Academic preparation, particularly in mathematics and science, is consistently identified as a primary predictor of STEM success. Prior research indicates that high school GPA, standardized test performance, and the rigor of prior mathematics and science preparation are important indicators of whether a student will persist in a STEM major ([7], [8]; [9]; [17]; [23]). For example, early performance in mathematics courses and math proficiency are critical filters; students who enter college with stronger quantitative readiness are more likely to remain in technical fields ([7], [8]; [18]).

Demographic factors also play a significant role in attrition patterns. Studies have repeatedly shown that persistence in STEM varies across gender, race/ethnicity, first-generation status, and socioeconomic circumstances ([2]; [7]; [23]). Socioeconomic status (SES) remains an important contextual factor because access to academic preparation, enrichment opportunities, and institutional resources is not evenly distributed across students ([10]; [12]).

Beyond academic and demographic variables, psychological and motivational factors are essential to understanding the STEM pipeline. Self-efficacy—the belief in one’s ability to succeed in specific tasks—is an important component of persistence, while expectancy-value beliefs shape students’ plans to continue in STEM ([2]; [24]). Students with a strong sense of STEM or science identity are also more likely to sustain engagement in STEM environments ([16]; [21]).

Interest and utility value—the perception that STEM is useful and relevant to one’s goals—also contribute to persistence ([3]; [13]). Conversely, a weak sense of belonging can undermine persistence even among academically capable students. Recent evidence highlights the importance of belonging, academic hope, identity, and cultural resources for students from groups historically underrepresented in STEM ([10]; [12]).

Recent advancements in educational data mining and machine learning have shifted the focus toward predictive analytics. Studies have employed logistic regression, neural networks, random forests, and other machine-learning approaches to identify students at risk of STEM attrition or non-retention ([1]; [4]; [9]; [17]; [20]).

Although predictive models can achieve strong overall classification performance, such results can be misleading when the outcome classes are imbalanced. A model may correctly classify many persisters while failing to identify the smaller group of students who actually leave STEM. Consequently, attrition studies increasingly emphasize class-specific performance and early-alert usefulness rather than overall accuracy alone ([9]; [20]). Longitudinal approaches further show that attrition risk develops over time and can be informed by academic preparation, engagement, and motivational trajectories ([18]; [20]).

The data were obtained from the High School Longitudinal Study of 2009 (HSLS:09), conducted by the National Center for Education Statistics ([11]). HSLS:09 followed a cohort of students from their 9th-grade year through postsecondary education and into the workforce. The original dataset contained 23,503 observations.

The analytical cohort was restricted to students meeting the initial STEM-intention criterion (`S3FIELD_STEM = 1`) and having valid later undergraduate-major information (`X4RFDGMJSTEM`). Of the 23,503 original records, 2935 met the STEM-intention criterion. After requiring valid later-major information, 2331 students remained; 604 were not retained at this later-major eligibility stage. The final cohort included 1572 students (67.44%) with a later STEM major and 759 (32.56%) with a later non-STEM major. The outcome does not measure continuous STEM enrollment or degree completion. The cohort construction is summarized in Figure 1.

1) Indicated an intention to pursue a STEM major during their high school years.

2) Had valid information regarding their later undergraduate major status.

The final sample consisted of 2331 students. These students were classified into two groups:

Persisters: Students whose later reference-degree major remained in a STEM field.Attriters: Students whose later major was in a non-STEM field.

Note. The 604 records were not retained at the valid-later-major eligibility step; the available analysis does not subdivide them into specific missingness categories.

Figure 1. Analytical cohort construction.

The predictor set comprised mathematics achievement (`X1TXMTH`), socioeconomic status (`X1SES`), sex (`X1SEX`), race/ethnicity (`X1RACE`), parent educational expectations (`X1PAREDEXPCT`), mathematics identity (`X1MTHID`), mathematics interest (`X1MTHINT`), mathematics self-efficacy (`X1MTHEFF`), science identity (`X1SCIID`), science interest (`X1SCIINT`), science self-efficacy (`X1SCIEFF`), and science utility (`X1SCIUTI`). Sex, race/ethnicity, and parent educational expectations were treated as categorical predictors; the remaining listed predictors were treated as numeric. GPA was not included in the model predictor set.

1) Academic achievement: standardized mathematics achievement score (`X1TXMTH`).

2) Socioeconomic status: HSLS:09 composite measure (`X1SES`).

3) Demographics: sex (`X1SEX`) and race/ethnicity (`X1RACE`).

