Search

An important goal of censored quantile regression is to provide reliable predictions of survival quantiles, which are often reported in practice to offer robust and comprehensive biomedical summaries. However, formal methods for evaluating and comparing working quantile regression models in terms of their performance in predicting survival quantiles have been lacking, especially when the working models are subject to model mis-specification. In this article, we proposes a sensible and rigorous framework to fill in this gap. We introduce and justify a predictive performance measure defined based on the check loss function. We derive estimators of the proposed predictive performance measure and study their distributional properties and the corresponding inference procedures. More importantly, we develop model comparison procedures that enable thorough evaluations of model predictive performance among nested or non-nested models. Our proposals properly accommodate random censoring to the survival outcome and the realistic complication of model mis-specification, and thus are generally applicable. Extensive simulations and a real data example demonstrate satisfactory performances of the proposed methods in real life settings., (© 2016, The International Biometric Society.)

Chemical analysis adopts a calibration curve to establish the relationship between the measuring technique's response and the target analyte's standard concentration. The calibration equation is established using regression analysis to verify the response of a chemical instrument to the known properties of materials that served as standard values. An adequate calibration equation ensures the performance of these instruments. There are two kinds of calibration equations: classical equations and inverse equations. For the classical equation, the standard values are independent, and the instrument's response is dependent. The inverse equation is the opposite: the instrument's response is the independent value. For the new response value, the calculation of the new measurement by the classical equation must be transformed into a complex form to calculate the measurement values. However, the measurement values of the inverse equation could be computed directly. Different forms of calibration equations besides the linear equation could be used for the inverse calibration equation. This study used measurement data sets from two kinds of humidity sensors and nine data sets from the literature to evaluate the predictive performance of two calibration equations. Four criteria were proposed to evaluate the predictive ability of two calibration equations. The study found that the inverse calibration equation could be an effective tool for complex calibration equations in chemical analysis. The precision of the instrument's response is essential to ensure predictive performance. The inverse calibration equation could be embedded into the measurement device, and then intelligent instruments could be enhanced. [ABSTRACT FROM AUTHOR]

Objective: We aimed to develop a Radiological-Radiomics (R-R) based model for predicting the high-grade pattern (HGP) of lung adenocarcinoma and evaluate its predictive performance. Methods: The clinical, pathological, and imaging data of 374 patients pathologically confirmed with lung adenocarcinoma (374 lesions in total) were retrospectively analyzed. The 374 lesions were assigned to HGP (n = 81) and non-high-grade pattern (n-HGP, n = 293) groups depending on the presence or absence of high-grade components in pathological findings. The least absolute shrinkage and selection operator (LASSO) method was utilized to screen features on the United Imaging artificial intelligence scientific research platform, and logistic regression models for predicting HGP were constructed, namely, Radiological model, Radiomics model, and R-R model. Also, receiver operating curve (ROC) curves were plotted on the platform, generating corresponding area under the curve (AUC), sensitivity, specificity, and accuracy. Using the platform, nomograms for R-R models were also provided, and calibration curves and decision curves were drawn to evaluate the performance and clinical utility of the model. The statistical differences in the performance of the models were compared by the DeLong test. Results: The R-R model for HGP prediction achieved an AUC value of 0.923 (95% CI: 0.891-0.948), a sensitivity of 87.0%, a specificity of 83.4%, and an accuracy of 84.2% in the training set. In the validation set, this model exhibited an AUC value of 0.920 (95% CI: 0.887-0.945), a sensitivity of 87.5%, a specificity of 83.3%, and an accuracy of 84.2%. The DeLong test demonstrated optimal performance of the R-R model among the three models, and decision curves validated the clinical utility of the R-R model. Conclusion: In this study, we developed a fusion model using radiomic features combined with radiological features to predict the high-grade pattern of lung adenocarcinoma, and this model shows excellent diagnostic performance. The R-R model can provide certain guidance for clinical diagnosis and surgical treatment plans, contributing to improving the prognosis of patients. [ABSTRACT FROM AUTHOR]

