Do Intentions Really Predict Behavior? Self-Generated Validity Effects in Survey Research Studies of the relationship between purchase intentions and purchase behavior have ignored the possibility that the very act of measurement may inflate the association between intentions and behavior, a phenomenon called “self-generated validity. In this research, the authors develop a latent model of the reactive effects of measurement that is applicable to intentions, attitude, or satisfaction data, and they show that this model can be estimated with a two-stage procedure.
In the first stage, the authors use data from surveyed consumers to predict the presurvey latent purchase intentions of both surveyed and nonsurveyed consumers. In the second stage, they compare the strength of the association between the presurvey latent intentions and the postsurvey behavior across both groups.
The authors find large and reliable self-generated validity effects across three diverse large-scale field studies. On average, the correlation between latent intentions and purchase behavior is 58% greater among surveyed consumers than it is among similar nonsurveyed consumers.
One study also shows that the reactive effect of the measurement of purchase intentions is entirely mediated by self-generated validity and not by social norms, intention modification, or other measurement effects that are independent of presurvey latent intentions. nsumers’ self-reported intentions have been used widely in academic and commercial research because they represent easy-to-collect proxies of behavior. For example, most academic studies of satisfaction use consumers’ intentions to repurchase as the criterion variable (for an exception, see Bolton 1998), and most companies rely on consumers’ purchase intentions to forecast their adoption of new products or the repeat purchase of existing ones (Jamieson and Bass 1989).
However, it is well known that consumers’ self-reported purchase intentions do not perfectly predict their future purchase behavior, nor do these differences cancel each other out when intentions and behavior are aggregated across consumers. In a metaanalysis of 87 behaviors, Sheppard, Hartwick, and Warshaw (1988) find a frequency-weighted average correlation between intentions and behavior of .53, with wide varia- C
Pierre Chandon is Assistant Professor of Marketing, INSEAD, and currently he is Visiting Assistant Professor of Marketing, Kellogg School of Management, Northwestern University (e-mail: pierre. [email protected] edu). Vicki G. Morwitz is Associate Professor of Marketing and Robert Stansky Faculty Research Fellow, Stern School of Business, New York University (e-mail: [email protected] nyu. edu). Werner J. Reinartz is Associate Professor of Marketing, INSEAD (e-mail: werner. [email protected] edu).
The authors acknowledge the helpful input of the anonymous JM reviewers, John Lynch, Gilles Laurent, and Albert Bemmaor, as well as those who participated when the authors presented this research at INSEAD; at the Association for Consumer Research Conference, Portland; at the Marketing Science Conference, College Park, Md. ; and at the ESSEC-HEC-INSEAD conference. In addition, the authors thank the French online grocer Prodigy Services Company and Allison Fisher for providing data and INSEAD for its financial assistance. tions across measures of intentions and types of behavior (for a review, see Morwitz 2001).
To improve the ability to forecast behavior from intentions, researchers have tested alternative scales (Reichheld 2003; Wansink and Ray 2000) and have developed models that account for biases in the measurement and reporting of intentions, the heterogeneity across customers, changes in true intentions between the time of the survey and the time of the behavior, and the stochastic and nonlinear nature of the relationship between intentions and behavior (Bemmaor 1995; Hsiao, Sun, and Morwitz 2002; Juster 1966; Kalwani and Silk 1982; Manski 1990; Mittal and Kamakura 2001; Morrison 1979).
In practice, the studies adjust the intention scores by analyzing the actual purchase behavior of consumers whose purchase intentions have been measured previously. For example, the popular ACNielsen BASES model forecasts aggregate purchase rates by applying conversion rates to measured purchase intentions (e. g. , it assumes that 75% of consumers who checked the top purchase-intentions box will actually purchase the product). To obtain these conversion rates, BASES uses previous studies that measured the purchase intentions of consumers and then tracked their actual purchases.
However, a limitation of these studies is that they focus on the internal rather than the external accuracy of purchase-intention measures. That is, the studies measure the improvement in the ability to forecast the behavior of consumers whose intentions they previously measured, not the behavior of consumers whose intentions they did not measure. Therefore, the studies assume that they can extrapolate the intention–behavior relationship of nonsurveyed consumers on the basis of the relationship that surveyed consumers exhibit. In doing so, the studies ignore the Intentions and Predicting Behavior / 1
Journal of Marketing Vol. 69 (April 2005), 1–14 potentially important problem that the measurement of intentions itself might self-generate some of the association between the intentions and the behavior of a particular consumer (Feldman and Lynch 1988). Finding that part of the predictive power of purchase intentions is an artifact of the measurement would have serious implications for researchers and managers. It would suggest that studies that measure the strength of the association between intentions and behavior on the same sample of consumers overstate the external predictive accuracy of purchase intentions.
