Predictors Of Missingness Research Articles

Predicting 5G throughput with BAMMO, a boosted additive model for data with missing observations

Abstract To deliver on the promise of 5G, network providers and application developers need to understand the factors impacting millimetre wave (mmWave) 5G throughput. Missing data, however, pose significant challenges for modelling throughput. Even in controlled settings, signal strength data may be only intermittently observed when a device’s connection is weak, leading to missing predictor values in model training. In addition, users may choose not to share their data once the model is deployed, meaning that key predictors may be missing when we want to predict throughput for their devices. To address these challenges, we introduce boosted additive model for data with missing observation (BAMMO), a novel additive model estimator obtained via a componentwise boosting algorithm that naturally incorporates data with missing values in model fitting. We validate BAMMO’s approach to handling missing data by comparing it with competing methods on real 5G network data with a high proportion of missing values and in simulations, finding that it delivers more accurate predictions and takes less time to compute. To identify key predictors of mmWave 5G throughput, we develop a novel extension of sparsity oriented importance learning for BAMMO, giving us a measure of variable importance based on the entire boosting solution path rather than a single selected model.

Multiple imputation with competing risk outcomes

AbstractIn time-to-event analyses, a competing risk is an event whose occurrence precludes the occurrence of the event of interest. Settings with competing risks occur frequently in clinical research. Missing data, which is a common problem in research, occurs when the value of a variable is recorded for some, but not all, records in the dataset. Multiple Imputation (MI) is a popular method to address the presence of missing data. MI uses an imputation model to generate M (M > 1) values for each variable that is missing, resulting in the creation of M complete datasets. A popular algorithm for imputing missing data is multivariate imputation using chained equations (MICE). We used a complex simulation design with covariates and missing data patterns reflective of patients hospitalized with acute myocardial infarction (AMI) to compare three strategies for imputing missing predictor variables when the analysis model is a cause-specific hazard when there were three different event types. We compared two MICE-based strategies that differed according to which cause-specific cumulative hazard functions were included in the imputation models (the three cause-specific cumulative hazard functions vs. only the cause-specific cumulative hazard function for the primary outcome) with the use of the substantive model compatible fully conditional specification (SMCFCS) algorithm. While no strategy had consistently superior performance compared to the other strategies, SMCFCS may be the preferred strategy. We illustrated the application of the strategies using a case study of patients hospitalized with AMI.

Imputation-Based Variable Selection Method for Block-Wise Missing Data When Integrating Multiple Longitudinal Studies

When integrating data from multiple sources, a common challenge is block-wise missing. Most existing methods address this issue only in cross-sectional studies. In this paper, we propose a method for variable selection when combining datasets from multiple sources in longitudinal studies. To account for block-wise missing in covariates, we impute the missing values multiple times based on combinations of samples from different missing pattern and predictors from different data sources. We then use these imputed data to construct estimating equations, and aggregate the information across subjects and sources with the generalized method of moments. We employ the smoothly clipped absolute deviation penalty in variable selection and use the extended Bayesian Information Criterion criteria for tuning parameter selection. We establish the asymptotic properties of the proposed estimator, and demonstrate the superior performance of the proposed method through numerical experiments. Furthermore, we apply the proposed method in the Alzheimer’s Disease Neuroimaging Initiative study to identify sensitive early-stage biomarkers of Alzheimer’s Disease, which is crucial for early disease detection and personalized treatment.

