Choosing between random forest vs. logistic regression is not simply a contest to identify the algorithm with the highest accuracy. Predictive analytics teams also need to consider the structure and quality of their data, the outcome they are predicting, how model results will be used, and whether stakeholders can interpret the output. A University of Hawai’i study comparing the methods to predict admissions yield illustrates these trade-offs.
Although the application focused on student enrollment rather than industrial operations, its evaluation approach has broader relevance for teams building classification models: compare results on held-out data, examine performance by outcome class, and involve the people expected to act on model outputs.
How the two approaches differ
Logistic regression is a statistical method for modeling the probability of a binary outcome. It represents relationships between input variables and the outcome through a mathematical equation and coefficients. Those coefficients indicate the direction of an association, while measures such as likelihood-based fit statistics and pseudo R-squared help assess the model. Analysts typically evaluate assumptions, investigate outliers and collinearity, and refine the model toward a useful, relatively parsimonious form.
Random forest is an ensemble learning method. It builds many decision trees using randomized samples and feature selection, then combines their predictions. Rather than describing the relationship through one equation, the forest aggregates decisions across trees. This structure can accommodate complex relationships and correlated inputs without relying on the same statistical assumptions as logistic regression. Its behavior and results still need evaluation; the method is not automatically the right choice for every dataset.
Study design: predicting enrollment yield
The study aimed to identify non-resident first-year applicants who had been admitted but might still be undecided about enrolling. Models used information available early in the recruitment cycle, including application and contact activity, geographic and demographic characteristics, academic information, application timing, and whether a FAFSA had been submitted.
Historical cohorts from fall 2012 through fall 2016 formed the training dataset, with 16,420 records. The fall 2017 cohort, containing 4,270 records, served as a holdout validation set. Its baseline enrollment yield was 18%. The researchers compared logistic regression with two random forest versions and assessed classification results and ROC curve area, adjusting classification cut values as part of the evaluation.
Random forest vs. logistic regression: what the results showed
On the same dataset, the first random forest version achieved 78.9% overall accuracy on the holdout set, compared with 76.4% for logistic regression. The random forest correctly classified 83.9% of non-enrollees, versus 80.9% for logistic regression. For enrollees, however, the results were nearly identical: 54.4% for random forest and 54.5% for logistic regression.
The ROC area was also close: 0.798 for the first random forest model and 0.792 for logistic regression. These results indicate broadly comparable predictive performance in this application, rather than a decisive advantage for one method. They also show why overall accuracy alone can be misleading. The models were substantially better at classifying non-enrollees than enrollees, so decision-makers needed to consider the class-specific results as well as the aggregate score.
A second random forest version, prepared differently for that analysis, achieved 76.7% overall accuracy, with correct classification rates of 83.7% for non-enrollees and 42.4% for enrollees. Its ROC area was 0.791. The difference between the two forest versions reinforces the importance of documenting data preparation and evaluating the specific model that will be deployed, rather than treating an algorithm name as a guarantee of performance.
Interpretability matters to the people using predictions
Model performance was only part of the evaluation. Enrollment managers found logistic regression’s coefficient direction and Delta P statistic intuitive, helping them understand how variables related to predicted enrollment. They also valued practical cut points from random forest for institutional grant aid and total offer amount, but wanted a side-by-side comparison of the forest’s effect scores and logistic regression results.
This is a useful lesson for operational analytics teams. A prediction has limited practical value if its intended users cannot understand it well enough to assess or apply it. In a plant or automation context, stakeholders may need to understand why a system flags an event, how to prioritize the result, and what threshold triggers a response. The appropriate model therefore depends not only on classification quality but also on the decision workflow and the explanations users need.
Limitations and model-selection considerations
The researchers noted that the data had relatively little collinearity, randomness, or complexity, which may not have provided the strongest test of random forest. Variables with low correlation to the outcome were largely excluded because the analysis began from a regression-oriented perspective; some might have helped the forest. They also identified imbalanced outcome data as a possible influence on random forest results.
These limitations caution against assuming that a result from one dataset will transfer directly to another. When comparing candidate models, teams should:
- Use a holdout set to assess how models perform on data not used for fitting.
- Review class-specific metrics alongside overall accuracy, particularly when outcome classes are uneven.
- Check the effects of data preparation and record how variables, missing values, and thresholds are handled.
- Assess usability by asking intended users what explanations and outputs support their decisions.
The study also reported separate retention analyses in which the relative results varied by prediction timing and outcome class. That variation further underscores that method choice is tied to the particular prediction task, available data, and point at which a forecast is needed.
Conclusion
In this enrollment-yield comparison, random forest and logistic regression delivered similar overall predictive performance, while differing in class-level results and interpretability. Logistic regression offered coefficients and fit measures that users found easier to explain; random forest provided a different way to capture patterns and useful decision cut points. For industrial automation and other operational analytics applications, the practical choice is the model that performs adequately on relevant validation data and produces outputs people can use. Compare alternatives in context, examine more than one accuracy measure, and treat data preparation and user feedback as integral parts of model selection.
Access Official Technical PDF