New Books in Review - American Marketing Association
Contents
1. based questionnaires Traditional conjoint analysis studies use paper and pencil techniques to collect either full profile data card sorts or 2 X 2 factor trade off matrices The ACA procedure requires the respondent to evaluate attributes and attribute levels separately make choices between pairs of partial profiles two to five attributes at a time and to indi cate a likelihood of purchase of several profiles defined on up to eight attributes A hybrid conjoint algorithm devel oped by SSI utilizes data from these three tasks to estimate individual factor level utilities which can then be used along with separately collected demographic data as input into an integrated choice simulator This version of ACA 4 0 contains significant improve ments over earlier versions It is designed to easily and quickly create a conjoint questionnaire for distribution to several sites and computers This compiled version of the questionnaire is used to collect the data Field generated data sets utilities and demographics are merged and then used for input to the simulator New with this version is the ability to use different rating scale ranges two to nine or ranks for attribute evaluation The researcher can also spec ify equal or optimal weighting of the self explicated and paired comparison sections of the questionnaire This latter 118 JOURNAL OF MARKETING RESEARCH FEBRUARY 1995 option could lead to a better fit with the calibrati2. can be measured at the group level That is the researcher does not have to guess beforehand which interactions may be sig nificant because all interactions can be tested This claim is questionable because of the potential confounding pooling bias that does not occur with prespecified fractional facto rial designs Estimation of three way interactions and high er must be done outside CBC which is limited to two way interactions The system is menu driven with on line help From the Main Menu a word processor of choice can be linked up to edit the questionnaire and other scripts identify the study to 119 be run compose and test the questionnaire prepare field disks and conduct the data analysis Compose and Test the Questionnaire The initial steps in CBC are similar to those in CVC ex cept they are more user friendly and comprehensive The re sult is a PC based questionnaire that can be shipped to the field to collect the data The format of the questionnaire con sists of several sets maximum of 50 of choice scenarios two to eight product profiles plus a none option that are randomly selected for each respondent Each profile can be defined on a maximum of six attributes with up to nine lev els per attribute A test mode enables the researcher to check that the con straints imposed on the attributes which can appear togeth er in the questionnaire in the random design do not inad vertently prevent the estima
3. above assumes equal effectiveness of the algorithms used in each of the packages of course such an assumption is not agreed upon by all researchers Instead the selection of one of these programs for a particular conjoint analysis project requires that the researcher recognize the philosophy implicit in each of the programs and select the program with which he or she feels most philosophically comfortable For example se lecting ACA requires data collection by computer not by paper and pencil using a hybrid conjoint analysis algorithm not OLS on full profile evaluations paired comparisons not card sorts and so on Each researcher has his or her own opinion about which of these options is better In recognition of the importance of these personal opin ions we hope our comments in this review on how each program works will help the researcher make his or her se lection Unfortunately the reader will have to watch the pro fessional journals for a definitive article on the relative ef fectiveness and accuracy of these programs We don t see anything being published in the near future and probably never will given all the parameters that would have to be included to satisfy all researchers that will clearly identify the best way to do a conjoint analysis study This controversial issue continues to intrigue researchers as they make choices for their projects F J CARMONE JR Wright State University C M SCHAFFER Universi
4. and parameters used in CVA are similar to those in ACA In fact utilities calculated by ACA New Books in Review could also be run in CVA and vice versa Segmentation vari ables up to 30 can also be used in the simulation CVA is not as self contained as ACA but it fills a niche for researchers who want to collect data in a full profile paired comparison or single concept mode but if there are more than about six or seven attributes ACA is strongly rec ommended by SSI instead of CVA They even suggest using ACA which is capable of handling many attributes and lev els and CVA to study a few selected attributes and price in detail in tandem which almost requires something like CONJOINT BRIDGER to combine the resultant utilities The manual is very helpful in explaining how to construct the several files required by CVA and in interpreting the out put It also offers guidelines and suggestions for setting the parameter that determines the number of questions This in volves a trade off between time and accuracy The question naire construction is designed to be most helpful for those trade offs involving prices for each attribute as in selecting a meal with each item appetizer entre and dessert having a different price or no price if it is not included for exam ple no dessert Researchers must feel confident in being able to set the number of questions to be asked the key pa rameter in this approach to conjoint analysi
