# PREDICTIVE MODELING WITH MAJOR DONORS

Document Sample

```					PREDICTIVE MODELING
WITH MAJOR DONORS
The 2002 CARA Summer Workshop
Peter Wylie, Margolis Wylie Associates
PREDICTIVE MODELING: AN
OVERVIEW
WHAT IS IT?
WHY DO IT?
HOW DO YOU DO IT?
DOES IT REALLY WORK?
SHOULD YOU DO IT YOURSELF OR
HAVE IT DONE FOR YOU?
WHAT IS IT?
A WAY TO USE THE RICHNESS OF
IDENTIFY GOOD PROSPECTS

CAN GET TECHNICALLY
COMPLICATED BUT CONCEPTUALLY
SIMPLE
WHY DO IT?
1. You can learn huge amounts about who
2. You can save big money on appeals
3. You can generate lots more money for
How Do You Do It?
1.   DECIDE WHAT YOU WANT TO PREDICT
2.   PICK A LIMITED NUMBER OF POSSIBLE
PREDICTORS
3.   BUILD A FILE (RANDOM SAMPLE FROM YOUR
DATABASE)
4.   IMPORT THE FILE INTO A STAT SOFTWARE
APPLICATION
5.   SPLIT THE FILE IN HALF AT RANDOM
6.   SEARCH FOR PREDICTORS ON ONE HALF OF THE
FILE AND BUILD A MODEL
7.   CHECK THE MODEL OUT ON THE OTHER SAMPLE
8.   TEST THE MODEL
9.   IMPLEMENT THE MODEL
Let’s Walk Through An
Example
from The U of Minnesota Annual Fund
Step 1:
Decide What You Want To Predict

Randy Bunney & Pete Wylie decide on:

– Life to date giving
Step 2:
Pick a Limited Number of Possible
Predictors

These are some of the ones we chose:
– Job Title
– Gender
– Birth Date
– Marital Status
– Degree Count
– Bus Phone
– Email
Step 3:
Build A Random Sample
IS folks built an Excel file of 10,000
random records from a database with over
700,000 living alumni and friends
Steps 4 & 5:
Importing And Splitting
Working over the phone, we imported
the excel file into the stat application
file into two halves of 5,000 records
each
Step 6: (on 1/2 of the file )
Find predictors. Build a model.
Some of promising predictors we found:
– Job title listed (Yes/No)
– Marital status listed as “married” (Yes/No)
– Born before 1948 (Yes/No)
Job Title Status

90 MEAN (AVERAGE) NUMBER OF GIFTS GIVEN BY
80 WHETHER OR NOT JOB TITLE WAS LISTED IN
DATABASE
70
60                                         6.2

50              7.0                              East
40              6.0
West
5.0
30              4.0
North
20              3.0

2.0
10              1.0
0             0.0
NOT LISTED   JOB
1st Qtr 2nd Qtr 3rd Qtr 4thTITLE LISTED
Qtr
Marital Status

90       MEAN (AVERAGE) NUMBER OF GIFTS GIVEN BY
WHETHER OR NOT LISTED AS "MARRIED" IN DATABASE
80
70                                              4.5

60            4.5

50            4.0                                       East
3.5
40            3.0
West
30
2.5
1.2                        North
2.0
20            1.5
1.0
10            0.5

0           0.0
NOT LISTED      LISTED AS MARRIED
1st Qtr 2nd Qtr 3rd Qtr 4th Qtr
Age as a Factor

90 MEAN (AVERAGE) NUMBER OF GIFTS GIVEN BY
WHETHER OR NOT BORN BEFORE 1948
80
70
6.2
60
7.0
50            6.0
East
40            5.0                                  West
30
4.0
1.8
North
3.0
20            2.0

10            1.0

0           0.0
OTHER         BORN BEFOR '48
1st Qtr 2nd Qtr 3rd Qtr 4th Qtr
The Model We Came Up With

Score = (Bus Phone Good) + (Home
Phone Good) + (Job Title Listed) +
(Married) + (Born Before 1948)
Step 7:
Check Model Against the Other Sample
MEAN (AVERAGE) NUMBER OF GIFTS GIVEN BY SCORE LEVEL ON
90                   UMINN CROSS VALIDATION SAMPLE
80
16.0
70                                                             14.3
14.0
60
12.0
50        10.0
9.7
10.1         East
40
West
30         6.0                                                     North
3.5
20         4.0

2.0             1.4
10                0.4
0.0
0              S0       S1       S2       S3       S4       S5

1st Qtr 2nd Qtr 3rd Qtr 4th Qtr
SCORE LEVEL
Step 8:
Test the Model
90   UNIVERSITY OF MINNESOTA MAIL
80    CAMPAIGN: PERCENTAGE OF
70       GIVING BY SCORE LEVEL
60
10.0                                   8.6
50                                                East
8.0                             6.2
40                                                West
6.0                 3.5   4.2
30 %                                              North
4.0           1.7
20      2.0   0.2
10      0.0
S0    S1    S2    S3    S4    S5
0
1st Qtr 2nd Qtr SCORE LEVEL Qtr
3rd Qtr 4th
More Testing

UNIVERSITY OF MINNESOTA MAIL
CAMPAIGN: MEAN (AVERAGE)
LEVEL

40.00                                      35.42

30.00
\$ 20.00
8.02
10.00                        5.56
2.96
0.14   0.83
0.00
S0     S1     S2     S3     S4     S5
SCORE LEVEL
Step 9:
Implement The Model

UM decided to only re-appeal to
records scored 3 or above.
