Beyond valence: a meta-analysis of discrete emotions in firm-customer encounters

Appraisal theories of emotions (e.g., Ellsworth and Scherer 2003; Smith and Ellsworth 1985) and the appraisal-tendency framework (Han et al. 2007; Lerner and Keltner 2000; Roseman 1991; So et al. 2015) propose that each discrete emotion can be defined by a unique set of central dimensions that describe its core meaning (Lazarus 1991; Lerner and Keltner 2000; Smith and Ellsworth 1985). This set is referred to as the appraisal pattern of the emotion (cf., Han et al. 2007). Appraisals can be described as the meaning-making process that interprets the change in core affect with regard to the specifics of the event that triggered the emotional experience and thus gives rise to different discrete emotions (Keltner and Horberg 2015).

The labels and the total number of identified appraisal dimensions varies across studies (cf., Frijda et al. 1989; Lazarus 1991; Roseman 1996; Smith and Ellsworth 1985). Usually between six (e.g., Smith and Ellsworth 1985; So et al. 2015) and nine (e.g., Roseman 1996; Van Dijk and Zeelenberg 2002) separate appraisal dimensions are used to describe discrete emotions, often including appraisals that overlap with the core affect dimensions of valence and/or arousal (e.g., Frijda et al. 1989; Ruth et al. 2002). Following this tradition, we use the core affect dimensions valence and arousal together with the following four appraisal dimensions to describe and discuss the differences between discrete emotions and their effects in firm–customer encounters. First, certainty refers to the extent to which a person is certain about the consequences of an event. For instance, the emotion anger is associated with high certainty—when a person has clear expectations about the negative consequences of the event, anger about these consequences is elicited (e.g., Lerner and Keltner 2000; Smith and Ellsworth 1985). Fear, in contrast, is associated with low certainty about the negative consequences of the event. One does not know if, how, and when the event creates harm and is therefore fearful. Second, control describes whether an entity had control over a situation or whether it was caused by circumstances. For happiness, for instance, the appraisal of control is not defining. Happiness can arise from someone’s achievement (entity control) or from sunny weather (circumstantial control). However, only the appraisal that one is personally responsible for an accomplishment will lead to pride (i.e., the accomplishment has to be under entity control; Keltner and Horberg 2015). Third, responsibility indicates whether oneself or another entity was deemed responsible for a situation or event (e.g., Roseman 1996). For instance, people assess responsibility for an event (e.g., De Hooge et al. 2014; Frijda et al. 1989; Ortony et al. 1988) which separates emotions related to own actions (e.g., pride) or others’ actions (e.g., gratitude). Similarly, while anger and embarrassment both may signal failure, anger usually arises from an external attribution of responsibility for this failure and embarrassment from an internal attribution (cf., Keltner et al. 1993). Finally, legitimacy focusses on perceptions of own morality in the situation (Roseman 1996; Van Dijk and Zeelenberg 2002) and, for instance, differentiates the emotions guilt and embarrassment from each other. One may perceive both emotions after own failures, but only if these failures are associated with own moral wrongdoing, one would perceive guilt.

Different appraisal patterns (i.e., discrete emotions) cause different reactions with regard to physiology, judgement, choice, and behavior (cf., Frijda et al. 1989; Keltner and Horberg 2015; Lench et al. 2011; Lerner et al. 2003). For instance, while anger and fear show the same level of heart rate acceleration, anger has been found to show higher skin conductance increase, diastolic blood pressure (Roberts and Weerts 1982; Schwartz et al. 1981), as well as higher hand and head temperature (Schachter 1957; Stemmler 1989). This has been explained by anger being associated with a desire to change the situation and to fight the entity (i.e., person) that is deemed to control the situation (Frijda et al. 1989), which leads to physiological preparations for a fight, such as increased blood flow to the hands (Ekman et al. 1983). Consistently, anger has also been shown to be associated with increased activation in the left frontal hemisphere of the brain, a pattern typical for approach motivation (Harmon-Jones 2003; Harmon-Jones and Sigelman 2001). The underlying appraisals of emotions serve as signals of whether a situation threatens or serves a goal and cause specific coping reactions that either serve the goal or try to counter the threat (Frijda et al. 1989; Zeelenberg et al. 2008). Coping reactions may include, for example, approach or avoidance behavior, attention, rejection, or reactance, and have been found to differ significantly between emotions with different appraisal patterns (Duhachek 2005; Frijda et al. 1989). Importantly, although appraisals are tailored to help the individual respond to the event that evoked the emotion, they persist beyond the eliciting situation and become an “unconscious perceptual lens for interpreting subsequent judgments and choices” (Lerner and Tiedens 2006, p. 119), called appraisal tendency.

