1 Introduction
Recent advances in internet communication tools and online social networks makes "social learning" about products by consumers easier. Previous marketing studies define social learning as the learning process that is promoted by the exchange of information among diverse consumers in (online) communities
[1, 2]. A closely related concept to social learning is "consumer engagement" among advertising practitioners. Although there is no established definition, the concept of consumer engagement is described as the consumers' prospect of a brand idea, which is enhanced and stimulated by online interactions with other consumers, and not only through the offline one-way marketing communications. In practice, consumer engagement is measured, for example, by the duration, frequency, and/or recency of visiting, viewing high- or medium-value content, providing personal information, and posting customer reviews and comments in online communities. Firms have been conducting engagement marketing to enhance social learning, which leads to long-term customer loyalty, thereby maximizing purchase probabilities.
This study explores individual level correlations between traditional marketing activity, that is, TV advertising and public relations (PRs), and online consumer's social learning for videogame consoles, Wii and PS3, launched in 2006. An anecdotal story of a pre-release advertising campaign for the introduction of the Wii videogame console in the Japanese market provides an example of consumer engagement marketing and stimulating social learning in the videogame industry. Nintendo initially showed only the silhouette of the console at an industry preview event, and then provided additional information over time until the release of the Wii in December 2006. By revealing limited information at the early stages, Nintendo prompted consumers to search for product information and discuss their speculations regarding the console in online communities. Throughout the learning process, consumers became highly engaged so that Nintendo's strategy could result in higher sales immediately after Wii console released. We note that despite recent growth of smartphone games, TV console games such as Wii and Play Station still occupy more than quarter of the total videogame market revenue
[3].
According to the consumers' learning literature in marketing, it is believed that consumers learn about products through two processes (as outlined in Figure 1). The first process, traditional personal learning, is also called cognitive learning and reason-based learning. In this process, consumers individually access all possible information about products and cognitively evaluate this information. The second source of learning about product quality is social learning, in which information is derived from other consumers. In the process of social learning, consumers interact with other consumers to acquire product information. Social learning is expected to stimulate consumer engagement, leading to higher purchase probabilities.
Figure 1 Outline of the research and the proposed model |
Full size|PPT slide
We assume that pageviews of community-based websites act as an indicator of the level of consumer engagement by social learning. Personal learning is assumed to be indicated by the browsing behaviors of all videogame-related websites, including both community-based and non-community-based websites. Furthermore, we anticipate that the two processes are correlated with each other.
In summary, our working research questions are (a) do traditional advertising and PR campaigns stimulate learning by consumers about products and (b) what is the relative importance of the two types of learning processes-personal learning and social learning-on consumer purchasing decisions? By quantifying the impact of traditional marketing media, firms can manage to enhance consumers' learning and promote greater engagement with products, potentially leading to increased product purchases.
The rest of the paper is organized as follows. Section 2 provides a review of the literature and a discussion of our position in the literature. In Section 3, we describe the data. The model is discussed in Section 4. Section 5 summarizes our results and counterfactual simulations. Section 6 is the conclusion.
2 Literature Review
There has been a long history of studying social learning in many fields such as social, biological, and computational sciences. Starting from Galton's "wisdom of crowds"
[4], theories of collective intelligence have shown that the aggregated estimate by a group of many individuals can be more accurate than the estimates of individual experts, when people make their judgements and decisions. This wisdom of crowd effect was supported by various examples of networks
[5], financial markets
[6, 7], political issues
[8], entertainments
[9] and human cultural evolutions
[10, 11]. However, some experimental studies have suggested that social influence undermines the wisdom of crowds, because individuals' estimates became more similar and less accurate when subjects observed each other's opinions and actions then adapt their judgments
[12]. In contrast, other analytical and experimental studies of social learning have found that the accuracy of group estimates may improve, in such cases that the structure of the interaction network varies the effects of social influence on collective intelligence
[13, 14]. Becker, et al.
[15] studied experimental results showing that, especially in decentralized networks group estimates become more accurate than in centralized networks, due to intensive information exchanges. In addition, not only accuracies of estimating about the matters of facts, but also those about the matters of taste found to be improved by collective social learning compared with individual learning
[16, 17]. Following the results of experimental studies
[9, 18], our study applies the assumption that collective social learning is biased from individual learning by exchanging information among social interactions, then our proposed empirical model enables to quantify the magnitude of the bias as initial shift and proportional effects.
