Bayesian Additive Regression Trees Research Articles

Significant variables affecting the performance of concrete panels impacted by wind-borne projectiles: A global sensitivity analysis

A prominent threat to key structures and components in critical facilities, especially during a tornado or hurricane, is impact by wind-borne projectiles. Identifying uncertain variables in these phenomena is key for developing cost-effective design formulae and efficient reliability analysis. This task however is challenging due to the scarcity of data and the large set of uncertain factors that contribute to the impact phenomenon. The present study performs a robust global sensitivity analysis (GSA) using Sobol's indices on the response of a concrete panel impacted by a Schedule 40 pipe. A high-fidelity, nonlinear Finite Element (FE) model is developed using the Smooth Particle Hydrodynamics (SPH) formulation in LS-DYNA. The developed model was validated with the experiments conducted by the Electric Power Research Institute (EPRI). Due to the high computational demand of the SPH model, a machine learning-based surrogate model called Bayesian Additive Regression Trees (BART) is applied to emulate the computational model. This surrogate model that is trained with a limited number of generated SPH simulations is capable of accurately predicting the behavior of concrete panels subjected to projectile impact over the space of predictors. The predictors here include uncertain variables associated with concrete and steel material properties. GSA is subsequently performed by integrating the constructed surrogate model into Sobol's algorithm. Results indicate that concrete tensile strength plays a significant role in panel damage, while concrete mass density and compressive strength and mass density of the steel pipe are also significant. These findings suggest that the development of new empirical design formulae and experimental studies on the impact of concrete panels should include the identified significant variables.

Prenatal exposure to a mixture of persistent organic pollutants (POPs) and child reading skills at school age.

Prenatal exposure to persistent organic pollutants (POPs) may affect child neurobehavior; however, exposures to mixtures of POPs have rarely been examined. We estimated associations of prenatal serum concentrations of 17 POPs, namely 5 polybrominated diphenyl ethers (PBDEs), 6 polychlorinated biphenyls (PCBs), dichlorodiphenyldichloroethylene (DDE), dichlorodiphenyltrichloroethane (DDT), and 4 per- and polyfluoroalkyl substances (PFAS), with Wide Range Achievement Test-4 reading composite scores at age 8 years in 161 children from a pregnancy and birth cohort (Health Outcomes and Measures of the Environment [HOME] Study, 2003-present) in Cincinnati, OH. We applied 6 statistical methods: least absolute shrinkage and selection operator (LASSO), elastic net (ENET), Sparse Principal Component Analysis (SPCA), Weighted Quantile Sum (WQS) regression, Bayesian Kernel Machine Regression (BKMR), and Bayesian Additive Regression Trees (BART), to estimate covariate-adjusted associations with individual and their mixtures in multi-pollutant models. Both LASSO and ENET models indicated inverse associations with reading scores for BDE-153 and BDE-28, and positive associations for CB-118, CB-180, perfluoroctanoate (PFOA), and perfluorononanoate (PFNA). The SPCA identified inverse associations for BDE-153 and BDE-100 and positive associations for perfluorooctane sulfonate (PFOS), PFOA, and PFNA, as parts of different principal component scores. The WQS regression showed the highest weights for BDE-100 (0.35) and BDE-28 (0.16) in the inverse association model and for PFNA (0.29) and CB-180 (0.21) in the positive association model. The BKMR model identified BDE-100 and BDE-153 for inverse associations and CB-118, CB-153, CB-180, PFOA, and PFNA for positive associations. The BART method found dose-response functions similar to the BKMR model. No interactions between POPs were identified. Despite some inconsistency among biomarkers, these analyses revealed inverse associations between prenatal PBDE concentrations and children's reading scores. Positive associations of PCB congeners and PFAS with reading skills were also found.