Summary: The G arvan Fracture Risk Calculator predicts risk of osteoporotic fractures. We evaluated its predictive performance in 16,682 women and 2839 men from Manitoba, Canada, and found significant risk stratification, with a strong gradient across scores. The tool outperformed clinical risk factors and bone mineral density for fracture risk stratification. Introduction: The optimal model for fracture risk estimation to guide treatment decision-making remains controversial. Our objective was to evaluate the predictive performance of the Garvan Fracture Risk Calculator (FRC) in a large clinical registry from Manitoba, Canada. Methods: Using the population-based Manitoba Bone Mineral Density (BMD) registry, we identified women and men aged 50–95 years undergoing baseline BMD assessment from September 1, 2012, onwards. Five-year Garvan FRC predictions were generated from clinical risk factors (CRFs) with and without femoral neck BMD. We identified incident non-traumatic osteoporotic fractures (OFs) and hip fractures (HFs) from population-based healthcare data sources to March 31, 2018. Fracture risk was assessed from area under the receiver operating characteristic curve (AUROC). Cox regression analysis and calibration ratios (5-year observed/predicted) were assessed for risk quintiles. All analyses were sex stratified. Results: We included 16,682 women (mean age 66.6 + / − SD 8.7 years) and 2839 men (mean age 68.7 + / − SD 10.2 years). During a mean observation time of 2.6 years, incident OFs were identified in 681 women and 140 men and HFs in 199 women and 22 men. AUROC showed significant fracture risk stratification with the Garvan FRC. Tool predictions without BMD were better than from age or decreasing weight, and the tool with BMD performed better than BMD alone. Garvan FRC with BMD performed better than without BMD, especially for HF prediction (AUROC 0.86 in women, 0.82 in men). There was a strong gradient of increasing risk across Garvan FRC quintiles (highest versus lowest, hazard ratios women 5.75 and men 3.43 for any OF; women 101.6 for HF). Calibration differences were noted, with both over- and underestimation in risk. Conclusions: Garvan FRC outperformed CRFs and BMD alone for fracture risk stratification, particularly for HF, but may require recalibration for accurate predictions in this population. [ABSTRACT FROM AUTHOR]

Support vector regression (SVR) is able to consider the nonlinear relationship between explanatory variables X and a target variable y to build a regression model with high predictive accuracy. Additionally, y values predicted with SVR models for new samples can exceed the actual y values in training data. However, because the Gaussian kernel, which is a kernel function generally used in SVR, is based on the Euclidean distance between samples, it is unable to consider the importance of X when building the regression model. Therefore, in this study, the focus was on the importance of X that can be calculated by random forests (RF), and a novel SVR method, called variable importance‐considering support vector regression (VI‐SVR), was proposed based on this importance. Because X is weighted based on importance, the greater the importance of X, the greater its contribution to the predicted value. Analysis using the spectral, quantitative structure–property relationship (QSPR), and quantitative structure–activity relationship (QSAR) datasets confirmed that the predictive accuracy of VI‐SVR was better than that of SVR. VI‐SVR Python code is available at https://github.com/hkaneko1985/dcekit. We proposed a novel SVR method, called variable importance‐considering support vector regression (VI‐SVR), based on this importance calculated with random forests. Because X is weighted based on importance, the greater the importance of X, the greater its contribution to the predicted value. Analysis using the spectral, QSPR, and QSAR datasets confirmed that the predictive accuracy of VI‐SVR was better than that of SVR. [ABSTRACT FROM AUTHOR]