Dirichlet process mixture models to impute missing predictor data in counterfactual prediction models: an application to predict optimal type 2 diabetes therapy

BackgroundThe handling of missing data is a challenge for inference and regression modelling. A particular challenge is dealing with missing predictor information, particularly when trying to build and make predictions from models for use in clinical practice.MethodsWe utilise a flexible Bayesian approach for handling missing predictor information in regression models. This provides practitioners with full posterior predictive distributions for both the missing predictor information (conditional on the observed predictors) and the outcome-of-interest. We apply this approach to a previously proposed counterfactual treatment selection model for type 2 diabetes second-line therapies. Our approach combines a regression model and a Dirichlet process mixture model (DPMM), where the former defines the treatment selection model, and the latter provides a flexible way to model the joint distribution of the predictors.ResultsWe show that DPMMs can model complex relationships between predictor variables and can provide powerful means of fitting models to incomplete data (under missing-completely-at-random and missing-at-random assumptions). This framework ensures that the posterior distribution for the parameters and the conditional average treatment effect estimates automatically reflect the additional uncertainties associated with missing data due to the hierarchical model structure. We also demonstrate that in the presence of multiple missing predictors, the DPMM model can be used to explore which variable(s), if collected, could provide the most additional information about the likely outcome.ConclusionsWhen developing clinical prediction models, DPMMs offer a flexible way to model complex covariate structures and handle missing predictor information. DPMM-based counterfactual prediction models can also provide additional information to support clinical decision-making, including allowing predictions with appropriate uncertainty to be made for individuals with incomplete predictor data.

FedNI: Federated Graph Learning with Network Inpainting for Population-Based Disease Prediction.

Graph Convolutional Neural Networks (GCNs) are widely used for graph analysis. Specifically, in medical applications, GCNs can be used for disease prediction on a population graph, where graph nodes represent individuals and edges represent individual similarities. However, GCNs rely on a vast amount of data, which is challenging to collect for a single medical institution. In addition, a critical challenge that most medical institutions continue to face is addressing disease prediction in isolation with incomplete data information. To address these issues, Federated Learning (FL) allows isolated local institutions to collaboratively train a global model without data sharing. In this work, we propose a framework, FedNI, to leverage network inpainting and inter-institutional data via FL. Specifically, we first federatively train missing node and edge predictor using a graph generative adversarial network (GAN) to complete the missing information of local networks. Then we train a global GCN node classifier across institutions using a federated graph learning platform. The novel design enables us to build more accurate machine learning models by leveraging federated learning and also graph learning approaches. We demonstrate that our federated model outperforms local and baseline FL methods with significant margins on two public neuroimaging datasets.

Developing prediction models when there are systematically missing predictors in individual patient data meta-analysis.

Clinical prediction models are widely used in modern clinical practice. Such models are often developed using individual patient data (IPD) from a single study, but often there are IPD available from multiple studies. This allows using meta-analytical methods for developing prediction models, increasing power and precision. Different studies, however, often measure different sets of predictors, which may result to systematically missing predictors, that is, when not all studies collect all predictors of interest. This situation poses challenges in model development. We hereby describe various approaches that can be used to develop prediction models for continuous outcomes in such situations. We compare four approaches: a "restrict predictors" approach, where the model is developed using only predictors measured in all studies; a multiple imputation approach that ignores study-level clustering; a multiple imputation approach that accounts for study-level clustering; and a new approach that develops a prediction model in each study separately using all predictors reported, and then synthesizes all predictions in a multi-study ensemble. We explore in simulations the performance of all approaches under various scenarios. We find that imputation methods and our new method outperform the restrict predictors approach. In several scenarios, our method outperformed imputation methods, especially for few studies, when predictor effects were small, and in case of large heterogeneity. We use a real dataset of 12 trials in psychotherapies for depression to illustrate all methods in practice, and we provide code in R.

Machine-learning Models Predict 30-Day Mortality, Cardiovascular Complications, and Respiratory Complications After Aseptic Revision Total Joint Arthroplasty.

Machine-learning Models Predict 30-Day Mortality, Cardiovascular Complications, and Respiratory Complications After Aseptic Revision Total Joint Arthroplasty.

