Statistics in Medicine最新文献

Evaluating classifications is crucial in statistics and machine learning, as it influences decision-making across various fields, such as patient prognosis and therapy in critical conditions. The Matthews correlation coefficient (MCC), also known as the phi coefficient, is recognized as a performance metric with high reliability, offering a balanced measurement even in the presence of class imbalances. Despite its importance, there remains a notable lack of comprehensive research on the statistical inference of MCC. This deficiency often leads to studies merely validating and comparing MCC point estimates-a practice that, while common, overlooks the statistical significance and reliability of results. Addressing this research gap, our paper introduces and evaluates several methods to construct asymptotic confidence intervals for the single MCC and the differences between MCCs in paired designs. Through simulations across various scenarios, we evaluate the finite-sample behavior of these methods and compare their performances. Furthermore, through real data analysis, we illustrate the potential utility of our findings in comparing binary classifiers, highlighting the possible contributions of our research in this field.

Generalized estimating equations (GEE) are of great importance in analyzing clustered data without full specification of multivariate distributions. A recent approach by Luo and Pan jointly models the mean, variance, and correlation coefficients of clustered data through three sets of regressions. We note that it represents a specific case of the more general estimating equations proposed by Yan and Fine which further allow the variance to depend on the mean through a variance function. In certain scenarios, the proposed variance estimators for the variance and correlation parameters in Luo and Pan may face challenges due to the subtle dependence induced by the nested structure of the estimating equations. We characterize specific model settings where their variance estimation approach may encounter limitations and illustrate how the variance estimators in Yan and Fine can correctly account for such dependencies. In addition, we introduce a novel model selection criterion that enables the simultaneous selection of the mean-scale-correlation model. The sandwich variance estimator and the proposed model selection criterion are tested by several simulation studies and real data analysis, which validate its effectiveness in variance estimation and model selection. Our work also extends the R package geepack with the flexibility to apply different working covariance matrices for the variance and correlation structures.

High-dimensional regression problems, for example with genomic or drug exposure data, typically involve automated selection of a sparse set of regressors. Penalized regression methods like the LASSO can deliver a family of candidate sparse models. To select one, there are criteria balancing log-likelihood and model size, the most common being AIC and BIC. These two methods do not take into account the implicit multiple testing performed when selecting variables in a high-dimensional regression, which makes them too liberal. We propose the extended AIC (EAIC), a new information criterion for sparse model selection in high-dimensional regressions. It allows for asymptotic FWER control when the candidate regressors are independent. It is based on a simple formula involving model log-likelihood, model size, the total number of candidate regressors, and the FWER target. In a simulation study over a wide range of linear and logistic regression settings, we assessed the variable selection performance of the EAIC and of other information criteria (including some that also use the number of candidate regressors: mBIC, mAIC, and EBIC) in conjunction with the LASSO. Our method controls the FWER in nearly all settings, in contrast to the AIC and BIC, which produce many false positives. We also illustrate it for the automated signal detection of adverse drug reactions on the French pharmacovigilance spontaneous reporting database.

Hazard models are the most commonly used tool to analyze time-to-event data. If more than one time scale is relevant for the event under study, models are required that can incorporate the dependence of a hazard along two (or more) time scales. Such models should be flexible to capture the joint influence of several time scales, and nonparametric smoothing techniques are obvious candidates. $P$ -splines offer a flexible way to specify such hazard surfaces, and estimation is achieved by maximizing a penalized Poisson likelihood. Standard observation schemes, such as right-censoring and left-truncation, can be accommodated in a straightforward manner. Proportional hazards regression with a baseline hazard varying over two time scales is presented. Efficient computation is possible by generalized linear array model (GLAM) algorithms or by exploiting a sparse mixed model formulation. A companion R-package is provided.

In medical research, individual-level patient data provide invaluable information, but the patients' right to confidentiality remains of utmost priority. This poses a huge challenge when estimating statistical models such as a linear mixed model, which is an extension of linear regression models that can account for potential heterogeneity whenever data come from different data providers. Federated learning tackles this hurdle by estimating parameters without retrieving individual-level data. Instead, iterative communication of parameter estimate updates between the data providers and analysts is required. In this article, we propose an alternative framework to federated learning for fitting linear mixed models. Specifically, our approach only requires the mean, covariance, and sample size of multiple covariates from different data providers once. Using the principle of statistical sufficiency within the likelihood framework as theoretical support, this proposed strategy achieves estimates identical to those derived from actual individual-level data. We demonstrate this approach through real data on 15 068 patient records from 70 clinics at the Children's Hospital of Pennsylvania. Assuming that each clinic only shares summary statistics once, we model the COVID-19 polymerase chain reaction test cycle threshold as a function of patient information. Simplicity, communication efficiency, generalisability, and wider scope of implementation in any statistical software distinguish our approach from existing strategies in the literature.