5. de scribed This description is very brief a short paragraph to two pages but it is generally adequate to illustrate the pro cedure Next example data are provided although they are often in summarized or partially analyzed form and the major calculations are illustrated Each test description is done in a open inviting format This section contains virtu ally any test one could encounter although a few are missing The next section consists of the 39 statistical tables need ed to interpret the results of the tests described in the book Clearly this is quite valuable The seventh section contains the references which the preface describes a guide to further reading Unfortunate ly this section consists of 26 citations that are not referred to in the descriptions of the tests nor are they arranged or categorized in this section except alphabetically Thus one is not sure why one should refer to books such as Statistical Analysis in Chemistry and the Chemical Industry Overall I think this is a book that with 20 more effort could have been twice as valuable It needs a much more useful classification scheme or set of schemes The less common tests need explicit references to more detailed treatments because the descriptions although very well done will often leave a question or two unanswered for many users of statistics Having said this I think it is still a valuable book because it does clearly yet briefly desc
6. it is often difficult to select the best solution Even within a technique it is often not clear what is the best num ber of clusters for final selection CONJOINT SEGMENTER assumes that input files are from either CONJOINT ANALYZER or CONJOINT LIN MAP with the associated design and data files The first step is to calculate distances between all pairs of respondents using both the individual utilities and transformations of the original data This is followed by the calculation and display of the 2 to 15 cluster solutions that is the display of the pooled within cluster variability This measure of vari ability is used to help with the selection of the number of clusters segments to analyze To aid in the selection of the number of clusters for detailed analysis it would be helpful to have the first differences of the pooled variances between adjacent cluster solutions displayed next to the number of clusters and or be able to toggle a scree type plot Once the number of clusters for detailed analysis has been determined a display of the initial mean utility values for each cluster is obtained by selecting that cluster solution These data are refined that is adjusted for limitations in the clustering solution The results can be saved in a new seg mentation file and or another detailed clustering solution can be investigated At this point it would be helpful to have a log file that contained the options selected values of the
7. of the BC approach because the researcher can check the goodness of fit to the original and or holdout data using part worth vector or quadratic models and either OLS or optimization algo rithms Hopefully the selection of models is theory driven rather than simply a search for better fits but the capability is there for exploration The breadth of the model simulation capability is also impressive Almost all the tasks implemented by BC IMS and SSI in their software could be done piecemeal with a word proces sor a Statistical package and or a spreadsheet An exception is the hybrid adaptive procedure used by ACA and the ad justment to cluster membership in CONJOINT SEG MENTER For a one time project this may be all that is needed if the user has the necessary computer skills The advantage of the BC IMS and SSI packages is that they are designed to appeal to the researcher who either does not have the expertise or time or both to develop his or her own software these packages are specially designed for conjoint analysis studies only Once the basic parameters of the study are formalized the researcher can select one of these packages for example if there will be about 20 to 25 attributes in the study then ACA or CONJOINT BRIDGER should be used or computers cannot be used to collect the data In such an instance CVA CONSURV or CONJOINT ANALYZER should be used and so on The interchangeability of the software mentioned
8. New Books in Review initions Thus one is left to wonder why the t test of a re gression coefficient which is designed to investigate the significance of the regression coefficient of y on x is clas sified as a two sample parametric classic test for central tendency At this stage Professor Kanji could have added signifi cant value by providing several classification schemes and subschemes so that one could start with a data and problem situation and identify one or more appropriate statistical tests This weakens the value of the text noticeably For ex ample I decided to look for a test that would compare the rank assigned a brand under one treatment condition with the rank assigned to that brand by different subjects inde pendent samples facing a different treatment I expected to find the traditional Mann Whitney U test which is not in cluded In fact I had some difficulty determining which of the tests described would be appropriate The fifth section is the heart of the text One hundred sta tistical tests are briefly described Each test is named for ex ample the median test of two populations and its objec tive is provided To test if two random samples could have come from two populations with the same frequency distri bution Its limitations are briefly described The two sam ples are assumed to be reasonably large The methodology of conducting the test including the formula is then
9. RV proceeds to overwrite the file without warning After requesting a design the user is transferred to a screen with one or more designs from CON SURV s table of designs The unique aspect of these designs is that some depending on the number of profiles in the de sign and the number of attributes and levels enable the user to estimate a limited number of interactions To the extent that a problem definitely includes interactions this is the only commercially available package that provides these kinds of designs Other packages either ignore the problem or suggest ways of getting around the issue of interactions If none of the designs generated by CONSURV are ac ceptable in the sense that the experimental design is too large the option exists to redefine the number of attributes and levels and generate a new set of designs This step would be easier if it was not necessary to type all the at tributes and levels first check the designs available and then add or subtract attributes and or levels That is if only the number of attributes and levels were entered first time could be saved because the attributes and levels would have to be typed only once after the design had been selected CONSURYV allows the user to convert BC designs to a for mat that it can use Once a design has been selected how one should do this is not discussed in detail and the user is left to his or her own expertise in experimental design there appea
