Model-based Imputation Research Articles

Two-stage strategy using denoising autoencoders for robust reference-free genotype imputation with missing input genotypes.

Widely used genotype imputation methods are based on the Li and Stephens model, which assumes that new haplotypes can be represented by modifying existing haplotypes in a reference panel through mutations and recombinations. These methods use genotypes from SNP arrays as inputs to estimate haplotypes that align with the input genotypes by analyzing recombination patterns within a reference panel, and then infer unobserved variants. While these methods require reference panels in an identifiable form, their public use is limited due to privacy and consent concerns. One strategy to overcome these limitations is to use de-identified haplotype information, such as summary statistics or model parameters. Advances in deep learning (DL) offer the potential to develop imputation methods that use haplotype information in a reference-free manner by handling it as model parameters, while maintaining comparable imputation accuracy to methods based on the Li and Stephens model. Here, we provide a brief introduction to DL-based reference-free genotype imputation methods, including RNN-IMP, developed by our research group. We then evaluate the performance of RNN-IMP against widely-used Li and Stephens model-based imputation methods in terms of accuracy (R2), using the 1000 Genomes Project Phase 3 dataset and corresponding simulated Omni2.5 SNP genotype data. Although RNN-IMP is sensitive to missing values in input genotypes, we propose a two-stage imputation strategy: missing genotypes are first imputed using denoising autoencoders; RNN-IMP then processes these imputed genotypes. This approach restores the imputation accuracy that is degraded by missing values, enhancing the practical use of RNN-IMP.

Filling data gaps in long-term solar UV monitoring by statistical imputation methods

Knowledge of long-term time trends of solar ultraviolet (UV) radiation on ground level is of high scientific interest. For this purpose, precise measurements over a long time are necessary. One of the challenges solar UV monitoring faces is the permanent and gap-free data collection over several decades. Data gaps hamper the formation and comparison of monthly or annual means, and, in the worst case, lead to incorrect conclusions in further data evaluation and trend analysis of UV data. For estimating data to fill gaps in long-term UV data series (daily radiant exposure and highest daily irradiance), we developed three statistical imputation methods: a model-based imputation, considering actual local solar radiation conditions using predictors correlated to the local UV values in an empirical model; an average-based imputation based on a statistical approach of averaging available local UV measurement data without predictors; and a mixture of these two imputation methods. A detailed validation demonstrates the superiority of the model-based imputation method. The combined method can be considered the best one in practice. Furthermore, it has been shown that the model-based imputation method can be used as an useful tool to identify systematic errors at and between calibration steps in long-term erythemal UV data series.

In-Database Data Imputation

Missing data is a widespread problem in many domains, creating challenges in data analysis and decision making. Traditional techniques for dealing with missing data, such as excluding incomplete records or imputing simple estimates (e.g., mean), are computationally efficient but may introduce bias and disrupt variable relationships, leading to inaccurate analyses. Model-based imputation techniques offer a more robust solution that preserves the variability and relationships in the data, but they demand significantly more computation time, limiting their applicability to small datasets. This work enables efficient, high-quality, and scalable data imputation within a database system using the widely used MICE method. We adapt this method to exploit computation sharing and a ring abstraction for faster model training. To impute both continuous and categorical values, we develop techniques for in-database learning of stochastic linear regression and Gaussian discriminant analysis models. Our MICE implementations in PostgreSQL and DuckDB outperform alternative MICE implementations and model-based imputation techniques by up to two orders of magnitude in terms of computation time, while maintaining high imputation quality.