Partially linear models provide a valuable tool for modeling failure time data with nonlinear covariate effects. Their applicability and importance in survival analysis have been widely acknowledged. To date, numerous inference methods for such models have been developed under traditional right censoring. However, the existing studies seldom target interval-censored data, which provide more coarse information and frequently occur in many scientific studies involving periodical follow-up. In this work, we propose a flexible class of partially linear transformation models to examine parametric and nonparametric covariate effects for interval-censored outcomes. We consider the sieve maximum likelihood estimation approach that approximates the cumulative baseline hazard function and nonparametric covariate effect with the monotone splines and $B$ -splines, respectively. We develop an easy-to-implement expectation-maximization algorithm coupled with three-stage data augmentation to facilitate maximization. We establish the consistency of the proposed estimators and the asymptotic distribution of parametric components based on the empirical process techniques. Numerical results from extensive simulation studies indicate that our proposed method performs satisfactorily in finite samples. An application to a study on hypobaric decompression sickness suggests that the variable TR360 exhibits a significant dynamic and nonlinear effect on the risk of developing hypobaric decompression sickness.

Multicancer detection (MCD) tests use blood specimens to detect preclinical cancers. A major concern is overdiagnosis, the detection of preclinical cancer on screening that would not have developed into symptomatic cancer in the absence of screening. Because overdiagnosis can lead to unnecessary and harmful treatments, its quantification is important. A key metric is the screen overdiagnosis fraction (SOF), the probability of overdiagnosis at screen detection. Estimating SOF is notoriously difficult because overdiagnosis is not observed. This estimation is more challenging with MCD tests because short-term results are needed as the technology is rapidly changing. To estimate average SOF for a program of yearly MCD tests, I introduce a novel method that requires at least two yearly MCD tests given to persons having a wide range of ages and applies only to cancers for which there is no conventional screening. The method assumes an exponential distribution for the sojourn time in an operational screen-detectable preclinical cancer (OPC) state, defined as once screen-detectable (positive screen and work-up), always screen-detectable. Because this assumption appears in only one term in the SOF formula, the results are robust to violations of the assumption. An SOF plot graphs average SOF versus mean sojourn time. With lung cancer screening data and synthetic data, SOF plots distinguished small from moderate levels of SOF. With its unique set of assumptions, the SOF plot would complement other modeling approaches for estimating SOF once sufficient short-term observational data on MCD tests become available.

To complement the conventional area under the ROC curve (AUC) which cannot fully describe the diagnostic accuracy of some non-standard biomarkers, we introduce a transformed ROC curve and its associated transformed AUC (TAUC) in this article, and show that TAUC can relate the original improper biomarker to a proper biomarker after a non-monotone transformation. We then provide nonparametric estimation of the non-monotone transformation and TAUC, and establish their consistency and asymptotic normality. We conduct extensive simulation studies to assess the performance of the proposed TAUC method and compare with the traditional methods. Case studies on real biomedical data are provided to illustrate the proposed TAUC method. We are able to identify more important biomarkers that tend to escape the traditional screening method.

Clinical decisions are often guided by clinical prediction models or diagnostic tests. Decision curve analysis (DCA) combines classical assessment of predictive performance with the consequences of using these strategies for clinical decision-making. In DCA, the best decision strategy is the one that maximizes the net benefit: the net number of true positives (or negatives) provided by a given strategy. Here, we employ Bayesian approaches to DCA, addressing four fundamental concerns when evaluating clinical decision strategies: (i) which strategies are clinically useful, (ii) what is the best available decision strategy, (iii) which of two competing strategies is better, and (iv) what is the expected net benefit loss associated with the current level of uncertainty. While often consistent with frequentist point estimates, fully Bayesian DCA allows for an intuitive probabilistic interpretation framework and the incorporation of prior evidence. We evaluate the methods using simulation and provide a comprehensive case study. Software implementation is available in the bayesDCA R package. Ultimately, the Bayesian DCA workflow may help clinicians and health policymakers adopt better-informed decisions.

In epidemiological studies, evaluating the health impacts stemming from multiple exposures is one of the important goals. To analyze the effects of multiple exposures on discrete or time-to-event health outcomes, researchers often employ generalized linear models, Cox proportional hazards models, and machine learning methods. However, observational studies are prone to unmeasured confounding factors, which can introduce the potential for substantial bias in the multiple exposure effects. To address this issue, we propose a novel outcome model-based sensitivity analysis method for non-Gaussian and time-to-event outcomes with multiple exposures. All the proposed sensitivity analysis problems are formulated as linear programming problems with quadratic and linear constraints, which can be solved efficiently. Analytic solutions are provided for some optimization problems, and a numerical study is performed to examine how the proposed sensitivity analysis behaves in finite samples. We illustrate the proposed method using two real data examples.