Analyzing Nonresponse in Longitudinal Surveys Using Bayesian Additive Regression Trees: A Nonparametric Event History Analysis

Increasing nonresponse rates is a pressing issue for many longitudinal panel studies. Respondents frequently either refuse participation in single survey waves (temporary dropout) or discontinue participation altogether (permanent dropout). Contemporary statistical methods that are used to elucidate predictors of survey nonresponse are typically limited to small variable sets and ignore complex interaction patterns. The innovative approach of Bayesian additive regression trees (BART) is an elegant way to overcome these limitations because it does not specify a parametric form for the relationship between the outcome and its predictors. We present a BART event history analysis that allows identifying predictors for different types of nonresponse to anticipate response rates for upcoming survey waves. We apply our novel method to data from the German National Educational Panel Study including N = 4,559 students in Grade 5 that observed nonresponse rates of up to 36% across five waves. A cross-validation and comparison with logistic regression models with least absolute shrinkage and selection operator penalization underline the advantages of the approach. Our results highlight the potential of Bayesian discrete-time event modeling for the long-term projection of panel stability across multiple survey waves. Finally, potential applications of this approach for operational use in survey management are outlined.

Estimation of causal effects of multiple treatments in observational studies with a binary outcome.

There is a dearth of robust methods to estimate the causal effects of multiple treatments when the outcome is binary. This paper uses two unique sets of simulations to propose and evaluate the use of Bayesian additive regression trees in such settings. First, we compare Bayesian additive regression trees to several approaches that have been proposed for continuous outcomes, including inverse probability of treatment weighting, targeted maximum likelihood estimator, vector matching, and regression adjustment. Results suggest that under conditions of non-linearity and non-additivity of both the treatment assignment and outcome generating mechanisms, Bayesian additive regression trees, targeted maximum likelihood estimator, and inverse probability of treatment weighting using generalized boosted models provide better bias reduction and smaller root mean squared error. Bayesian additive regression trees and targeted maximum likelihood estimator provide more consistent 95% confidence interval coverage and better large-sample convergence property. Second, we supply Bayesian additive regression trees with a strategy to identify a common support region for retaining inferential units and for avoiding extrapolating over areas of the covariate space where common support does not exist. Bayesian additive regression trees retain more inferential units than the generalized propensity score-based strategy, and shows lower bias, compared to targeted maximum likelihood estimator or generalized boosted model, in a variety of scenarios differing by the degree of covariate overlap. A case study examining the effects of three surgical approaches for non-small cell lung cancer demonstrates the methods.

Significant variables for leakage and collapse of buried concrete sewer pipes: a global sensitivity analysis via Bayesian additive regression trees and Sobol’ indices

Identifying key variables that influence crack formation in buried concrete sewer pipes and impact their structural integrity is crucial for evaluating the leakage and collapse vulnerability of these underground structures. The present study performs a comprehensive global sensitivity analysis for these critical pipe responses considering various uncertainties associated with material properties and loading. A high-fidelity three-dimensional nonlinear Finite Element (FE) model of a typical pipe is developed and used to capture pipe behavior. Considering the significantly high computational demand of the generated FE model, a surrogate model using Bayesian Additive Regression Trees (BART) is developed to accurately approximate input-output relationships in this high-dimensional problem. Significant variables are subsequently determined via a global sensitivity analysis based on Sobol’s theorem through the application of Markov Chain Monte Carlo simulation to the trained BART model. Results indicate that concrete and backfill soil properties and variables associated with truckloads contribute significantly to the variation of key pipe responses, thereby affecting the likelihood of leakage and collapse. Findings of this study can benefit the design and management of sewer pipes in large wastewater networks and significantly reduce the complexity of assessing the reliability of sewer pipes.

Problems with products? Control strategies for models with interaction and quadratic effects

AbstractModels testing interactive and quadratic hypotheses are common in Political Science but control strategies for these models have received little attention. Common practice is to simply include additive control variables, without relevant product terms, into models with interaction or quadratic terms. In this paper, we show in Monte Carlos that interaction terms can absorb the effects of other un-modeled interaction and non-linear effects and analogously, that included quadratic terms can reflect omitted interactions and non-linearities. This problem even occurs when included and omitted product terms do not share any constitutive terms. We show with Monte Carlo experiments that regularized estimators, the adaptive Lasso, Kernel Regularized Least Squares (KRLS), and Bayesian Additive Regression Trees (BART) can prevent the misattribution of interactive/quadratic effects, minimize the problems of efficiency loss and overfitting, and have low false-positive rates. We illustrate how inferences drawn can change when relevant product terms are used in the control strategy using a recent paper. Implementing the recommendations of this paper would increase the reliability of conditional and non-linear relationships estimated in many papers in the literature.