Other stream of studies on social learning has investigated social learning strategies. Some research worked on a question about from whom people learn and others dealt with a question about how they can form their collective judgments. The former type of studies focused on the effect of perceived similarity among individuals, that is, people are swayed more by similar individuals. The large-scale simulation study by Analytis, et al.
[16] found the best social learning strategy that experienced individuals should rely mostly on seemingly similar people, in contrast, inexperienced individuals have to follow the mainstream option despite differences in their similarity. Experimental results by Gershman, et al.
[19] suggested that people infer groups of similar individuals in their network, and they use these groups to decide whose choices to follow. On the other hand, the latter type of studies has suggested four classes of social learning strategies; frequency-dependent rules, payoff-based rules, confidence-based rules and prestige-based rules. They found that people use different social learning strategies and tend to keep using the same learning strategy irrespective of the interaction context
[20, 21]. Rendell, et al.
[22] found that payoff-based social learning is remarkably successful, because the highest-payoff options are so salient as to enable individuals to filter irrelevant information and copy the best decisions easily. Vostroknutov, et al.
[23] experimentally studied the relationship between social learning and cognitive abilities, and found that individuals with low cognitive abilities tend to use social learning, compared to the individuals with high cognitive abilities. Some studies assumed that individuals are Bayesian, who observe past actions and sequentially update their expectations according to incoming information
[13, 24]. In the same manner as these studies, this study proposes an extended Bayesian learning model in which consumers learn form opinions of the similar peers according to the frequency-dependent rule.
In the marketing literature, many studies have explored how consumers learn about the quality of products from available information in the market, and these studies have suggested that consumer's learning occurs through dual or multiple processes. Petty and Cacioppo
[25] conceptualized the dual learning processes of systematic and heuristic routes. Other papers have introduced emotional or experiential responses in addition to cognitive processing (e.g., [
26,
27]).
On top of these well-known learning processes, social learning has also played an important role in consumers' purchasing decisions. Social learning is promoted by the exchange of information among diverse consumers in problem-solving communities. Jayanti and Singh
[1] defined and examined social learning process, which is generated by interactive cycles among communities and is motivated by actions aimed at solving problems, for healthcare via online BBSs (bulletin board systems). Calder, et al.
[2] discussed the effectiveness of advertising on consumer engagement by using experiments with eight different online experiences on websites; they found that two types of engagement with online media, Personal Engagement and Social-Interactive Engagement are both positively associated with advertising effectiveness. Moreover, Social-Interactive Engagement was strongly correlated with advertising after controlling for Personal Engagement. Therefore, this study follows these previously established definitions of social learning and personal learning processes, as represented by web-browsing behaviors to gather information about products.
Erdem, et al.
[28] closely shared our research interests. They investigated product learning processes involving word-of-mouth communication. However, their study considered only a single learning process and did not differentiate between social learning and personal learning. In addition, they used aggregated product-level datasets and assumed a structure to replicate unobserved individual consumer's learning process. In contrast, this study uses disaggregated data to directly infer individual learning processes.
Quantitative models of new product adoption and diffusion have studied the underlying social interactions among consumers in the diverse disciplines of economics, marketing, and sociology. Young
[29] reviewed social interactions in the extant theory of diffusion models and distinguished social learning from the similar concepts of social contagion and social influence, defining social learning as follows: "People adopt once they see enough empirical evidence to convince them that the innovation is worth adopting, where the evidence is generated by the outcomes among prior adopters.'' He also noted that the key difference among these three concepts was the assumption of utility maximization in social learning rather than the notion of exposure in the other two. For example, in cases of social contagion and social influence, a consumer may adopt a new product merely because he or she observes other consumers who have already adopted it or notices that the large enough number of peers adopted the product in his or her close community. In contrast, in the social learning case, consumers make their decisions on adopting a product by considering costs and benefits according to their utility through their prior beliefs, amount of information gathered, and idiosyncratic factors of social interactions with prior adopters.
Table 1 shows a list of selected studies on social interactions. Our study occupies a unique position in terms of using individual-level data on browsing behaviors across multiple websites to identify consumers' social learning processes during the pre-release period before the launch of products. Several recent studies have shed light on the role of social interactions in understanding the transmission of information across social networks (e.g., [
30-
34]). Similar to our study, Park, et al.