This would explain why so many new products fail even after they perform well in purchaseintention tests. In general, researchers who are interested in measuring the true association between two constructs (in this case, for consumers whose behavior was not influenced by the measurement) would need a method that detects and corrects for the effects of measurement. In this research, we develop a comprehensive latent framework to conceptualize the reactive effects of the measurement of purchase intentions.
This framework distinguishes between two sources of measurement reactivity. The first is self-generated validity effects, which we define as a strengthened relationship between latent intentions and behavior due to the measurement of intentions. The second source includes all measurement effects that are independent of latent intentions, such as those that social norms or postsurvey intention modifications create. We also describe a two-stage procedure to detect whether the act of measurement alters the strength of he relationship between a latent construct that is measured through surveys, experiments, or observations and its consequence (e. g. , intentions–behavior, attitudes–intentions, attitudes–behavior, satisfaction–behavior) and to determine the true magnitude of the relationship in the absence of measurement. We demonstrate three empirical applications of this method using large-scale data sets that contain purchase or profitability data from both consumers whose purchase intentions were measured and similar consumers whose purchase intentions were not measured.
In the three applications (groceries, automobiles, and personal computers [PCs]), we show that the strength of the relationship between latent intentions and behavior is stronger for surveyed consumers than for similar nonsurveyed consumers. In the final section, we discuss the managerial and research implications of our results. In competitive markets in which most existing customers have positive attitudes toward a product category, the measurement of purchase intentions increases purchasing in the category of accessible and preferred brands.
Research has shown these effects for both hypothetical and real brands, for financially important and relatively inconsequential behaviors, and for short (a few minutes) and long (six months) delays between the measurement and the behavior (Chandon, Morwitz, and Reinartz 2004; Dholakia and Morwitz 2002b; Fitzsimons and Morwitz 1996; Morwitz and Fitzsimons 2004). In a related stream of research, studies show that asking consumers to predict their future behavior influences the likelihood that they will engage in that behavior (Sherman 1980; Spangenberg 1997; Spangenberg and Greenwald 1999; Sprott et al. 003). Focusing on socially normative behavior, these studies demonstrate that if respondents are asked to predict the likelihood that they will perform a behavior in the future, they are more likely to engage in socially desirable behaviors, such as voting or recycling, and less likely to engage in socially undesirable behaviors, such as singing “The Star-Spangled Banner” over the telephone. Self-Generated Validity Theory The self-generated validity theory (Feldman and Lynch 1988), the most popular explanation of the reactive effects of measurement, uses two lines of argument.
First, preexisting intentions may become more accessible in memory when the researcher asks the question. (It is also possible that consumers have no preexisting intentions and form them only in response to the researcher’s question. ) The measurement process thereby leads survey respondents to form judgments that they otherwise would not access in their memory or that they otherwise would not form. Second, higher relative accessibility and diagnosticity of intentions, compared with other inputs for purchase decisions (e. g. tastes, mood, competitive environment), may make subsequent purchase behavior more consistent with prior intentions. Several studies provide indirect evidence in support of the self-generated validity theory for public opinion (Simmons, Bickart, and Lynch 1993) and marketing research (Fitzsimons and Morwitz 1996; Morwitz and Fitzsimons 2004; Morwitz, Johnson, and Schmittlein 1993). However, none has examined the core prediction of self-generated validity theory directly, namely, that the association between prior intentions and behavior is stronger among surveyed consumers than among similar nonsurveyed consumers.
The studies have been unable to test this prediction because they have not estimated the purchase intentions of consumers who were not surveyed. Consistent with Feldman and Lynch’s (1988) predictions, Fitzsimons and Morwitz (1996) find that the measurement of general intentions to purchase automobiles increases the likelihood that buyers will repurchase the automobile brand that they previously owned and that firsttime buyers will purchase brands with large market shares. Under the assumption that prior purchase rates or market shares are proxies for latent, brand-specific purchase inten-
Self-Generated Validity and Other Sources of Measurement Reactivity Reactive Effects of Measurement Ample evidence indicates that measurement can influence both the intensity of a measured construct and its association with other constructs. In intentions research, the reactive effects of measurement have been called the “mere measurement effect,” “the self-erasing error of prediction,” and “self-prophecy. ” We refer to the behavioral differences between surveyed and nonsurveyed consumers as the “reactive effects of measurement” or simply as “measurement reactivity. 2 / Journal of Marketing, April 2005 tions, Fitzsimons and Morwitz’s results suggest that the measurement of general intentions increases the association between latent, brand-specific intent and brand choice. Similarly, in a series of laboratory studies, Morwitz and Fitzsimons (2004) find that the measurement of general purchase intentions for candy bars makes consumers more likely to choose brands that they like and less likely to choose those that they dislike.