Cannabis use is associated with a small increase in the risk of postoperative nausea and vomiting: a retrospective machine-learning causal analysis

BackgroundCannabis legalization may contribute to an increased frequency of chronic use among patients presenting for surgery. At present, it is unknown whether chronic cannabis use modifies the risk of postoperative nausea and vomiting (PONV).MethodsThis study was a retrospective cohort study conducted at 2 academic medical centers. Twenty-seven thousand three hundred eighty-eight adult ASA 1–3 patients having general anesthesia for non-obstetric, non-cardiac procedures and receiving postoperative care in the Post Anesthesia Care Unit (PACU) were analyzed in the main dataset, and 16,245 patients in the external validation dataset. The main predictor was patient reported use of cannabis in any form collected during pre-anesthesia evaluation and recorded in the chart. The primary outcome was documented PONV of any severity prior to PACU discharge, including administration of rescue medications in PACU. Relevant clinical covariates (risk factors for PONV, surgical characteristics, administered prophylactic antiemetic drugs) were also recorded.Results10.0% of patients in the analytic dataset endorsed chronic cannabis use. Using Bayesian Additive Regression Trees (BART), we estimated that the relative risk for PONV associated with daily cannabis use was 1.19 (95 CI% 1.00–1.45). The absolute marginal increase in risk of PONV associated with daily cannabis use was 3.3% (95% CI 0.4–6.4%). We observed a lesser association between current, non-daily use of cannabis (RR 1.07, 95% CI 0.94–1.21). An internal validation analysis conducted using propensity score adjustment and Bayesian logistic modeling indicated a similar size and magnitude of the association between cannabis use and PONV (OR 1.15, 90% CI 0.98–1.33). As an external validation, we used data from another hospital in our care system to create an independent model that demonstrated essentially identical associations between cannabis use and PONV.ConclusionsCannabis use is associated with an increased relative risk and a small increase in the marginal probability of PONV.

Personalized Network Modeling of the Pan-Cancer Patient and Cell Line Interactome.

PURPOSEPersonalized network inference on diverse clinical and in vitro model systems across cancer types can be used to delineate specific regulatory mechanisms, uncover drug targets and pathways, and develop individualized predictive models in cancer.METHODSWe developed TransPRECISE (personalized cancer-specific integrated network estimation model), a multiscale Bayesian network modeling framework, to analyze the pan-cancer patient and cell line interactome to identify differential and conserved intrapathway activities, to globally assess cell lines as representative models for patients, and to develop drug sensitivity prediction models. We assessed pan-cancer pathway activities for a large cohort of patient samples (> 7,700) from the Cancer Proteome Atlas across ≥ 30 tumor types, a set of 640 cancer cell lines from the MD Anderson Cell Lines Project spanning 16 lineages, and ≥ 250 cell lines’ response to > 400 drugs.RESULTSTransPRECISE captured differential and conserved proteomic network topologies and pathway circuitry between multiple patient and cell line lineages: ovarian and kidney cancers shared high levels of connectivity in the hormone receptor and receptor tyrosine kinase pathways, respectively, between the two model systems. Our tumor stratification approach found distinct clinical subtypes of the patients represented by different sets of cell lines: patients with head and neck tumors were classified into two different subtypes that are represented by head and neck and esophagus cell lines and had different prognostic patterns (456 v 654 days of median overall survival; P = .02). High predictive accuracy was observed for drug sensitivities in cell lines across multiple drugs (median area under the receiver operating characteristic curve > 0.8) using Bayesian additive regression tree models with TransPRECISE pathway scores.CONCLUSIONOur study provides a generalizable analytic framework to assess the translational potential of preclinical model systems and to guide pathway-based personalized medical decision making, integrating genomic and molecular data across model systems.