[30] investigated the effect of social contagion on purchasing in-game items on a specific online videogame platform and found that social interactions, that is, playing online games together with someone who spent a lot on in-game items-encourage the other players to purchases those items. However, they only considered social contagion in terms of the exposure to the players' spending rather than persuasive information gathered from social interactions. On the other hand, the proposed model accommodates consumers' dynamic beliefs about the quality of products so as to investigate the effects of social learning. Except for our research, all the studies listed in
Table 1 were limited to examining social interactions after the product has launched. As explained above, in practice, it is common for companies to execute engagement marketing even before releasing their products into markets in order to increase buzz and stimulate potential consumers to search eagerly for information about the products. Finally, the previous studies used data from a single community. In contrast, our research investigated browsing behaviors across a large number of multiple websites, because consumers want to learn about product quality not only from the single source but from various websites, and they then compare the informative contents.
Table 1 Selected studies and the position of this study |
| | Interaction periods | Social interactions | Communities | Products | Analysis unit level |
| This study | Pre-release Dynamic | Social learning | Multiple websites | Videogame consoles | Disaggregate |
| Park, et al.[30] | Post-release Dynamic | Social contagion | Online game platform | In-game items | Disaggregate |
| Hu and Van den Bulte[31] | Post-release Static | Social contagion | Academic collaboration | Experiment kits | Disaggregate |
| Bollinger and Gillingham[32] | Post-release Static | Social influence | Geographic proximity | Solar panels | Aggregate |
| Iyengar, et al.[33] | Post-release Static | Social contagion | Physicians network | Prescription drugs | Disaggregate |
| Van den Bulte and Joshi[34] | Post-release Dynamic | Social influence | 33 products | Aggregate | |
| Erdem, et al.[28] | Post-release Dynamic | NO (Single learning) | NO | PCs | Aggregate |
3 Data
In this study, we used user-centric internet clickstream data collected by Video Research Interactive, Inc., which has maintained a panel of approximately 12, 000 Japanese individual subjects whose website browsing behaviors were recorded over time by a firm's proprietary software installed on their computers at home. The collected data include URLs for websites that were accessed by the subjects and times of visits1.
1Details and the descriptive statistics for the datasets will be provided upon request to the author.
3.1 Videogame Console Ownership
The company also conducted annual written surveys for a randomly selected subset of its existing panelists. In total, 7, 053 subjects responded to the annual survey in November 2007 (around one year after the release of Wii and PS3 in December 2006).
As shown in Table 2, 24% of the subjects owned one of the available videogame consoles at the date of the survey in 2007. Wii was owned by 25.5% of all videogame users. In contrast, the share of PS3 was 4.5%, and 1.6% of the videogame users owned both consoles.
Table 2 Data description: Videogame console possession |
| | # Users | Percentage | Share |
| Wii owners | 434 | 6.2% | 25.5% |
| PS3 owners | 77 | 1.1% | 4.5% |
| Both (Wii & PS3) owners | 28 | 0.4% | 1.6% |
| Other console owners | 1, 160 | 16.4% | 68.3% |
| All videogame owners | 1, 699 | 24.1% | 100.0% |
3.2 Pre-Purchase Browsing Behaviors of Videogame Websites and Classification of Community-Based Websites
This study centered on the pre-purchase website browsing behaviors of subjects who bought one or both of the videogame consoles Wii and PS3. For this purpose, we translated their website browsing records into the daily number of pageviews of videogame-related websites from the date that the product information was first announced (April 28, 2006) to the date of the product launch (December 2, 2006). We selected 49 major videogame-related Japanese websites based on the large volume of their pageviews.
Three individual raters classified the videogame-related websites as community-based when the website had community features, such as a BBS or systems for posting user reviews. The final classification was determined by the majority votes of the three raters' responses for each videogame-related website. Consequently, we chose nine websites that were considered community-based websites. Note that in order to validate the classifications, we examined the inter-rater agreement and found "substantial agreement" among raters using Fleiss's kappa
[35].
Next, we counted pageviews of the community-based websites and used them as an indicator of social learning process. For the empirical analysis, we selected and used the observations by the panelists who owned any of the available videogame consoles and visited the videogame-related websites more than twice during the analysis period from April to December 2006. These criteria resulted in 1, 078 panelists remaining for the analysis.
3.3 TV GRPs and PRs (Public Relations)
The company has also reported the gross rating points (GRPs) for all TV commercials aired in the Japanese market. We used the aggregated TV GRPs for videogame ads, segmented according to gender and age groups (i.e., teen/20s/30s/40s/50s/60s), and matched the segmented GRPs with the pageview data for subjects in the same demographic segment according to the gender and the age groups. In addition, we classified the types of the videogame TV ads into console ads and software ads by Nintendo, Inc. and SCE (Sony Computer Entertainment, Inc.).