10. ant topics for example reducing the number of attributes in a study and incomplete block designs CONJOINT ANALYZER The software used to enter respondent data estimate indi vidual utilities and perform simulations is CONJOINT AN ALYZER This version contains several methodological as well as programming improvements Probably the most im portant is the ability to compare models using an adjusted r It is now possible to statistically test the difference between an ideal point and a vector model to see which provides a better fit to the data This of course is done at the group level but at some point in time software may become avail able to do this at the individual level The famous camera study continues to be used in the tu torial It clearly illustrates the various functions and capabil ities of the software One function that is retained for this version and has always fascinated these reviewers is the ability to clean the data This is interesting because the ana lyst can second guess the respondent he she reversed the scales smaller is better than larger Carefully used the cleaning function can prevent difficult to explain results Less carefully used the researcher is sure to have trouble justifying the results Assuming there is a theory to justify the direction of preference for the levels of the attributes CONJOINT ANALYZER generates some useful diagnostics on reversals Then respondents with suspect pat
11. e with a more sophisticated and flexible logit program CBC can create an export file to be used as input to such software Once the calculations have been completed the final task is to run simulations The CBC simulation module requires that the user first define the products to be simulated Using the parameters from the logit analysis including interac tions a total utility for each product can be calculated and then adjusted for various user specified reasons for exam ple external effect and then normalized to generate an esti mate of share of preference CBC is the first SSI program to use multinomial logit analysis to estimate attribute level utilities at the group level Although this iterative procedure takes longer to run than an equivalent size OLS estimation it can exhaustively investigate interaction effects In some problems it is ex tremely important to be able to test for the significance of 120 JOURNAL OF MARKETING RESEARCH FEBRUARY 1995 selected interactions For researchers who have this as a pri ority CBC is much easier to use than a program from a gen eralized statistical package Summary and Conclusion After spending several person days playing with this soft ware that is reading manuals generating designs and mod ifying them constructing questionnaires and testing them and running simulations on two different machines an IBM compatible 486 50 desktop and a Toshiba 486 33 lap top we are u
12. els which appears to have been gen erated from one sort of the cards containing all the at tributes These utilities can then be used as input to SIM GRAF for the simulation component of the project Al though the idea of breaking up a large conjoint study into two parts is somewhat appealing from a data collection per spective we are unaware of any published empirical evi dence that supports the use of this approach In summary CONJOINT LINMAP and BRIDGER were designed to fill perceived niches in the market for conjoint software that is the need for a program using optimization procedures for utility estimation and for a procedure to han dle a very large number of attributes and levels in a paper and pencil conjoint study For researchers who have these needs CONJOINT LINMAP and BRIDGER are the only games in town Intelligent Marketing Systems Inc CONSURYV is a program used to implement several of the tasks required to complete a conjoint analysis project Once the attributes and levels have been defined CONSURV is used to 1 generate the experimental design that is utilized to compose profiles for respondent evaluation 2 design the actual questionnaire used to evaluate these profiles 3 cal culate the utilities for each respondent using various models and 4 estimate shares of choice based on selected product profiles The only tasks that must be completed outside the CONSURV program are the definition of the approp
13. g Analysis Menu The functions on this menu are used for setting up one or more simulation runs The product definitions are easily en tered for a base case that consists of one or more product profiles All subsequent simulations are modifications of this base case with product profiles added deleted and or modified An available option is to do a sensitivity analysis automatically Instead of running several simulations manu ally or with a batch file the product profile and attributes to be systematically changed are identified and all simulations are run automatically According to the manual sensitivity analysis on all attributes can be requested but we were un able to get this option to work with the tutorial data We later learned that this feature had not been implemented in the current version ACA has gone through several versions since its intro duction to the marketplace These have incorporated modi fications requested by users and improvements suggested by other researchers At this point the system is highly param eterized to increase flexibility and control by the user Sev eral different simulation models are available to handle both new products and product line extensions The easy to learn user interface and on line help facilitate learning the system quickly Appendices now include a detailed description of the analytic procedure used to estimate individual utilities including the new option to differentially we
