that relies on the differences between the geometric medians among various groups is introduced. The distribution of the new test statistic is derived under the null hypothesis using Gaussian approximations, and its consistency under the alternative hypothesis is established. To approximate the distribution of the new statistic in high dimensions, a wild bootstrap algorithm is proposed and theoretically justified. Through simulation studies conducted across a variety of dimensions, sample sizes, and data-generating models, we demonstrate the finite-sample performance of our geometric median-based MANOVA method. Additionally, we implement the proposed approach to analyze a breast cancer gene expression dataset.

UNASSIGNED: Mega-trials can provide large-scale evidence on important questions.
UNASSIGNED: To explore how the results of mega-trials compare with the meta-analysis results of trials with smaller sample sizes.
UNASSIGNED: ClinicalTrials.gov was searched for mega-trials until January 2023. PubMed was searched until June 2023 for meta-analyses incorporating the results of the eligible mega-trials.
UNASSIGNED: Mega-trials were eligible if they were noncluster nonvaccine randomized clinical trials, had a sample size over 10 000, and had a peer-reviewed meta-analysis publication presenting results for the primary outcome of the mega-trials and/or all-cause mortality.
UNASSIGNED: For each selected meta-analysis, we extracted results of smaller trials and mega-trials included in the summary effect estimate and combined them separately using random effects. These estimates were used to calculate the ratio of odds ratios (ROR) between mega-trials and smaller trials in each meta-analysis. Next, the RORs were combined using random effects. Risk of bias was extracted for each trial included in our analyses (or when not available, assessed only for mega-trials). Data analysis was conducted from January to June 2024.
UNASSIGNED: The main outcomes were the summary ROR for the primary outcome and all-cause mortality between mega-trials and smaller trials. Sensitivity analyses were performed with respect to the year of publication, masking, weight, type of intervention, and specialty.
UNASSIGNED: Of 120 mega-trials identified, 41 showed a significant result for the primary outcome and 22 showed a significant result for all-cause mortality. In 35 comparisons of primary outcomes (including 85 point estimates from 69 unique mega-trials and 272 point estimates from smaller trials) and 26 comparisons of all-cause mortality (including 70 point estimates from 65 unique mega-trials and 267 point estimates from smaller trials), no difference existed between the outcomes of the mega-trials and smaller trials for primary outcome (ROR, 1.00; 95% CI, 0.97-1.04) nor for all-cause mortality (ROR, 1.00; 95% CI, 0.97-1.04). For the primary outcomes, smaller trials published before the mega-trials had more favorable results than the mega-trials (ROR, 1.05; 95% CI, 1.01-1.10) and subsequent smaller trials published after the mega-trials (ROR, 1.10; 95% CI, 1.04-1.18).
UNASSIGNED: In this meta-research analysis, meta-analyses of smaller studies showed overall comparable results with mega-trials, but smaller trials published before the mega-trials gave more favorable results than mega-trials. These findings suggest that mega-trials need to be performed more often given the relative low number of mega-trials found, their low significant rates, and the fact that smaller trials published prior to mega-trial report more beneficial results than mega-trials and subsequent smaller trials.

OBJECTIVE: Data imbalance is a pervasive issue in medical data mining, often leading to biased and unreliable predictive models. This study aims to address the urgent need for effective strategies to mitigate the impact of data imbalance on classification models. We focus on quantifying the effects of different imbalance degrees and sample sizes on model performance, identifying optimal cut-off values, and evaluating the efficacy of various methods to enhance model accuracy in highly imbalanced and small sample size scenarios.
METHODS: We collected medical records of patients receiving assisted reproductive treatment in a reproductive medicine center. Random forest was used to screen the key variables for the prediction target. Various datasets with different imbalance degrees and sample sizes were constructed to compare the classification performance of logistic regression models. Metrics such as AUC, G-mean, F1-Score, Accuracy, Recall, and Precision were used for evaluation. Four imbalance treatment methods (SMOTE, ADASYN, OSS, and CNN) were applied to datasets with low positive rates and small sample sizes to assess their effectiveness.
RESULTS: The logistic model\'s performance was low when the positive rate was below 10% but stabilized beyond this threshold. Similarly, sample sizes below 1200 yielded poor results, with improvement seen above this threshold. For robustness, the optimal cut-offs for positive rate and sample size were identified as 15% and 1500, respectively. SMOTE and ADASYN oversampling significantly improved classification performance in datasets with low positive rates and small sample sizes.
CONCLUSIONS: The study identifies a positive rate of 15% and a sample size of 1500 as optimal cut-offs for stable logistic model performance. For datasets with low positive rates and small sample sizes, SMOTE and ADASYN are recommended to improve balance and model accuracy.