4 Model
We propose a unique model consisting of two components — A bivariate learning processes for product quality over time before the product launch and individual purchase decisions post-launch.
4.1 Learning Process of Product Quality
We assume that consumers have two different modes of learning processes. As a first basic case, we describe a mere single process model that only considers the personal learning process. Note that our unique formulation of signal information enables a simpler estimation procedure. Then, in a second dual learning process, the model is extended from the conventional bivariate Bayesian learning model
[36]. Our proposed model considers the possible correlation between the two learning processes.
4.1.1 Basic Case: A Single Learning Process
Personal learning is expected to capture cognitive aspects of products and contribute to cognitive beliefs about the quality of products over time. Thus, denotes consumer 's personal (cognitive) belief about product at date . denotes the true quality of product for consumer . Following the standard context of the Bayesian learning process, consumers have uncertainty about the true quality of the product but have a belief about its value.
At the date of , consumer 's initial prior belief about the quality of product is assumed to be normally distributed. We also assume that the mean initial belief is also normally distributed across consumers.
We expected the videogame-related website browsing behaviors of consumers to represent their level of engagement or the strength of their interest in products, driven by beliefs about the product quality. Therefore, we treat the number of pageviews as an indicator of a consumer's personal belief. The indicator signal is assumed to follow a normal distribution.
However, pageviews can be biased indicators of the signal of product quality, since they are influenced by other factors that may influence website browsing, such as seasonal factors, like weekends and holidays. In addition, we are interested in the impact of firms' marketing activities, such as TV advertising and PRs. Thus, we formulated the log number of pageviews as the additive combination of the signal of the product belief, seasonal factors, the company's marketing actions, and white noise, as described by Equation (3).
This equation can be rewritten by substituting Equation (2) as follows.
This leads to a consistent estimate of a consumer's product belief, which can be expressed by Equation (5), given the information available by date . , and denote consistent estimates of the covariates via linear regression and , , and are means of the log number of pageviews, seasonal factors, -type of TV ads (console and software), and PRs up to date .
Finally, by combining the prior belief and the information from the signals, the posterior belief about product quality is normally distributed with an updating formulation for dates .
The advantage of this formulation is that it does not require simulations in its estimation procedure. In standard learning models, it is necessary to infer the mean value of the signal information by simulating to recover its distribution (e.g., [
28]). In our model, consistent estimates of the signals are simply obtained from the linear regression of the pageviews, as in Equation (5).
4.1.2 Extended Case: Bivariate Learning Processes
As discussed, we assume that there is another mode of learning in addition to personal learning. Social learning occurs through interactions with other consumers via community websites and helps consumers construct social beliefs about product quality. Social belief is assumed to be correlated with the individual's personal belief. Ackerberg
[36] proposed bivariate learning processes. However, the two beliefs were considered to be correlated only at the initial belief in his model. In contrast, our model accounts for the correlation both at the initial time point and throughout the entire updating processes, as below.
Equation (7) assumes that consumers' true perceptions of product quality through social learning during the entire updating process are systematically correlated with the value for personal learning, shifted by an amount and proportional to the ratio of .
According to the formulation of social belief shown in Equation (7), we can rewrite Equation (1), consumer 's initial personal and social beliefs about the quality of product on the date and the distribution of different signals.
Similar to the case of the single learning process, we assume that the pageviews of community-based websites are related to the signal from the social belief and other confounding factors. Then, we can describe consistent estimates of personal and social beliefs as follows.
Finally, similar to Equation (6) for the single learning case, the posterior beliefs about product quality follow a bivariate Bayesian learning process for based on initial beliefs and signal information updates.
where
4.2 Purchase Choices Based on Cumulative Product Beliefs
When there are two new products available on the market, consumers choose one of the following options
: where
denotes buying neither product and
means buying both products. By using the complementary bundle choice model of Gentzkow
[37], the expected mean utility functions are as follows.
where the mean of overall quality beliefs, , is assumed to be a convex combination of the mean personal beliefs and the mean social beliefs on the date of the product launch, , that is, , where . and denote cumulative summations of all TV ads and all PRs from the dates to . The parameter shows the complementarity (if it is positive) between two products. The error terms, , assume that consumers' persistent taste follows a bivariate normal distribution.
More specifically, denoting as the element of the choice options , the indirect expected utility can be defined over the choice options as follows.