PNS25 METHODS FOR DETECTING TREATMENT EFFECT MODIFIERS IN ANCHORED INDIRECT TREATMENT COMPARISONS: AN OVERVIEW AND SIMULATION STUDY OF MACHINE LEARNING ALGORITHMS

Anchored matching-adjusted indirect treatment comparisons (ITC) or propensity score adjustment methods (e.g. matching, reweighting) leverage individual patient data (IPD) to compare treatments across trials by anchoring through a common comparator. Since anchored ITCs adjust for imbalances in prognostic factors via the common comparator, health technology agencies (HTAs) such as NICE have recommended adjusting only for treatment effect modifiers that exhibit large imbalances between studies. Although many automated and semi-automated methods (e.g. high dimensionality propensity score, random forest, Bayesian Additive Regression Trees) for variable selection are available in open-source software packages, most practical applications of these methods have focused on detecting prognostic factors rather than treatment effect modifiers, and do not factor in imbalances between studies. This study provides a general overview of common machine-learning variable selection routines often employed for prognostic applications and explore their feasibility for identifying imbalanced treatment effect modifiers. We use Monte Carlo simulation studies to evaluate the finite sample performance of several supervised machine learning algorithms, in terms of their ability to detect true treatment effect modifiers with large imbalances, bias, and efficiency. Simulations are performed using simulated IPD that include one or more of the following characteristics: a) small samples; b) data sparsity relative to the number of potential treatment effect modifiers; and c) collinearity among variables. The findings from this overview and simulation study can inform data-driven variable selection methods for HTA and regulatory submissions involving indirect treatment comparisons which require the identification of variables that are important effect modifiers and differ between studies.

Embarcadero: Species distribution modelling with Bayesian additive regression trees in r

Abstract embarcadero is an r package of convenience tools for species distribution modelling (SDM) with Bayesian additive regression trees (BART), a powerful machine learning approach that has been rarely applied to ecological problems. Like other classification and regression tree methods, BART estimates the probability of a binary outcome based on a set of decision trees. Unlike other methods, BART iteratively generates sets of trees based on a set of priors about tree structure and nodes, and builds a posterior distribution of estimated classification probabilities. So far, BARTs have yet to be applied to SDM. embarcadero is a workflow wrapper for BART species distribution models, and includes functionality for easy spartial prediction, an automated variable selection procedure, several types of partial dependence visualization and other tools for ecological application. The embarcadero package is an open source and available on Github. To show how embarcadero can be used by ecologists, I illustrate a BART workflow for a virtual species distribution model. The supplement includes a more advanced vignette showing how BART can be used for mapping disease transmission risk, using the example of Crimean–Congo haemorrhagic fever in Africa.

Comparing the performance of statistical methods that generalize effect estimates from randomized controlled trials to much larger target populations

Policymakers use results from randomized controlled trials to inform decisions about whether to implement treatments in target populations. Various methods—including inverse probability weighting, outcome modeling, and Targeted Maximum Likelihood Estimation—that use baseline data available in both the trial and target population have been proposed to generalize the trial treatment effect estimate to the target population. Often the target population is significantly larger than the trial sample, which can cause estimation challenges. We conduct simulations to compare the performance of these methods in this setting. We vary the size of the target population, the proportion of the target population selected into the trial, and the complexity of the true selection and outcome models. All methods performed poorly when the trial size was only 2% of the target population size or the target population included only 1,000 units. When the target population or the proportion of units selected into the trial was larger, some methods, such as outcome modeling using Bayesian Additive Regression Trees, performed well. We caution against generalizing using these existing approaches when the target population is much larger than the trial sample and advocate future research strives to improve methods for generalizing to large target populations.

Bayesian Additive Regression Trees: A Review and Look Forward

Bayesian additive regression trees (BART) provides a flexible approach to fitting a variety of regression models while avoiding strong parametric assumptions. The sum-of-trees model is embedded in a Bayesian inferential framework to support uncertainty quantification and provide a principled approach to regularization through prior specification. This article presents the basic approach and discusses further development of the original algorithm that supports a variety of data structures and assumptions. We describe augmentations of the prior specification to accommodate higher dimensional data and smoother functions. Recent theoretical developments provide justifications for the performance observed in simulations and other settings. Use of BART in causal inference provides an additional avenue for extensions and applications. We discuss software options as well as challenges and future directions.

