Presentation on theme: "Introduction to Adaptive Conjoint Analysis (ACA) Copyright Sawtooth Software, Inc."— Presentation transcript:
Introduction to Adaptive Conjoint Analysis (ACA) Copyright Sawtooth Software, Inc.
Background: What is Conjoint? A way to learn how people make decisions about products or services that are made up of different features. A questioning method which mimics real world: we show people products and let them tell us how much they like them. By varying the features of the product and analyzing responses, we can quantify how aspects of the product drive preference. We assume that the value (utility) of the product is equal to the sum of its parts (attributes): Computer = (Brand) + (Processor Speed) + (RAM) + (Monitor) + (Price)
Background: What is Conjoint? (cont.) The attributes we measure must be levelable –Brand: Dell, Compaq, IBM –Processor Speed: 200 MHz, 300 MHz –RAM: 32 Mbytes, 64 Mbytes, 128 Mbytes –Price: $1,500, $2,000, $3,000 If we learn how much value (utility) people have for each of these levels, we can add them up and predict how much they would like potential PCs we could present: Hypothetical Utilities for Individual i Computer = (Dell) + (300 MHz) + (64 Mbytes RAM) + ($1,500) (45) = (5) + (10) + (5) + (25)
Background: Card-Sort Conjoint Conjoint method used in industry since the 1970s. Full-profile approach: describe products using one level from each attribute. Carefully choose a set of products and print each on separate card. Combinations of attributes are chosen so we can estimate the impact of each feature independent of the others. Ask respondents to sort cards from best to worst (or rate cards based on preference). Use Ordinary Least Squares (OLS) regression for ratings or non- parametric techniques (LINMAP, Linear programming, or Monotone Regression) for rankings to analyze results.
Traditional Conjoint: Card-Sort Method (CVA) ( Three Attributes in Full-Profile) Using a 100-pt scale where 0 means definitely would NOT and 100 means definitely WOULD… How likely are you to purchase… 1997 Honda Accord Automatic transmission $18,900 Your Answer:___________
Full-Profile Conjoint Works very well as long as the number of attributes and number of questions (cards) is kept to a reasonable number. What is a reasonable number?
Three Attribute Full-Profile Conjoint Easy to process three attributes per card Assuming 3 levels per attribute, 9 total levels: –(#Levels - #Attributes + 1) is minimum number of cards –3x minimum is recommended number to achieve stable estimates for each individual –( ) = 7 cards minimum, 21 cards recommended 21 cards described on 3 attributes each is relatively easy task for respondents
Traditional Conjoint: Card-Sort Method (CVA) (Six Attributes) Using a 100-pt scale where 0 means definitely would NOT and 100 means definitely WOULD… How likely are you to purchase… 1997 Honda Accord Automatic transmission No antilock brakes Driver and passenger airbag Blue exterior/Black interior $18,900 Your Answer:___________
Six Attribute Full-Profile Conjoint Challenging, but typically manageable for most respondents to process six attributes per card, as long as attribute text is concise Assuming 3 levels per attribute, 18 total levels, ( ) = 13 cards minimum, 39 cards recommended 39 cards described on 6 attributes each is relatively difficult, but doable for many respondents
Traditional Conjoint: Card-Sort Method (CVA) (Fifteen Attributes) Using a 100-pt scale where 0 means definitely would NOT and 100 means definitely WOULD How likely are you to purchase… 1997 Honda Accord Automatic transmission No antilock brakes Driver and passenger airbag Blue exterior/Black interior 50,000 mile warranty Leather seats optional trim package 3-year loan 5.9% APR financing CD-player No cruise control Power windows/locks Remote alarm system $18,900 Your Answer:___________
Fifteen Attribute Full-Profile Conjoint Very challenging to process fifteen attributes per card Assuming 3 levels per attribute, 45 total levels, ( ) = 31 cards minimum, 93 cards recommended 93 cards described on 15 attributes each is nearly impossible for most respondents to do without adopting simplification strategies
Possible Solutions to the Problem of Too Many Attributes Respondents only evaluate a subset of cards, and analyst computes utilities across individuals (block or randomized design) –Assumes respondent homogeneity, lose richness of individual-level data –Respondents still need to evaluate cards described on many attributes Show only a subset of attributes (partial profiles) in a block or randomized design and compute utilities across individuals –Assumes respondent homogeneity, lose richness of individual-level data
Possible Solutions to the Problem of Too Many Attributes Self-explicated Approach –Respondents rank-order (or rate) levels within the attributes –Respondents state how important each attribute is –Multiply level preferences by attribute importances to compute utilities –Self-explicated data are easy for the respondent to provide and very informative (a lot of information per respondent effort) –But this sort of task is not very realistic –Possible double counting of attributes which are not independent and low discrimination between attribute importances are common problems Hybrid Conjoint –Collect both self-explicated data and additional conjoint judgments –Combine data from both sources to compute utilities
Possible Solutions to the Problem of Too Many Attributes ACA (Adaptive Conjoint Analysis), Richard M. Johnson, 1985 –Computerized, adaptive interview –Respondents rank-order (or rate) levels within the attributes –Respondents state how important each attribute is –Pairwise conjoint questions show only a subset of attributes at a time (usually 2 or 3) –Pairwise conjoint questions focus on attributes of most importance to respondents, and are customized to be relevant and informative
ACA Pros Good choice for measuring many attributes (more than about six) Very efficient at capturing a lot of information in short time Respondents typically find ACA more enjoyable and perceive it to take less time to complete than traditional conjoint ACA utilities are robust and have fewer reversals (out of order part-worths) than traditional conjoint methods Has advantages for predicting preferences for high-involvement categories Attribute prohibitions can be specified, and part-worth estimation is more stable in the face of attribute prohibitions than traditional conjoint or choice methods
ACA Cons Must be a computerized survey. Potential double-counting of attributes that are not truly independent. Respondents may have difficulty keeping in mind that all other attributes not involved in the current question are assumed to be equal. May flatten importances (particularly for low-involvement categories) due to spreading respondents attention across individual attributes--but the jury is still out. Can underestimate the importance of price (especially if many attributes included). CBC, Adaptive CBC (ACBC), and CVA considered better for pricing research.
ACA Best Practices Show only 2 or 3 attributes at a time in the pairs section. More than that causes respondent fatigue, which outweighs the modest amount of additional information. ACA can measure up to 30 attributes, but users should streamline studies to have as few attributes as necessary for the business decision. Pretest the questionnaire to make sure it is not too long. If too long, reduce number of attributes, levels, number of pairs questions, or complexity of pairs questions. Examine pretest data to make sure results are logical and conform to general expectations. Make sure respondents are engaged in the task: understanding the attributes and levels and being in the market/having an interest in the category.
ACA Use Today Since about 2000, the use of ACA is declining. Why? Many researchers have shifted to CBC, as choices are viewed as more realistic and accurate than concept ratings. Sawtooth Softwares newest Adaptive form of CBC (ACBC) has many of the benefits of ACA, but without the weakness in pricing research.