Modeling Active Learning Classrooms

By analyzing data from over 10,000 students across diverse disciplines, this study develops a predictive model identifying that specific combinations of lecture time, group worksheets, clicker questions, and student questions reliably predict conceptual learning gains, revealing that exceptional outcomes are achieved either through a balanced mix of these activities or by dedicating 30% or more of class time to group worksheets, while noting that active learning strategies without group worksheets yield results comparable to traditional lectures.

The Gist

Imagine you are a chef trying to figure out the perfect recipe for a dish that makes everyone at the table smarter. For decades, the standard recipe for university science classes was simple: The Lecture. The professor stands at the front, talks for 90 minutes, and the students sit quietly and take notes.

But researchers have long known that this "all-lecture" recipe often results in students failing the test or not actually understanding the material. So, educators started adding "active learning" ingredients—things like group work, clicker quizzes, and asking questions.

The problem? The recipe book was a mess. Everyone was throwing in different ingredients in different amounts, and no one knew exactly which combination would make the dish a masterpiece. Some said, "More group work!" Others said, "More clickers!"

This paper is like a massive cooking competition analysis. The authors gathered data from over 10,000 students across 24 universities (like a huge tasting panel) to build a predictive model. Think of this model as a super-smart AI chef that can taste a class's "recipe" (how much time is spent on what activity) and predict exactly how well the students will learn.

Here is what their "AI Chef" discovered:

1. The "No-Worksheet" Trap

The model found a hard rule: If you don't have group worksheets, you can't get a great result.
It doesn't matter if you use clickers, ask questions, or do other fun activities. If the students aren't sitting in groups working on a worksheet together, the learning gains are as low as a boring, all-lecture class. It's like trying to bake a cake without flour; you can add all the fancy sprinkles you want, but it won't rise.

2. The Two "Golden Recipes"

The model identified two specific combinations of ingredients that consistently produce "exceptional" learning (the culinary equivalent of a Michelin-star dish):

• Recipe A (The Balanced Mix):

  • 10–20% Group Worksheets: Students working together on paper problems.
  • 20–40% Group Clickers: Students using clickers to answer questions in teams.
  • 2+ Student Questions per Hour: Students actually raising their hands and asking questions.
  • The rest of the time? The professor can still lecture! You don't need to ban lectures entirely, just keep them to a reasonable amount.

• Recipe B (The Worksheet Heavy-Hitter):

  • 30% or more Group Worksheets.
  • If you spend a third of the class or more on group worksheets, you get great results, even if you don't use clickers.

3. The "Clicker-Only" Illusion

This was a surprising discovery. Many teachers think, "If I just use clickers (multiple-choice questions) instead of lecturing, that's active learning!"
The model says: Nope.
Classes that only use clickers (and no worksheets) perform just as poorly as fully lecture-based classes. It's like a car that has a really nice radio (clickers) but no engine (no deep group work). It looks active, but it doesn't go anywhere.

4. The Power of Student Questions

The model found that student questions are a secret sauce. When students ask questions in front of the whole class (about 2 or more per hour), learning skyrockets. The researchers aren't 100% sure why yet (maybe it wakes everyone up, or maybe it clears up confusion for the whole room), but the data shows it works.

How They Did It (The "Kitchen" Logic)

• The Ingredients: They used a tool called COPUS, which is like a stopwatch that ticks every 2 minutes. Observers marked what was happening: Was the teacher lecturing? Were students doing worksheets? Were they using clickers?
• The Taste Test: They measured learning using "Concept Inventories" (standardized tests given before and after the course). They calculated how much the students improved.
• The Prediction: They fed all this data into a computer model. Instead of just saying "Group work is good," the model created a map. It showed exactly how much of each ingredient is needed to hit the "sweet spot" of learning.

The Bottom Line for Teachers

You don't need to reinvent the wheel or stop lecturing entirely. To get the best results:

1. Must have: Group worksheets.
2. Great addition: Group clicker questions.
3. Secret weapon: Encourage students to ask questions.
4. Avoid: Relying only on clickers or doing no group work.

This study gives teachers a clear, testable recipe to follow, moving away from guesswork and toward a proven method for helping students learn science more effectively and fairly.

Technical Summary

1. Problem Statement

Despite extensive research demonstrating that active learning (non-lecture instruction) improves undergraduate science learning outcomes compared to traditional lectures, the field lacks specificity. The term "active learning" is an umbrella term encompassing diverse strategies (e.g., clickers, worksheets, peer instruction) without clear definitions regarding:

• Which specific strategies are most effective.
• The optimal duration or proportion of class time each strategy should occupy.
• How different combinations of strategies interact to influence learning.

Current literature often relies on descriptive analyses or compares broad pedagogical frameworks (e.g., SCALE-UP vs. traditional lecture) rather than providing testable, quantitative predictions for specific classroom configurations. This ambiguity hinders instructors from making evidence-based decisions on how to structure their courses.