14. gn The BC software designers feel that from an applications per spective interactions generate more heat than light and they have decided not to include other more complex designs Once the software has been installed and the security key is in place one can easily and quickly create designs with a maximum of 30 attributes 16 levels per attribute and 81 cards With this software it is easy to see the effect on the design of adding and or deleting attributes and levels To generate a design one must specify the type of estimation model for each factor part worth linear or quadratic so the number of parameters to be estimated can be calculated This is also used to calculate the condition number an es timate of how analyzable the design is that is when the ratio of the number of parameters to the number of cards data points used to estimate those parameters approaches 1 0 the degrees of freedom approach zero and the parame ter estimates are unstable This feature and its discussion in the manual are welcome additions to the software CONJOINT DESIGNER automatically generates up to three designs for each set of specifications The manual of fers advice on how to select a design which in general is the one with the fewest cards but with several degrees of free dom At this point it is useful to check the design for dupli cate cards and unreasonable combinations of levels Re searchers know that it is better
15. ies The manual is very helpful in demonstrating how to use the software as well as how to interpret the results Sug gestions are made both for selection of the appropriate num ber of clusters and for analyzing the various output of each step However the usefulness of the package could be im proved by the aforementioned additions CONJOINT LINMAP AND BRIDGER Both of these packages have been reviewed previously Albaum 1989 Albaum and Carmone 1991 but for the sake of completeness they are briefly discussed here For re searchers who have rank order data and or prefer an opti mization approach to estimating individual utilities CON JOINT LINMAP is the only commercially available non metric PC package that does this There is some evidence that constrained estimation yields a better prediction of choice than unconstrained estimation Although this may be correct there does not appear to be as much use of this pro cedure as there is of OLS Utilities estimated with CON JOINT LINMAP can be used as input to SIMGRAF for more complex simulation studies BRIDGER was designed to fill a niche in paper and pen cil conjoint studies that is how do you handle a large num ber of attributes The BRIDGER approach requires separat ing the attributes into two overlapping sets Each set is eval uated separately by respondents and then merged in a statis tically reasonable way by BRIDGER The result is a set of utilities for all factor lev
16. ify the attributes and levels the num ber of questions to produce the prices for the attribute lev els and an indication of which levels should not appear to gether Note that the researcher decides on the number of questions to be asked which is essentially specifying the de sign to be used subsequently for calculating the utilities CVA s design is not necessarily orthogonal but it is gener ated to balance the number of times attribute levels appear together File outputs at this step specify the design being used and the questionnaire to be edited on a word processor Note that these files automatically overwrite previous copies so that files from earlier studies or previous versions of the same study must be renamed if they are to be used at some later date This also applies to the input files which keep the same name independent of the study being conducted Calculating the Utilities Using the design file the respondent data collected out side CVA constraints on the order of the levels of the at tributes and if necessary identification of data blocks with in respondent data data from trade off matrices CVA can estimate utilities using either multiple linear regression or monotone regression Statistical measures of goodness of fit are included for isolating respondents whose choices were not well predicted by the model The utility files then are input to the simulation Running Simulations The simulation models
17. igh parts of the data self explicated and paired comparisons Statistical tests used by ACA are explained along with the rationale for their use and details on how to interpret them Overall Version 4 0 of ACA is a significant improvement over earlier versions and it continues to be responsive to new research results and user concerns CONJOINT VALUE ANALYSIS CVA CVA is designed specifically to handle conjoint studies where price is an attribute to be explored in detail Ques tionnaires generated by CVA can be of the pencil and paper type or computer based for use with SSI s Ci2 or Ci3 inter viewing system Unlike ACA which is a hybrid analytic procedure and uses partial product profiles CVA uses either ordinary least squares or monotone regression to calculate the individual utilities from full profile evaluations Respon dents are shown pairs of full profiles selected from a CVA generated design single concepts can also be shown The maximum number of attributes and attribute levels is 10 and 15 respectively These steps are accomplished using three main modules in CVA not all of which must be used in each study Each of the three modules requires one or more files to be pre pared outside CVA usually with a word processor to pro duce an ASCII file The Main Menu items modules are briefly discussed below Composing the Questionnaire Before putting together a questionnaire several files must be constructed to spec
18. ill however attempt to present the ad vantages and disadvantages of each of the packages Bretton Clark The BC suite of programs was designed to handle most of the tasks in a paper and pencil conjoint study for example profile design utility estimation and simulations The soft ware for the actual construction of the instrument which would normally be done on a word processor or in a desk top publishing environment is not included Each program s user manual contains a description of the algorithms being used Recommendations are made as to how to parameterize the software to meet varying research needs Appendices de scribe the file layouts and specifications for the program Some of the manuals have an index to the major topics dis cussed All programs use pull down menus with on line help available most of the time System requirements rec ommended by the publisher are an IBM PC or compatible with 512K of memory DOS 2 0 or higher and a floppy disk A math coprocessor is not required but it can be used if it is installed The software requires a security key to run CONJOINT DESIGNER This version of CONJOINT DESIGNER is even easier to use than the earlier version which was the first commercial attempt to provide experimental designs for use in conjoint analysis studies All designs generated by CONJOINT DE SIGNER continue to be only orthogonal array it is not pos sible to estimate interactions with this class of desi
