StatMatch: Statistical Matching Challenges SMOTE for Imbalanced Data

A new study reports that a data-balancing method borrowed from statistical matching can rival the widely used SMOTE technique on some imbalanced classification problems, even though SMOTE still wins most head-to-head comparisons. The method, called StatMatch, was developed by Francis Fuller Bbosa and published in BMC Artificial Intelligence on 3 November 2026. It tackles a stubborn problem in machine learning: when one class of examples vastly outnumbers another, standard classifiers tend to favour the abundant class and overlook the rare cases that often matter most in fraud detection, medical diagnosis and risk prediction.

StatMatch adapts statistical matching, a technique traditionally used to merge data from different sources when no shared unique identifier exists. In classical statistical matching, two datasets are combined to produce a richer dataset containing variables that were never jointly observed. Bbosa repurposed this idea for a different goal: instead of integrating sources, the method uses matching principles to select which majority-class observations to keep when shrinking a lopsided dataset down to a balanced one.

The procedure treats the minority class as a recipient dataset and the majority class as a donor dataset. Each minority observation is paired with a majority observation using a shared matching variable, and the degree of agreement is scored with an exact matching indicator: zero when the two observations agree on the chosen variable, one when they differ. The donor observation with the smallest distance is retained, and if several donors tie at the same minimum distance, a deterministic rule selects the one with the smallest unique identifier so the match is repeatable. Once a majority observation has been chosen, it is excluded from further consideration. The result is a balanced dataset built entirely from observed data, with no synthetic points generated.

This distinguishes StatMatch from nearest-neighbour-based undersampling methods such as NearMiss, Condensed Nearest Neighbour and Tomek Links, which rely on geometric distances in feature space, and from oversampling methods such as SMOTE, which create synthetic minority observations by interpolating between existing minority instances. StatMatch instead selects majority-class observations through an exact matching criterion based on categorical variable agreement. The study implemented a single-variable version of the matching framework, setting the number of matching variables to one, and matched without replacement. The authors argue this makes the approach attractive in settings such as healthcare, finance and official statistics, where data provenance and interpretability are crucial and fabricated observations are undesirable.

To test the idea, StatMatch was benchmarked against SMOTE using two classifiers with different learning paradigms: Logistic Regression, a parametric linear model, and Artificial Neural Networks (ANN), a non-parametric nonlinear model. Four real-world imbalanced datasets from the UCI repository were used: Blood Transfusion Service Centre, Haberman’s Survival, Breast Cancer Wisconsin and Parkinson’s Disease. Each dataset was preprocessed by removing missing data points, converting class labels into a binary outcome variable, identifying the minority and majority groups, and transforming variables where appropriate. The data were split into 70 percent training and 30 percent testing subsets using stratified sampling, and the training portion was evaluated with repeated stratified 10-fold cross-validation, repeated ten times with different random partitions. Performance was measured with the F-measure and the Area Under the Receiver Operating Characteristic Curve (AUC), both considered insensitive to imbalanced data, and differences between methods were assessed with paired Wilcoxon signed-rank exact tests because the paired differences did not satisfy the normality assumption required for a paired t-test.

The results showed SMOTE generally achieving higher performance. On the Blood dataset, SMOTE reached F-measure values of 68.7 ± 3.7 percent with Logistic Regression and 68.8 ± 2.0 percent with ANN, compared with 54.2 ± 5.6 percent and 57.4 ± 4.9 percent for StatMatch, differences that were statistically significant with p-values of 0.004 and 0.002. The AUC gap on Blood was even starker: SMOTE achieved 76.3 ± 3.1 percent and 75.8 ± 3.1 percent against 57.6 ± 5.1 percent and 56.5 ± 5.3 percent for StatMatch, significant at p = 0.002 for both classifiers. On the Haberman dataset, SMOTE again led on F-measure, with Logistic Regression attaining 63.3 ± 6.1 percent versus 56.5 ± 8.9 percent for StatMatch, and ANN reaching 67.0 ± 8.0 percent versus 53.7 ± 9.2 percent; the ANN difference was significant at p = 0.006 while the Logistic Regression difference was not, at p = 0.106.