The appraisal tendency framework (e.g., Keltner and Horberg 2015; Lerner and Keltner 2001) helps understanding when and how specific emotions influence subsequent judgement and behavior. Most importantly, it argues that discrete emotions affect judgment “in a manner consistent with the emotion’s underlying appraisal tendency, but only in domains related to the appraisal” (Keltner and Horberg 2015, p. 641, emphasis added). For instance, the judgement of risk should predominantly be influenced by emotions that are appraised with certainty and control (Han et al. 2007). And indeed, fear, an emotion appraised with uncertainty and lack of individual control, was related to pessimistic risk assessments (Lerner and Keltner 2000, 2001). On the other hand, fear should not affect fairness judgements, as fairness is not associated with certainty or control appraisals (Han et al. 2007). According to the appraisal tendency framework, these appraisals influence judgement and decision making in two distinct processes. On the one hand, appraisals affect the content of thought such that the appraisals that are characteristic for the emotional experience spill over and drive the salience of these appraisal dimensions in subsequent situations. For instance, anger as compared to sadness, is defined by high entity control and low circumstantial control and therefore triggers the tendency to also expect and perceive strong entity control in new situations, which in turn increases the likelihood of blaming others (Keltner et al. 1993). On the other hand, appraisals are also predicted to affect the depth of thought, or in other words the motivation to process information either heuristically or systematically (Lerner and Tiedens 2006; Tiedens and Linton 2001). For instance, Tiedens and Linton (2001; see also Bodenhausen et al. 1994) found evidence that emotions associated with certainty appraisals (e.g., anger, happiness) made people also feel more certain in subsequent situations than emotions associated with uncertainty appraisals (e.g., fear, sadness). In turn, the perception of certainty was associated with heuristic processing, whereas the perception of uncertainty fostered systematic processing.

The first goal of the present meta-analysis is to investigate the effects of different underlying appraisals of emotions on relevant outcomes of firm–customer encounters (such as evaluation of the firm, purchase behavior, and sharing behavior; see below for a discussion of relevant outcome variables). Building on appraisal theories of emotion and the appraisal tendency framework, we expect that emotions during firm–customer encounters will particularly affect outcome variables, if these are consistent with the domain of the underlying appraisal of the emotion. Thus, emotions that share the same association with a specific appraisal dimension, should differently affect outcome variables of firm–customer encounters compared to emotions that are reversely (or not at all) associated with this appraisal dimension. Further, because each discrete emotion uniquely arises from a distinct appraisal pattern (e.g., Ruth et al. 2002; Smith and Ellsworth 1985; So et al. 2015) we expect that interaction patterns among multiple appraisal dimensions may give rise to varying effects of different discrete emotions beyond the main effects of appraisal dimensions (see Han et al. 2007 for a discussion of the need to understand such interaction patterns). We expect that a more fine-grained investigation of discrete emotions in firm–customer encounters beyond mere valence and even beyond the additive main effects of valence, arousal, and the appraisal dimensions adds valuable insights and explanatory power to the understanding of the effects of emotions in firm–customer encounters. To summarize, we formulate the first three research questions (RQ) that we investigate in this meta-analysis as follows:

The focus of most consumer research on emotions lies in investigating how consumers’ emotions impact firm-relevant outcomes such as satisfaction, loyalty, or word-of-mouth. Although correlated, we distinguish between three categories of outcomes to be able to offer more fine-grained advice to managers depending on their most important performance indicator: evaluation, purchase behavior, and sharing behavior. While consumers’ evaluations (e.g., satisfaction) and purchase behaviors (e.g., repeat purchases, loyalty) toward firms have been the focus of most consumer research and managerial attention in the past, a wide range of recent studies have specifically focused on the relationship between emotions and sharing behaviors such as word-of-mouth (e.g., Berger 2014; Watson et al. 2015). Interestingly, these three outcome categories differ in a number of important characteristics which may be more or less congruent with the specific appraisal pattern of a discrete emotion. For instance, evaluations almost naturally occur after any firm–customer encounter and do not require a specific motivation or involve additional effortful action from the consumer. Because evaluations are no decisions nor behaviors (although may lead to both), they are also not associated with a certain risk position or have direct monetary consequences. Both purchase and sharing behaviors, on the other hand, involve decision-making (i.e., whether or not to buy, what to pay, whether or not to share information about the experience, what to share, etc.) and the anticipation of immediate or future action. However, they differ with regard to the associated perceived risk position—only a purchase has direct monetary consequences and is thus associated with financial and performance risks (cf., Kaplan et al. 1974), whereas sharing behavior is predominantly associated with privacy or social risks, or may not be perceived as risky at all (e.g., Eisingerich et al. 2015).