2. Methodology

Data Collection and Scope

The study utilizes a comprehensive dataset comprising 69 undergraduate science courses across 24 institutions (US and Canada), involving over 10,000 students and 40 unique instructors.

• Sources: Data was aggregated from three existing studies (Sundstrom et al. 2025; Connell et al. 2016; Weir et al. 2019) and new data collected by the authors (12 introductory physics/astronomy courses).
• Disciplines: Physics, Astronomy, and Biology.
• Class Sizes: Ranging from 11 to 576 students.

Variables and Metrics

• Independent Variables (Instructional Strategies): Classroom activities were quantified using the Classroom Observation Protocol for Undergraduate STEM (COPUS). The study focused on four specific codes representing the fraction of class time (2-minute intervals):
  1. Lec: Instructor lecturing.
  2. WG: Students working on group worksheets.
  3. CG: Students working on group clicker questions.
  4. SQ: Student questions asked during class.
     Note: Other COPUS codes were excluded due to broad definitions or high covariance with the selected variables.
• Dependent Variable (Learning Outcome): Student learning was measured using Cohen's d (effect size), calculated as the standardized mean difference between pre-test and post-test scores on discipline-specific concept inventories.

Modeling Approach

The authors employed a predictive modeling framework rather than a purely explanatory one.

• Algorithm: Multiple Linear Regression using Ordinary Least Squares (OLS).
• Feature Engineering: To capture non-linear relationships and interactions, the model included linear terms, squared/cubed terms, and cross-terms (e.g., $WG \times CG$, $WG^2 \times CG$) for the four COPUS codes.
• Feature Selection: A rigorous selection process was performed 100,000 times with randomized feature removal orders. The final model retained the feature sets with the 10 lowest Akaike Information Criterion corrected for small sample sizes (AICc) to prevent overfitting.
• Validation and Error Estimation:
  • Bootstrap Sampling: The model was trained on 1,000 subsets (85% of data) to generate a distribution of predictions. The mean of this distribution served as the prediction, and the standard deviation served as the prediction error.
  • Reliability Criteria: Predictions were deemed unreliable if the error exceeded 0.3 (approx. 10% of the effect size range) or if they required extrapolation beyond the training data's parameter space.
  • Out-of-Sample Testing: Leave-One-Out (LOO) cross-validation and testing on the 12 new classes confirmed the model's robustness, yielding a mean absolute error of 0.17 for reliable predictions.

3. Key Results

The predictive model identified two distinct instructional configurations that consistently yield exceptional learning gains (effect sizes > 2):

1. The "Balanced Active" Model:

   • 10–20% of class time on Group Worksheets (WG).
   • 20–40% of class time on Group Clicker Questions (CG).
   • Average of 2+ Student Questions (SQ) per hour.
   • Note: The remaining time can be lecture, yet these classes still achieve high gains.

2. The "Worksheet-Heavy" Model:

   • 30% or more of class time on Group Worksheets (WG).
   • Note: These classes achieve high gains even without group clicker questions, though the data for this specific combination is less abundant, requiring further validation.

Critical Negative Findings:

• No Worksheets = Low Gains: Classes that utilize zero group worksheets consistently show small effect sizes, even if they employ other active strategies like clicker questions or student questions.
• Clickers Alone are Insufficient: Relying solely on group clicker questions (without worksheets) yields learning outcomes comparable to fully lecture-based classes.
• Lecture is Not Inherently Bad: While 100% lecture is detrimental, "majority-lecture" classes can still achieve high learning outcomes if they incorporate the specific active learning ratios identified above.

4. Significance and Contributions

• Shift from Descriptive to Predictive: This study moves beyond describing that active learning works to predicting how to implement it. It provides a quantifiable map linking specific time allocations to learning outcomes.
• Actionable Recommendations: The study offers testable, specific guidelines for instructors. For example, it suggests that simply adding clickers is insufficient; the inclusion of group worksheets and student questioning is critical.
• Nuance in "Active Learning": It challenges the binary view of "lecture vs. active learning," demonstrating that the combination and proportion of activities matter more than the mere presence of non-lecture time.
• Methodological Rigor: By using bootstrap sampling to quantify prediction error and explicitly defining "unreliable" regions, the authors provide a framework for future studies to know where their data falls within the model's confidence bounds.

5. Limitations and Future Work

• Scope of Variables: The model does not account for out-of-class activities (homework, labs) or other COPUS codes.
• Coding Granularity: COPUS uses 2-minute intervals, which may overestimate the duration of short activities (e.g., a single student question).
• Selection Bias: The dataset may overrepresent instructors who are education researchers or high-performing students who consent to research participation.
• Future Directions: The authors call for controlled studies to test the "optimal" combinations identified (specifically the 10-20% WG / 20-40% CG / 2+ SQ configuration) and to investigate the mechanisms behind the impact of student questions.