Journal of Data Science

The rapid accumulation and release of data have fueled research across various fields. While numerous methods exist for data collection and storage, data distribution presents challenges, as some datasets are restricted, and certain subsets may compromise privacy if released unaltered. Statistical disclosure control (SDC) aims to maximize data utility while minimizing the disclosure risk, i.e., the risk of individual identification. A key SDC method is data perturbation, with General Additive Data Perturbation (GADP) and Copula General Additive Data Perturbation (CGADP) being two prominent approaches. Both leverage multivariate normal distributions to generate synthetic data while preserving statistical properties of the original dataset. Given the increasing use of machine learning for data modeling, this study compares the performance of various machine learning models on GADP- and CGADP-perturbed data. Using Monte Carlo simulations with three data-generating models and a real dataset, we evaluate the predictive performance and robustness of ten machine learning techniques under data perturbation. Our findings provide insights into the machine learning techniques that perform robustly on GADP- and CGADP-perturbed datasets, extending previous research that primarily focused on simple statistics such as means, variances, and correlations.

Analysis of nonprobability survey samples has gained much attention in recent years due to their wide availability and the declining response rates within their costly probabilistic counterparts. Still, valid population inference cannot be deduced from nonprobability samples without additional information, which typically takes the form of a smaller survey sample with a shared set of covariates. In this paper, we propose the matched mass imputation (MMI) approach as a means for integrating data from probability and nonprobability samples when common covariates are present in both samples but the variable of interest is available only in the nonprobability sample. The proposed approach borrows strength from the ideas of statistical matching and mass imputation to provide robustness against potential nonignorable bias in the nonprobability sample. Specifically, MMI is a two-step approach: first, a novel application of statistical matching identifies a subset of the nonprobability sample that closely resembles the probability sample; second, mass imputation is performed using these matched units. Our empirical results, from simulations and a real data application, demonstrate the effectiveness of the MMI estimator under nearest-neighbor matching, which almost always outperformed other imputation estimators in the presence of nonignorable bias. We also explore the effectiveness of a bootstrap variance estimation procedure for the proposed MMI estimator.

Predictive modeling often ignores interaction effects among predictors in high-dimensional data because of analytical and computational challenges. Research in interaction selection has been galvanized along with methodological and computational advances. In this study, we aim to investigate the performance of two types of predictive algorithms that can perform interaction selection. Specifically, we compare the predictive performance and interaction selection accuracy of both penalty-based and tree-based predictive algorithms. Penalty-based algorithms included in our comparative study are the regularization path algorithm under the marginality principle (RAMP), the least absolute shrinkage selector operator (LASSO), the smoothed clipped absolute deviance (SCAD), and the minimax concave penalty (MCP). The tree-based algorithms considered are random forest (RF) and iterative random forest (iRF). We evaluate the effectiveness of these algorithms under various regression and classification models with varying structures and dimensions. We assess predictive performance using the mean squared error for regression and accuracy, sensitivity, specificity, balanced accuracy, and F1 score for classification. We use interaction coverage to judge the algorithm’s efficacy for interaction selection. Our findings reveal that the effectiveness of the selected algorithms varies depending on the number of predictors (data dimension) and the structure of the data-generating model, i.e., linear or nonlinear, hierarchical or non-hierarchical. There were at least one or more scenarios that favored each of the algorithms included in this study. However, from the general pattern, we are able to recommend one or more specific algorithm(s) for some specific scenarios. Our analysis helps clarify each algorithm’s strengths and limitations, offering guidance to researchers and data analysts in choosing an appropriate algorithm for their predictive modeling task based on their data structure.