The appraisal tendency framework makes clear suggestions about how emotions affect judgement and decision making by evoking implicit cognitive predispositions when processing subsequent events that are congruent with the appraisal pattern of the emotion. Consistent with the dynamic discussed above, we for instance expect emotions that are defined by appraisals of certainty and/or control to differently affect outcome categories that are directly associated with risks (e.g., purchase behavior) than such that are not (or at least less prominently, e.g., evaluation, sharing behavior). Further, the notion that appraisal patterns spill over to subsequent situations and affect the content of thought (e.g., Han et al. 2007, see also above) suggests that the researched outcome categories may be differently affected by emotions with appraisal patterns that are congruent with the character of the judgement/decision at hand. For instance, concern for other consumers has been found to be one of the most important driving motivations behind sharing (electronic) word-of-mouth (Hennig-Thurau et al. 2004). This other-focus matches with the appraisals of other-responsibility of the emotion anger, which in turn may increase its effect on sharing behavior. And indeed, in a study of Gelbrich (2010) anger was found to increase negative word-of mouth behavior. As another example, purchase behavior requires effortful cost-benefit considerations, while mere evaluation does not. It thus seems reasonable, that the two categories of outcome variables should be differently affected by appraisal patterns that foster or hinder systematic processing (i.e., differently influence the depth of thought; Han et al. 2007; Tiedens and Linton 2001).

Given the simultaneous occurrence of these different appraisal tendency related processes and the notion that each emotion is defined by a unique appraisal pattern (Lazarus 1991; Lerner and Keltner 2000; Smith and Ellsworth 1985), we expect that discrete emotions, even with largely (but not completely) similar appraisal patterns, show deviating influences with regard to the different outcome variables of firm–customer encounters. While the former set of research questions focused on comparing different appraisal dimensions (RQ2) or discrete emotions (RQ3) regarding their effects across all outcome variables combined, the following set of research questions thus focusses on comparing effects of one specific appraisal dimension manifestation (RQ4) or discrete emotion (RQ5) on different individual outcome variables:

In addition to investigating if and how discrete emotions differently affect firm-relevant outcomes (beyond the effects of valence alone), the second major goal of this meta-analysis is to identify moderators of those effects to explain the heterogeneity in effect sizes. It has been argued that different contextual influences of a decision situation may interact with cognitive appraisals of an emotion such that these become more or less activated, salient, or relevant in a situation (So et al. 2015). We argue that the specific marketing context studied in this meta-analysis—firm–customer encounters—provides such situational characteristics that can interact with the appraisal pattern of specific emotions to influence firm-relevant outcomes. We searched our set of included primary studies for potential situational characteristics such as other entities (interactions with employees), the history that led to the formation of the emotion (e.g., recovery efforts), or the context (business-to-business; search, experience, or credence goods or services). The interactions are represented in Fig. 1 and described below.

Taken together, we formulate the following research question to further address potential moderating effects in our meta-analysis:

To identify an encompassing set of studies investigating the effects of discrete emotions on consumer evaluations, purchase behavior, and sharing behavior, we conducted the following steps. First, we searched multiple databases (EBSCO Business Source Premier, Web of Science, Google Scholar) for studies containing the terms emotion*, feeling*, affect*, mood*, or any of the discrete emotions investigated in this meta-analysis in combination with terms suggesting a context of firm–customer encounters (customer*, consumer*, service, shopping, retail, store). Second, we checked the identified studies for potentially relevant cross-references. Third, we also checked the reference lists of previous studies offering an overview of emotions and affect in marketing (e.g., Bagozzi et al. 1999; Erevelles 1998; Laros and Steenkamp 2005; Richins 1997) for any relevant empirical studies. Fourth, we inquired published and unpublished studies from researchers in the field via a relevant mailing list (AMA-ELMAR).

