Presented university scoring using Multi Criteria Decision Analysis (MCDA) & clustering to improve efficiency of Teach for America recruitment efforts.
2. Keep 3 tiers in 2018, but choose the tiers built by our AUSS model.
By implementing the proposed AUSS model in 2017, TFA could have
decreased the resource utilization by 16%
If TFA had 1 additional recruiter in 2017, as per the AUSS model it could
have gained additional 200 accepted students.
Recommendations
3. Decline in the applicant volume by 35 %
1. Are we targeting the students who are genuinely motivated to work for TFA ?
2. Have we analyzed the past data to find which parameters bring in most number of students ?
To maximize the volume of accepted students :
4. Framework for identifying optimal tiers
Step 1. Exploratory data
analysis (EDA)
o Understand student
motivation
o Which parameters in
the past had major
impact on recruitment?
Step 2. Develop
Student Score card
o Develop a score card
which rates every
student’s
attractiveness
Step 3. Develop
Aggregate University
Student Score( AUSS )
Step 4. Develop
University tiers
o Aggregate the
generated student
scores and given
University parameters
to form AUSS model
o Cluster the
Universities with
similar features
5. Step 1. Exploratory data
analysis (EDA)
o Understand student
motivation
o Which parameters in
the past had major
impact on recruitment?
6. Select impactful variables
For Interpretation:
‘Met’ – Students who network with the recruiter are the one’s who are motivated to join TFA.
Percentage of accepted ‘ Met students ‘ is 65%
‘Undergrad’ – Seniors being in their final must be more inclined to start with a job. Percentage of
accepted Senior students is around 90%
7. Only 7 Qualitative variables important to develop model
Discarded
Variables
Included
Variables
8. Step 2. Develop Student Score
card
o Develop a score card which
rates every student’s
attractiveness
9. What student scorecard will look like
Weightage – Stresses on parameters which can bring more students
Universal Ratings – Stresses on parameters which have traditionally brought more students in the past.
These 2 features of our model will work towards maximizing student acceptance rate
10. Relative importance of student criteria determined by Analytic
Hierarchy Process (AHP)
TFA Selection Criteria
• Demonstrated past achievement
• Perseverance in the face of challenges
• Strong critical thinking skills
• Ability to influence and motivate others
• Organizational ability
• Understanding of and desire to work
relentlessly in pursuit of our vision
• Respect for students and families in
low-income communities
Selection process subjective
Experienced Recruiter Input essential
Connection between recruiter
judgement and outcomes necessary
Use a Multi Criteria
Decision Method
(MCDM) suitable for
qualitative parameters
Source : TFA Program Continuum Presentation
11. Use AHP to generate weightage of each variable
Analytical Hierarchy Process (AHP) is a type of Multi Criterion Decision Method (MCDM) developed by
Thomas L. Saaty
We took perspective of 1 TFA volunteer and 2 college recruiters
1. Build Model & Run it 2. Results 3. Check model consistency
12. Use Ratios to develop a standardized universal scale for ratings
For Example: Consider the Qualitative variable ‘Met’
Step 1: 65% accepted students- Yes; 35 % accepted
students– No
Step 2: Convert into 10 point scale
13. Student Score obtained
Weightage – Stresses on parameters which can bring more students
Universal Ratings – Stresses on parameters which have traditionally brought more
students in the past.
14. Step 3. Develop Aggregate
University Student Score(AUSS )
o Aggregate the generated
student scores and given
University parameters to form
AUSS model
15. University Score obtained
Individual Student Scores were mapped to
each of the respective universities they
belonged to.
Clusters to be formed based on the above University parameters
16. Step 4. Develop University
tiers
o Cluster the Universities with
similar features
17. k Means clustering: Tiers through
Euclidian distance between Universities
Cluster the Universities offering similar features
1. Build model and Run
2. Clusters formed
3. Optimal Cluster selection
18. Results obtained by AUSS methodology on 2016 data:
The three tiers obtained from the above will be tested on 2017 actual data to validate the
model.
University distribution as per AUSS model
20. It turns out that TFA excluded 203 universities from the 3rd tier of our AUSS model. The Universities
considered least important by AUSS were considered least important by TFA too.