BART with targeted smoothing: An analysis of patient-specific stillbirth risk

This article introduces BART with Targeted Smoothing, or tsBART, a new Bayesian tree-based model for nonparametric regression. The goal of tsBART is to introduce smoothness over a single target covariate $t$ while not necessarily requiring smoothness over other covariates $x$. tsBART is based on the Bayesian Additive Regression Trees (BART) model, an ensemble of regression trees. tsBART extends BART by parameterizing each tree’s terminal nodes with smooth functions of $t$ rather than independent scalars. Like BART, tsBART captures complex nonlinear relationships and interactions among the predictors. But unlike BART, tsBART guarantees that the response surface will be smooth in the target covariate. This improves interpretability and helps to regularize the estimate. After introducing and benchmarking the tsBART model, we apply it to our motivating example—pregnancy outcomes data from the National Center for Health Statistics. Our aim is to provide patient-specific estimates of stillbirth risk across gestational age $(t)$ and based on maternal and fetal risk factors $(x)$. Obstetricians expect stillbirth risk to vary smoothly over gestational age but not necessarily over other covariates, and tsBART has been designed precisely to reflect this structural knowledge. The results of our analysis show the clear superiority of the tsBART model for quantifying stillbirth risk, thereby providing patients and doctors with better information for managing the risk of fetal mortality. All methods described here are implemented in the R package tsbart.

Effect of Physician Gender and Race on Simulated Patients’ Ratings and Confidence in Their Physicians

Women and black physicians encounter workplace challenges because of their gender and race. It is unclear whether these individuals are assessed with lower patient satisfaction or confidence ratings compared with white male physicians. To examine whether physician gender and race affect participant ratings in scenarios in which physician competence is challenged. This randomized trial enrolled a geographically diverse sample of 3592 online respondents in the United States who were recruited from 2 crowdsourcing platforms: Amazon Mechanical Turk (n = 1741) and Lucid (n = 1851). A 2 × 2 factorial design for the gender and race of simulated physicians was conducted between March 9 and July 25, 2018. Participants were excluded before intervention if they were younger than 18 years, were pregnant, or had a history of cancer or abdominal surgical procedures. A clinical vignette was presented to the participant with a picture of the emergency department physician. Participants were randomly assigned to physicians with different gender and race, with 823 assigned to black women, 791 to black men, 828 to white women, and 835 to white men. A contradictory diagnosis from an online symptom checker introduced doubt about the clinical diagnosis. A composite outcome (range, 0-100, with 0 representing low patient confidence and satisfaction and 100 representing the maximum on the composite scale) measured participant (1) confidence in the physician, (2) satisfaction with care, (3) likelihood to recommend the physician, (4) trust in the physician's diagnosis, and (5) likelihood to request additional tests. Among 3277 adult participants, complete data were available for 3215 (median age, 49 years [range, 18-89 years]; 1667 [52%] female; 2433 [76%] white). No significant differences were observed in participant satisfaction and physician confidence for the white male physician control physicians (mean composite score, 66.13 [95% CI, 64.76-67.51]) compared with white female (mean composite score, 66.50 [95% CI, 65.19-67.82]), black female (mean composite score, 67.36 [95% CI, 66.03-68.69]), and black male (mean composite score, 66.96 [95% CI, 65.55-68.36]) physicians. Machine learning with bayesian additive regression trees revealed no evidence of treatment effect heterogeneity as a function of participants' race, gender, racial prejudice, or sexism. No significant differences were observed for simulated patients' evaluations of female or black physicians, suggesting that bias in favor of white male physicians is negligible in survey-based measures of patient satisfaction. ClinicalTrials.gov Identifier: NCT04190901.