Next, we checked the identified set of studies for eligibility. Given that our research focuses on the impact of discrete emotions on consumer responses to firm–customer encounters, we defined three specific criteria for inclusion: First, we only included studies of firm–customer encounters. Specifically, we only included studies that either (a) were conducted during or directly after a real firm–customer encounter, (b) asked the subjects to recall a real firm–customer encounter, or (c) featured a vignette of a fictional firm–customer encounter. Second, studies had to feature a measurement of an attitudinal or behavioral reaction to the firm–customer encounter or the organization associated (e.g., satisfaction, purchase, word-of-mouth, store, or brand image). Third, studies had to manipulate or measure discrete emotions. Studies that merely reported negative and positive affect were included as well if the applied measurement scale featured discrete emotions as items within the scale (e.g., the PANAS scale, Watson et al. 1988) and effect sizes of these single items could be derived from the information provided in the article. If this was not the case, we emailed the authors and inquired this information (we emailed 58 authors in total and also inquired if they had any unpublished work on the topic). Based on these criteria, our dataset consists of 84 manuscripts featuring 112 independent studies that yielded 1035 effect sizes (see Web Appendix A for a list of papers that form the dataset).

We chose correlation coefficients as the measure of effect size, because the majority of studies in our dataset featured them. Further, most studies did not manipulate emotions but measured them on continuous scales. Thus, correlation coefficients constitute suitable effect sizes for our meta-analysis as they capture the continuous nature of discrete emotions and their consequences (i.e., evaluation, purchase behavior, and sharing behavior). If studies reported regression coefficients only, we transformed them to correlation coefficients based on the formula by Peterson and Brown (2005). Other conversions were based on Borenstein et al. (2013). Next, we corrected all correlation coefficients for systematic variance caused by unreliable measures and transformed them to Fisher z-scores. The reliability-adjusted Fisher z-scores were then weighed with their inverse variance in order to give more weight to more precise effect size estimates (Borenstein et al. 2013). Finally, effect sizes were transformed back to correlation coefficients for presentation.

Studies in our dataset largely varied in terms of participants, settings, and methods. We thus chose to estimate random-effects models as those models do not assume identical true effect sizes across studies. Instead, random-effects models assume effect sizes to be distributed around a true mean value (Borenstein et al. 2013). Further, the majority of manuscripts produced more than one effect size: often manuscripts featured several studies that, in turn, produced effect sizes for several discrete emotions and per discrete emotion correlation coefficients for different outcome variables or moderators. To account for this nested nature of effect sizes (first level: effect sizes; second level: studies; third level: manuscripts), we conducted our meta-analysis based on multilevel (i.e., three-level) models (e.g., Konstantopoulos 2011). We followed the procedure for multilevel meta-analyses as described in Knoll and Matthes (2017). The resulting equations per level look as follows:

Combined into one equation our models are described by:

Observed effect sizes (γ_i) are thus predicted from the average population effect (β_i) as well as a sampling variance component (e_i), a within-manuscript / between-study error (u_i), and a between-manuscript error (a_i). We used the METAFOR package to estimate our models in R. We performed sets of analyses on an overall level, per appraisal category, and per discrete emotion.

First, we assessed the overall average effect size of all emotions on the outcome variables in our dataset. Based on a multilevel meta-analytical model of the absolute values of the correlation coefficients, we find a significant average absolute effect size of all emotions combined of 0.308 (95% CI: 0.109; 0.482). The results showed high and significant heterogeneity among effect sizes (Q (1,034) = 37,606.082, p < .001). Thus, it is valuable to explore this heterogeneity further (Borenstein et al. 2013). We do so by subsequently accounting for differences in (a) manifestations of appraisals (e.g., all emotions that are described by high certainty), (b) discrete emotions, (c) firm-relevant outcome variables, and (d) additional situational characteristics that are relevant to the marketing situation.

To address RQ1 and examine whether an analysis of discrete emotions has more explanatory power than purely valence- (and arousal- and appraisal-) based approaches, we estimated several nested multilevel meta-analytical models on the entire dataset (i.e., for all outcome variables combined). Model 1 included only the valence of emotions as a predictor, while Model 2 included valence and arousal and the interaction of the two variables. Model 3 additionally included the appraisal profiles and Model 4 finally also the discrete emotions as predictors. Likelihood ratio tests suggest that Model 4 (which includes the discrete emotions as predictors) outperforms the valence model (Model 1; χ²(11) = 75.45 > critical value (CV) 31.26, p < .001), the valence and arousal model (Model 2; χ²(8) = 60.39 > CV 26.13, p < .001), as well as the valence, arousal, and appraisal model (Model 3; χ²(2) = 19.22 > CV 13.82, p < .001).