19. ll profile studies prevented the use of all possible combinations of attribute levels Some re searchers favored a 2 X 2 attribute data collection procedure and they used heuristics to limit the number of 2 X 2 tables seen by each respondent i e each attribute had to only ap pear in at most three tables Once the data was collected and utilities estimated software such as that needed to esti mate shares of choice and perform sensitivity analysis was developed on a project by project basis Enter the entrepreneurial marketing researchers at Bret ton Clark BC with CONJOINT DESIGNER and CON JOINT ANALYZER Sawtooth Software SSI with ACA and followed much later by Intelligent Marketing Systems IMS with CONSERV The result of their efforts was the first generation of PC based commercially available con joint analysis software Each of the vendors approached the problem from a slightly different perspective which result ed in software that is algorithmically and procedurally different Reviews of earlier versions of several of these programs have appeared in the Journal of Marketing Research Al baum 1989 Albaum and Carmone 1991 Carmone 1986 1987 Green 1987 1992 To conserve space we will com ment mainly on the improvements made in these programs The interested reader is referred to the original reviews for more detailed comments Users of BC and or SSI products as well as other re searchers who are interested in conj
20. nable to recommend any of these packages based solely on its ease of use or user friendliness Although it is true that there are features of one package that we may prefer more than another for example the three ring binders of SSI documentation are easy to lay flat and read while en tering data or commands as well as maintaining updates we thought that overall the user friendliness of each suite of programs is well above the minimum expected level and therefore not an important differentiating characteristic We thought the manuals for these programs were well written filled with suggestions on how to parameterize the runs and contained detailed examples of how to interpret the output CONSURV has less of this and BC and SSI the most The manuals for the latter two packages also contain a little hype for example reasons why their approach is bet ter than others or the software designer s personal bias for a certain statistical measure or estimate of the number of at tributes a consumer can handle which the user has to isolate from the less controversial factual material The SSI manuals are especially well done with a very professional orientation and feel It appears they have been very responsive to other researchers academics and practi tioners and have incorporated their suggestions and com ments for improving the software Although we did not test the various models for effective ness we like the breadth and flexibility
21. oint analysis are aware of the controversy about the relative effectiveness and valid ity of the two different algorithms and procedures full pro file conjoint versus adaptive hybrid like conjoint they use for analyzing consumers choice behavior Both companies have sent position papers to their users and others inter ested in conjoint analysis To the best of our knowledge IMS has not been directly involved in this controversy However because CONSURYV is a paper and pencil full profile ap proach to conjoint analysis it is in the BC camp by default 114 JOURNAL OF MARKETING RESEARCH FEBRUARY 1995 Others who are not privy to these position papers are re ferred to the papers by Herman and Shocker 1993 and Johnson 1991 The major differences in the two points of view concern effectiveness efficiency validity and time to administer Questions have been raised and vigorously de bated such as Does one get better predictions of choice be havior using a full profile or an ACA approach Does a full profile study take more time to administer than ACA Very little has been said about measures of reliability accu racy stability of choice predictions over time This review of the software from these vendors will not address these controversial issues or results Instead it will comment on other aspects of purchasing and using these products mainly from the perspective of the researcher or study designer It w
22. on likeli hood of purchase data System requirements include an IBM PC or compatible a hard disk and 640K of memory A math coprocessor is not required but it can be used if installed The system is designed with pull down menus and on line help The five major items on the Main Menu bar are briefly discussed below Study Menu This menu is used to create a new study select a previous study delete and or back up a study A previously backed up study can also be restored a very useful file management capability Compose and Options Menus Using these menus the questionnaire can be constructed and screen colors can be set for example defining the at tributes and levels setting parameters such as scale limits and time preventing unreasonable pairing of attributes sample wording for sections of the questionnaire and the monitor s colors A nice feature is the ability to test the ques tionnaire before compiling and distributing it A hard copy of the frames and parameters can be obtained for documen tation With the options menu the word processor to be used in constructing the questionnaire can be identified so foreign language frames can be easily accommodated Field Menu The functions in this menu are 1 to facilitate production of distribution disks and 2 to combine these field disks into a single data set for further processing Additional data can be included for segmentation analysis and respondent weightin
23. or distributing shares of choice when using the first choice model Unfor tunately the use of these options is not for the faint hearted To the best of the reviewers knowledge no theory exists to suggest how to set the values of the various parameters e g thresholds and exponents The manual presents a detailed discussion of these op tions and justification for their use but several of them se lect the one that feels right For example the manual sug gests relating the choice threshold value to the rating scale being used similar to the top box versus all others approach used in marketing and advertising studies In addition the levels of a factor can be swept which means that a simulation does not have to be manually run for each level of the factor which can become quite tedious Unfortunately too much flexibility e g several models and multiple parameters can lead the naive analyst down the wrong road In this instance BC feels that analysts need this power and didn t prespecify their best estimates of the pa rameters It is up to the analyst to set the value of these pa rameters using guidelines offered by BC Bretton Clark has provided the capability for researchers to acquire an execution only version of SIMGRAF for a par ticular project This run time option is priced separately 250 for each project but it enables clients to run simula tions at their leisure In summary SIMGRAF is for the analyst who i