Bayesian negative binomial regression model with unobserved covariates for predicting the frequency of north atlantic tropical storms

Predicting the annual frequency of tropical storms is of interest because it can provide basic information towards improved preparation against these storms. Sea surface temperatures (SSTs) averaged over the hurricane season can predict annual tropical cyclone activity well. But predictions need to be made before the hurricane season when the predictors are not yet observed. Several climate models issue forecasts of the SSTs, which can be used instead. Such models use the forecasts of SSTs as surrogates for the true SSTs. We develop a Bayesian negative binomial regression model, which makes a distinction between the true SSTs and their forecasts, both of which are included in the model. For prediction, the true SSTs may be regarded as unobserved predictors and sampled from their posterior predictive distribution. We also have a small fraction of missing data for the SST forecasts from the climate models. Thus, we propose a model that can simultaneously handle missing predictors and variable selection uncertainty. If the main goal is prediction, an interesting question is should we include predictors in the model that are missing at the time of prediction? We attempt to answer this question and demonstrate that our model can provide gains in prediction.

Analysis of local sensitivity to nonignorability with missing outcomes and predictors.

The ISNI (index of sensitivity to local nonignorability) method quantifies local sensitivity of parametric inferences to nonignorable missingness in an outcome variable. Here we extend ISNI to the situations where both outcomes and predictors can be missing and where the missingness mechanism can be either parametric or semi-parametric. We define the quantity MinNI (minimum nonignorability) to be an approximation to the norm of the smallest value of the transformed nonignorability that gives a nonnegligible displacement of the estimate of the parameter of interest. We illustrate our method in a complete data set from which we synthetically delete observations according to various patterns. We then apply the method to real-data examples involving the normal linear model and conditional logisticregression.

Prognostic factors and prediction models for acute aortic dissection: a systematic review

ObjectiveOur study aimed to systematically review the methodological characteristics of studies that identified prognostic factors or developed or validated models for predicting mortalities among patients with acute aortic dissection (AAD),...

Real-time imputation of missing predictor values improved the application of prediction models in daily practice

ObjectivesIn clinical practice, many prediction models cannot be used when predictor values are missing. We, therefore, propose and evaluate methods for real-time imputation. Study Design and SettingWe describe (i) mean imputation (where missing values are replaced by the sample mean), (ii) joint modeling imputation (JMI, where we use a multivariate normal approximation to generate patient-specific imputations), and (iii) conditional modeling imputation (CMI, where a multivariable imputation model is derived for each predictor from a population). We compared these methods in a case study evaluating the root mean squared error (RMSE) and coverage of the 95% confidence intervals (i.e., the proportion of confidence intervals that contain the true predictor value) of imputed predictor values. Results–RMSE was lowest when adopting JMI or CMI, although imputation of individual predictors did not always lead to substantial improvements as compared to mean imputation. JMI and CMI appeared particularly useful when the values of multiple predictors of the model were missing. Coverage reached the nominal level (i.e., 95%) for both CMI and JMI. ConclusionMultiple imputations using either CMI or JMI is recommended when dealing with missing predictor values in real-time settings.

A Hybrid Approach for the Prediction of Relative Permeability Using Machine Learning of Experimental and Numerical Proxy SCAL Data

Summary Accurate estimation of relative permeability is one of the key parameters for decision making in upstream applications from project appraisal to field development and evaluation of various field development options. In this study, we identify Euler number (Arns et al. 2001) (a quantitative measure of fluid connectivity/distribution) and saturation as being the first-order predictors of relative permeability and develop a reliable correlation between them using machine learning of experimental special core analysis (SCAL) data and pore network simulation results. In order to achieve our objective, first, we developed a machine-learning model based on the random forest algorithm (Breiman 2001) to analyze specific SCAL data that indicates a key missing feature in the traditional saturation-based relative permeability prediction. We identified this missing feature and proposed the Euler characteristic as a potential first-order predictor of relative permeability in combination with in-situ fluid saturations. We generated “artificial” relative permeability data using pore network simulation (Valvatne and Blunt 2004) by systematically varying a set of key parameters such as pore geometry, wettability, and saturation history. Subsequently, we used machine learning to rank the importance of each parameter and identify possible correlative responses to those selected variables. At a fixed saturation (zero-dimensional volumetric abundance) and Euler number coordinates, the relative permeability is very consistent and varies insignificantly across different cases, suggesting these two parameters as first-order predictors. Euler number characterizes the fluid connectivity/distribution, while saturation represents the net volumetric fluid quantity. We believe that Euler number could be the missing first-order predictor in traditional saturation-based predictive relative permeability models, especially for connected pathway dominated flow regime. Finally, we identified the quantitative relationship between relative permeability and Euler characteristic, and present a reliable correlation to determine the relative permeability on the basis of Euler number and saturation.