Abstract TMP95: Context Matters: Geographic Decision-Making Could Improve Prehospital Stroke Triage

Introduction: Emergency medical services (EMS) play a key role in stroke outcomes, as triage decisions made under diagnostic uncertainty impact availability of time sensitive treatments. Quantifying both time varying effects and uncertainty allows for optimized algorithms to improve destination decisions for stroke patients. We developed a Bayesian model of neurological outcomes and an algorithm for hospital triage based on clinical trials data. Methods: We used data from Virtual International Stroke Trials Archive (VISTA), RACECAT, and FAST-MAG. To model prehospital diagnosis, we included patients with any stroke-like illness (ischemic/hemorrhagic stroke, and mimics). Primary outcome was 90-day modified Rankin scale (mRS). Due to variability in reported measures from each trial, model-based Bayesian imputation was used. Our model accepts time-from symptoms, location, and capability of all local hospitals as inputs. Results: We included 9048 participants with 8600 (95%) ischemic strokes, 1801 (20%) had large vessel occlusion, and 4064 (45%) had a 90-day mRS of 0-2 (good neurologic outcome). We evaluated the impact of our algorithm using 1,320 stroke events from RACECAT. The estimated median absolute difference in chance of good mRS score between bypass and non local hospital bypass was 2.3% (ranging from 0 to 7.4%) among 1,320 stroke events in the RACECAT data set; covariate-adjusted patient scenarios predicted a preference for local hospital bypass of -2.9% to 7.4%. Destination maps for Catalonia show important patterns of bypass optimality. Conclusion: We found that no uniform destination selection approach is likely to be optimal under all circumstances. This finding is important, because it suggests that a personalized approach considering diagnostic uncertainty, patient characteristics, and geospatial factors could outperform simpler approaches. Predicted time-to-treatment targets may be insufficient for appropriate system optimization.

Unlocking the Potential of Autism Detection: Integrating Traditional Feature Selection and Machine Learning Techniques

The diagnostic process for Autism Spectrum Disorder (ASD) typically involves time-consuming assessments conducted by specialized physicians. To improve the efficiency of ASD screening, intelligent solutions based on machine learning have been proposed in the literature. However, many existing ML models lack the incorporation of medical tests and demographic features, which could potentially enhance their detection capabilities by considering affected features through traditional feature selection approaches. This study aims to address the aforementioned limitation by utilizing a real dataset containing 45 features and 983 patients. To achieve this goal, a two-phase methodology is employed. The first phase involves data preparation, including handling missing data through model-based imputation, normalizing the dataset using the Min-Max method, and selecting relevant features using traditional feature selection approaches based on affected features. In the second phase, seven ML classification techniques recommended by the literature, including Decision Trees (DT), Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), AdaBoost, Gradient Boosting (GB), and Neural Network (NN), are utilized to develop ML models. These models are then trained and tested on the prepared dataset to evaluate their performance in detecting ASD. The performance of the ML models is assessed using various metrics, such as Accuracy, Recall, Precision, F1-score, AUC, Train time, and Test time. These metrics provide insights into the models' overall accuracy, sensitivity, specificity, and the trade-off between true positive and false positive rates. The results of the study highlight the effectiveness of utilizing traditional feature selection approaches based on affected features. Specifically, the GB model outperforms the other models with an accuracy of 87%, Recall of 87%, Precision of 86%, F1-score of 86%, AUC of 95%, Train time of 21.890, and Test time of 0.173. Additionally, a benchmarking analysis against five other studies reveals that the proposed methodology achieves a perfect score across three key areas. By considering affected features through traditional feature selection approaches, the developed ML models demonstrate improved performance and have the potential to enhance ASD screening and diagnosis processes.

Risk Prediction and Machine Learning: A Case-Based Overview.

Risk Prediction and Machine Learning: A Case-Based Overview.

A Systematic Literature Review On Missing Values: Research Trends, Datasets, Methods and Frameworks

Handling of missing values in data analysis is the focus of attention in various research fields. Imputation is one method that is commonly used to overcome this problem of missing data. This systematic literature review research aims to present a comprehensive summary of the relevant scientific literature that describes the use of the imputation method in overcoming missing values. The literature search method is carried out using various academic databases and reliable sources of information. Relevant keywords are used to find articles that match the research question. After selection and evaluation, 40 relevant articles were included in this study. The findings of this study reveal a variety of imputation approaches and methods used in various research fields, such as social sciences, medicine, economics, and others. Commonly used imputation methods include single imputation, multivariate imputation, and model-based imputation methods. In addition, several studies also describe a combination of imputation methods to deal with more complex situations. The advantage of the imputation method is that it allows researchers to maintain sample sizes and minimize bias in data analysis. However, the research results also show that the imputation method must be applied with caution, because inappropriate imputation decisions can lead to biased results and can affect the accuracy of the research conclusions. In order to increase the validity and reliability of research results, researchers are expected to transparently report the imputation method used and describe the considerations made in the imputation decision-making process. This systematic review of the literature review provides an in-depth view of the use of the imputation method in handling missing values. In the face of the challenge of missing data, an understanding of the various imputation methods and the context in which they are applied will be key to generating meaningful findings in various research fields.