24. parameters used and the mean values that are displayed on the various screens Two other types of output would also be tremendously helpful The first would be a cross tabulation of the results of cluster solution i versus cluster solution j this would en able the analyst to see the changes in cluster membership at each level for example when comparing the three and four cluster solutions is the fourth cluster formed by adding respondents from all three clusters or do all the clusters re main intact except for one that is splitting to form the fourth cluster This has an obvious impact on the qualitative inter pretation of the clusters Cross tabulations of the cluster so lution with selected demographics would also be very help ful if they were included in the basic data file The second type of output that would help with the inter pretation of the cluster solutions is some form of hit matrix 116 JOURNAL OF MARKETING RESEARCH FEBRUARY 1995 which is found in most discriminant analysis software For example are the clusters from the three cluster solution tighter than those of the four cluster solution This is anoth er way of asking the question Are the mean utility values statistically different among the various solutions This is an ad hoc procedure that is frequently used to select the best cluster solution In summary CONJOINT SEGMENTER is a fast easy to use procedure for forming segments based on individual utilit
25. riate at tributes and levels and data collection Installing CONSURV to run on a hard disk is reasonably straightforward The Getting Started section of the manual involves typing the word INSTALL and answering a few questions If the user is in doubt about the appropriate re sponse accepting the defaults will usually suffice Comput er requirements include an IBM PC AT or compatible 640k RAM DOS 3 3 or higher and a hard drive with 1 5 MB of free space The software requires a security key to run Once the program is loaded it is easy to follow the ex ample in the manual to complete a demonstration problem Discussion of these steps follows CONSURYV is designed with pull down menus and is quite easy to use Once the program is started the Main Menu lists the four major components of a conjoint analysis study design instrument construction analysis and market simulations To begin the attributes and levels must be spec ified to select an appropriate experimental design the labels for both the attributes and levels must be specified A maxi mum of 200 attributes at 10 levels each can be entered CONSURYV itself can generate a design with a maximum of 30 attributes After entering the attributes and levels the file can be saved and the user can proceed to the generation of the ex perimental design It is important to remember that when a file is saved there is no notification whether another file of the same name exists CONSU
26. ribe 100 statistical tests These descriptions coupled with the statistical tables provided make it a useful reference DEL I HAWKINS University of Oregon CONJOINT DESIGNER Version 3 1990 CONJOINT ANALYZER Version 3 1992 SIMGRAF Version 2 1992 CONJOINT LINMAP 1989 CONJOINT SEG 113 MENTER 1993 BRIDGER 1988 New York Bretton Clark 500 to 900 per program student versions 5 per program limited problem size CONSURV Conjoint Analysis Software Version 3 0 April 8 1993 Edmonton Alberta Intelligent Marketing Systems Inc 549 student version 15 ADAPTIVE CONJOINT ANALYSIS SYSTEM ACA Version 4 0 1994 CHOICE BASED CONJOINT SYS TEM CBC 1994 CONJOINT VALUE ANALYSIS SYSTEM CVA Version 1 1 1994 1 500 to 3 000 student version 300 for ACA Since the late 1960s when it was first introduced to mar keting researchers conjoint analysis has become one of their favorite tools for understanding and predicting the choice behavior of consumers Early users of the technique either had to either develop their own software or adapt mainframe statistical packages to perform the calculations for conjoint analysis A knowledge of experimental design also was nec essary to select fractional factorial designs that met certain constraints e g orthogonality This was extremely important because the number of at tributes e g features factors and dimensions and attribute levels included in most fu
27. rketing Association and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder s express written permission However users may print download or email articles for individual use
28. rs to be no way to easily check for duplicate profiles assuming they can occur and factor level combinations that do not make sense The formatting option on the main menu can be used to generate the survey instrument However this is somewhat tedious the first time through but after doing it a few times New Books in Review it probably becomes much easier The exact terminology to be used must be selected to collect the data along with the particular type of scale for example five or seven points The number of profiles to be shown on a page and their form for example row or column format must be defined as well as the order in which they are to be shown After these choices have been made a print file containing the questionnaire can be created Various files are created using these steps and then brought together for the final instrument file CONSURV creates up to eight files in the process of completing the several tasks in a conjoint study After the data are collected they can be entered using the CONSURYV data entry facility or by creating an ASCII data file with some other application program Entering the data directly into CONSURV is relatively straightforward The profile fields appear in the order in the questionnaire so one just enters the data as they appear Checking for out of range values is possible but it is also possible to enter a blank without any warning from the system it is then treat ed as missing da