It confirms the accuracy of AUSS model !
Universities eliminated by TFA
Finding 1 : Why did TFA exclude 209 universities in 2017 ?
21. Number of resources required to
recruit the same number of
students decreased by 16 %.
Acceptance Ratio more equally
distributed
Tier 1 has a maximum Aggregate
Student Score of 809 implying
best students get recruited
Model Advantages
Testing the AUSS tier distribution on actual TFA 2017 tier distribution
Finding 2 : 16% less resources required for the same job
22. Capture 200 more students with same number of resources by including the excluded 209 universities
In 2017, TFA reduced the number of Universities from 503 to 294, reducing the number of accepted students
by 200. Earlier we found that 203 of those excluded Universities were from AUSS tier 3.
Using the AUSS model , if TFA had included those universities, the number of resources required would have
been 113. i.e if TFA had added one more recruiter ( 112 +1 =113 ) in 2017 as per AUSS model , it would have
gained additional 200 students in 2017
Finding 3 : Add 1 get 200 offer
23. Next steps: Going forward
Use the tier distribution proposed by AUSS for 2018 recruitment.
This is a dynamic model and can use fresh data to capture changing recruitment trends.
TFA can itself generate the tiers anytime it wants. Steps to do the same:
1. Ask the recruiters to fill form which runs on AHP - Generates weights to track student behaviour
2. Analyse the previous years acceptance percentage - Generates Universal ratings which signifies
parameters which have brought most numbers in the past.
3. Fill the values in the excel model and perform clustering.
The weights and Universal rating are the two features of AUSS model which will work towards
maximum resource utilization !
25. Prospect Type
Students in the final year must hunt for jobs as compared to juniors hence they
should be the target applicants.
Met ?
Students who are motivated to get the job are the ones who do networking with
the recruiter.
Source
The ones who make efforts to reach out to TFA for info sessions
are highly motivated.
GPA
It is possible that the students with higher GPA tend to go for corporate jobs . Also
in general less students will have GPA > 3.6
Deadline
We can observe two peaks in the graph. This means we have students who are
inclined to apply early and others who finish just before the deadline.
Major 1
Students having high interest in Psychology, Political Science, English
etc are more inclined towards education unequity
University Selectivity
Students in more selective university tend to apply more for TFA. This could be
because students in most selective university tend to go for corporate jobs.
Exhibit 1: Variable Assumptions
Why motivation imp - More opportunities available to students due to growth in economy
Total volume of applicants decreased from 3022 in AY 2016 to 2388 in AY 2017. Volume of accepted candidates also reduced by 21%
Why motivation imp - More opportunities available to students due to growth in economy
Total volume of applicants decreased from 3022 in AY 2016 to 2388 in AY 2017. Volume of accepted candidates also reduced by 21%
Combine each of the important parameters to form a ‘Student Scorecard’
Score will be an indicator of attractiveness as well as motivation.
It will need two things:
Weights: for relative importance, Points Standardization: to stress on params which traditionally brought more students
We observed that the process was subjective and hence we use a method called Analytic Hierarchy Process which was suitable
We took the perspective of 1 TFA volunteer and college recruiters and asked them to rate each pair of parameters based on their exp
And we obtained the results as seen above: Sr/Jr is most imp, U.S news selectivity was 2nd in importance etc.
To standardize the scale
Here, since the objective is to maximize student acceptance rate, we took the historical acceptance as a percentage of total acceptance for each category and converted it to a scale of 10
Like in the Eg: Qualitative variable ‘Met’ has 2 values: yes and no. To assign a score to ‘yes’, we looked at the % accepted in ‘yes’ and converted it to a 10 point scale
By the process and using the two methods, student score for each student was obtained as seen
We needed to combine the individual scores & map them to of the student to his/her univ and added their score to obtain an aggregate score
This indicated a combined student quality level in each university and now it was ready for harvesting: that is deciding the university tiers
The more the number of students accepted from a university, the higher its score will be
To decide on the number of tiers and tiers for each university
Combine AUSS, # applicants, Region and Alumni to get clusters
No more going to schools to find 50 students but getting only 5