4) Motivational measures: mathematics identity, interest, and self-efficacy; science identity, interest, self-efficacy, and utility; and parent educational expectations (`X1PAREDEXPCT`).

Missing numeric predictor values were imputed using medians and categorical values using modes estimated from the training data to avoid data leakage. Preprocessing also included factor handling, dummy encoding, novel-level handling, zero-variance removal, and normalization, with training-derived preprocessing applied to the test data.

Four predictive models were evaluated and compared:

1) Logistic Regression: A standard statistical baseline for binary classification.

2) Elastic Net Regularized Logistic Regression: A model that incorporates both L1 (Lasso) and L2 (Ridge) penalties to handle potential multicollinearity and perform feature selection ([25]).

3) Random Forest: A non-linear ensemble method that builds multiple decision trees to improve stability and predictive power ([5]).

4) Extreme Gradient Boosting (XGBoost): An advanced gradient boosting framework known for high performance and efficiency in structured data tasks ([6]).

The data were divided into stratified training and test samples using an 80:20 split (n = 1864 and n = 467, respectively). Model development used repeated stratified five-fold cross-validation with two repetitions and random seed 123. ROC-AUC was the primary model-selection metric. The models were logistic regression, Elastic Net, random forest, and XGBoost. The reported models did not incorporate NCES sampling weights or complex-survey design adjustments.

ROC-AUC: Area under the Receiver Operating Characteristic curve.PR-AUC: Area under the Precision-Recall curve.Sensitivity (Recall): The ability to correctly identify students who will attrite.Specificity: The ability to correctly identify students who will persist.Precision: The accuracy of attrition predictions.F1-Score: The harmonic mean of precision and recall.

For the Elastic Net model, candidate classification thresholds from 0.10 to 0.60 in increments of 0.01 were evaluated using cross-validation predictions. Because repeated cross-validation produced multiple predictions per observation, probabilities were averaged by observation. The threshold maximizing cross-validated balanced accuracy (approximately 0.604) was 0.30. The selected cutoff was then applied to the held-out test sample.

The final cohort (N = 2331) included 1572 students (67.44%) with a later STEM major and 759 (32.56%) with a later non-STEM major. The analysis therefore concerns the realization of an initial STEM intention in later major status. The distribution of these two outcome groups is shown in Figure 2.

The predictive discrimination was moderate. Elastic Net had the highest cross-validation ROC-AUC (0.643; SE = 0.008). The independent test-set comparison is summarized in Table 1. Classification metrics at the selected Elastic Net threshold are presented separately in Table 2. Figure 3 provides a visual comparison of the independent test-set performance across the four models.

Note. The bars show the number and proportion of students with a later non-STEM major (attriters) and a later STEM major (persisters) in the eligible analytical cohort.

Figure 2. Distribution of STEM-intention outcomes in the analytical cohort.

Table 1. Cross-validation and default-threshold test performance by model.

Table 2. Elastic Net held-out test performance at the selected threshold (0.30).

Note. Sensitivity and specificity in this figure are based on the default classification cutoff, not the optimized threshold.

Figure 3. Independent test-set performance by model.

Performance Metrics for the Optimized Elastic Net Model(Threshold= 0.3)

The corresponding classification results at the selected threshold are illustrated in Figure 4.

Note. Rows represent predicted class and columns represent observed class; the selected cutoff is 0.30.

Figure 4. Elastic net confusion matrix at threshold 0.30.

Notably, regularized regression (Elastic Net) performed competitively with the more complex tree-based models (Random Forest and XGBoost), suggesting that the predictive signal in this dataset is effectively captured by an additive model.

Feature importance was examined for the tree-based models. The values are model-specific measures of predictive contribution and should not be interpreted as causal effects. Figure 5 presents the XGBoost feature-gain values, while Table 3.

Note. Bars show XGBoost feature gain, a model-specific predictive importance measure, not a causal effect.

Figure 5. Top XGBoost predictors of STEM attrition.

Table 3. Leading feature-importance values from XGBoost and random forest.

compares the leading predictors identified by XGBoost and Random Forest. The XGBoost feature-gain results are visualized in Figure 5.

Top Predictors Ranked by XGBoost and Random Forest

Mathematics achievement emerged as the strongest predictor across all models, emphasizing the foundational role of quantitative skills in STEM persistence ([7]; [8]; [14]). Motivational variables, specifically math identity and science self-efficacy, also ranked highly, indicating that how students perceive their own abilities is a critical indicator of their future academic path ([14]; [22]).