Assuming that the error term, , follows the type-Ⅰ extreme distribution, then the probability that the consumers choose to purchase the products given the consumers' persistent taste vector can be written as a multinomial logit formulation.
where is the set of parameters consisting of and is the set of parameters determined through the dual learning processes .
4.3 Identification and Estimation
The main identification issue is based on differentiating the latent social belief from the personal belief. The first source of identification comes from the assumption described in Equation (7). The size parameter, , accounts for a shifting bias to the social belief from the personal belief. The proportional parameter, , is the correlation between the two beliefs. In addition, data may be able to support the identification of these two parameters. When the difference between two beliefs varies over time, those changes are considered by the proportional parameter. However, if a one-time shock causes a stationary difference across the entire period, it is accommodated by the size parameter.
In the estimation, we need to normalize the first element of the variance of the initial beliefs as to for identification purpose in Equation (12). Finally, we can integrate out error terms, , from Equation (16) and derive the conditional likelihood function, as in Equation (17). To replicate the error distribution, we apply the simulated maximum likelihood method to our empirical parameter estimation.
5 Results and Discussion
5.1 Results of the Bivariate Learning Processes
Table 3 presents the parameter estimates for the bivariate learning processes. The variance of posterior beliefs in personal learning () and social learning () are estimated to be larger than the variance of initial belief (). These results suggest that the updates in product beliefs by social learning are more precise than the updates by personal learning because the estimated variance is smaller. In other words, the information signal from social learning is more informative for updating consumers' quality perceptions.
Table 3 Estimates of the bivariate learning processes |
| | Estimate | Std.Err. | |
| Initial social belief () | | | |
| Wii | 0.937 | 0.550 | * |
| PS3 | 0.865 | 0.418 | * |
| Correlation of beliefs () | 2.131 | 0.034 | ** |
| Variance of initial belief () | 0.443 | 0.089 | ** |
| Variance of posterior beliefs | | | |
| Personal learning () | 10.708 | 0.883 | ** |
| Social learning () | 8.200 | 0.713 | ** |
| Significance code: 0.01 '**', 0.05 '*', 0.1 '.' |
As shown in Table 3, we also find that initial social beliefs () are small compared to the zero means of initial personal beliefs (). According to Equation (7), this means that the biases of social beliefs from personal beliefs are significant but small at the initial learning periods. However, the proportional relationship of beliefs () is reasonably large, suggesting that as the number of updates increases, the biases toward social beliefs become larger in proportion to the correlation between beliefs.
5.2 Results of the Pageview Equations
The estimated results from the pageview equations, that is, Equation (10), are illustrated in Tables 4 and 5. Overall, we find that parameter estimates for the constant, daily trend, and all pageviews are positive and significant, in all cases. In addition, the results suggest that the coefficients for two types of TV ads and PRs are positive and significant in the social pageview equations. In contrast, only TV GRPs of Wii software ads are marginally significant in the personal pageview equations.
Table 4 Summary of personal pageview equation: |
| | Wii | | PS3 |
| Estimate | Std.Err. | | Estimate | Std.Err. | |
| Constant | 2.180 | 0.271 | ** | | 2.130 | 0.336 | ** |
| Trend | 0.031 | 0.002 | ** | | 0.034 | 0.002 | ** |
| Holiday | 0.442 | 0.473 | | | 0.139 | 0.598 | |
| All pageviews | 0.003 | 0.000 | ** | | 0.001 | 0.000 | |
| PRs | 0.164 | 0.334 | | | 0.043 | 0.390 | |
| TV GRPs Console | 0.004 | 0.004 | | | 0.003 | 0.007 | |
| TV GRPs Software | 0.003 | 0.002 | . | | 0.004 | 0.005 | |
| Significance code: 0.01 '**', 0.05 '*', 0.1 '.' |
Table 5 Summary of the social (community-based) pageview equation: |
| | Wii | | PS3 |
| Estimate | Std.Err. | | Estimate | Std.Err. | |
| Constant | 5.461 | 0.235 | ** | | 6.935 | 0.269 | ** |
| Trend | 0.058 | 0.001 | ** | | 0.049 | 0.001 | ** |
| Holiday | 1.094 | 0.410 | ** | | 1.259 | 0.479 | ** |
| All pageviews | 0.003 | 0.000 | ** | | 0.004 | 0.000 | ** |
| PRs | 0.428 | 0.290 | . | | 0.665 | 0.312 | * |
| TV GRPs Console | 0.018 | 0.003 | ** | | 0.011 | 0.006 | * |
| TV GRPs Software | 0.015 | 0.002 | ** | | 0.022 | 0.004 | ** |
| Significance code: 0.01 '**', 0.05 '*', 0.1 '.' |
The estimated constants for the social pageview equations are larger than those for the personal pageview equations. They are two times larger for Wii and three times larger for PS3. This is consistent with the above results of the bivariate learning processes, since, according to Equation (11), the estimated constants should be the consistent estimators of personal and social beliefs. The updates of both learning processes should approach the values of the estimated constants.