If we consider prior brand preference a proxy for intention to purchase the brand, this study suggests that the measurement of consumers’ intentions to buy from the category increases the association between their latent intentions to buy the brand and the likelihood of subsequently choosing this brand. Finally, Morwitz, Johnson, and Schmittlein (1993) examine the effects of repeated measurements of intentions (and behavior) on people with high and low initial measured purchase intentions.
They find that the repeated measurement of intentions and behavior increases the association between behavior and the initial measure of intentions. However, their analysis is restricted to consumers whose purchase intentions have been measured at least once. Other Sources of Measurement Reactivity Ignoring obvious alternative explanations, such as selection biases, that violate our definitional assumption that surveyed and nonsurveyed consumers are identical, we consider at least two other explanations for the reactive effects of measurement in purchase-intentions surveys: social norms and intention modification.
Both differ from selfgenerated validity in that they operate independently of consumers’ intentions at the time of the survey. In the context of socially normative behaviors, Sherman (1980) shows that asking people to predict their future behavior biases their reported intentions toward a social norm (e. g. , donating to charities, not singing over the telephone). Consumers then act according to their newly reported intentions, not according to their prior unreported intentions, to reduce the cognitive dissonance between their reported intentions and their behavior (Spangenberg and Greenwald 1999; Sprott et al. 003). With regard to intention modification, consumers tend to evaluate market research surveys positively because they either find the survey informative or enjoy being asked their opinion (Dholakia and Morwitz 2002a; Sudman and Wansink 2002). In a subsequent stage, this positive evaluation of the survey carries over to the evaluation of the company and its products. Consumers also regard the survey as a signal of the firm’s customer orientation, which directly improves their evaluation of the company and its products. In both ases, the positive attitude triggered by the survey leads to greater purchasing by surveyed consumers. Both explanations share the view that the measurement of purchase intentions modifies consumers’ purchase intentions rather than makes prior intentions more accessible in memory or more diagnostic of future purchase decisions. In the context of purchase-intention surveys for common products and services, the measurement effects make consumers more likely to report positive purchase intentions and then actually purchase the product, regardless of their purchase intentions at the time of the survey.
Summary To better understand the differences between the possible sources of measurement reactivity, in Figure 1 we plot hypothetical purchase behavior (e. g. , purchase quantity) as a function of presurvey, latent (i. e. , unmeasured) purchase intentions for both consumers whose intentions were not measured (control group) and those whose intentions were measured.
In Figure 1, we show that the different sources of measurement reactivity have markedly different effects on purchase behavior and on the link between intentions and behavior. Intention modification leads to a consistent upward shift in purchase behavior but leaves the slope of the relationship between presurvey intentions and behavior unchanged. In contrast, self-generated validity effects do not lead to a general increase in purchase behavior but strengthen the association between intentions and behavior.
If measurement reactivity is due to self-generated validity, intention measurement makes consumers with positive purchase intentions more likely to purchase but also makes consumers with negative purchase intentions less likely to purchase, which increases the steepness of the slope between intentions and behavior. In Figure 1, we also show that in contrast to intention modification, self-generated validity effects do not necessarily lead to measurement reactivity. For example, the FIGURE 1 Self-Generated Validity and Other Sources of Measurement Reactivity High Low –– – 0 + ++ Behavior
Prior Latent Intentions Survey group (with other measurement effects) Survey group (with selfgenerated validity effects) Control group Intentions and Predicting Behavior / 3 measurement of intentions does not change the purchase behavior of consumers who have neutral purchase intentions, that is, those who are undecided about purchasing and not purchasing. Similarly, self-generated validity effects cancel out if there are as many positively inclined consumers as there are negatively inclined ones (i. e. , if the distribution of purchase intentions is symmetric around the neutral point).
