Conjoint Analysis
Research question: To what extent does each component (factor) contribute to the total utility of a product? Assumption: Total utility = Sum of all partial utilities
Data base of the Conjoint Analysis are preferences of the interviewed subjects Important application: Design of a new product according to the requirements of the market
Conjoint Analysis
Factors and their values are defined by the researcher in advance The various combinations of the factor values yield fictive products that are being ranked by the interviewed persons With Conjoint Analysis it is possible to derive metric partial utilities from the ranking results The summation of these partial utilities therefore results in metric total utilities
Conjoint Analysis
Independent variables: Object attributes Dependent variable: Preferences of the interviewed person for the fictive products The utility structure of a number of persons can be computed through aggregation of the single results
Conjoint Analysis
Factors and Factor Values
Important for the choice of factors and their values are: Relevance Interference Independence Realisability Compensatory relationships of the various factor values They do not constitute exclusion criteria Terminableness
Conjoint Analysis
Example: Attitudes towards dishwashing products 1. Clean: glass/dishes clean 2. Shiny: glass/dishes shiny 3. Smell: Non-perfumed/lemon fresh/intensive lemon fresh 4. Quantity: small/medium/x-large 5. Packaging: loose in box/tab in plastic/tab in dissolving plastic 6. Design: single/multi-colored/multi-colored + ball
Conjoint Analysis
Survey Design
1.
Definition of the incentives Incentive = Combination of factor values that is presented for rating Profile method: incentive results from the combination of the different factor values of each factor Two-factor method (Trade-off Analysis): Only two factors are used at a time. For each of these pairs a Trade-off-Matrix is set up Which method is used depends on how demanding the interview is for the respondent, the expenditure of time and the closeness to reality
Conjoint Analysis
Survey Design
2.
Number of incentives Can increase very fast depending on the number of factors and their values. It is advisable to choose a subset (reduced design) out of all theoretically possible incentives (full design). There shouldnt be more than 20 fictive products. Choice of the number can be based on random selection (not very common) or systematic selection Possible are symmetric (= all factors have the same number of values) and asymmetric (divergent number of values) designs Addelman developed several Basic Plans for the construction of reduced asymmetric designs
Conjoint Analysis
Example (cont.): Attitudes towards dishwashing products The full design consists of 2x2x3x3x3x3=324 different incentives The reduced design consists of 16 incentives/cards, therefore we get a ranking from 116 for each person
Conjoint Analysis
The column Card_ shows the numbering of the cards The column Status_ can show the values 0, 1 or 2. incentives that are part of the reduced design get the number 0 A value of 1 tells us that the corresponding card is a so-called holdout card that isnt used for the estimation of the utility values, but for validation. A value of 2 indicates a simulation card that isnt presented to the interviewed persons
Conjoint Analysis
Rating of the incentives
Possibilities: Ranking Rough classification into groups of different utility with succeeding ranking within these groups. Aggregation of these results leads to a total ranking. Used when there is a large number of incentives. Rating scales Paired comparison
Conjoint Analysis
Estimation of the utility values
Conjoint Analysis is used to determine partial utilities (partworths) for all factor values based upon the ranked data Furthermore, with this partworths it is possible to compute the metric total utilities y of all incentives and the relative importance of the single object attributes Individual Conjoint Analysis: For each person utility values are computed Combined Conjoint Analysis: Only one value for each factor category
Conjoint Analysis
Estimation of the utility values
The additive model of the Conjoint Analysis is defined as:
yk =
J j =1 m=1
Mj
jm x jm
with
jm
yk : estimated total utility for incentive k
: partial utility for value (category) m of factor j
x jm = 1 if incentive k has value m of factor j
0 else
Conjoint Analysis
Estimation of the utility values
Target criterion for the determination of the partial utilities: The resulting total utilities should yield a good representation of the empirically ranked data Related procedure for the determination of the partial utilities: monotonous analysis of variance Metric solution metric ANOVA Non-metric solution monotonous ANOVA (not covered here)
Conjoint Analysis
Estimation of the utility values
Model for the metric solution
yk = +
J j =1 m=1
Mj
jm x jm
Steps needed to get the solution:
1. 2.
Determination of (= average of all ranks) Determination of the mean rank for each factor value (= which ranks did the interviewed person assign on the average in conjunction with this attribute?)
Conjoint Analysis
Estimation of the utility values
3.
4.
The average utility of a factor value then equals the difference of its average empirical rank value and The partial utilities are OLS Now it is possible to compute the total utility of the incentives
Missing values are being replaced by zeroes There shouldnt be too much missing values in order to prevent the analysis from breaking down
Conjoint Analysis
Aggregation of the utility values
For the comparability of the individual analyses of the single persons it is necessary to carry out a standardisation: 1. The factor value that yields the lowest utility contribution is set to zero 2. The differences between the single partial utility values and the lowest partial utility value are computed:
Min * = jm jm j
Conjoint Analysis
Aggregation of the utility values
3.