The picture changed on the Breast Cancer dataset, where both methods performed strongly. Both balancing approaches produced F-measure scores above 94 percent for both classifiers, with SMOTE obtaining 97.4 ± 0.8 percent using Logistic Regression and 97.4 ± 1.2 percent with ANN, while StatMatch achieved 95.5 ± 1.2 percent and 94.6 ± 1.8 percent. Both comparisons were statistically significant despite the small numerical gaps. For AUC, both methods exceeded 96 percent, and Logistic Regression produced scores of 99.5 ± 0.3 percent with SMOTE and 99.2 ± 0.3 percent with StatMatch, a difference that was not statistically significant at p = 0.065. With ANN, however, SMOTE reached 99.2 ± 0.7 percent against StatMatch’s 96.9 ± 1.3 percent, a significant improvement at p = 0.002. The study notes that when classes are already clearly distinguishable, the choice of balancing strategy may have little impact on the classifier’s ability to discriminate.

The most intriguing result came from the Parkinson dataset, where StatMatch marginally outperformed SMOTE with the ANN classifier on both metrics. StatMatch attained an F-measure of 76.2 ± 15.6 percent versus 75.0 ± 9.1 percent for SMOTE, and an AUC of 84.4 ± 12.7 percent versus 82.9 ± 9.2 percent, though neither difference reached statistical significance, with p-values of 0.695 and 0.557. The author suggests that statistical matching might retain meaningful structures within the minority class that support the performance of certain nonlinear models, while cautioning that Parkinson data are defined by nonlinear interactions among biomedical voice metrics, so creating representative observations through matching alone may be harder than synthetic interpolation. With Logistic Regression on Parkinson, SMOTE produced a notably higher AUC of 90.0 ± 4.1 percent compared with 72.7 ± 11.5 percent for StatMatch, a significant difference at p = 0.002, underscoring how strongly outcomes depend on the classifier as well as the dataset.

Across the eight dataset-classifier combinations, SMOTE produced higher mean F-measure and AUC values in seven. The study concludes that SMOTE remains a strong baseline method for improving classification performance on imbalanced datasets, while positioning StatMatch as a promising complementary strategy rather than a universal replacement. Because StatMatch preserves only observed data, the author argues it may serve as a practical alternative in applications where maintaining data authenticity is an important consideration. The findings also reinforce the broader lesson that no single imbalance-handling method works best everywhere: performance varies with class overlap, dataset size, number of features, classifier complexity and the extent of imbalance.

The study acknowledges several limitations. The evaluation used only four publicly available binary classification datasets with differing imbalance ratios, and StatMatch was compared only with SMOTE rather than with other oversampling, undersampling or hybrid techniques. The deterministic design of the algorithm leaves open scenarios where multiple majority observations share a distance of zero, and future work could explore a stochastic variant. Only Logistic Regression and ANN were tested, so decision trees, random forests, support vector machines and gradient boosting methods may reveal different interactions between sampling strategy and classifier. Future research directions include extending the framework to multiclass imbalanced problems, assessing it on larger and more varied datasets, refining the matching algorithm, and investigating computational efficiency, scalability and high-dimensional applications. The analysis was performed in R version 4.3.1 using packages including dplyr, caret, ggplot2, pROC and smotefamily, and the program scripts are available from the author upon request.

Subject of Research: A statistical matching approach for balancing imbalanced datasets in binary classification

Article Title: A statistical matching approach to handling imbalanced datasets using binary classifiers

Article References: Bbosa, F. F. (2026). A statistical matching approach to handling imbalanced datasets using binary classifiers. BMC Artificial Intelligence, 2(1), Article 23. https://doi.org/10.1186/s44398-026-00042-9

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00042-9

Keywords: class imbalance, StatMatch, SMOTE, statistical matching, undersampling, binary classification, machine learning, Logistic Regression, artificial neural networks, F-measure, AUC, UCI datasets