Prediction and Inference With Missing Data in Patient Alert Systems

We describe the Bedside Patient Rescue (BPR) project, the goal of which is risk prediction of adverse events for non-intensive care unit patients using ∼100 variables (vitals, lab results, assessments, etc.). There are several missing predictor values for most patients, which in the health sciences is the norm, rather than the exception. A Bayesian approach is presented that addresses many of the shortcomings to standard approaches to missing predictors: (i) treatment of the uncertainty due to imputation is straight-forward in the Bayesian paradigm, (ii) the predictor distribution is flexibly modeled as an infinite normal mixture with latent variables to explicitly account for discrete predictors (i.e., as in multivariate probit regression models), and (iii) certain missing not at random situations can be handled effectively by allowing the indicator of missingness into the predictor distribution only to inform the distribution of the missing variables. The proposed approach also has the benefit of providing a distribution for the prediction, including the uncertainty inherent in the imputation. Therefore, we can ask questions such as: is it possible this individual is at high risk but we are missing too much information to know for sure? How much would we reduce the uncertainty in our risk prediction by obtaining a particular missing value? This approach is applied to the BPR problem resulting in excellent predictive capability to identify deteriorating patients. Supplementary materials for this article, including a standardized description of the materials available for reproducing the work, are available as an online supplement.

Risks and predictive biomarkers of sudden unexpected death in epilepsy patient.

The current review updates our knowledge regarding sudden unexpected death in epilepsy patient (SUDEP) risks, risk factors, and investigations of putative biomarkers based on suspected mechanisms of SUDEP. The overall incidence of SUDEP in adults with epilepsy is 1.2/1000 patient-years, with surprisingly comparable figures in children in recently published population-based studies. This risk was found to decrease over time in several cohorts at a rate of -7% per year, for unknown reasons. Well established risk factors include frequency of generalized tonic-clonic seizures, while adding antiepileptic treatment, nocturnal supervision and use of nocturnal listening device appear to be protective. In contrast, recent data failed to demonstrate the predictive value of heart rate variability, periictal cardiorespiratory dysfunction, and postictal generalized electroencephalography suppression. Preliminary findings suggest that brainstem and thalamic atrophy may be associated with a higher risk of SUDEP. Novel experimental and human data support the primary role of generalized tonic-clonic seizure-triggered respiratory dysfunction and the likely contribution of altered brainstem serotoninergic neurotransmission, in SUDEP pathophysiology. Although significant progress has been made during the past year in the understanding of SUDEP mechanisms and investigation of numerous potential biomarkers, we are still missing reliable predictors of SUDEP beyond the well established clinical risk factors.

Duration of Electrocardiographic Monitoring of Emergency Department Patients With Syncope.