In this case, the average purchase behavior of surveyed consumers may be the same as the average purchase behavior of similar nonsurveyed consumers, though the purchase behavior of each consumer is more extreme. However, self-generated validity effects are a sufficient condition for measurement reactivity when the majority of consumers have positive purchase intentions—the most common case in field studies of actual products in competitive markets—because the measurement of purchase intentions makes these consumers more likely to follow their intentions (i. . , more likely to purchase). Conceptualizing and Estimating the Reactive Effects of Measurement A Latent Model of the Effects of the Measurement of Purchase Intentions The framework we present in Figure 2 relates purchase behavior (B) to measured (self-reported) purchase intentions (MI), prior latent (unmeasured) purchase intentions (LI), and the survey that measures purchase intentions (S).
In line with conventional representations of structural equation models, we use rectangles to represent observed variables, ovals for latent variables, arrows between constructs for causal relations, an arrow pointing to another arrow for an interaction effect, and a double arrow for a correlation. We consider LI an unobserved hypothetical construct that captures, without error, consumers’ determination to purchase just before the time of the survey. Thus, B is a function of LI (with regression coefficient ? 1) and random error (? ).
In the model, we assume that all consumers, both surveyed and nonsurveyed, have some latent purchase intentions at the time of the survey. However, this assumption does not imply that consumers have decided whether to buy before the survey, because prior latent intentions can be neutral; rather, it implies that consumers do not form intentions only when they are surveyed (we explore the implications of this assumption in the “General Discussion” section). By definition, these prior latent intentions are independent of whether the consumers’ intentions are surveyed or not.
If S is randomly administered, LI are identical for surveyed and nonsurveyed consumers, as we show in Figure 2 by excluding a link between S and LI. We present the observed measures of LI on the left-hand side of Figure 2. Purchase intentions measured by the survey constitute one such measure, but this is not the only one. We can also measure latent intentions by other reflective indicators (denoted RI1, RI2, …, RIn), including indirect measures, such as physiological measures or implicit tests, and behavioral measures, such as information search or the purchase of complementary products.
Both LI and the measurement error (? RI) influence these reflective indicators. Other indicators of LI may be formative (e. g. , prior purchase behavior, demographics), in which case LI is a function of the m formative indicators (denoted FI1, FI2, …, FIm) and a random disturbance term (? FI). We assume that these other indicators are independent of intention measurement (no correlation with S), whereas MI exist only for surveyed consumers (the correlation between MI and S is one). To identify the latent model, we must scale it by choosing one indicator for which the factor loading is set to one and the intercept is zero.
Choosing MI as the scaling indicator enables us to scale the LI to the familiar units of MI. In doing so, we assume that there are no systematic reporting biases and that surveyed consumers retrieve their prior LI from memory. (We subsequently report simulation studies in which we examine what happens if MI are systematically biased upward because of social norms or intention modification. ) With the latent model, we can define self-generated validity effects more broadly. Originally, Feldman and Lynch (1988) studied the effects of measurement on the FIGURE 2 A Latent Model of the Reactive Effects of the Measurement of
Purchase Intentions 1 Survey (S) ? 2 ?MI Measured intentions (MI) Reflective indicator (RI) Formative indicator (FI) 1 ?FI Prior latent intentions (LI) ?3 ?RI ? RI ? FI ?1 Behavior (B) ? 4 / Journal of Marketing, April 2005 observed correlations among constructs. For example, Simmons, Bickart, and Lynch (1993) asked specific questions about the strength of election candidates before or after they measured general voting intentions. They then measured the impact of question order on the observed correlation between answers to specific questions and general voting intentions.
We argue that the measurement of intentions makes presurvey latent intentions relatively more accessible and diagnostic than it does other antecedents of behavior, which strengthens the relationship between presurvey latent intentions and postsurvey behavior. Therefore, we represent self-generated validity in Figure 2 by the ? 3 parameter, or the effect that S has on the link between LI and B. This broader definition enables us to test for selfgenerated validity effects among latent (nonmeasured) and observed (measured) constructs and not only between observed constructs, as in Simmons, Bickart, and Lynch’s (1993) study.
It also excludes social norms and intentionmodification effects, both of which imply that the surveying of intentions increases purchase behavior independent of prior latent intentions and that the relationship between prior latent intentions and behavior remains the same. However, these other sources of measurement reactivity lead to an increase in purchase behavior, regardless of prior latent intentions. Therefore, in Figure 2, we represent their effects by the ? 2 parameter, which captures the effect of S on B and is not mediated by the strengthening of the relationship between LI and B.