1. The Value of Purchase History Data in Target Marketing (1996)
Peter Rossi - Robert McCulloch - Greg Allenby

2. Authors

3. Peter Rossi
James Collins Professor of Marketing, Statistics, and Economics at UCLA
PhD from University of Chicago
Professor at Chicago when this paper was published

4. Robert McCulloch Professor of Statistics, Arizona State University
PhD from the University of Minnesota
Professor at Chicago when this paper was published

5. Greg Allenby
Professor of Marketing/Statistics at Ohio State University
PhD from University of Chicago

6. Rossi, Allenby, McCulloch (2005)

7. Motivation

8. Motivation Increased availability of consumer-level purchase data
Few firms take advantage of data
Data processing/storage costs are rapidly declining
Authors believe that large gains in revenue can be obtained by implementing a customer-centric targeted marketing strategy

9. Goals of Paper
Assess how targeted marketing strategies, each utilizing different information content, affect expected revenue
Propose Hierarchical Bayesian model to accomplish this

10. Random Coefficients Model for Consumer Heterogeneity

11. Random Coefficients Model For Costumer Heterogeneity
Goal of proposed model
The authors outline a few goals for their model
Accommodate household-specific inferences from individual-level parameter estimates
Allow for both observed and unobserved heterogeneity

12. Random Coefficients Model For Costumer Heterogeneity
Recall random Intercept model1
Illustration in two dimensions
Model
Distribution for individual-level intercepts
1: Fahrmeir, Ludwig, Thomas Kneib, Stefan Lang, and Brian D.. Marx. "Mixed Models." Regression: Models, Methods and Applications. Berlin: Springer, N. pag. Print.

13. Random Coefficients Model For Costumer Heterogeneity
Recall random coefficient and slope model2
Illustration in two dimensions
Model
Distribution for individual-level slope
2: Fahrmeir, Ludwig, Thomas Kneib, Stefan Lang, and Brian D.. Marx. "Mixed Models." Regression: Models, Methods and Applications. Berlin: Springer, N. pag. Print.

14. Random Coefficients Model For Costumer Heterogeneity
Hierarchical Bayesian Model
Demographics
N(Δzh,Vβ)
Betas β1 β2 … βh
Utility …
y11
y12
y1t … …
y21
y22
y2t
yh1
yh2
yht …
Choice …
I11 …
I12
I1t
I21
I22
I2t
Ih1
Ih2
Iht

15. Random Coefficients Model For Costumer Heterogeneity
Hierarchical Bayesian Model
Demographics
zh: dx1 vector of demographics variables
Δh: effect of demographic variables on β
βh: model mean of β as function of z
Betas
yht: Utility (linear predictor) follows a multivariate regression
Utility
X contains vector of product features, brand loyalties, and log of prices
Choice
Iht: Choice is defined as the maximum of utilities on each choice occasion
Probit link function

16. Random Coefficients Model For Costumer Heterogeneity
Specification of Priors
Three parameters across households:
Λ: Covariance matrix of utility errors, ε
Δ: Effect of demographics on mean β
Vβ: Covariance matrix of β
Conjugate priors for Δ, Vβ. Very non-informative for all:
Δ: Normal
Vβ: Inverse Wishart
Λ: Independent inverted gamma distribution
Λ = Lambda

17. Random Coefficients Model For Costumer Heterogeneity
Why use a Bayesian Model?
Reasons to use hierarchical bayesian approach:
We can make household-specific inferences from individual-level parameter estimates
We need a model that can effectively estimate parameters given only a few (in some cases one) observations for each individual (Gibbs Sampling)
Characterize uncertainty around household-level parameters

18. Assessing Value of Information Sets (Model Estimation)

19. Alternative Information Sets
Goal: Assess information content of various data/information sets available for targeted marketing
Variance in data:
Short/long
Causal data/No Causal Data
Authors propose typology for various information sets (table 1)

20. Description of Data
Scanner panel dataset of tuna purchases in Springfield Missouri
400 of 775 households randomly selected
At least 1.5 years in data set
Five tuna brands

21. Predictor Variables Causal: Demographics (Δ): Brands/Features:
Chicken of the sea (water) [REFERENCE]
Starkist (water)
House Brand (water)
Chicken of the sea (oil)
Starkist (oil)
Price (log)
Demographics (Δ):
Household income
Family size
Retirement (dummy)
Unemployed (dummy)
Female head of household (dummy)
Causal:
In-store display (dummy)
Feature ads (dummy)

22. Effect of demographics (Δ) on βh
Low income, unemployed
High variance. Demographics
offer limited explanation
Retired, unemployed

23. Effects of brand (βh) on choice
Strong brand pref.
Weak brand pref.
Individual-level estimates for 10 households
Marg refers to marginal predictive dist. or the overall effect. This would be used in the ‘base information set
One observation is model used with data on first observed purchase
Note how informative the parameters estimates from ‘full information’ model are. Decreases as you remove information
Demos only: Demographic data only
Very limited value
Similarity of top two plots: Adding causal data yields little information gain
Increased variation. Wider point estimates

24. Effects of price (βh) on choice
Individual-level price sensitivities
Full information model gives best estimates of the price coefficient
Even one observation yields decent estimation of price coefficient

25. Model Estimation
Conclusion: Our analysis of model parameters suggests that individual-level purchase history data could be of great use in customizing marketing activities. Further, demographic information is of limited value.
We will formalize this analysis next.

26. Assessing Value of Information Sets (Formal Metric)

27. Metrics For Couponing in Targeted Marketing
F is face value of coupon
Pr(i) is purchase probability of ith choice occasion
M is manufacturer margin
Goal: find F that maximizes revenue (π)
‘Plugging in’ model point estimates will overstate revenue
Want to incorporate uncertainty
Choose F to maximize expected net revenue, averaged over distribution of β (decision-theoretic approach)
Incremental Sales:
Expected Net Revenue (π)

28. Targeted Marketing Optimization
Expected revenue from one household for values of F
Dots: Parameter estimates from posterior means (plug- in)
Bars: Mean from decision- theoretic calculation
Note optimal price is the same here, but not for all households
High level of uncertainty from small number of household observations
Optimal F

29. What is the value of targeted marketing?
We can compare by computing expected revenue per household from different targeting strategies:
i.e. Full targeting, choices only, blanket, etc…
We use optimal F, F*, to calculate expected revenue (Π), then average over all households
Gain Relative to Blanket:
(Revenuetarget - RevenueBlanket)/(RevenueBlanket - Revenuenone)
This allows us to compare how the implementation of a targeting strategy compares to the implementation of a blanket strategy
Large gains: even the one obs. information set yields a 56% increase in revenue

30. Conclusion
Targeted couponing strategies based on customer purchase history can have large, positive effects on revenue
Even strategies that use short purchase histories, can have positive impact
As data processing/storage costs continue to decline, this type of modeling will be increasingly easier to for managers to implement