The optimal duration of cardiac rhythm monitoring after emergency department (ED) presentation for syncope is poorly described. We sought to describe the incidence and time to arrhythmia occurrence to inform decisions regarding duration of monitoring based on ED risk stratification. We conducted a prospective cohort study with enrolled adult patients (≥16 years old) presenting within 24 hours of syncope at 6 EDs. We collected baseline characteristics, time of syncope and ED arrival, and the Canadian Syncope Risk Score (CSRS) risk category. We followed subjects for 30 days, and our adjudicated primary outcome was serious arrhythmic conditions (arrhythmias, interventions for arrhythmias, and unexplained death). After excluding patients with an obvious serious condition on ED presentation and those with missing CSRS predictors, we used Kaplan-Meier analysis to describe the time to serious arrhythmic outcomes. A total of 5581 patients (mean age, 53.4 years; 54.5% females; 11.6% hospitalized) were available for analysis, including 346 (6.2%) for whom the 30-day follow-up was incomplete and who were censored at the last follow-up time. A total of 417 patients (7.5%) experienced serious outcomes, 207 of which (3.7%; 95% CI, 3.3%-4.2%) were arrhythmic (161 arrhythmias, 30 cardiac device implantations, 16 unexplained deaths). Overall, 4123 (73.9%) were classified as CSRS low risk, 1062 (19.0%) medium risk, and 396 (7.1%) high risk. The CSRS accurately stratified subjects as low risk (0.4% risk for 30-day arrhythmic outcome), medium risk (8.7% risk), and high risk (25.3% risk). One-half of arrhythmic outcomes were identified within 2 hours of ED arrival in low-risk patients and within 6 hours in medium- and high-risk patients, and the residual risk after these cut points were 0.2% for low-risk, 5.0% for medium-risk, and 18.1% for high-risk patients. Overall, 91.7% of arrhythmic outcomes among medium- and high-risk patients, including all ventricular arrhythmias, were identified within 15 days. None of the low-risk patients experienced ventricular arrhythmia or unexplained death, whereas 0.9% of medium-risk patients and 6.3% of high-risk patients experienced them ( P<0.0001). Serious underlying arrhythmia was often identified within the first 2 hours of ED arrival for CSRS low-risk patients and within 6 hours for CSRS medium- and high-risk patients. Outpatient cardiac rhythm monitoring for 15 days for selected medium-risk patients and all high-risk patients discharged from the hospital should also be considered.

A comparison of joint species distribution models for presence–absence data

Abstract Joint species distribution models (JSDMs) account for biotic interactions and missing environmental predictors in correlative species distribution models. Several different JSDMs have been proposed in the literature, but the use of different or conflicting nomenclature and statistical notation potentially obscures similarities and differences among them. Furthermore, new JSDM implementations have been illustrated with different case studies, preventing direct comparisons of computational and statistical performance. We aim to resolve these outstanding issues by (a) highlighting similarities among seven presence–absence JSDMs using a clearly defined, singular notation; and (b) evaluating the computational and statistical performance of each JSDM using six datasets that vary widely in numbers of sites, species, and environmental covariates considered. Our singular notation shows that many of the JSDMs are very similar, and in turn parameter estimates of different JSDMs are moderate to strongly, positively correlated. In contrast, the different JSDMs clearly differ in computational efficiency and memory limitations. Our framework will allow ecologists to make educated decisions about the JSDM that best suits their objective, and enable wider uptake of JSDM methods among the ecological community.

Behavioral predictors of improved speech output in minimally verbal children with autism.

We investigated the relationship between eight theoretically motivated behavioral variables and a spoken-language-related outcome measure, after 25 sessions of treatment for speech production in 38 minimally verbal children with autism. After removing potential predictors that were uncorrelated with the outcome variable, two remained. We used both complete-case and multiple-imputation analyses to address missing predictor data and performed linear regressions to identify significant predictors of change in percent syllables approximately correct after treatment. Baseline phonetic inventory (the number of English phonemes repeated correctly) was the most robust predictor of improvement. In the group of 17 participants with complete data, ADOS score also significantly predicted the outcome. In contrast to some earlier studies, nonverbal IQ, baseline levels of expressive language, and younger age did not significantly predict improvement. The present results are not only consistent with previous studies showing that verbal imitation and autism severity significantly predict spoken language outcomes in preschool-aged minimally verbal children with autism, but also extend these findings to older minimally verbal children with autism. Autism Res 2018, 11: 1356-1365. © 2018 International Society for Autism Research, Wiley Periodicals, Inc. LAY SUMMARY: We wished to understand what baseline factors predicted whether minimally verbal children with autism would improve after treatment for spoken language. The outcome measure was change in percentage (%) syllables approximately correct on a set of 30 two-syllable words or phrases. Fifteen were both practiced in treatment and tested; the remainder were not practiced in treatment, but only tested, to assess how well children were able to generalize their new skills to an untrained set of words. Potential predictors tested were sex, age, expressive language, phonetic inventory (the number of English speech sounds repeated correctly), autism severity, and nonverbal IQ. Phonetic inventory and (for some children) autism severity predicted children's posttreatment improvement. Nonverbal IQ and expressive language ability did not predict improvement, nor did younger age, suggesting that some older children with autism may be candidates for speech therapy.