An Other Experiment
Oklahoma State University

SCORE = (Bus Phone Yes) + (Oc-Tit Listed) +
(Emplr Listed) + (Bus City Listed) + (Stud Org
Listed) + (Alum Member) + (Mrtl Code Listed)
+ (Child Fir Nam Listed) + (Child Birth Date
Listed)
Oklahoma State University
Pledges By Score

OKLAHOMA STATE UNIVERSITY PHONE CAMPAIGN:
PERCENTAGES OF ALUMS MAKING PLEDGES BY
SCORE LEVEL

25                                                    23

19
20
16        17
15

15               12                                         13
%
10          9

5
5

0
S0   S1   S2    S3       S4        S5    S6    S7    S8
SCORE LEVEL
Oklahoma State University
Dollars Pledged by Score

OKLAHOMA STATE UNIVERSITY PHONE CAMPAIGN:
MEAN (AVERAGE) DOLLARS PLEDGED BY SCORE
LEVEL

18.00                                                               16.41
15.97
16.00                                               15.29

14.00

12.00

10.00
\$                                                7.74
8.00                  6.16   6.66   6.28
6.00

4.00           2.89
2.00    1.15

0.00
S0     S1     S2     S3     S4      S5     S6      S7      S8
SCORE LEVEL

Will modeling work as well
as it does for the annual fund?
Data From 5 Other Schools

Large samples of records noting if a
person:
– Had given a total of \$1,000 or more or not to
the school
– Had an age of 52 or older listed or not
PERCENTAGES OF DONORS GIVING \$1000 OR MORE BY WHETHER OR
NOT A BUSINESS PHONE IS LISTED

20     19

18
16
16

14
12
12

% 10                                                          BUS PHONE LISTED
9
NOT LISTED
8
6                     6
6                                           5
4
4                     3

2                                                      1

0
SCHOOL A   SCHOOL B   SCHOOL C   SCHOOL D   SCHOOL E
E-mail Status

PERCENTAGES OF DONORS GIVING \$1000 OR MORE BY WHETHER OR
NOT AN E-MAIL ADDRESS IS LISTED

25                24

20

15                           14

%          12                                                   E-MAIL LISTED
NOT LISTED
10                                       9
7
6          6
5
5                      4

1
0
SCHOOL A   SCHOOL B   SCHOOL C   SCHOOL D   SCHOOL E
Giving and Age
PERCENTAGES OF DONORS GIVING \$1000 OR MORE BY WHETHER OR
NOT THEY ARE LISTED AS 52 OR OLDER

30

26
25
23

20

15
% 15      14                                                   52 OR OLDER
OTHER

10

5           4                     4          4      4
3
1
0
SCHOOL A    SCHOOL B   SCHOOL C   SCHOOL D   SCHOOL E
LET’S LOOK AT AGE AT ONE OF
THESE SCHOOLS
PERCENTAGE OF RECORDS GIVING \$50,000 OR MORE BY BIRTH YEAR

35          33

30

25

20                                                               17
%
15

9
10
4
5
1

0
1953 OR    1954-1962   1963-1969   1970 OR LATER   NOT LISTED
EARLIER
Multiple Factors
PERCENTAGES OF DONORS GIVING \$1000 OR MORE BY WHETHER
NONE, ONE, TWO, OR THREE ATTRIBUTES (BUSPHONE, E-MAIL, AND
52 OR OLDER) ARE LISTED

45                                     44

40                                                         38

35                                33                                         33

30
27
NONE
25
%                  22                                                                                  ONE
20                                                                                                 TWO
16                                      15
15                           14                                         14                         THREE
11                                     11
10                                                                  9
5
5     3                                      4                  3                     3
2                                                        1
0
SCHOOL A            SCHOOL B           SCHOOL C            SCHOOL D          SCHOOL E
Modeling:
In-house Or Have It Done For You?

• Doing it all by yourself isn’t feasible.

• Besides, there are excellent products and
services out there that shouldn’t be
ignored.
A NEW KIND OF RESEARCHER
• Without an inside specialist, the data
enhancement products and services you
purchase are less likely to be used
effectively. A blunt question. In the past five years, have you
spent more than \$25,000 on enhancing your database with
estimates of wealth, capacity to give, and so on, only to have the
information untapped and unused by your development officers?
Why buy the stuff if you’re not going to use it? An inside data analyst
can not only help you use data effectively, he or she also can be a
persistent thorn in your side until you do use it.

• A good data analyst is worth the effort and
cost of creating a new position.
In-house Modeling & Analysis

Worth the consideration because of:
•   The richness of info in your database
•   Speed
•   Continuity
•   The “Big Picture”
•   Vendor screening