We next analyzed whether the average effect sizes on all outcome variables vary depending on different manifestations of valence, arousal, and the included four appraisals (i.e., control, responsibility, certainty, legitimacy). For this, we conducted several multilevel meta-analyses (including one variable at a time as predictor) on the absolute effect sizes.³ Importantly, we find that the average effect size for positive emotions (0.464; 95% CI: 0.289, 0.610) is larger than that for negative emotions (0.203; 95% CI: −0.025, 0.412; omnibus Wald-type test⁴: Q (1) = 7.386, p = .007; see Table 3). Further, a significant positive average effect size was found for emotions that are high in certainty (0.401; 95% CI: 0.226, 0.552) and emotions high in control (0.352; 95% CI: 0.204, 0.485) but not for emotions low in certainty (0.153; 95% CI: −0.132, 0.414; Q (1) = 6.119, p = .013) or control (0.127; 95% CI: −0.173, 0.405; Q (1) = 4.551, p = .003). Lastly, significant positive average effect sizes were only found for medium- (0.370; 95% CI: 0.149, 0.556) and high-arousal emotions (0.333; 95% CI: 0.130, 0.509) but not for low-arousal emotions (0.162; 95% CI: −0.102, 0.404; Q (2) = 5.841, p = .054).

First, in order to inspect the average effect sizes per discrete emotion, we split the data into sub-samples for each discrete emotion and conducted multiple multilevel meta-analyses. Average effects vary substantially even among emotions with the same valence (Table 4). Effect sizes for positive emotions range (highest to lowest) from 0.688 (95% CI: −0.023; 0.937) for gratitude to 0.174 (95% CI: 0.027; 0.314) for pride. Similarly, effect sizes for negative emotions range from −0.278 (95% CI: −0.427; −0.115) for anger to 0.074 (95% CI: −0.100; 0.242) for guilt. Interestingly, we only found consistent significant effects for four out of ten emotions (i.e., love, happiness, pride, anger), suggesting the importance of moderator analyses. In line with this, even after splitting the sample into sub-samples the effects of all discrete emotions are still characterized by high levels of residual heterogeneity with significant Q-statistics (see Table 4 for a full overview).

A very high fail-safe N (967,593; p < .001; calculated with the Rosenthal approach) suggests that publication bias might not be a serious issue in our dataset. Calculations of the fail-safe Ns per discrete emotion (see Table 4) suggest the same. However, these results should be interpreted with caution, as no other methods to assess publication bias for multilevel meta-analyses are available yet.

In the next step we further investigate the heterogeneity of effect sizes between and within studies, and analyze which marketing relevant characteristics of the emotion-eliciting situation might moderate the effects of discrete emotions. To do so, we first assess whether emotions affect the three classes of outcome variables differently (evaluation, purchase behavior, sharing behavior).

To address RQ4 and test whether the average effect sizes of emotions vary between different outcome variables, we started with estimating a meta-analytical model on the appraisal level. As Fig. 2 shows, we mainly find significant average effects for evaluation and sharing behavior, but only a few significant effects on purchase behavior on the level of discrete emotions. The same key finding applies to the appraisal level. Specifically, we find a significant difference for negative emotions, which have significant average effects on sharing behavior (0.387; 95% CI: 0.269, 0.492) and evaluation (0.231; 95% CI: −0.001, 0.440), but not on purchase behavior (0.138; 95% CI: −0.049, 0.315). Furthermore, our results suggest the same pattern for emotions described by low levels of arousal as well as emotions with low certainty and low legitimacy. All three types of emotions have significant effects for sharing behavior (low arousal: 0.343; 95% CI: 0.208, 0.466; low certainty: 0.336; 95% CI: 0.210, 0.453; low legitimacy: 0.543; 95% CI: 0.343, 0.697) and evaluation (low arousal: 0.181; 95% CI: 0.002, 0.349; low certainty: 0.170; 95% CI: −0.015, 0.345; low legitimacy: 0.189; 95% CI: 0.039, 0.331) but not for purchase behavior (low arousal: 0.112; 95% CI: −0.035, 0.253; low certainty: 0.115; 95% CI: −0.030, 0.256; low legitimacy: 0.125; 95% CI: −0.087, 0.327). Table 5 shows the results for all other appraisal manifestations.