On sliced inverse regression with missing values

ABSTRACTTo deal with predictors missing at random in sufficient dimension reduction, IPW-SIR (Li and Lu (2008), ‘Sufficient dimension reduction with missing predictors’, Journal of American Statistical Association, 103, 882–831) combines inverse probability weighting and sliced inverse regression (Li (1991), ‘Sliced inverse regression for dimension reduction’ (with discussion). Journal of the American Statistical Association, 86, 316–342). IPW-SIR is extended to handle response missing at random in this paper. The -consistency and asymptotic normality of the proposed estimator are established. Furthermore, IPW-SIR is adapted to deal with the challenging case when both the response and the predictors are missing at the same time. The superior performances of the proposed estimators over existing methods are demonstrated through simulation studies as well as analysis of a HIV clinical trial data.

Graphical diagnostics for occupancy models with imperfect detection

Summary Occupancy‐detection models that account for imperfect detection have become widely used in many areas of ecology. As with any modelling exercise, it is important to assess whether the fitted model encapsulates the main sources of variation in the data, yet there have been few methods developed for occupancy‐detection models that would allow practitioners to do so. In this paper, a new type of residual for occupancy‐detection models is developed according to the method of Dunn &amp; Smyth (Journal of Computational and Graphical Statistics, 5, 1996, 236–244). Residuals are separately constructed to diagnose the occupancy and detection components of the model. Because the residuals are quite noisy, we suggest fitting a smoother through plots of residuals against predictors of fitted values, with 95% confidence bands, to diagnose lack‐of‐fit. The method is illustrated using Swiss squirrel data, and evaluated using simulations based on that dataset. Plotting residuals against predictors or against fitted values performed reasonably well as methods for diagnosing violations of occupancy‐detection model assumptions, particularly plots of residuals against a missing predictor. Relatively high false positive rates were sometimes observed, but this seems to be controlled reasonably well by fitting smoothers to these plots and being guided in interpretation by 95% confidence bands around the smoothers.

Methods for Handling Missing Variables in Risk Prediction Models.

Prediction models should be externally validated before being used in clinical practice. Many published prediction models have never been validated. Uncollected predictor variables in otherwise suitable validation cohorts are the main factor precluding external validation. We used individual patient data from 9 different cohort studies conducted in the United States, Europe, and Latin America that included 7,892 patients with chronic obstructive pulmonary disease who enrolled between 1981 and 2006. Data on 3-year mortality and the predictors of age, dyspnea, and airflow obstruction were available. We simulated missing data by omitting the predictor dyspnea cohort-wide, and we present 6 methods for handling the missing variable. We assessed model performance with regard to discriminative ability and calibration and by using 2 vignette scenarios. We showed that the use of any imputation method outperforms the omission of the cohort from the validation, which is a commonly used approach. Compared with using the full data set without the missing variable (benchmark), multiple imputation with fixed or random intercepts for cohorts was the best approach to impute the systematically missing predictor. Findings of this study may facilitate the use of cohort studies that do not include all predictors and pave the way for more widespread external validation of prediction models even if 1 or more predictors of the model are systematically missing.