Implementing Multiple Imputation for Missing Data in Longitudinal Studies When Models are Not Feasible: An Example Using the Random Hot Deck Approach.

Researchers often use model-based multiple imputation to handle missing at random data to minimize bias. However, constraints within the data may sometimes result in implausible values, making model-based imputation infeasible. In these contexts, we illustrate how random hot deck imputation can allow for plausible multiple imputation in longitudinal studies. Our motivating example is the Childhood Health, Activity, and Motor Performance School Study Denmark (CHAMPS-DK), a prospective cohort study that measured weekly sports participation for 1700 Danish schoolchildren. Using observed data on 4 variables (pain, activity frequency, sport, sport counts), we created a gold-standard data set without missing data. We then created a synthetic data set by setting some variable values to missing based on a prediction model that mimicked real-data missingness patterns. To create 5 imputed data sets, we matched each record with missing data to several fully observed records, generated probabilities from matched records, and sampled from these records based on the probability of each occurring. We assessed variability and agreement (kappa) between the imputed data sets and the gold-standard data set. We compare results to common model-based imputation methods. Variability across data sets appeared reasonable. The range of kappa for the random hot deck approach was moderate for activity frequency (0.65 to 0.71) and sport (0.59 to 0.85), and poor for common model-based approaches (range 0.00 to 0.11). The range of kappas for sport count was strong (0.87 to 0.97) for random hot deck imputation and weak to moderate (0.55 to 0.71) for common model-based imputation. Agreement was higher when more information was present, and when prevalence was higher for our binary variable sport. Random hot deck imputation should be considered as an alternative method when model-based approaches are infeasible, specifically where there are constraints within and between covariates.

Data imputation and comparison of custom ensemble models with existing libraries like XGBoost, CATBoost, AdaBoost and Scikit learn for predictive equipment failure

This paper presents comparison of a custom ensemble models with the models trained using existing libraries like XGBoost, CATBoost, AdaBoost and Scikit learn, for predictive equipment failure for the case of oil extracting equipment setup. The dataset used in this paper, contains many missing values and the paper proposes different model-based data imputation strategies to impute the missing values. Various machine-learning based pre-processing techniques are discussed to tackle missing values in the data efficiently without harming the data. Custom ensemble models that are proposed in the paper are a combination of a base estimator and a meta classifier. The base estimator is an ensemble model which is trained on the pre-processed data and the outputs derived from the base estimator are used as inputs for the meta classifier. The meta classifier helps in better generalization of data as it helps in pruning the base estimator. The architecture and the training and testing process of the custom ensemble models are explained in detail.

ScMTD: a statistical multidimensional imputation method for single-cell RNA-seq data leveraging transcriptome dynamic information

BackgroundSingle-cell RNA sequencing (scRNA-seq) provides a powerful tool to capture transcriptomes at single-cell resolution. However, dropout events distort the gene expression levels and underlying biological signals, misleading the downstream analysis of scRNA-seq data.ResultsWe develop a statistical model-based multidimensional imputation algorithm, scMTD, that identifies local cell neighbors and specific gene co-expression networks based on the pseudo-time of cells, leveraging information on cell-level, gene-level, and transcriptome dynamic to recover scRNA-seq data. Compared with the state-of-the-art imputation methods through several real-data-based analytical experiments, scMTD effectively recovers biological signals of transcriptomes and consistently outperforms the other algorithms in improving FISH validation, trajectory inference, differential expression analysis, clustering analysis, and identification of cell types.ConclusionsscMTD maintains the gene expression characteristics, enhances the clustering of cell subpopulations, assists the study of gene expression dynamics, contributes to the discovery of rare cell types, and applies to both UMI-based and non-UMI-based data. Overall, scMTD’s reliability, applicability, and scalability make it a promising imputation approach for scRNA-seq data.

