Adaptive discrete choice designs Presentation held at 2008 · Adaptive discrete choice designs...

Adaptive discrete choice designs

Presentation held at 2008

Boris Vaillantwww.quantitative-consulting.eu

www.quantitative-consulting.eu 2

Product attributes

The goal of discrete choice analysis in marketing is to assess the influence of product- and service-attributes on customers’ choice behaviour

Brand

Engine

Price Consumption

HP

Transmission... Flexibility

Insulation Shielding

Control fibres

Fibres

Brand

Cross-section

Mechanical resistance

Price

Marking

Packaging

...

Efficacy

Brand

Price Application

Dosage

Side effects

Treatment duration

...


Product comparisons

In product comparison questionnaires, respondents indicate their preferred choice or purchase pattern in a series of ‘choice tasks’

Typical example:Doctors are asked to

indicate the prescription share for each of the shown treatments (e.g. for the last

10 patients with the corresponding indication)


The preference of individual customers or customer groups for the different product elements is parametrised and can be used to calculate the preference for existing or new products

Car example:

Sedan segment

Utility parameters

MB

BMW

Audi

Brand

20

0

10Utility model 1Utility model 1

-10 0 10 20 30 40 50 60

Total

Tiptronic

Medium

7 L

130 HP

34 KEUR

Audi

51

7

4

18

0

12

10

10 L

9 L

8 L

7 L

Consumption

0

7

14

18

42 KEUR

38 KEUR

34 KEUR

30 KEUR

Price

0

4

12

24

High

Medium

Base

Equipment

10

4

0

Tiptronic

Manual

Gearshift

7

0

190 HP

170 HP

150 HP

130 HP

HP

7

6

3

0

Utility model 2Utility model 2

-10 0 10 20 30 40 50 60

Total

Manual

Base

9 L

150 HP

42 KEUR

MB

30

0

0

7

3

0

20


In the marketing practice, discrete choice data often feeds larger market models in which companies can test the outcome of different strategies

Customer preferences

Demand curve and profit-optimal price

Product information● Target segments● Description of existing and

future products

Price information● Most important

products in the market● Changes in time

Market information● Market size● Market shares

Market simulationSimulation of customers

choice in scenarios Price changes New products Competitive reaction

● Customer preferences as measured via the discrete choice questionnaire

● Either individual or group preferences

● Group-wise results or aggregation of individual results

● Obtain market effects e.g. for a price change

● Determine optimal new price

Additional information from the survey

● Price barriers● Brand value● Strengths and weaknesses

Market simulation


Marketing applications: Example of a simulation tool


● Substitution effects strongly influence the optimal price decision

● Complexity of substitution decisions is best captured by models based on individual choice data

One practical challenge is the estimation of substitution (or portfolio-) effects. This often requires that the parameters of the model be estimated for each individual customer

Marketing applications: Substitution effects

50 100 150 200

050

100

150

Demand including substitutes

Price index

Dem

and

inde

x

Sales new product aloneEffects from substitute AEffects from substitute BEffects from substitute C

50 100 150 200

6080

100

120

Profit including substitutes

Price indexP

rofit

inde

x

Profit new product aloneEffects from substitute AEffects from substitute BEffects from substitute CProfit effect portfolio


Creating optimal bundles is another area which requires knowledge of the individual preferences

● Testing of bundling scenarios leads to the best bundles to sell and their optimal price

● Individual preferences and WTP for the different products are estimated in a discrete choice study

● From this, one can estimate which products customers will purchase and in which combination

Marketing applications: Bundling

Demand for equipment X and Y

WTP equipment X

WTP

equ

ipm

ent Y

Effect of bundlingPrice of XPrice of YPrice of bundle (X+Y)

Customer buys moreCustomer buys lessNo change

0 2000 4000 6000 8000

050

010

0015

00

Profit of equipment bundle

Price of bundle

Pro

fit o

f bun

dle

Profit equipment XProfit equipment YPrice X + Price YPrice of bundle (X+Y)


Goals of optimal discrete choice design

The goal of discrete choice design is to minimise the error of parameter estimates while keeping respondent burden at a minimum

Respondents● Reduce task complexity● Reduce number of choice

tasks● Avoid respondent fatigue

Consultant Obtain reliable and valid results Have a variety of design options

Study sponsor Reduce cost and

obtain reliable and valid results


Discrete choice modelling


Classical design criteria (1)


Classical design criteria (2)


Bayesian adaptive design (1)


Bayesian adaptive design (2)

H(u)

H(u|Z)

H(Z)

H(Z|u)I(Z,u)


Sequential Monte Carlo (1)


Sequential Monte Carlo (2)

(U1)Start

(U2)Reweight

(U3)Resample

(U4)Move

(U2)Reweight

...


R-implementation

We created a test-suite in R to compare the different design strategies in simulations. Our tests are based on a typical scenario from pharmaceutical market research

This is a typical scenario from pharmaceutical market research. Here, doctors are asked to estimate the prescription share of each of the presented treatments. Compared to simple choice questions, this provides sufficient information for individual-level estimation


Results (1)

B1 Det B2 Max B3 Ent B5 Simple Fixed Random

0.0

0.1

0.2

0.3

0.4

0.5

RMSE for Utilities

M=0.248P=2e-04

M=0.254P=0.002

M=0.253P=0.001

M=0.269P=0.2

M=0.282P=NA

M=0.32P=3e-05

5 10 15 20

0.3

0.4

0.5

0.6

0.7

0.8

RMSE vs. # questions

# of questions

RM

SE

B1 DetB2 MaxB3 EntB5 SimpleFixedRandom


Results (2)

Det Max Ent Sim Fix Rnd

0.10

0.15

0.20

0.25

0.30

0.35

0.40

ML D-criterion: Avg SE

-2 -1 0 1 2

-3

-2

-1

0

1

2

3Bias of estimates

'True' values

Est

imat

ed v

alue

s

Bias factorsB1 Det : 1.108B2 Max : 1.088B3 Ent : 1.102B5 Simple : 1.102Fixed : 1.067Random : 1.016


Performance

1000 5000 10000 20000 40000

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

RMSE vs. MC sample size and processing time

MC sample size

RM

SE

Update=0.45 s Select=0.51 s



Update=11 s Select=13 s

Update=22 s Select=24 s

M=0.278P=NA

M=0.248P=0.002

M=0.263P=0.1

M=0.241P=1e-04

M=0.242P=2e-04


References (1)


References (2)

Adaptive discrete choice designs Presentation held at 2008 · Adaptive discrete choice designs...

Documents

Transcript of Adaptive discrete choice designs Presentation held at 2008 · Adaptive discrete choice designs...