As a next step and to address RQ5, we repeated the analyses on the discrete emotion level. As can be obtained from Table 6 and Fig. 2, the average effect sizes of gratitude, love, and pride were significantly stronger for evaluation (gratitude: 0.752; 95% CI: 0.142, 0.948; love: 0.689; 95% CI: 0.577, 0.775; pride: 0.302; 95% CI: 0.123, 0.462) than for purchase behavior (gratitude: 0.500; 95% CI: −0.275, 0.881; Q (2) = 5.141, p = .076; love: 0.433; 95% CI: 0.261, 0.578; Q (1) = 7.627, p = .006; pride: 0.065; 95% CI: −0.123, 0.248; Q (1) = 5.487, p = .006). The estimates for sharing behaviors did not significantly differ from both evaluation and purchase behavior. We further found that guilt showed stronger average effect sizes for sharing (0.543; 95% CI: 0.282, 0.730) than for purchase behavior (−0.013; 95% CI: −0.273, 0.249) and evaluation (0.044; 95% CI: −0.119, 0.205; Q (1) = 10.937, p = .004).

To analyze the moderating effects of the different situational characteristics, we conducted a series of subgroup analyses per discrete emotion (Table 7; subgroup analyses on the appraisal level can be obtained from Web Appendix C).

The core premise of this research was that emotions are distinct and differ from each other on various important levels and thus cannot be effectively categorized into valence or valence and arousal alone. We therefore suggest that it is valuable to explicitly consider discrete emotions when studying firm–customer encounters. This notion is based on a large body of literature in psychology (e.g., Keltner et al. 1993; Lench et al. 2011; Tiedens and Linton 2001) and has also received support in more managerial work (e.g., Magids et al. 2015). Our meta-analytical synthesis of 1035 effect sizes representing 40,777 research subjects indeed reveals that the effect sizes of discrete emotions of the same valence vary substantially in magnitude, and that the consideration of discrete emotions adds significant explanatory power and precision over the sole use of valence and arousal.

One interesting outcome of our analysis on the appraisal level is that we find a larger general average effect size for positive than for negative emotions. Thus, positive emotions in our study play a stronger and more consistent role in determining firm-relevant outcomes than negative emotions do. Interestingly, these results contribute to the scholarly discussion on whether in general positive or negative emotions have stronger effects on human judgment and behavior. Drawing from prospect theory (Kahneman and Tversky 1979), many scholars have argued and found that negative emotions have stronger effects on humans than positive emotions (e.g., Baumeister et al. 2001). Other studies have found a reversal of this effect and show the occurrence of a positivity bias (for a review see Skowronski and Carlston 1989), based on the notion that humans perceive positive information to be more diagnostic when judging the ability of an entity. Thus, positive emotions might also be perceived as being more diagnostic when evaluating a firm and, in turn, have stronger effects on firm-relevant outcome variables. The mobilization-minimization hypothesis (Taylor 1991) can serve as another possible explanation for those reversed effects: while negative events usually evoke strong immediate cognitive and emotional responses, they are also—unlike positive events—followed by other responses the organism uses to minimize the effects of those events. Thus, the immediate effects of negative events might be strong, while those effects often fade away quickly and can result in less pronounced responses compared to positive emotions. Our findings are in line with the latter perspective, although it is of course relevant to add that there is considerable variation among the different positive emotions.

Further, the strongest overall effects of emotions stem from the medium-arousal emotion gratitude. More generally, we do not find that the effects of medium-arousal emotions significantly differ from high-arousal emotions. This seemingly contradicts research on advertising which has demonstrated that high arousal levels polarize the effect of emotions (e.g., Gorn et al. 2001) and that high-arousal emotions have a stronger impact on word of mouth (e.g., Berger 2011; Berger and Milkman 2012). In light of this finding, it is also noteworthy that medium-level arousal emotions like gratitude seem to attract much less attention in research and practice than many high-arousal emotions (see Roberts 2005 for a discussion of brand love). Generally, the strong effect sizes of medium-arousal emotions might be explained by the tendency of humans to negatively value high levels of arousal (Skowronski and Carlston 1989). As a consequence, humans engage in counteracting processes to minimize the physiological reactions of high arousal which also makes emotions fade away rather quickly and have smaller effects on outcome variables. Marketing research and practitioners would thus do well to more strongly emphasize the importance of medium-arousal emotions.

We further find that only emotions that result from a situation that is perceived to be under the control of an entity (not under circumstantial control) show significant average effect sizes in our dataset. This suggests that attributions of responsibility do not only drive the elicitation of emotions but also their effects on firm-relevant outcomes. Incidental emotions seem to only carry over to the unrelated firm if they are appraised as being under entity control. We also find a similar pattern for emotions that are appraised with high certainty.