Big Data for Finite Population Inference: Applying Quasi-Random Approaches to Naturalistic Driving Data Using Bayesian Additive Regression Trees

Abstract Big Data are a “big challenge” for finite population inference. Lack of control over data-generating processes by researchers in the absence of a known random selection mechanism may lead to biased estimates. Further, larger sample sizes increase the relative contribution of selection bias to squared or absolute error. One approach to mitigate this issue is to treat Big Data as a random sample and estimate the pseudo-inclusion probabilities through a benchmark survey with a set of relevant auxiliary variables common to the Big Data. Since the true propensity model is usually unknown, and Big Data tend to be poor in such variables that fully govern the selection mechanism, the use of flexible non-parametric models seems to be essential. Traditionally, a weighted logistic model is recommended to account for the sampling weights in the benchmark survey when estimating the propensity scores. However, handling weights is a hurdle when seeking a broader range of predictive methods. To further protect against model misspecification, we propose using an alternative pseudo-weighting approach that allows us to fit more flexible modern predictive tools such as Bayesian Additive Regression Trees (BART), which automatically detect non-linear associations as well as high-order interactions. In addition, the posterior predictive distribution generated by BART makes it easier to quantify the uncertainty due to pseudo-weighting. Our simulation findings reveal further reduction in bias by our approach compared with conventional propensity adjustment method when the true model is unknown. Finally, we apply our method to the naturalistic driving data from the Safety Pilot Model Deployment using the National Household Travel Survey as a benchmark.

Simultaneous transformation and rounding (STAR) models for integer-valued data

We propose a simple yet powerful framework for modeling integer-valued data, such as counts, scores, and rounded data. The data-generating process is defined by Simultaneously Transforming and Rounding (STAR) a continuous-valued process, which produces a flexible family of integer-valued distributions capable of modeling zero-inflation, bounded or censored data, and over- or underdispersion. The transformation is modeled as unknown for greater distributional flexibility, while the rounding operation ensures a coherent integer-valued data-generating process. An efficient MCMC algorithm is developed for posterior inference and provides a mechanism for adaptation of successful Bayesian models and algorithms for continuous data to the integer-valued data setting. Using the STAR framework, we design a new Bayesian Additive Regression Tree model for integer-valued data, which demonstrates impressive predictive distribution accuracy for both synthetic data and a large healthcare utilization dataset. For interpretable regression-based inference, we develop a STAR additive model, which offers greater flexibility and scalability than existing integer-valued models. The STAR additive model is applied to study the recent decline in Amazon river dolphins.

Heteroscedastic BART via Multiplicative Regression Trees

Bayesian additive regression trees (BART) has become increasingly popular as a flexible and scalable nonparametric regression approach for modern applied statistics problems. For the practitioner dealing with large and complex nonlinear response surfaces, its advantages include a matrix-free formulation and the lack of a requirement to prespecify a confining regression basis. Although flexible in fitting the mean, BART has been limited by its reliance on a constant variance error model. Alleviating this limitation, we propose HBART, a nonparametric heteroscedastic elaboration of BART. In BART, the mean function is modeled with a sum of trees, each of which determines an additive contribution to the mean. In HBART, the variance function is further modeled with a product of trees, each of which determines a multiplicative contribution to the variance. Like the mean model, this flexible, multidimensional variance model is entirely nonparametric with no need for the prespecification of a confining basis. Moreover, with this enhancement, HBART can provide insights into the potential relationships of the predictors with both the mean and the variance. Practical implementations of HBART with revealing new diagnostic plots are demonstrated with simulated and real data on used car prices and song year of release. Supplementary materials for this article are available online.

