Adaptive discrete choice designs Presentation held at - PowerPoint PPT Presentation

Adaptive discrete choice designs Presentation held at 2008 Boris Vaillant www.quantitative-consulting.eu

Product attributes The goal of discrete choice analysis in marketing is to assess the influence of product- and service-attributes on customers’ choice behaviour Shielding Insulation Consumption Price Packaging Fibres HP Price Marking Engine Brand Mechanical ... resistance Brand Cross- section Flexibility Control ... fibres Transmission Application Price Dosage Brand Side effects Efficacy ... Treatment duration 2 www.quantitative-consulting.eu

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) 3 www.quantitative-consulting.eu

Utility parameters 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 C a Brand Consumption r e S x e Utility model 1 Utility model 1 a d m a 7 L 18 n p 10 Audi l e s e : Audi 10 g m 14 8 L e n 34 KEUR 12 0 t BMW 7 9 L 130 HP 0 7 L 18 MB 20 0 10 L Medium 4 Tiptronic 7 Price Equipment Total 51 30 KEUR 24 0 Base -10 0 10 20 30 40 50 60 12 34 KEUR 4 Medium 4 38 KEUR Utility model 2 Utility model 2 High 10 0 42 KEUR MB 20 42 KEUR 0 HP Gearshift 150 HP 3 0 130 HP 9 L 7 0 Manual 150 HP 3 Base 0 Manual 0 6 170 HP Total 30 Tiptronic 7 7 190 HP -10 0 10 20 30 40 50 60 4 www.quantitative-consulting.eu

Market simulation In the marketing practice, discrete choice data often feeds larger market models in which companies can test the outcome of different strategies Customer preferences Product information Additional information from the survey Target segments ● Price barriers ● Description of existing and ● future products Brand value ● Strengths and weaknesses ● Demand curve and profit-optimal price Market simulation Simulation of customers choice in scenarios  Price changes  New products  Competitive reaction ● Group-wise results or Market information Price information aggregation of individual Market size ● Most important ● results Market shares ● products in the market ● Obtain market effects e.g. Changes in time ● for a price change ● Determine optimal new price ● Customer preferences as measured via the discrete choice questionnaire ● Either individual or group preferences 5 www.quantitative-consulting.eu

Marketing applications: Example of a simulation tool 6 www.quantitative-consulting.eu

Marketing applications: Substitution effects 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 Profit including substitutes Demand including substitutes 150 120 100 100 Demand index Profit index 80 50 Profit new product alone Effects from substitute A Sales new product alone 60 Effects from substitute B Effects from substitute A Effects from substitute C Effects from substitute B Profit effect portfolio Effects from substitute C 0 50 100 150 200 50 100 150 200 Price index Price index ● Substitution effects strongly influence the optimal price decision ● Complexity of substitution decisions is best captured by models based on individual choice data 7 www.quantitative-consulting.eu

Marketing applications: Bundling Creating optimal bundles is another area which requires knowledge of the individual preferences Demand for equipment X and Y Profit of equipment bundle Effect of bundling Profit equipment X Price of X Customer buys more Profit equipment Y Price of Y Customer buys less Price X + Price Y 1500 Price of bundle (X+Y) No change Price of bundle (X+Y) WTP equipment Y Profit of bundle 1000 500 0 0 2000 4000 6000 8000 WTP equipment X Price of bundle ● Individual preferences and WTP for the ● Testing of bundling scenarios leads to different products are estimated in a the best bundles to sell and their optimal discrete choice study price ● From this, one can estimate which products customers will purchase and in which combination 8 www.quantitative-consulting.eu

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 Consultant Study sponsor Respondents  Obtain reliable and valid results ● Reduce task complexity  Reduce cost and  Have a variety of design options obtain reliable and ● Reduce number of choice valid results tasks ● Avoid respondent fatigue 9 www.quantitative-consulting.eu

Discrete choice modelling 10 www.quantitative-consulting.eu

Classical design criteria (1) 11 www.quantitative-consulting.eu

Classical design criteria (2) 12 www.quantitative-consulting.eu

Bayesian adaptive design (1) 13 www.quantitative-consulting.eu

Bayesian adaptive design (2) H(Z) H(u|Z) I(Z,u) H(Z|u) H(u) 14 www.quantitative-consulting.eu

Sequential Monte Carlo (1) 15 www.quantitative-consulting.eu

Sequential Monte Carlo (2) (U1) (U2) (U3) (U4) (U2) ... Start Reweight Resample Move Reweight 16 www.quantitative-consulting.eu

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 17 www.quantitative-consulting.eu

Results (1) RMSE for Utilities RMSE vs. # questions 0.5 B1 Det 0.8 B2 Max B3 Ent B5 Simple 0.4 0.7 Fixed Random 0.3 0.6 RMSE 0.5 0.2 0.4 0.1 M=0.248 M=0.254 M=0.253 M=0.269 M=0.282 M=0.32 P=2e-04 P=0.002 P=0.001 P=0.2 P=NA P=3e-05 0.3 0.0 # of questions B1 Det B2 Max B3 Ent B5 Simple Fixed Random 5 10 15 20 18 www.quantitative-consulting.eu

Results (2) ML D-criterion: Avg SE Bias of estimates 3 0.40 2 0.35 Estimated values 0.30 1 0.25 0 Bias factors B1 Det : 1.108 0.20 -1 B2 Max : 1.088 B3 Ent : 1.102 0.15 B5 Simple : 1.102 -2 Fixed : 1.067 Random : 1.016 0.10 'True' values -3 Det Max Ent Sim Fix Rnd -2 -1 0 1 2 19 www.quantitative-consulting.eu

Performance RMSE vs. MC sample size and processing time 0.7 Update=2.5 s Update=11 s Select=2.7 s Select=13 s Update=0.45 s Update=5.2 s Update=22 s 0.6 Select=0.51 s Select=5.8 s Select=24 s 0.5 0.4 RMSE 0.3 0.2 0.1 M=0.278 M=0.248 M=0.263 M=0.241 M=0.242 P=NA P=0.002 P=0.1 P=1e-04 P=2e-04 0.0 MC sample size 1000 5000 10000 20000 40000 20 www.quantitative-consulting.eu

References (1) 21 www.quantitative-consulting.eu

References (2) 22 www.quantitative-consulting.eu