Multiple imputation of ordinal missing not at random data

We introduce a selection model-based imputation approach to be used within the Fully Conditional Specification (FCS) framework for the Multiple Imputation (MI) of incomplete ordinal variables that are supposed to be Missing Not at Random (MNAR). Thereby, we generalise previous work on this topic which involved binary single-level and multilevel data to ordinal variables. We apply an ordered probit model with sample selection as base of our imputation algorithm. The applied model involves two equations that are modelled jointly where the first one describes the missing-data mechanism and the second one specifies the variable to be imputed. In addition, we develop a version for hierarchical data by incorporating random intercept terms in both equations. To fit this multilevel imputation model we use quadrature techniques. Two simulation studies validate the overall good performance of our single-level and multilevel imputation methods. In addition, we show its applicability to empirical data by applying it to a common research topic in educational science using data of the National Educational Panel Study (NEPS) and conducting a short sensitivity analysis. Our approach is designed to be used within the R software package mice which makes it easy to access and apply.

An Adaptive Classification Model for Predicting Epileptic Seizures Using Cloud Computing Service Architecture

Data science techniques have increasing importance in medical data analysis, including detecting and predicting the probability of contracting a disease. A large amount of medical data is generated close to the patients in the form of a stream, such as data from sensors and medical devices. The distribution of these kinds of data may change from time to time; adaptive Machine Learning (ML) consists of a continuous training process responding to the distribution’s change. Adaptive ML models require high computational resources, which can be provided by cloud computing. In this work, a classification model is proposed to utilize the advantages of cloud computing, edge computing, and adaptive ML. It aims to precisely and efficiently classify EEG signal data, thereby detecting the seizures of epileptic patients using Adaptive Random Forest (ARF). It includes a global adaptive classifier in the cloud master node and a local light classifier in each edge node. In this model, the delayed labels consider missing values, and the Model-based imputation method is used to handle them in the global classifier. Implementing the proposed model on a real huge dataset (CHB-MIT) showed an accurate performance. It has a 0.998 True Negative Rate, a 0.785 True Positive Rate, and a 0.0017 False Positive Rate, which overcomes much of the research in the state-of-the-art.

Combining farm and household surveys with modelling approaches to improve post-harvest loss estimates and reduce data collection costs

While there is growing awareness of the issue of food losses at the political level, official post-harvest loss data for informing policymaking and reporting on SDG Indicator 12.3.1. (a) Food Loss Index is scarce. Representative sample-based surveys are necessary to obtain information on on-farm losses at the country level, but due to the issue’s complexity, a loss module covering several key questions is needed. One main strategy proposed by the 50x2030 Initiative for optimizing data collection is sub-sampling for some of the survey modules. This paper examines whether modelling approaches can be combined with sub-sampling to improve harvest and post-harvest loss estimates and allow for further sample and cost reduction. The paper first presents the loss models generated on four selected surveys conducted in Malawi, Zimbabwe, and Nigeria, which were built using the Classification and Regression Tree (CART) method. The performance of each model is assessed for different sizes of sub-samples to improve the sample-based estimates, either by model-based estimates or by model-based imputation. The research concludes that the model-based estimates improve the loss estimates of the sub-samples due to post-stratification implied in the CART method, whereby they can constitute a cost-effective complement to sub-sampling strategies, while model-based imputations should only be used on a reduced number of missing observations. The models perform best when the survey invests in obtaining more detailed on-farm loss data and considers some key variables identified as relevant for on-farm loss models. Sub-sampling allows for investment in more detailed questionnaires and some considerations are derived for its design.

