Conjoint Analysis (328053-M-6)
Fall 2020 Summary
Course taught by Max Pachali and Francesca Verna.
Includes materials of seven lectures and four tutorials.
Examples are given in italics.
Sacha Lena, Fall 2020
,Lecture 1 - Introduction
Conjoint analysis: survey-based technique that allows the analyst to understand people’s
preferences for a product, service, brand, medical treatment, job, course, and especially the
trade-offs they make in making choices
→ In direct surveys, respondents might consider all attributes important (not informative)
→ Enforces tradeoffs between attributes as in a real purchase occasion
→ All attributes evaluated at once
→ Respondents evaluate complete products with both strong and weak attributes
→ Reduces problem of socially desirable answers
→ Adds realism: in real-life, consumers evaluate products rather than isolated attributes (do they
consciously know which attributes matter?)
→ Straightforward (suitable software (Sawtooth) available)
Why should we run hypothetical choice experiments if firms can increasingly use large amounts of
transactional purchase data?
1. Lack of experimental price variation required to learn consumers’ preferences
2. Conjoint allows to measure consumer preferences for products or attribute levels not yet
introduced in the marketplace
3. Especially relevant (pricing of new product innovations)
Field experiments are prominent alternatives for the goal of learning consumer preferences.
However, field experiments are often difficult to conduct and not feasible in high ticket product
categories (cars, laptops). Field experiments are limited to products already existing in the
marketplace.
Conjoint Design
Ranking-based conjoint: choose the most-preferred product, then the second most-preferred
product, until the least-preferred product
Rating-based conjoint: give a score to each product in turn
→ (Problem) Not realistic: in real-life, we buy rather than rate products
→ (Problem) Not clear whether spread in ratings is due to real preferences or due to response style (small
spread; weak preferences or cautious answers?)
→ (Problem) Implications for sales levels and market shares are not clear
→ Sales and shares result from consumer choices, not ratings
→ What would be the rating threshold?
→ Why don’t we ask the respondents to choose a product directly, rather than asking them to rate
products?
Choice-based conjoint: choice between different variants (choose most preferred product only)
→ If you do this repeatedly, you see what consumers trade off (wine example: price, region)
→ Record choices made by every customers during n tasks (choice sets)
→ In every choice set, a different combination of attribute levels is used; we can derive the effect of different
combination of attribute levels on choice
→ Preferences = attribute part-worths
→ When repeating the conjoint exercise across many customers, it can be detected whether different
customers have different preferences (customer-specific preferences)
→ Range: 10 and 50 choices; ideally (often not feasible due to time, intellectual constraints of respondents)
Sacha Lena, Fall 2020
, Choice-based conjoint advantages:
1. Trade-offs are enforced even more
2. Realistic: the choice-setting mimics real-life
3. No-choice option (‘’None of the offered alternatives is attractive’’, ‘’I would like to stick to
my current product’’) → sales proxy (market share and related sales)
→ Car dealer: rather sticking to current car than buying a new one
→ Supermarket: no organic milk available, then rather no milk at all
4. Avoids the need of ad-hoc rules to predict market shares
5. No subjective scaling (no rescaling of numerical scales; can be done mathematically, but it
isn’t intuitive)
6. Choice is cognitively less demanding than ratings
Choice-based conjoint disadvantages:
1. Hypothetical bias: respondents’ product choices (potentially at very large prices) might be
influenced by the experimental setting (with no consequences for actual purchase behavior
in the real world)
→ No real consequences for monetary value, well-being
→ Consumers tend to be less price-sensitive (‘’It is just the experiment’’)
→ Might bias optimal price you get from choice-based conjoint
2. Small individual level data: respondents become fatigue if exposed to a large number of
choice tasks that are actually required if the experiment includes large numbers of attribute
levels
3. Bayesian methods and prior specifications: Bayesian statistical methods help as they
efficiently pool information across respondents (shrinkage); however, some analysts regard
the inclusion and specification of priors as subjective
Conjoint Analysis
A statistical model is needed to translate/estimate results of consumer response data → logistic
regression
→ DV: product chosen (or not)
→ IV: product attributes
→ DV = logit(IV)
→ Choice = f(price, quality, color, speed, discount)
What if consumers have different tastes? (extremely price sensitivity, medium price sensitivity) →
Latent Class Analysis and Hierarchical Bayes
Aggregate level: same part-worths for all respondents
→ (Advantage) High precision (all respondents are combined)
→ (Disadvantage) Assuming same preferences may give misleading results
→ Method of analysis: logit
Segment level: different part-worths for different segments
→ (Advantage) Realistic (takes into account different preferences)
→ (Advantage) High precisions if all respondents are used in one big analysis
→ Method of analysis: LCA
Sacha Lena, Fall 2020
Fall 2020 Summary
Course taught by Max Pachali and Francesca Verna.
