Alternative Models For Capturing the Compromise Effect

EXTENDED ABSTRACT - The compromise effect denotes the finding that brands gain share when they become the intermediate rather than extreme options in choice sets (Simonson 1989). It has substantial implications for consumer choice, and represents a significant violation of standard microeconomic theory. Despite the robustness and importance of this phenomenon, choice modelers have neglected to incorporate the compromise effect within formal choice models and to test whether such models outperform the standard value maximization model. In this article, we suggest four context-dependent models that capture the compromise effect within a heterogeneous multiattribute logit framework.

Citation:

Ran Kivetz, Oded Netzer, and V. Srinivasan (2003) ,"Alternative Models For Capturing the Compromise Effect", in NA - Advances in Consumer Research Volume 30, eds. Punam Anand Keller and Dennis W. Rook, Valdosta, GA : Association for Consumer Research, Pages: 111.

Advances in Consumer Research Volume 30, 2003     Page 111

ALTERNATIVE MODELS FOR CAPTURING THE COMPROMISE EFFECT

Ran Kivetz, Columbia University

Oded Netzer, Stanford University

V. Srinivasan, Stanford University

EXTENDED ABSTRACT -

The compromise effect denotes the finding that brands gain share when they become the intermediate rather than extreme options in choice sets (Simonson 1989). It has substantial implications for consumer choice, and represents a significant violation of standard microeconomic theory. Despite the robustness and importance of this phenomenon, choice modelers have neglected to incorporate the compromise effect within formal choice models and to test whether such models outperform the standard value maximization model. In this article, we suggest four context-dependent models that capture the compromise effect within a heterogeneous multiattribute logit framework.

All of these models are motivated by theory from decision and consumer behavior research. A key characteristic of all models is that they view choice as a constructive process (e.g., Payne, Bettman, and Johnson 1992), whereby consumers construct their preferences based on the local choice set. Further, following the suggestion of Hardie, Johnson, and Fader (1993), the four alternative models incorporate both relative (i.e., reference-dependent) and absolute (or global) elements of consumer choice. In particular, our modeling approach assumes that the utilities (partworths) of attribute levels are known and have been measured at a global (context-independent) level; these global partworth functions can assume any shape (linear, concave, convex, etc.). The models then transform these partworth utilities according to the local context (i.e., based on the relationships among options in a given choice set). Thus, the alternative models are illustrated using equations and graphs that operate at the subjective utility space rather than the objective attribute level space. Moreover, the alternative models consist of individual-level utility functions, and therefore, account for heterogeneity through the estimated context-independent partworths. The proposed models are in no way limited to a particular method of preference modeling (e.g., partworth function vs. vector model) or preference measurement technique (e.g., full profile vs. self-explicated approach). That is, when discussing the models, we use the term "partworth" in the most general sense, to denote the utility or worth of a specific attribute level.

The alternative models differ with respect to the particular mechanism that underlies the compromise effect. In particular, the first two models, termed the Contextual Concavity Model (hereafter, CCM) and the Normalized Contextual Concavity Model (NCCM) account for the compromise effect by combining the notions of diminishing sensitivity and context-dependence; in other words, via "contextual concavity." A third model, called the Relative Advantage Model (RAM), is based on a modeling framework proposed by Tversky and Simonson (1993). A fourth model, entitled the Loss Aversion Model (hereafter, LAM), captures the compromise effect by incorporating two major principles that have emerged from behavioral decision research: reference-dependence and loss aversion (e.g., Tversky and Kahneman 1991).

Using two empirical applications and experimentally generated choice data (in two product categories), we contrast these alternative models among themselves and relative to the standard value maximization model. We employ a partworth function preference model (see, e.g., Green and Srinivasan 1978, 1990) and estimate individual-level partworths using the self-explicated approach (e.g., Srinivasan and Park 1997). The self-explicated task covers the entire range of attribute levels used in the choice study; thus, the self-explicated partworths are independent of the local choice context. We also employ choices calibrated by the multinomial logit model.

The calibration and validation results indicate that accounting for the local choice context, or the relative positions of options within the choice set, can significantly improve the predictive validity and fit of the standard choice model. Indeed, all four alternative models provided a significant improvement in fit over the nested value maximization model (hereafter, VMM), as indicated by the likelihood ratio test. More importantly, the validation results indicated that incorporating the compromise effect can lead to significantly higher predictive validity. However, while the contextual concavity and loss aversion models (i.e., CCM, NCCM, and LAM) exhibited superior predictive validity relative to the standard VMM, the RAM did not. In particular, the aggregate-level cross choice set predictions showed that the contextual concavity and loss aversion models yielded substantial improvements in predictive validity (ranging from 50 to 67%) over the VMM. Conversely, averaged across the two product categories, the RAM provided no improvement. Further, these tests suggested that the contextual concavity models had increased predictive validity relative to the LAM, and that the NCCM was somewhat better than the CCM. The individual-level cross choice set predictions, as well as the BIC measures of fit, also supported the notion that the NCCM, CCM, and LAM were superior to the VMM and RAM. In addition to their superior predictive validity and fit, the three leading models also predicted the enhanced shares of the compromise options and the substantial compromise effects observed in the validation data (ranging from 15% to 34%). By contrast, the RAM was unable to predict these compromise effects, and the VMM even predicted reversed compromise effects, consistent with betweenness inequality and the similarity hypothesis (Tversky and Simonson 1993).

A second empirical application generalizes the compromise effect to larger choice sets with five alternatives and four product attributes (instead of the traditional design of 3 options x 2 attributes). The results indicate strong compromise effects. Moreover, as in the first empirical application, the NCCM, CCM, and LAM are superior to the VMM and RAM both in terms of fit and predictive validity.

In addition to empirically testing the alternative models, we discuss the conceptual similarities and differences between the models and their ability to predict other context effects (e.g., asymmetric dominance). We conclude by suggesting that the models proposed in this paper have practical implications for applications that utilize choice models, such as conjoint choice simulators and for predicting choice shares for any given product line portfolio. In addition, we believe that the present research can be seen as part of the ongoing fruitful attempt to bridge the consumer behavior and marketing science disciplines.

----------------------------------------