This paper invents the Adaptive Logit Model (ALM) and discovers that the logistic regression or logit model can perform better prediction if it is trained in an adaptive way. For each observation (m), ALM uses the weights (wk m), to adjust parameters (ßk). If a weight equals to 1, there is no adjustment of that parameter. By iteration of f rounds for each observation, the training will find the final weights (wk f, m), that lead to the match between that particular observation's true value of the dependent variable, Ym whose value is binary of either 0 or 1, and its estimated value, Ym calculated from the predictive probability, Pr(Ym = 1) derived from the traditional logit model and compared to the threshold of 0.50 which means if the predictive probability exceeds 0.50 the estimated value becomes 1. For each observation, the iterative process keeps running until Ym = Ym. The final weights are then collected. Completing this procedure for all observations. The average weights (wk f, m) can be obtained by arithmetic and geometric mean. For the prediction, these average weights are used as denominators for the related parameters from the traditional logit model in calculating the predictive probability. The prediction is based on the threshold of 0.50 again. The predictive performance after the modification of the parameters are more accurate for overall accuracy along with the True positive, true negative, false positive, and false negative rates. These improvements are true for all tests against the training and testing sets with larger portion of "one," balanced between "one" and "zero," and larger portion of "zero." Comparing to the traditional logit model, this ALM with arithmetic mean yields 3.09 percentage points higher of the overall accuracy. Besides, the ALM with geometric mean improves the overall accuracy by 3.47 percentage points. However, this study is still limited to a case of two explanatory variables without severe multicollinearity. [ABSTRACT FROM AUTHOR]

Abstract Semi-supervised learning (SSL) is a method for learning the relationship between X and y, and the essential structure of the corresponding dataset, using both labeled and unlabeled data. In this paper, an approach to use a combination of labeled and unlabeled samples to reduce the dimension, then perform regression analysis using the labeled samples in a low-dimensional space is focused in SSL methods. While various SSL methods for regression have been developed, there has been insufficient discussion as to why SSL is effective in regression analysis. Therefore, in this study, the merits of SSL in regression analysis are discussed in terms of the stability or the robustness and applicability domains of regression models and prior distribution of X-variables. The superiorities of SSL methods over fully supervised methods in regression are demonstrated using data from numerical simulations, quantitative structure–activity relationships and quantitative structure–property relationships. Highlights • The merits of semi-supervised learning (SSL) in regression analysis are discussed. • Regression models are stabilized by the use of unlabeled data in SSL. • SSL enables the applicability domains of the regression models to be expanded. • SSL clarifies the prior distribution of X-variables. • The superiorities of SSL methods over fully supervised methods in regression are demonstrated. [ABSTRACT FROM AUTHOR]

The fields of chemoinformatics and chemometrics require regression models with high prediction performance. To construct predictive regression models by appropriately detecting outlier samples, a new outlier detection and regression method based on ensemble learning is proposed. Multiple regression models are constructed and y-values are estimated based on ensemble learning. Outlier samples are then detected by comprehensively considering all regression models. Furthermore, it is possible to detect outlier samples robustly and independently by repeated calculations. By analyzing a numerical simulation dataset, two quantitative structure-activity relationship datasets and two quantitative structure-property relationship datasets, it is confirmed that automatic outlier sample detection can be achieved, informative compounds can be selected, and the estimation performance of regression models is improved. [ABSTRACT FROM AUTHOR]

When making predictions and inferences, data analysts are often faced with the challenge of selecting the best model among competing models as a result of large number of regressors that cumulate into large model space. Bayesian model averaging (BMA) is a technique designed to help account for uncertainty inherent in model selection process. In Bayesian analysis, issues of the choice of prior distribution have been quite delicate in data analysis and posterior model probabilities (PMP) in the context of model uncertainty under model selection process are typically sensititve to the specification of prior distribution. This research identified a set of eleven candidate default priors (Zellner's g-priors) prominent in literature and applicable in Bayesian model averaging. A new robust g-prior specification for regression coefficients in Bayesian Model Averaging is investigated and its predictive performance assessed along with other g-prior structures in literature. The predictive abilities of these g-prior structures are assessed using log predictive scores (LPS) and log maximum likelihood (LML). The sensitivity of posterior results to the choice of these g-prior structures was demonstrated using simulated data and real-life data. The simulated data obtained from multivariate normal distribution were first used to demonstrate the predictive performance of the g-prior structures and later contaminated for the same purpose. Similarly for the same purpose, the real life data were normalized before using the data as obtained. Empirical findings reveal that under different conditions, the new g-prior structure exhibited robust, equally competitive and consistent predictive ability when compared with identified g-prior structures from the literature. The new g-prior offers a sound, fully Bayesian approach that features the virtues of prior input and predictive gains that minimise the risk of misspecification. [ABSTRACT FROM AUTHOR]