Principal Canonical Correlation Analysis with Missing Data in Small Samples

Missing data occur in various fields, such as clinical trials and social science. Canonical correlation analysis often used to analyze the correlation between two random vectors, cannot be performed on a dataset with missing data. Canonical correlation coefficients (CCCs) can also be calculated from a covariance matrix. When the covariance matrix can be estimated by excluding (complete-case and available-case analyses) or imputing (multivariate imputation by chained equations, k-nearest neighbor (kNN), and iterative robust model-based imputation) missing data, CCCs are estimated from this covariance matrix. CCCs have bias even with all observation data. Usually, estimated CCCs are even larger than population CCCs when a covariance matrix estimated from a dataset with missing data is used. The purpose is to bring the CCCs estimated from the dataset with missing data close to the population CCCs. The procedure involves three steps. First, principal component analysis is performed on the covariance matrix from the dataset with missing data to obtain the eigenvectors. Second, the covariance matrix is transformed using first to fourth eigenvectors. Finally, the CCCs are calculated from the transformed covariance matrix. CCCs derived using with this procedure are called the principal CCCs (PCCCs), and simulation studies and numerical examples confirmed the effectiveness of the PCCCs estimated from the dataset with missing data. There were many cases in the simulation results where the bias and root-mean-squared error of the PCCC estimated from the missing data based on kNN were the smallest. In the numerical example results, the first PCCC estimated from the missing data based on kNN is close to the first CCC estimated from the dataset comprising all observation data when the correlation between two vectors is low. Therefore, PCCCs based on kNN were recommended.

Compatibility in imputation specification.

Missing data such as data missing at random (MAR) are unavoidable in real data and have the potential to undermine the validity of research results. Multiple imputation is one of the most widely used MAR-based methods in education and behavioral science applications. Arbitrarily specifying imputation models can lead to incompatibility and cause biased estimation. Building on the recent developments of model-based imputation and Arnold's compatibility work, this paper systematically summarizes when the traditional fully conditional specification (FCS) is applicable and how to specify a model-based imputation model if needed. We summarize two Compatibility Requirements to help researchers check compatibility more easily and a decision tree to check whether the traditional FCS is applicable in a given scenario. Additionally, we present a clear overview of two types of model-based imputation: the sequential and separate specifications. We illustrate how to specify model-based imputation with examples. Additionally, we provide example code of a free software program, Blimp, for implementing model-based imputation.

Robust estimation of traffic density with missing data using an adaptive-R extended Kalman filter

Traffic density is a crucial indicator of traffic congestion, but measuring it directly is often infeasible and hence, it is usually estimated based on other measurements. However, a challenge in measuring traffic parameters is the high probability of sensor failure, which results in missing measurement or missing data. To overcome this difficulty, in this paper, we propose a novel adaptive-R extended Kalman filter (AREKF) combined with a model-based data imputation technique to estimate traffic density. We show analytically that the AREKF is able to accurately estimate the density even when the noise covariance matrices are not accurately known. Microscopic traffic simulations demonstrated the efficacy of the AREKF, where the estimated density is fed into a real-time ramp metering control algorithm to control vehicle flow on a merging road, which is highly susceptible to traffic congestion. The results show that the proposed AREKF with data imputation is able to accurately estimate the traffic density even when data is missing, and the ramp-metering controller significantly improves the traffic flow and thus, alleviates congestion.

Attribute Sentiment Scoring with Online Text Reviews: Accounting for Language Structure and Missing Attributes

The authors address two significant challenges in using online text reviews to obtain fine-grained, attribute-level sentiment ratings. First, in contrast to methods that rely on word frequency, they develop a deep learning convolutional–long short-term memory hybrid model to account for language structure. The convolutional layer accounts for spatial structure (adjacent word groups or phrases), and long short-term memory accounts for the sequential structure of language (sentiment distributed and modified across nonadjacent phrases). Second, they address the problem of missing attributes in text when constructing attribute sentiment scores, as reviewers write about only a subset of attributes and remain silent on others. They develop a model-based imputation strategy using a structural model of heterogeneous rating behavior. Using Yelp restaurant review data, they show superior attribute sentiment scoring accuracy with their model. They identify three reviewer segments with different motivations: status seeking, altruism/want voice, and need to vent/praise. Surprisingly, attribute mentions in reviews are driven by the need to inform and vent/praise rather than by attribute importance. The heterogeneous model-based imputation performs better than other common imputations and, importantly, leads to managerially significant corrections in restaurant attribute ratings. More broadly, the results suggest that social science research should pay more attention to reducing measurement error in variables constructed from text.