The total utility value of the incentive that shows the strongest preference among all persons is set to one. The partial utility values are computed accordingly:
jm
* jm
max { }
J J =1 m * jm
The importance of an attribute for the preference change results from the utility range of the several values of the attribute
Conjoint Analysis
Aggregation of the utility values
A high range value means that by varying the concerning attribute a meaningful change of the total utility can occur The relative importance of a factor can be obtained by weighting on the sum of the ranges
Conjoint Analysis
Aggregation of the utility values
The largest standardised partial utility values allow to make a statement about the relative importance of the attributes at the same time Aggregation always implies the loss of information If the aggregated utility structures are highly heterogeneous, this loss of information is substantial In this case, more homogeneous subgroups should be formed by performing a Cluster Analysis
Conjoint Analysis
Example (cont.): Attitudes towards dishwashing products There is no graphical user interface available in SPSS that would allow the performance of a Conjoint Analysis. The SPSS-Syntax has to be used in order to retrieve the required procedure CONJOINT. After the Syntax-window has been opened, the following commands should be entered:
Conjoint Analysis
Example (cont.): Attitudes towards dishwashing products
Conjoint Analysis
The commands in the syntax have the following meaning:
With the TITLE statement it is possible to define a title for the results in the output window The actual Conjoint Analysis is performed with help of the procedure CONJOINT. There are various subcommands within this procedure:
-
The PLAN subcommand tells CONJOINT which file contains the data for the reduced design Each of the following commands is preceded by a slash (/):
Conjoint Analysis
-
The subcommand DATA tells CONJOINT which file contains the rankings of the respondents FACTORS is used to define the relationship between the factors and the preference values (ranks) of the interviewed persons. The relationships can be DISCRETE, LINEAR, IDEAL or ANTIIDEAL. In this example, all of the relationships are DISCRETE because only categorical data is used. It means that no assumptions are made concerning the relationship between variables and ranks. The subcommand RANK tells CONJOINT that the data is coded in such a way that the sequence of the variables corresponds to the sequence of the cards
Conjoint Analysis
-
Alternatively, the subcommands SCORE or SEQUENCE can be used. By the UTILITY subcommand, a file with the name [Link] is being created, where various information is saved. This information contains the following variables: CONSTANT: The constant term of the conjoint estimation Several partial utility values SCORE1 to SCORE16: Total utility values of the reduced design
With the SUBTITLE command, a new subsection of the output is introduced With the get file statement the contents of the [Link] file are printed in the output window
For our example, the results would be the following:
Conjoint Analysis
Interpretation of the results
The term Averaged importance score under the box with the importance values tells us that a Combined Conjoint Analysis has been performed. These values give us a measure (in percent) of the relative importance of the single factors for the determination of the utilities. We can see that quantity is the most important factor whereas shiny is the least important. The Correlation coefficients are a measure for the quality of reproduction of the empirical data by the results of the Conjoint analysis. The estimated partial values for the factor values are given in the column Utility Estimate
Conjoint Analysis
Interpretation of the results
Summaries of Importance and Utility in form of bar charts are also given The partial utility values allow the computation of the total utility values for all the incentives. For example, the total utility value for card 1 would be: 8,96-0,23-0,23-0,08+0,83+0,43-0,62=9,06
The constant term can be interpreted as base utility, and the other factor values contrast with in positive or negative direction
Conjoint Analysis
Interpretation of the results
Following this, the total utility values for all the cards can be computed similarly. Like it was said above, the results of this computations can be found in the [Link] file (Score1 to Score16):
Conjoint Analysis
Interpretation of the results
We can see that card 2 has the highest total utility (10,82). A product that makes the dishes clean, the glass shiny, has a lemon-fresh smell, a multicolored design and is packaged loose in a medium sized box therefore is the most preferred product out of the 16 alternatives. Here, just the results for the reduced design are given. But it is also possible to compute total utilities for all the products of the complete design, even though they were not part of the interview.
Conjoint Analysis
Interpretation of the results
It is therefore possible to find out that, for example, a product that is similar to card 2, except that the box contains single packaged tabs, would have an even higher total utility than card 2.
Conjoint Analysis
Clustering the partial utility values
As aforementioned, a Cluster Analysis should be performed before aggregating the results if the aggregated utility structures are not homogeneous To get the data that is needed to do this, a new line has to be inserted into the syntax:
Conjoint Analysis
Clustering the partial utility values
The results for each person are computed and saved into the file Person [Link]:
Conjoint Analysis
Clustering the partial utility values
Now the formulae given in the section Aggregation of the utility values are used to calculate the standardised partial utility values (Transform > Compute). It is reasonable to do this in two steps: 1. Compute the differences between the single partial utility values and the lowest partial utility value:
Conjoint Analysis
Clustering the partial utility values
2.
Set the most preferred incentive to 1 and compute the standardised partial utility values accordingly:
These two steps have to be repeated for each factor value. After that, the partial utility values and the standardised partial utility values can be attached to the original data file
Conjoint Analysis
Clustering the partial utility values
To do this, use Data > Merge files > Add variables If both data files are opened and the original data file is the active one, the option An open dataset can be chosen. Otherwise there is the possibility to browse for the corresponding file:
Conjoint Analysis
Clustering the partial utility values
We can exclude variables from the dataset that arent needed anymore (i.e., the SClean1 to SDesign3 variables). The variables that arent excluded form the new data set:
Conjoint Analysis
Clustering the partial utility values
The next step is the actual performance of the Cluster Analysis using the menu Analyze > Classify > K-means Cluster
Conjoint Analysis
Clustering the partial utility values
All the standardised variables from NClean1 to NDesign3 are used as input To save the results of the Cluster Analysis in the data file, the box Cluster Membership in the Save-menu has to be checked It is necessary to try different numbers of clusters to find out what the (as regards content) best solution is Cross-classifying the variable for the Cluster membership with certain demographic variables is needed to assess the best Cluster solution
Conjoint Analysis
Clustering the partial utility values
If the Cluster solution has finally been found, it is possible to compute an Combined Conjoint Analysis for each Cluster separately (using Data > Select Cases)