The coefficient of determination and the root-mean-squared error (RMSE) evaluate regression models for test samples without considering the applicability domains (ADs) of the models. In this study, we propose a new measure for evaluating the predictive performance of regression models that considers their ADs. The purpose is not selecting the best regression model among various competing models, but determining an appropriate model group corresponding to the AD of each model. The proposed measure is the area under coverage and RMSE curve for coverage less than p % ( p % - AUCR). It is confirmed that some regression models have global predictive ability and others have local predictive ability, and p %-AUCR is an appropriate indicator for selecting between local and global regression models depending on the coverage and considering the AD. Selecting a regression model for each sample or each chemical structure using p % - AUCR can improve the prediction accuracy of data sets. [ABSTRACT FROM AUTHOR]

Joint modeling methods have become popular tools to link important features extracted from longitudinal data to a primary event. While most modeling strategies have focused on the association between the longitudinal mean trajectories and risk of an event, we consider joint models that incorporate information from both long-term trends and short-term variability in a longitudinal submodel. We also consider both shared random effect and latent class (LC) approaches in the primary-outcome model to predict a binary outcome of interest. We develop simulation studies to compare and contrast these two modeling strategies; in particular, we study in detail the effects of the primary-outcome model misspecification. Among other findings, we note that when we analyze data from a shared random-effect using a LC model while the information from the longitudinal data is weak, the LC approach is more sensitive to such a model misspecification. Under this setting, the LC model has a superior performance in within-sample prediction that cannot be duplicated when predicting new samples. This is a unique feature of the LC approach that is new as far as we know to the existing literature. Finally, we use the proposed models to study how follicle stimulating hormone (FSH) trajectories are related to the risk of developing severe hot flashes for participating women in the Penn Ovarian Aging Study. [ABSTRACT FROM AUTHOR]

Abstract: Improving predictions of the location of suitable environmental conditions for species using species distribution models (SDM) is at the core of biodiversity/climate change research, but modelling species abundance, rather than distribution, is proving particularly challenging. Using data from more than 200,000 forest plots in eastern North America and Random Forest, we evaluated the performance of species abundance models (SAM) in predicting the relative abundance (measured as importance value) of each of 105 tree species in relation to climate, edaphic, and topographic variables. We calculated the coefficient of determination between observed and predicted abundances as a measure of model performance for each species. We also performed multiple linear regressions to explain variation of among species using five biogeographical or spatial attributes of species as explanatory variables. Predictive performances of SAM were generally low, ranging from 0.016 to 0.815 (mean=0.258). Black spruce (Picea mariana) had the best predictive model and Florida maple (Acer barbatum) and American chestnut (Castanea dentata) the worst. Thirty-seven of the 41 best performing species had climate ranked as the best and/or second best predictor. Species with the best performance tended to be those that could reach dominance, showed aggregated distribution of abundance, and/or had high latitudinal limits in the study area. Climate change is likely to affect patterns of dominance in communities by altering patterns of co-occurrences, but for many species that constitute the bulk of tree diversity, predictions based solely on the current distribution of relative abundances may not be reliable enough to inform conservation or management decisions. Predicting tree abundance in a warming climate using SAM remains a challenge, but it is only by reporting performances across a range of climate and statistical models, regions and species, as well as by highlighting model limitations and strengths, that we will improve the reliability of predictions and in turn better inform forest conservation and management. [Copyright &y& Elsevier]

The article presents a study that calculates standardized mortality ratio (SMR) of 30 equally-sized consecutive subsets from a total of 12,143 ICU patients aged 80 years or older and plotted them on a PreControl Chart. It calculates individuaI hospital SMRs in 2009, with and without repeated recalibration of the models on earlier data. It is mentioned that the tree-model has seemingly a longer shelf life than the logistic regression model.