Exploring Item Bank Stability through Live and Simulated Datasets

LanguageCert manages the construction of its tests, exams and assessments using a sophisticated item banking system which contains large amounts of test material that is described, inter alia, in terms of content characteristics such as macroskills, grammatical and lexical features and measurement characteristics such as Rasch difficulty estimates and fit statistics. In order to produce content and difficulty equivalent test forms, it is vital that the items in any LanguageCert bank manifest stable measurement characteristics. The current paper is one of two linked studies exploring the stability of one of the item banks developed by LanguageCert [Note 1]. This particular bank has been used as an adaptive test bank and comprises 820 calibrated items. It has been administered to over 13,000 test takers, each of whom have taken approximately 60 items. The purpose of these two exploratory studies is to examine the stability of this adaptive test item bank from both statistical and operational perspectives. The study compares test taker performance in the live dataset with over 13,000 test takers (where each test taker takes approximately 60 items) with a simulated ‘full’ dataset generated using model-based imputation. Simulation regression lines showed a good match and Rasch fit statistics were also good: thus indicating that items comprising the adaptive item bank are of high quality both in terms of content and statistical stability. Potential future stability was confirmed by results obtained from a Bayesian ANOVA. As mentioned above, such item bank stability is important when item banks are used for multiple purposes, in this context for adaptive testing and the construction of linear tests. The current study therefore lays the ground work for a follow-up study where the utility of this adaptive test item bank is verified by the construction, administration and analysis of a number of linear tests.

DataSifter II: Partially synthetic data sharing of sensitive information containing time-varying correlated observations.

There is a significant public demand for rapid data-driven scientific investigations using aggregated sensitive information. However, many technical challenges and regulatory policies hinder efficient data sharing. In this study, we describe a partially synthetic data generation technique for creating anonymized data archives whose joint distributions closely resemble those of the original (sensitive) data. Specifically, we introduce the DataSifter technique for time-varying correlated data (DataSifter II), which relies on an iterative model-based imputation using generalized linear mixed model and random effects-expectation maximization tree. DataSifter II can be used to generate synthetic repeated measures data for testing and validating new analytical techniques. Compared to the multiple imputation method, DataSifter II application on simulated and real clinical data demonstrates that the new method provides extensive reduction of re-identification risk (data privacy) while preserving the analytical value (data utility) in the obfuscated data. The performance of the DataSifter II on a simulation involving 20% artificially missingness in the data, shows at least 80% reduction of the disclosure risk, compared to the multiple imputation method, without a substantial impact on the data analytical value. In a separate clinical data (Medical Information Mart for Intensive Care III) validation, a model-based statistical inference drawn from the original data agrees with an analogous analytical inference obtained using the DataSifter II obfuscated (sifted) data. For large time-varying datasets containing sensitive information, the proposed technique provides an automated tool for alleviating the barriers of data sharing and facilitating effective, advanced, and collaborative analytics.

Foundations of Machine Learning-Based Clinical Prediction Modeling: Part III-Model Evaluation and Other Points of Significance.

Various available metrics to describe model performance in terms of discrimination (area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 Score) and calibration (slope, intercept, Brier score, expected/observed ratio, Estimated Calibration Index, Hosmer-Lemeshow goodness-of-fit) are presented. Recalibration is introduced, with Platt scaling and Isotonic regression as proposed methods. We also discuss considerations regarding the sample size required for optimal training of clinical prediction models-explaining why low sample sizes lead to unstable models, and offering the common rule of thumb of at least ten patients per class per input feature, as well as some more nuanced approaches. Missing data treatment and model-based imputation instead of mean, mode, or median imputation is also discussed. We explain how data standardization is important in pre-processing, and how it can be achieved using, e.g. centering and scaling. One-hot encoding is discussed-categorical features with more than two levels must be encoded as multiple features to avoid wrong assumptions. Regarding binary classification models, we discuss how to select a sensible predicted probability cutoff for binary classification using the closest-to-(0,1)-criterion based on AUC or based on the clinical question (rule-in or rule-out). Extrapolation is also discussed.