We provide an overview of the differing overall effect sizes of 10 discrete emotions (see Table 4) and show that some emotions consistently affect relevant outcome variables, while others do not. We particularly find significant overall effect sizes for a variety of positive emotions. Based on Evans’ (1996) categorization of correlation coefficients, we find moderate positive average effect sizes for gratitude, love, and happiness, and significant, yet very weak effect sizes for pride. Among the negative emotions, we only find a significant yet weak effect size for anger. All other negative emotions show no significant overall effect sizes. In an absolute sense, the positive emotions gratitude, love, and happiness are the three emotions with the largest impact on firm-relevant outcomes in our meta-analysis.

Our results also provide insights in the prevalence of affect congruent and incongruent effects. While most scholars support the view that positive (negative) emotions lead to positive (negative) outcomes (e.g., Greifeneder et al. 2011; Schwarz and Clore 1983), others have found evidence for affect incongruent effects (Goldsmith et al. 2012). Distinguishing only between positive and negative affect, Puccinelli and colleagues (Puccinelli et al. 2016) find in their meta-analysis general support for affect congruent effects while they identify single affect incongruent effects. Our findings shed more light on this by distinguishing discrete emotions. Apart from single exceptions, we find mostly affect congruent (or no) effect sizes for discrete emotions. We only identify affect incongruent effect sizes for guilt on the outcome variable of sharing behavior and for interactions with an employee. Thus, based on our meta-analytical results for 10 discrete emotions, affect incongruent effects seem to be fairly uncommon.

Further, we find evidence that even a consideration of discrete emotions instead of valence and arousal leaves substantial heterogeneity among the observed effect sizes. By investigating potential sources of this heterogeneity, we find (a) that many emotions vary in their effects on different firm-relevant outcomes (evaluation, purchase behavior, and sharing behavior), and (b) that this dynamic is for some (but not all) emotions moderated by situational characteristics of the emotional experience, like the question whether the firm–customer encounter involves an employee of the firm, happens after a service recovery attempt, or whether the encounter happens in a B2B or B2C setting.

For managers, our results provide valuable insights into which discrete emotions are especially important in driving consumer responses. From a customer experience management perspective, it is important to consider a holistic view that explicitly includes emotional responses of consumers (Kranzbühler et al. 2018). Recent research also stresses the importance of such emotional responses in relation to factors that drive consumer engagement (e.g., Hollebeek et al. 2019; Kumar et al. 2019), to the point that they are considered more impactful than cognitive evaluations (Hollebeek et al. 2019). As Hollebeek et al. (2019) point out, emotions do not just drive purchase, but they are inherently communicative. We find substantial effects of emotions on sharing behavior, which is considered an important articulation of customer engagement (Kumar et al. 2019). In order for managers to utilize these insights it is important to adjust their marketing dashboards to include more fine-grained measures of emotions. Although it is not uncommon for firms to use emotion-related measures (e.g., smileys to express satisfaction), these measures are typically simplistic in nature and valence oriented.

Our results suggest that it is valuable to map discrete emotions across the customer journey. The customer journey describes different touchpoints that consumers encounter throughout the entire customer experience. Current customer journey mapping practices typically tend to focus on capturing these touchpoints and the extent to which they are considered satisfactory or not or elicit a particular NPS score. In line with research on customer mindset metrics that stresses the importance of a wider set of metrics including affect (Srinivasan et al. 2010), we argue that customer journey mapping should be extended to capture an emotional “heat map.” This will allow firms to gain a better understanding of how to manage the customer experience and where there are opportunities to promote specific discrete emotions or when to counter them. In formulating recommendations to manage emotions across the customer journey, we draw inspiration from Yih et al. (2019) by considering both emotion generation as well as emotion regulation as a way to influence consumer perceptions and actions.

Our first important insight in this regard is that we find consistent evidence that managers should prioritize emotion generation by triggering positive emotions over avoiding negative ones. For all outcome variables and situations, we find the strongest effect sizes stemming from the positive emotions of gratitude, love, and happiness. Emotion regulation is especially relevant with regard to negative emotions. Avoiding anger and sadness is also important for firm-relevant outcomes but as the detrimental effects of those emotions are smaller in magnitude, resources—if limited—should first and foremost be allocated to facilitating an experience that elicits gratitude, love, and happiness.

Table 8 provides hands-on guidelines for firms to manage the emotional customer journey. It presents not only the actions that can be taken to promote or counter certain emotions, but also presents cues that enable managers and employees to recognize these emotions as they occur.