Because moclobemide pharmacokinetics vary considerably among individuals, monitoring of plasma concentrations lends insight into its pharmacokinetic behavior and enhances its rational use in clinical practice. The aim of this study was to evaluate whether single concentration–time points could adequately predict moclobemide systemic exposure. Pharmacokinetic data (full 7-point pharmacokinetic profiles), obtained from 21 depressive inpatients receiving moclobemide (150 mg 3 times daily), were randomly split into development (n = 18) and validation (n = 16) sets. Correlations between the single concentration–time points and the area under the concentration–time curve within a 6-hour dosing interval at steady-state (AUC0-6) were assessed by linear regression analyses. The predictive performance of single-point sampling strategies was evaluated in the validation set by mean prediction error, mean absolute error, and root mean square error. Plasma concentrations in the absorption phase yielded unsatisfactory predictions of moclobemide AUC0-6. The best estimation of AUC0-6 was achieved from concentrations at 4 and 6 hours following dosing. As the most reliable surrogate for moclobemide systemic exposure, concentrations at 4 and 6 hours should be used instead of predose trough concentrations as an indicator of between-patient variability and a guide for dose adjustments in specific clinical situations. [ABSTRACT FROM PUBLISHER]

The United Nations (UN) Education Index (EI), a widely accepted measure of educational attainment within a country and a component of the Human Development Index, is measured by combining the average adult years of schooling with expected years of schooling for children. It is concerning that sub-Saharan African (SSA) countries tend to have a relatively low average EI of 0.45 compared to a global average of 0.65 for all countries. This suggests that SSA countries are underperforming in this regard. Insufficient attention has been devoted to identifying key factors that could explain the variability in educational attainment and, most especially, the underperformance in SSA countries. This study investigates the predictive power of female enrollment, child mortality, and corruption, among other variables, to explain the EI. The data comprised 165 countries and 47 explanatory variables, and one target variable (EI), obtained from the World Bank and Transparency International. Machine learning and variable selection techniques were utilized to identify the relevant indicators. Predictive performance (in-sample and out-of-sample using cross-validation) facilitated a deeper understanding of the indicators that best explain the variation in the EI. The variable selection process was accomplished using the least absolute shrinkage and selection operator (LASSO). The mortality rate of children under the age of five and the corruption perception index are two non-education-related indicators that contributed most to the prediction of the dependent variable (EI). The education-related predictor was identified as the net female enrollment in secondary schools. Afterward, a linear regression was performed on the selected variables. To ensure that the model can generalize, a cross-validation technique was applied using the leave one out (LOO) approach. The out-of-sample performance of R2 = 0.89 implies that the model explains much of the variability in the EI. The identified relationships offer a basis for policy recommendations. [ABSTRACT FROM AUTHOR]

A new measure of predictive performance for objective variables in regression analysis is proposed, enabling the y‐errors of new samples or test samples to be estimated in the applicability domains (ADs) of regression models. The proposed measure, based on y‐randomization, considers chance correlations and is calculated using only training data. This chance correlation‐excluded mean absolute error (MAECCE) can estimate the y‐errors of new samples considering the influence of chance correlations in the given dataset on the regression models. Experiments using numerical simulation, quantitative structure‐activity relationship, and quantitative structure‐property relationship datasets confirm that MAECCE can estimate the distribution of y‐errors of new samples in ADs for various training datasets, descriptor sets, and regression analysis methods, enabling chance correlations to be eliminated from data analysis. Python and MATLAB codes for the proposed algorithm are available at https://github.com/hkaneko1985/maecce. The proposed measure, based on y‐randomization, considers chance correlations and is calculated using only training data. This chance correlation‐excluded mean absolute error (MAECCE) can estimate the y‐errors of new samples or test samples in the applicability domains (ADs) of regression models, considering the influence of chance correlations in the given dataset on the regression models. [ABSTRACT FROM AUTHOR]