29. s Although it is not as user friendly as ACA CVA is still easy to use and of fers capabilities not provided in other packages CHOICE BASED CONJOINT CBC CBC the latest program in the SSI suite of conjoint anal ysis software is designed to capture consumer preferences in a simulated choice environment that is selection of a product from two or more products each defined by all the relevant attributes full profile presentation from a PC based questionnaire In this simulated choice context CBC allows the respondent the option of none of the products shown to more closely simulate an actual choice scenario than the more typical conjoint data collection of rating or ranking product profiles Because the analysis is done at the aggregate level no individual utilities at the attribute level are calculated interactions of attributes can be investigated The statistical algorithm used to calculate the group utili ties is multinomial logit analysis a technique frequently used in transportation research but less in marketing It is available in most mainframe statistical packages The design of the set of choices by CBC is done through random con struction rather than the more typical fractional factorial main effects only designs The manual claims the random approach is preferred in this context because all interactions not just the two way or three way interactions or the pre specified interactions as in a compromise type design
30. s looking for numerous ways to 1 analyze a priori segments of the data 2 test several traditional choice models i e first choice Bayesian probabilistic and 3 test modifica tions improvements of the traditional choice models All this can be done with a minimum of hassle If BC products are not used to generate the utilities data file layouts are pre sented so that other data can be put into the BC format to use SIMGRAF 2 115 CONJOINT SEGMENTER CONJOINT SEGMENTER is intended to provide an easy to use clustering program to group respondents who have similar patterns of utilities It is designed for re searchers who choose to treat the attributes in a conjoint analysis study as benefits and wish to identify benefit seg ments using the individual factor level utilities The claimed advantage of using CONJOINT SEG MENTER as the clustering procedure instead of some other mainframe or PC based clustering algorithm is 1 it is ca pable of clustering a very large data set both the number of respondents and the number of factor levels because of its use of virtual memory and 2 it uses modifications to Ward s hierarchical method of clustering which was sug gested by Srinivasan and Weir 1992 and is not readily available elsewhere We feel it is important to remind readers of the typical caveat when using clustering procedures Different algo rithms yield different results and unless we have external criteria
31. spondent where data is missing In addition there is no no tification that respondents with missing data are not ana lyzed Individual utilities are the input to the market simulator To run the simulations the products to be simulated are defined and then the appropriate choice model is selected All choice models are variations of the first choice model These are the traditional first choice maximum utility gets complete choice residual sampling error added to the total utility for each product for each respondent and normal sampling error added is from normal distribution A sensitivity analysis can easily be done by allowing all attributes to vary among their prespecified levels one factor changing at a time This is very convenient because sever al keystrokes are required to change a product concept de scription Interpolation between factor levels is permissible for all product concepts The results from the simulation 117 runs are not shown on the screen the user must go to the ed itor within CONSURV to check the results The format of the file is not very compact e g the product descriptions are listed sequentially instead of using the more compact side by side format As a result the output will be volumi nous for even a moderate number of products and runs The manual contains an index and a list of error messages that can occur when creating a survey These features as well as on line help are very usef
32. ta in the calculations Whereas the system permits entry of a blank for a given response it does not per mit going back and editing a response by changing a valid response to a blank So an error is made in entering data and a data point had to be changed to reflect that it is missing it is not clear how this would be possible One segmentation variable to be used later in subgroup analyses can be entered A weight for each respondent can be specified to adjust for sample differences from the popu lation Following data entry the factor levels must be recorded and the type of model used for estimating the util ities must be scheduled A decision also can be made to drop an attribute at this point based on its insignificance in a previous analysis The analyst can select from several coding options each with certain advantages and disadvan tages which are briefly discussed in the manual If the de sign selected will allow the estimation of interaction terms the attributes interacting at this point can be specified After completing these specifications the utilities for each re spondent can be calculated using multiple linear regression Several statistics for each type of estimation are available The significance of each of the attributes by segment and for the complete sample can be checked It is possible to see how well the selected model fits each respondent However it should be noted that no utilities are generated for a re