Spatial mapping of the provenance of storm dust: Application of data mining and ensemble modelling

Spatial modelling of storm dust provenance is essential to mitigate its on-site and off-site effects in the arid and semi-arid environments of the world. Therefore, the main aim of this study was to apply eight data mining algorithms including random forest (RF), support vector machine (SVM), bayesian additive regression trees (BART), radial basis function (RBF), extreme gradient boosting (XGBoost), regression tree analysis (RTA), Cubist model and boosted regression trees (BRT) and an ensemble modelling (EM) approach for generating spatial maps of dust provenance in the Khuzestan province, a main region with active sources for producing dust in southwestern Iran. This study is the first attempt at predicting storm dust provenance by applying individual data mining models and ensemble modelling. We identified and mapped in a geographic information system (GIS), 12 potential effective factors for dust emissions comprising two for climate (wind speed, precipitation), five soil characteristics (texture, bulk density, Ec, organic matter (OM), available water capacity (AWC)), a normalized difference vegetation index (NDVI), land use, geology, a digital elevation model (DEM) and land type, and used a mean decrease accuracy measure (MDAM) to determine the corresponding importance scores (IS). A multicollinearity test (including the variance inflation factor (VIF) and tolerance coefficient (TC)) was applied to assess relationships between the effective factors, and an existing map of dust provenance was randomly categorized into two groups consisting of training (70%) and validation (30%) data. The individual data mining models were validated using the area under the curve (AUC). Based on the TC and VIF results, no collinearity was detected among the 12 effective factors for dust emissions. The prediction accuracies of the eight data mining models and an EM assessed by the AUC were as follows: EM (with AUC = 99.8%) > XGBoost > RBF > Cubist > RF > BART > SVM > BRT > RTA (with AUC = 79.1%). Among all models, the EM was found to provide the highest accuracy for predicting storm dust provenance. Using the EM, areas classified as being low, moderate, high and very high susceptibility for storm dust provenance comprised 36, 13, 23 and 28% of the total mapped area, respectively. Based on MDAM results, the highest and lowest IS were obtained for the wind speed (IS = 23) and geology (IS = 6.5) factors, respectively. Overall, the modelling techniques used in this research are helpful for predicting storm dust provenance and thereby targeting mitigation. Therefore, we recommend applying data mining EM approaches to the spatial mapping of storm dust provenance worldwide. • Data mining algorithms and ensemble modelling used to map dust provenance spatially. • Mean decrease accuracy measure used to determine importance scores for the controls. • Ensemble model provided the highest accuracy for predicting storm dust provenance. • Techniques used are helpful for targeting mitigation.

Estimating treatment effects with machine learning

To demonstrate the performance of methodologies that include machine learning (ML) algorithms to estimate average treatment effects under the assumption of exogeneity (selection on observables). Simulated data and observational data on hospitalized adults. We assessed the performance of several ML-based estimators, including Targeted Maximum Likelihood Estimation, Bayesian Additive Regression Trees, Causal Random Forests, Double Machine Learning, and Bayesian Causal Forests, applying these methods to simulated data as well as data on the effects of right heart catheterization. In Monte Carlo studies, ML-based estimators generated estimates with smaller bias than traditional regression approaches, demonstrating substantial (69 percent-98 percent) bias reduction in some scenarios. Bayesian Causal Forests and Double Machine Learning were top performers, although all were sensitive to high dimensional (>150) sets of covariates. ML-based methods are promising methods for estimating treatment effects, allowing for the inclusion of many covariates and automating the search for nonlinearities and interactions among variables. We provide guidance and sample code for researchers interested in implementing these tools in their own empirical work.

Advanced wind speed prediction using convective weather variables through machine learning application

High precision and reliable wind speed forecasting is a challenge for meteorologists. We used multiple nonparametric tree-based machine learning techniques, for predicting the maximum wind speed at 10 m using selected convective weather variables. Analysis is based on 127 convective storms from 2005 to 2013. The study evaluated two error models - the Bayesian Additive Regression Trees (BART) and the Quantile Regression Forests (QRF) - and compares them in terms of point estimates and prediction intervals. The error model performances were evaluated based on different error metrics evaluating both the bias and random error of point estimates and the prediction intervals using ensemble verification statistics. The study showed that error modeling based on QRF is superior to BART, especially in terms of point estimate and prediction interval results. Wind speed prediction through QRF was successfully verified using systematic and random error metrics, and ensemble verification statistics of the corresponding prediction intervals. The model generated realizations of wind speed that successfully encapsulated the reference wind speed and notably reduced systematic and random error. The predicted wind speed from QRF can potentially support emergency preparedness efforts associated with severe weather impacts.