AT&T PowerPoint® Tutorial

Andrew Johnson

Description

Propensity Score Matching in Marketing Program EvaluationsApril 27, 2007NISS Affiliates Annual MeetingBackground: Common Marketing Analysis SituationDid our marketing program work, and how well? No controlled experimentStuck with Treated vs. non-Treated, or Buyers vs. non-BuyersRestrospective, observational study…But fortunately LOTS of data!!• Extensive contact, purchase, usage, switching, care calling • Demographics• Internal predictive modelsExample: Airline Miles Loyalty ProgramPage 2Background: Common Marketing Analysis SituationAre customers with X characteristic more likely to Y…all other factors being equal? • Difficult to isolate characteristic X• Important strategic question• Observational study…Again, LOTS of data!!Example: Word-of-Mouth ReferralPage 3Example 1: Airline Miles Loyalty Program or…what I wish I’d known• Optional, no cost to consumer• AT&T spend -> Airline Miles• No program-level control group• Very expensive• Unadjusted comparison possibly misleading:Self-selectionBig spenders,Occasional major highly targetedenrollment campaignsPage 4Previous ApproachIsolate program variable after “controlling for” other observable factorsModel:Leave AT&T = Program + [tenure vars] + [usage vars] + [purchase history] + [switching history] + [demographics] + [Program interactions]Unobservables an issue – survey research helpedUneasy – profile indicates profound differences between program members and ...

Informations

• Publié par: Fupip
• Nombre de lectures: 28
• Langue: English

Extrait

Propensity Score Matching in Marketing Program Evaluations

April 27, 2007 NISS Affiliates Annual Meeting

Background: Common Marketing Analysis Situation

Did our marketing program work, and how well?

No controlled experiment Stuck with Treated vs. non-Treated, or Buyers vs. non-Buyers Restrospective, observational study

…But fortunately LOTS of data!! • Extensive contact, purchase, usage, switching, care calling • Demographics • Internal predictive models

Example: Airline Miles Loyalty Program

Page 2

Background: Common Marketing Analysis Situation

Are customers with X characteristic more likely to Y…all other factors being equal?

• Difficult to isolate characteristic X • Important strategic question • Observational study

…Again, LOTS of data!!

Example: Word-of-Mouth Referral

Page 3

c   ma emtn sn

Example 1: Airline Miles Loyalty Program

apgi

Very expensive

Unadjusted comparison possibly misleading:

•

•

or…what I wish I’d known

Page 4

No program-level control group

Optional, no cost to consumer AT&T spend -> Airline Miles

•

• •

Self-selection

Big spenders, highly targeted

ccOoisal najoma en r  llro

Previous Approach

Isolate program variable af ter “controlling for” other observable factors

Model: Leave AT&T = Program + [tenure vars] + [usage vars] + [purchase history] + [switching history] + [demographics] + [Program interactions]

Unobservables an issue – survey research helped

Uneasy – profile indicates profound differences between program members and non-members.

Page 5

Despite huge non-member population – programming challenge!!

Page 6

Each member matched to non-member within +/- 5%

Re-estimated model predicting Leave AT&T Æ more conservative estimate of program impact.

ytisocs  rP er sopo cedtouszemicod  oPro (wo)man’spropen cu tdapoce hrpdob●High-tty to huorP•imix eT●eeruny onagusTe : •phle otsrf cano4 ee srollh enmatcely solc  ot dengised, upro gonisarmp l)demokil  noit (doohile

Example 2:  Word-of-Mouth Referral

or….An unexpected and very long digression

New, innovative AT&T service, early market entrant

No overt viral or buzz marketing initially – BUT we hoped that WOM would positively influence sales

Early exploratory analyses suggestive of viral activity.

Again, unadjusted comparison possibly misleading.

Page 7

Homophily

Example 2: Word-of-Mouth Referral, cont’d

How to estimate WOM effect on sales?

AT&T network provides WOM framework:

• Start with early subscribers

• Establish their WOM connections – people they talked to before buying

• Were these WOM-connected people more likely to buy, compared to similar WOM-less people?

Page 8

Early Subscriber, WOM Connections, and Matches

Page 9

ES

Initial Approach

Isolate WOM effect after “controlling for” other observable factors. Population is potential buyers.

Model: Buy = WOM-connection + [tenure vars] + [usage vars] + [purchase history] + [switching history] + [demographics] + [WOM interactions]

Page 10

Page 11

With WOM Connection

Initial Approach WOM connection associated with large differences in usage and demographics.

Usage Variable

WOM-less

Wealth Variable

With WOM Connection

WOM-less

 hto eraosla rts   g onedprtoic  o rs  fubiygn! B