33. terns can be New Books in Review sought and a decision made whether they were seriously playing the game Once the data are satisfactorily cleaned the final step is to perform market simulations A log file is maintained so that information on each run can be kept for later analysis The models available are first choice and probabilistic The man ual is very good at explaining how to interpret much of the output from the simulations and other tasks for example holdout samples the relative importance of attributes and interactions Although much can be gleaned from the basic simulation output more powerful capabilities are available only in SIMGRAF another BC product discussed below In summary CONJOINT ANALYZER is the basic OLS analytic engine in the BC suite of programs It performs all the tasks after selecting a design and of course collecting the data In addition it provides some interesting cleaning options that are designed to ensure further analysis of only those respondents that can be modeled very well the defini tion of very well is of course up to the analyst SIMGRAF For the researcher who is interested in serious simulation e g testing several different choice models sensitivity analysis and segment analysis SIMGRAF is BC s answer to your needs This version differs from earlier versions be cause it incorporates several more realistic decision rules e g a no buy option and a threshold value f
34. tion of some coefficients such as main effects or interactions It 1s strongly recommended that this test be performed Possibly the system should be modified so that it automatically does the test or checks to see whether it had been done This would prevent unpleas ant surprises later in the analysis phase and after the data have been collected Conduct Data Analysis This module enables the user to first combine data for a maximum of 2 000 respondents and a maximum of 50 seg mentation variables from the field that is to define the group before doing any calculations These variables allow for analysis by prespecified subgroups Then the user can choose one of two options for analyzing the data The sim plest procedure merely counts the number of times an at tribute level was included in the selected products from each choice scenario The second procedure is multinomial logit analysis The manual is very good at presenting the advantages and disad vantages of using this procedure and also at interpreting the output Not all effects can be estimated simultaneously a maximum of 90 parameters can be used so it is necessary to define them before running the logit analysis Because logit analyses are iterative procedures and there fore time consuming CBC allows the user to build logit files in binary form so that reads and writes during the calcula tions proceed much more quickly For the demanding statis tician who is comfortabl
35. to spot an unreasonable prod uct concept for example if a car with 500 hp gets 50 mpg in the city and costs 10 000 it s just too good to be true at this stage rather than when the study is in the field CON JOINT DESIGNER is very easy to use in this add sub tract inspect mode of the design phase which is one of the most important phases in a conjoint project One annoying feature that appears in this version of CONJOINT DESIGNER but not in earlier versions is the automatic randomizing of the labels of each factor This was an option in the earlier version that is automatically includ ed in this version The BC justification for forced random ization appears to be that analysts must be protected from themselves they don t always randomize and in the exper imental design course one was taught to always randomize The current version provides a cumbersome way of reorder ing the labels back to the original order if you so desire but this seems to be an unnecessary step In any truly user friendly environment the analyst should be given the option of randomizing or not even if the chosen option is incorrect CONJOINT DESIGNER is still probably one of the most frequently used packages in the BC suite of programs it is an important easy to use program every full profile con joint study requires a design The manual itself is useful to analysts as a reference because it contains discussions and recommendations on several import
36. ty of Denver ACKNOWLEDGMENT This review has benefitted from comments by the vendors and Dr P E Green Wharton School University of Pennsylvania REFERENCES Albaum Gerald 1989 Review BRIDGER and SIMGRAF Journal of Marketing Research 26 November 486 88 and Carmone Frank J 1991 Review CONJOINT LINMAP Journal of Marketing Research 28 February 117 19 Carmone Frank J 1986 Review CONJOINT DESIGNER Journal of Marketing Research 23 August 311 12 1987 Review ACA System for Adaptive Conjoint Analysis Journal of Marketing Research 24 August 325 27 Green Paul E 1987 Review CONJOINT ANALYZER Jour nal of Marketing Research 24 August 327 29 1992 Review CONSURV CONJOINT ANALYSIS SOFTWARE Journal of Marketing Research 29 August 387 90 Herman Steven J and Shocker Alan D 1993 The Effective ness of Alternative Preference Elicitation Procedures in Predict ing Choice A Comment Morristown NJ Bretton Clark Johnson Richard M 1991 Comments on Studies Dealing with ACA Validity and Accuracy With Suggestions for Future Re search working paper Ketchum ID Sawtooth Software Srinivasan V and Weir H 1992 A Conjoint Analysis Based Approach for Determining Benefit Segments June 1992 Ad vanced Research Techniques Forum Lake Tahoe NV Copyright of Journal of Marketing Research JMR is the property of American Ma
37. ul to first time and infre quent users For the more sophisticated users file layouts are also available Another useful feature is that the software remembers previously entered values for as long as the pro gram runs before turning off the machine or exiting to an other application and allows simple editing for changes It appears that one of the penalties of designing user friendly highly interactive interfaces is that the user cannot run CONSURYV in a batch mode That is the user cannot put together a batch file containing the control input for several runs of the simulation at one time and use this file as input to CONSURYV Instead every run except those using the all design levels option must be entered sequentially from the keyboard Analysis that requires several runs of the simula tor can become tedious As mentioned previously the save option overwrites ex isting files without warning Another slight annoyance is that in the pull down menus the first letters are not always unique but the user is supposed to be able to select an op tion by just typing the first letter For example typing M can result in either move or modify when they appear in the same menu A simple programming change could use move and modify for example when they appear together Sawtooth Software Inc ADAPTIVE CONJOINT ANALYSIS ACA ACA is the only commercially available software for adaptively collecting conjoint like data from computer
Download Pdf Manuals
Related Search