UNASSIGNED: Time-to-event (TTE) endpoints are evaluated as the primary endpoint in single-arm clinical trials; however, limited options are available in statistical software for sample size calculation. In single-arm trials with TTE endpoints, the non-parametric log-rank test is commonly used. Parametric options for single-arm design assume survival times follow exponential distribution or Weibull distribution.
UNASSIGNED: The exponential- or Weibull-distributed survival time assumption does not always reflect hazard pattern of real-life diseases. We therefore propose gamma distribution as an alternative parametric option for designing single-arm studies with TTE endpoints. We outline a sample size calculation approach using gamma distribution with a known shape parameter and explain how to extract the gamma shape estimate from previously published resources. In addition, we conduct simulations to assess the accuracy of the extracted gamma shape parameter and to explore the impact on sample size calculation when survival time distribution is misspecified.
UNASSIGNED: Our simulations show that if a previously published study (sample sizes ≥ 60 and censoring proportions ≤ 20 %) reported median and inter-quartile range of survival time, we can obtain a reasonably accurate gamma shape estimate, and use it to design new studies. When true survival time is Weibull-distributed, sample size calculation could be underestimated or overestimated depending on the hazard shape.
UNASSIGNED: We show how to use gamma distribution in designing a single-arm trial, thereby offering more options beyond the exponential and Weibull. We provide a simulation-based assessment to ensure an accurate estimation of the gamma shape and recommend caution to avoid misspecification of the underlying distribution.

暂无摘要。

This article reflects on the potential value and many pitfalls of underpowered studies to help authors and readers consider whether and how they contribute meaningfully to the published literature. A basic introduction to power and sample size calculations is provided. Several problems that can arise in analysis and publication of underpowered studies are described. In addition, features of underpowered studies that may provide value are proposed, including when the hypothesis test of interest is a limited part of the story, the data is rich enough to showcase interesting features of the population of interest, when the rarity or ubiquity of events is an important finding, and when the study is preregistered to reduce the impact of publication bias. Several reporting guidelines for underpowered studies are also suggested.

This research addresses the presence of substances of very high concern (SVHCs) confronting the drinking water sector. Responding adequately to the potential hazards by SVHCs, knowledge of emission pathways, toxicity, presence in drinking water sources, and removability during water production is crucial. As this information cannot be received for each compound individually, we employed a detailed clustering approach based on chemical properties and structures of SVHCs from lists with over 1,000 compounds. Through this process, 915 substances were divided into 51 clusters. We tested this clustering in risk assessment. To assess the risks, we developed toxicity prediction models utilizing random forests and multiple linear regression. These models were applied to make toxicity predictions for the list of compounds. This study shows that clustering is a viable approach to reducing sample size. In addition, the toxicity models provide insights into the potential human health risks. This research contributes to more informed decision-making and improved risk assessment in the drinking water sector, aiding in the protection of human health and the environment. This principle is generally applicable. If in a group a suitable representative is found, data from experiments with this compound can be used to gauge the behaviour of chemicals in this group.

The original idea of rejecting studies with low power and authorising them if their power is sufficiently high is reasonable and even an obligation, although in practice this reasoning is heavily constrained by the fact that the power of a study depends on several factors, rather than a single one. Furthermore, there is no threshold separating \'high\' power values from \'low\' power values\'. However, if the result is very significant, considering how powerful it was it makes little sense after the study has been carried out. It is only possible to take advantage of the result. Situations in which this result is not statistically significant warrant further consideration. Consideration of the power may be useful in these circumstances. This article focuses on the position that should be adopted in these cases, and it shows that in order to draw reasonable conclusions about the effect size of the population, calculating the confidence interval is more useful than calculating the power, and its interpretation is more easily understood by physicians who lack training in statistical analysis.
BACKGROUND: Potencia estadística de una investigación médica. ¿Qué postura tomar cuando los resultados de la investigación no son significativos?
La idea original de rechazar estudios con baja potencia y autorizarlos si es suficientemente alta es razonable e incluso obligada, aunque en la práctica este razonamiento se ve muy limitado por el hecho de que la potencia de un estudio depende de varios factores y, por tanto, no es única. Además, no hay un valor frontera que separe los valores ‘altos’ de potencia de los ‘bajos’. Pese a esto, una vez realizado el estudio, si su resultado es muy significativo, no tiene sentido preguntarnos por la potencia que tenía. Sólo cabe aprovechar su resultado. Consideración aparte merece el caso en que dicho resultado no sea estadísticamente significativo. Entonces sí puede ser pertinente considerar su potencia. A continuación, se hace una reflexión sobre qué postura adoptar en estos casos y se muestra que, para sacar conclusiones razonables sobre el efecto poblacional, el cálculo de su intervalo de confianza es más útil que el cálculo de la potencia y su interpretación más fácilmente entendible por el médico sin formación en análisis estadístico.

Early in the SARS-CoV2 pandemic, in this journal, Hou et al. (BMC Med 18:216, 2020) interpreted public genotype data, run through functional prediction tools, as suggesting that members of particular human populations carry potentially COVID-risk-increasing variants in genes ACE2 and TMPRSS2 far more often than do members of other populations. Beyond resting on predictions rather than clinical outcomes, and focusing on variants too rare to typify population members even jointly, their claim mistook a well known artifact (that large samples reveal more of a population\'s variants than do small samples) as if showing real and congruent population differences for the two genes, rather than lopsided population sampling in their shared source data. We explain that artifact, and contrast it with empirical findings, now ample, that other loci shape personal COVID risks far more significantly than do ACE2 and TMPRSS2-and that variation in ACE2 and TMPRSS2 per se unlikely exacerbates any net population disparity in the effects of such more risk-informative loci.