### slides

```Learning in MOOCs!
Evidence and Correlates
Dave Pritchard and //RELATE.MIT.edu
S. Rayyan, R. Teodorescu, A. Pawl, Y. Bergner, A. Barrantes,
Chen, D. Seaton, C. Fredericks, J. Champaign, K. Colvin,
A. Liu, J. Doucette
Evidence of Learning/Improved Learning?
What Activities Correlate with Learning?
What Behaviors Correlate with Learning?
Two MOOCs: our 8.MReV – Mechanics Review
• 6.002x – MIT Electronics and Circuits
Simple Way to Measure Learning ?
• Give Same Test pre- and post- instruction
• See if there is Improvement, Gain = (post-pre)
8MReV only
Gain and Normalized Gain (-slope)
100%
Gain (= Post – Pre) 
Forbidden Region:
More than 100% on posttest!
g is the fraction of
unknowns on
pretest learned on
post test
0%
0%
Pretest Percentage 
–100% Normalized Gain g =
Gain (= Post – Pre)
100% - Pre
100%
Gain (posttest – pretest) vs Pretest
From R. Hake’s study of
6545 students in 62
classes. HSTop College
Most Interactive Above
Force Questions Gain in 8.MReV MOOC
Gain vs Pre-Score: equal Learning for all cohorts
g = 0.30± 0.02
6 Items
Non-Force Concept Questions 8.MReV
g = 0.33 ± 0.02
5 Items
N = 343
Concept and Quantitative 8.MReV
g = 0.41± 0.03
7 Items (2 quantitative)
N = 176
What & Why Item Response Theory
• Measures ability or skill of student
– Independent of which Questions Answered
– Intrinsic, not extrinsic (like total score)
• Sophisticated grading on a curve
– In Standard Deviations from Class Average
• We use it Two Ways:
– Alternate way to analyze pre and post-test
– Measure Relative Improvement HW and Tests
IRT Skill Increase PrePost N =579
Skill Increase in Course 
The key finding here is
that the less skillful
students learn as much
as more skillful students
-2.0
-1.0
0.0
1.0
Average Skill in Course 
2.0
Summary – Conceptual Learning
• Conceptual Learning in 8.MReV slightly greater
• None of the various cohorts we studied
showed significantly less normalized gain
– HS students vs those with advanced degrees
– poor prerequisites: math or physics courses
– Students of low average skill
• Contrary to concerns, no evidence that
unskillful, less educated, or less prepared
students learn less
Teachers, Non-Teachers, and MIT Students
We use 253 questions in both 8.011 and MOOC
Weekly IRT Skill of 8.MReV Various Cohorts
versus on-campus students
• On-campus students have the advantage of a
flipped classroom with MAPS instruction
• Hypothesis: They should show steady
improvement relative to MOOC students
On-Campus vs 8.MReV Weekly Skills
-Does Class Improve Skill?
There is no significant relative improvement of the 8.011 students .
Relative Improvement 0.6 (Skill Average -0.50 )
8.MReV Where Students Spent Time
Students attempting more than 50% of problems (N=1080).
Note that cool colors indicate instruction and warm colors indicate assessment
What Correlates with Learning?
• Initial Knowledge?
• Study Habits?
The fractional division of time among the
various resources of 6.002x
Data are for XXXX certificate earners who spent an average of 95 hours on the entire
course. Note that cool colors indicate instruction and warm colors indicate assessment
Correlates of Weekly Improvement and Gain
• Based on weekly IRT skills (e.g. on a curve)
• Find the slope of these: Relative Improvement
• Correlate with time on various components
– eText, Video, Discussion (instructional)
– Checkpoint questions, Homework (assessment)
Correlation Coefficients Visualized
-0.62
+0.30
Color Sign
FractionNumber
8.MReV Where Students Spent Time
Students attempting more than 50% of problems (N=1080).
Note that cool colors indicate instruction and warm colors indicate assessment
8.MReV Measures of Skills and Log of Time on Tasks (N = 292)
Posttest-Pretest Gain
Average Skill
Initial Skill
Relative Improvement
“Score” in Course
Checkpoint Discussion
eText
Problems
Total Time
The fractional division of time among the
various resources of 6.002x
Data are for XXXX certificate earners who spent an average of 95 hours on the entire
course. Note that cool colors indicate instruction and warm colors indicate assessment
6.002x Measures of Skills and Log of Time on Tasks (n=5948)
Skill Avg
Skill Initial
Relative
Improve
Score
Homework Video Lecture
Questions
Do students who spend
more time on Homework
have higher skill?
No, negative correlation
Lab
Book
Tutorial Discussion Wiki
Total Time
Do students who spend
more time watching lecture
videos improve more?
No, they improve less
Why Negative correlations!?
• More time on HW or Labs more skill?
• More skill takes less time to do HW or Lab!
• Why do we suppose the same instruction will
benefit students widely different in skill?
• Maybe we can analyze particular cohorts to
find effective instruction for some!
Conclusions and Future
• 8.MReV
– Positive correlations with conceptual learning
– Weaker correlation with Relative Improvement
• 6.002x: Broad Range of Skills & Demographics
– Strong Negative Correlations with Skills
– No significant Correlation with Relative Improve’t
• Future:
– examine different cohorts
– Experimental/Control group experiments
– Student Habits & Clusters of Characteristics
Predicting (Classifying) Improvement 8.MReV
We used various Machine Learning Algorithms to predict
whether students would be above or below average in
relative improvement. (50% correct is pure guessing)
Algorithm
Accuracy %
Support Vector Machcine
Decision Tree Learner C4.5 J48 Weka
Multiple Regression
55 +/- 1
71 +/- 6
73 +/- 6
Like Quantum Mechanics, only worse
Closer Look
At Homework
Copying
Palazzo, D. et. al. Phys. Rev. ST
Phys. Educ. Res. vol. 6, (2010), p.
010104
2.4 Sigma
Learning!
But no help on
conceptual
Amount of Symbolic Homework Copied 
Symbolic vs. Conceptual Difference! ??
Physics Teacher Expectation
• Students Start Symbolic Problems from Conceptual
Analysis
• Answer Numerical Questions by Plugging in
• The problems cover
the same topics, so
This result is
Unexpected
 Students not
Experts
Homework
Copying
Palazzo, D. et. al. Phys. Rev. ST
Phys. Educ. Res. vol. 6, (2010), p.
010104
2.4 Sigma
Learning!
But no help on
conceptual
Amount of Symbolic Homework Copied 
LORE: Library of Open Research-based
Educational Resources
• National Research Council: “researchbased educational resources produce
dramatically better learning outcomes”
• Open edX.org MOOC platform
– Have content from ~50 universities &
organizations
– Rapid way to vet assessments
– Enables big-data analysis of learning
The LORE Library
• Catalog with informative and actionable
– Learning Objective
– Level & Difficulty, Time to Complete…
• Directly assignable and automatically graded
• Vetted by trusted process
Library of Research-Based Resources
Data Mining
Psychometrics
Student
MOOC
Vetted
Calibrated
Library
Testing
New
Teacher
MOOC
Classical Test vs. IRT – MIT data
Classical Test
Item Response Theory
Item Response Theory
MIT 8.01 Class
MasteringPhysics
Std. Dev. Above
Fraction Correct 
Classical Test Theory
Chapter 
Chapter 
Chapter
The IRT graph has less error and shows the trend better:
Students selected by SAT scores have an advantage until the fifth week of the course at
MIT (vs. second semester in most colleges as claimed by ETS).
34
```