Includes materials of seven lectures and four tutorials.
Examples are given in italics.
Sacha Lena, Fall 2020
,Lecture 1 - Introduction
Conjoint analysis: survey-based technique that allows the analyst to understand people’s
preferences for a product, service, brand, medical treatment, job, course, and especially the
trade-offs they make in making choices
→ In direct surveys, respondents might consider all attributes important (not informative)
→ Enforces tradeoffs between attributes as in a real purchase occasion
→ All attributes evaluated at once
→ Respondents evaluate complete products with both strong and weak attributes
→ Reduces problem of socially desirable answers
→ Adds realism: in real-life, consumers evaluate products rather than isolated attributes (do they
consciously know which attributes matter?)
→ Straightforward (suitable software (Sawtooth) available)
Why should we run hypothetical choice experiments if firms can increasingly use large amounts of
transactional purchase data?
1. Lack of experimental price variation required to learn consumers’ preferences
2. Conjoint allows to measure consumer preferences for products or attribute levels not yet
introduced in the marketplace
3. Especially relevant (pricing of new product innovations)
Field experiments are prominent alternatives for the goal of learning consumer preferences.
However, field experiments are often difficult to conduct and not feasible in high ticket product
categories (cars, laptops). Field experiments are limited to products already existing in the
marketplace.
Conjoint Design
Ranking-based conjoint: choose the most-preferred product, then the second most-preferred
product, until the least-preferred product
Rating-based conjoint: give a score to each product in turn
→ (Problem) Not realistic: in real-life, we buy rather than rate products
→ (Problem) Not clear whether spread in ratings is due to real preferences or due to response style (small
spread; weak preferences or cautious answers?)
→ (Problem) Implications for sales levels and market shares are not clear
→ Sales and shares result from consumer choices, not ratings
→ What would be the rating threshold?
→ Why don’t we ask the respondents to choose a product directly, rather than asking them to rate
products?
Choice-based conjoint: choice between different variants (choose most preferred product only)
→ If you do this repeatedly, you see what consumers trade off (wine example: price, region)
→ Record choices made by every customers during n tasks (choice sets)
→ In every choice set, a different combination of attribute levels is used; we can derive the effect of different
combination of attribute levels on choice
→ Preferences = attribute part-worths
→ When repeating the conjoint exercise across many customers, it can be detected whether different
customers have different preferences (customer-specific preferences)
→ Range: 10 and 50 choices; ideally (often not feasible due to time, intellectual constraints of respondents)
Sacha Lena, Fall 2020
, Choice-based conjoint advantages:
1. Trade-offs are enforced even more
2. Realistic: the choice-setting mimics real-life
3. No-choice option (‘’None of the offered alternatives is attractive’’, ‘’I would like to stick to
my current product’’) → sales proxy (market share and related sales)
→ Car dealer: rather sticking to current car than buying a new one
→ Supermarket: no organic milk available, then rather no milk at all
4. Avoids the need of ad-hoc rules to predict market shares
5. No subjective scaling (no rescaling of numerical scales; can be done mathematically, but it
isn’t intuitive)
6. Choice is cognitively less demanding than ratings
Choice-based conjoint disadvantages:
1. Hypothetical bias: respondents’ product choices (potentially at very large prices) might be
influenced by the experimental setting (with no consequences for actual purchase behavior
in the real world)
→ No real consequences for monetary value, well-being
→ Consumers tend to be less price-sensitive (‘’It is just the experiment’’)
→ Might bias optimal price you get from choice-based conjoint
2. Small individual level data: respondents become fatigue if exposed to a large number of
choice tasks that are actually required if the experiment includes large numbers of attribute
levels
3. Bayesian methods and prior specifications: Bayesian statistical methods help as they
efficiently pool information across respondents (shrinkage); however, some analysts regard
the inclusion and specification of priors as subjective
Conjoint Analysis
A statistical model is needed to translate/estimate results of consumer response data → logistic
regression
→ DV: product chosen (or not)
→ IV: product attributes
→ DV = logit(IV)
→ Choice = f(price, quality, color, speed, discount)
What if consumers have different tastes? (extremely price sensitivity, medium price sensitivity) →
Latent Class Analysis and Hierarchical Bayes
Aggregate level: same part-worths for all respondents
→ (Advantage) High precision (all respondents are combined)
→ (Disadvantage) Assuming same preferences may give misleading results
→ Method of analysis: logit
Segment level: different part-worths for different segments
→ (Advantage) Realistic (takes into account different preferences)
→ (Advantage) High precisions if all respondents are used in one big analysis
→ Method of analysis: LCA
Sacha Lena, Fall 2020