This is relevant input for the emotional journey mapping as well as for training of frontline personnel. Like consumers, employees may follow a “reasoning-forward” logic as proposed by Ong et al. (2019), and apply an appraisal-like process to inform themselves about consumers’ emotions. While this is typically considered an intuitive process, employees can be trained to recognize specific cues that foretell which emotions are being generated and consequently should be regulated given the outcome that emotion is likely to cause. Following this, we provide a set of concrete actions that can be used to manage these discrete emotions, in which the focus is first and foremost on promoting gratitude, love, and happiness, and second on countering negative ones.

It is important to nuance this discussion in relation to specific outcomes that firms may strive for. These outcomes also align well with different phases in the customer journey and provide further nuance to the emotional journey mapping. Evaluation is often considered part of the earlier phases of the journey, followed by the purchase phase that is related to the actual (transactional) behavior, and a post-purchase phase that tends to include consumer communications about the purchase to others (e.g., sharing; Voorhees et al. 2017). Our meta-analysis shows how these phases differ with regard to the impact of emotions. If the focal goal is to create positive consumer evaluations, purchases, and sharing behavior promoting the above-mentioned positive emotions seems most valuable. Negative emotions, specifically anger, should be regulated as they do impact the different outcomes, but less so than positive emotions. Sadness is particularly relevant in relation to sharing behavior.

The importance of these findings may vary with context. Of the emotions highlighted in Table 8, anger has a stronger effect in B2B than other situations. While aggravating customers to the point of anger is never a good strategy, it is especially important to avoid this emotion in B2B settings as its effect across all outcome variables is more pronounced. Given that B2B often characterizes itself by more personal, long-term relationships, firms also have an opportunity here to focus their training of account managers specifically on signaling early signs of this emotion and how to deal with this in their specific role.

Beyond the emotions with the strongest overall effect, it is worth noting that also pride and guilt should be managed differently throughout the customer journey. Especially in those touch points where there are customer-facing employees involved, guilt is an emotion that can have positive effects on the overall outcomes. While we should be wary of suggesting managers to stimulate feelings of guilt, it is also important to note that what is typically considered a negative emotion can in fact have positive effects.

Finally, pride has a stronger effect in recovery situations than otherwise. Inducing pride in a recovery situation might be challenging for firms, as a recovery implies that something went wrong in a preceding phase of the customer journey. However, firms could consider co-created solutions as part of their recovery strategy, in which they give consumers an active role in developing the solution and thus take ownership and pride in their role. Such a strategy would call for a more principle-based (rather than a rule-based) way of working, where employees have freedom to develop such solutions together with customers.

The present meta-analysis provides an encompassing overview of the scattered literature on the firm-relevant effects of discrete emotions. However, it comes with limitations that can serve as a basis for future research. First, up to now research has very much focused on high-arousal emotions such as happiness and anger. However, our results indicate that substantial effects on firm-relevant outcomes also stem from emotions with lower arousal levels such as gratitude or sadness. Thus, future research should investigate these emotions further and shed more light on their effects and how they come about (and thus also how firms can more strategically address these in the design of service encounters). Also, our investigation was restricted to the core consumption emotions featured in Ruth et al. (2002). It could thus be that other discrete emotions that were not included in this meta-analysis would provide stronger effect sizes. At this point, we can only report the dynamic with regard to the ten emotions featured in this analysis, but we do believe that further research on this interesting area would be valuable. Another limitation related to the availability of studies is that some analyses were based on rather small numbers of effect sizes (e.g., only 41 of 1035 effect sizes involved a service recovery attempt). Thus, non-significant results should be interpreted with caution as they might also stem from a lack of statistical power.

Second, future research should further investigate the identified affect-incongruent effect of guilt as it is only based on a low number of effect sizes. The positive effect of guilt might be explainable by its special nature as a self-focused other-directed emotion that can result in restoring behavior toward the other party (Burnett and Lunsford 1994).

Finally, future research should investigate the effects of discrete emotions over time. Based on the mobilization-minimization hypothesis (Taylor 1991), it would be interesting to see whether most or select negative emotions indeed affect short-term outcomes but that their impact quickly disappears over time. The majority of existing studies asks respondents to recall experiences or read vignettes. Those methodologies might misrepresent the levels of especially negative emotions directly after experiencing a firm–customer encounter. With the rise in popularity of mobile phones, such strong immediate emotions can easily be shared with a lot of people via social media.