Conjoint Analysis - Comparing Data Collection Approaches
Conjoint Analysis - Comparing Data Collection Approaches
The arguments in favor of the full-profile approach are that (1) the description of the concepts is more realistic since all aspects are considered at the same time, (2) the concept evaluation task can employ either a ranking or rating scale, and (3) fewer judgments have to be made by the respondent than if the two-attribute trade-off approach were used. Unfortunately, as the number of attributes increases the task of judging the individual profiles becomes very complex and demanding. With more than five or six attributes there is a strong possibility of information overload, which usually leads respondents to ignore variations in the less important factors. To get the flavor of this problem, look back to Figure 19-2 and see how difficult it is to choose one package holiday over another.
The pairwise trade-off approach is not a panacea either. Because more judgments are required, the task can be tedious and time consuming. Consequently, respondents may lose their place in the matrix or develop a standardized response pattern just to get the job done. Since only two attributes are being considered there is a potential loss of realism. This problem is most troublesome when there is substantial environmental correlation among attributes for technological or other reasons. For example, the 0-to-55 mph acceleration time, gas mileage, horsepower rating, and top speed of an automobile are not independent attributes. When only two of these four attributes are being considered, respondents may be unclear as to what should be assumed about the other factors. This problem also is encountered with price, because it may be used as an indicator of quality. Of. course, if environmental correlations are high it may be possible to create a composite factor. This means losing information about the component attributes.
Studies comparing the two methods typically find that the estimated utilities are roughly similar and that, for large numbers of factors which are not environmentally correlated, the trade-off approach yields somewhat higher predictive validity. In part because it is difficult to find factors that are not correlated, the full-profile approach has become increasingly preferred. Almost 70 percent of recent studies used this approach, and another 15 percent used a combination of full-profile and two factors at a time.
Analyzing and Interpreting the Data
The analysis of conjoint or trade-off studies, like all other marketing research, is guided by the research purpose. As an illustration, a manufacturer of automobile batteries with a lifetime guarantee wanted to know how much emphasis to place on the fact that the batteries never need water. A conjoint study was conducted in which respondents were asked to evaluate full-concept profiles made up of combinations of three attributes and three levels:
Attribute
Levels
Price
Length of guarantee Maintenance required
$30, $45, $60
Lifetime, 60 months, 48 months No water needed, add water once a year, add water as needed
The input to the analysis was a ranking of these stimulus profiles by preference. The data analysis problem is first to estimate a set of utilities (sometimes called part-worths) for the nine attribute levels such that
1. The sum of the attribute level utilities for each specific profile equals the total utility for the profile
2. The derived ranking of the stimulus profiles, based on the sum of estimated attribute level utilities, corresponds as closely as possible to the original ranking by the respondent.
8In the regression approach the rank ordering is the dependent variable and the independent variables are 0-1 variables for each level on an attribute less one. Thus, for the price attribute there would be a 0-1 variable for $45 (coded as " 1" only if the profile had a $45 price) and a 0-1 variable for $60 (coded as a " 1" only if the profile had a $60 price). The profile with a $30 price would be the reference level and therefore would not have its own 0-1 variable. If the $45 variable is coded "0" and the $60 variable is coded "0" then the level must be $30.
While the details of the techniques used to achieve this are beyond the scope of this discussion, the elements are straightforward. The part-worth utilities can be obtained with an iterative procedure which starts with an arbitrary set of utilities and systematically modifies them until the total utility of each profile correlates maximally to the original ranks. The procedure continues until no change in the utility of an attribute level will improve the correlation. As a practical matter most analysts use regression analysis (Chapter 20) to obtain the attribute weight utilities since it provides very similar results and is much easier and cheaper to use than an iterative procedure.
Once the utilities are estimated, they are displayed and the relative importance of each attribute is determined. In the case of automobile batteries, the following graph displays the results. According to these utility values, both price and length of guarantee are more important than the maintenance attribute. However, the relative difference in the total utility from a change in the level of maintenance (0.6 - 0.2 = 0.4) is greater than for a change in the length of guarantee (0.9 - 0.7 = 0.2). Clearly, the fact that the battery doesn't need water should be emphasized as an advertisinc. appeal, especially if potential buyers are not aware of this benefit.
Validity Issues
The conjoint model involves the assumptions that preference can be modeled by adding the utilities associated with attribute levels and that these utilities can be estimated from either full-profile or trade-off-data. There are several indications that the model has a substantial level of validity.
A bottom-line validity test is the remarkable acceptance of conjoin: analysis in industry. As already noted, the technique, only introduced in 1971, was estimated to have had over 1000 applications during the firs: seven years. It is now used routinely by many market research firms and their clients.
The conjoint model has been found to predict the market share of transportation modes between two cities and the type of fare in North Atlantic air travel.9 At the individual level, a conjoint model involving eight factors (including salary, location, need to travel, and so on) predicted subsequent job choice by M.B.A. students with an impressive 63 percent hit rate.
Another study used conjoint to successfully predict the acceptance of 13 new products introduced into supermarkets.
The reliability of conjoint analysis has been the subject of several studies. In general the reliability has been found to be high. One study exploring both the full profile and the trade-off approach using five product classes found that the output was not sensitive to which five of six attributes were included or whether three or five levels were used on the price attribute.10 Another study concluded that the reliability fell when the number of attributes was increased.
The arguments in favor of the full-profile approach are that (1) the description of the concepts is more realistic since all aspects are considered at the same time, (2) the concept evaluation task can employ either a ranking or rating scale, and (3) fewer judgments have to be made by the respondent than if the two-attribute trade-off approach were used. Unfortunately, as the number of attributes increases the task of judging the individual profiles becomes very complex and demanding. With more than five or six attributes there is a strong possibility of information overload, which usually leads respondents to ignore variations in the less important factors. To get the flavor of this problem, look back to Figure 19-2 and see how difficult it is to choose one package holiday over another.
The pairwise trade-off approach is not a panacea either. Because more judgments are required, the task can be tedious and time consuming. Consequently, respondents may lose their place in the matrix or develop a standardized response pattern just to get the job done. Since only two attributes are being considered there is a potential loss of realism. This problem is most troublesome when there is substantial environmental correlation among attributes for technological or other reasons. For example, the 0-to-55 mph acceleration time, gas mileage, horsepower rating, and top speed of an automobile are not independent attributes. When only two of these four attributes are being considered, respondents may be unclear as to what should be assumed about the other factors. This problem also is encountered with price, because it may be used as an indicator of quality. Of. course, if environmental correlations are high it may be possible to create a composite factor. This means losing information about the component attributes.
Studies comparing the two methods typically find that the estimated utilities are roughly similar and that, for large numbers of factors which are not environmentally correlated, the trade-off approach yields somewhat higher predictive validity. In part because it is difficult to find factors that are not correlated, the full-profile approach has become increasingly preferred. Almost 70 percent of recent studies used this approach, and another 15 percent used a combination of full-profile and two factors at a time.
Analyzing and Interpreting the Data
The analysis of conjoint or trade-off studies, like all other marketing research, is guided by the research purpose. As an illustration, a manufacturer of automobile batteries with a lifetime guarantee wanted to know how much emphasis to place on the fact that the batteries never need water. A conjoint study was conducted in which respondents were asked to evaluate full-concept profiles made up of combinations of three attributes and three levels:
Attribute
Levels
Price
Length of guarantee Maintenance required
$30, $45, $60
Lifetime, 60 months, 48 months No water needed, add water once a year, add water as needed
The input to the analysis was a ranking of these stimulus profiles by preference. The data analysis problem is first to estimate a set of utilities (sometimes called part-worths) for the nine attribute levels such that
1. The sum of the attribute level utilities for each specific profile equals the total utility for the profile
2. The derived ranking of the stimulus profiles, based on the sum of estimated attribute level utilities, corresponds as closely as possible to the original ranking by the respondent.
8In the regression approach the rank ordering is the dependent variable and the independent variables are 0-1 variables for each level on an attribute less one. Thus, for the price attribute there would be a 0-1 variable for $45 (coded as " 1" only if the profile had a $45 price) and a 0-1 variable for $60 (coded as a " 1" only if the profile had a $60 price). The profile with a $30 price would be the reference level and therefore would not have its own 0-1 variable. If the $45 variable is coded "0" and the $60 variable is coded "0" then the level must be $30.
While the details of the techniques used to achieve this are beyond the scope of this discussion, the elements are straightforward. The part-worth utilities can be obtained with an iterative procedure which starts with an arbitrary set of utilities and systematically modifies them until the total utility of each profile correlates maximally to the original ranks. The procedure continues until no change in the utility of an attribute level will improve the correlation. As a practical matter most analysts use regression analysis (Chapter 20) to obtain the attribute weight utilities since it provides very similar results and is much easier and cheaper to use than an iterative procedure.
Once the utilities are estimated, they are displayed and the relative importance of each attribute is determined. In the case of automobile batteries, the following graph displays the results. According to these utility values, both price and length of guarantee are more important than the maintenance attribute. However, the relative difference in the total utility from a change in the level of maintenance (0.6 - 0.2 = 0.4) is greater than for a change in the length of guarantee (0.9 - 0.7 = 0.2). Clearly, the fact that the battery doesn't need water should be emphasized as an advertisinc. appeal, especially if potential buyers are not aware of this benefit.
Validity Issues
The conjoint model involves the assumptions that preference can be modeled by adding the utilities associated with attribute levels and that these utilities can be estimated from either full-profile or trade-off-data. There are several indications that the model has a substantial level of validity.
A bottom-line validity test is the remarkable acceptance of conjoin: analysis in industry. As already noted, the technique, only introduced in 1971, was estimated to have had over 1000 applications during the firs: seven years. It is now used routinely by many market research firms and their clients.
The conjoint model has been found to predict the market share of transportation modes between two cities and the type of fare in North Atlantic air travel.9 At the individual level, a conjoint model involving eight factors (including salary, location, need to travel, and so on) predicted subsequent job choice by M.B.A. students with an impressive 63 percent hit rate.
Another study used conjoint to successfully predict the acceptance of 13 new products introduced into supermarkets.
The reliability of conjoint analysis has been the subject of several studies. In general the reliability has been found to be high. One study exploring both the full profile and the trade-off approach using five product classes found that the output was not sensitive to which five of six attributes were included or whether three or five levels were used on the price attribute.10 Another study concluded that the reliability fell when the number of attributes was increased.
Comments
Post a Comment