5.3 Results of the Purchase Choice Analysis
5.3.1 Parameter Estimates and Elasticity
From the purchase choice model results shown in Table 6, we find positive and significant parameter estimates for consumers' overall beliefs about product quality. Furthermore, the estimated coefficient of quadratic overall beliefs is positive and significant, suggesting that the relationship between consumers' overall beliefs about product quality and choice probability is U-shaped. Consequently, consumers tend to be risk-taking in terms of product quality. In particular, consumers are likely to buy videogame consoles in cases where their social belief and personal belief are at either high or low levels, but not at moderate levels.
Table 6 Summary of factors contributing to purchase choices |
| | Elasticity | Estimate | Std.Err. | |
| Weight of personal belief () | - | 0.410 | 0.045 | ** |
| Complementarity () | - | 0.001 | 7.567 | |
| Variance of persistent taste () | - | 1.910 | 0.010 | ** |
| Covariance of persistent taste () | - | 0.000 | 0.070 | |
| Overall belief | 0.051% | 0.064 | 0.000 | ** |
| Quadratic of overall belief | 3.659% | 0.465 | 0.000 | ** |
| Price (＄) | 910.768% | 0.443 | 0.000 | ** |
| Cumulative PRs | 2.355% | 0.284 | 0.000 | ** |
| Cumulative TV GRPs Console | 243.384% | 0.340 | 0.000 | ** |
| Cumulative TV GRPs Software | 754.314% | 0.796 | 0.000 | ** |
| Significance code: 0.01 '**', 0.05 '*', 0.1 '.' |
The weight of personal belief () is about , indicating that the composition ratio of the overall belief is to for personal belief and social belief. In other words, combining the results of the dual learning process, which shows that the average of the social belief is at least two times larger than that of personal belief, we can conclude that consumers weigh social beliefs to times more heavily than personal beliefs in their purchase choice.
Table 6 also shows that the price coefficient is negative and significant. The average price elasticity was calculated based on the parameter estimates, defined as the mean percentage increase in the choice probability relative to a one percent change in console price. The results suggest that own price elasticity has the largest effect size among all variables. In addition, the cumulative values of PRs and cumulative TV GRPs for console and software ads are positively and significantly correlated with purchase decisions. Overall, cumulative software ads have the largest elasticity.
5.3.2 Complementarity and Substitution
One advantage of the proposed model is that it allows us to investigate the complimentary pattern between purchase decisions for the Wii and PS3. Table 6 shows that the complementarity parameter () is not significant and is close to zero, indicating that there is no incremental utility in buying both Wii and PS3 consoles compared with purchasing a single console. This result is consistent with the common perception that the characteristics of the two products are distinct and the two products target different segments (our data also suggested that only of videogame users owned both consoles).
In addition, we examined substitution patterns between changes in the independent variables for the two products. Table 7 summarizes the cross-elasticity indicators of purchase probabilities between two products. The results suggest that the substitution patterns are asymmetric and the cross-elasticity of PS3 variables leads to a larger increase in the Wii purchase probability, in general. Exceptionally, a change in the price of the Wii has a greater effect on the PS3 purchase probability. We believe that this can be explained by the lower price of the Wii than the PS3; a decrease in the price of the Wii would therefore hurt demand for PS3 more than the opposite scenario would hurt demand for the Wii.
| | Wii to PS3 | PS3 to Wii |
| Overall belief | -0.026% | -0.085% |
| Quadratic of overall belief | -1.717% | -6.329% |
| Price (＄) | 1129.102% | 594.685% |
| Cumulative PRs | -2.364% | -2.283% |
| Cumulative TV GRPs Console | -98.935% | -434.092% |
| Cumulative TV GRPs Software | -544.172% | -1024.622% |
5.4 Model Comparisons
Two benchmark models were tested to assess the performance of the proposed model. The first benchmark model was a complementary bundle choice model
[37] without a product quality learning process. Instead of including consumers' social and personal beliefs about product quality, this model regarded exponential sums of the social pageviews and the entire personal (videogame-related) pageviews as proxies of the two beliefs. The second baseline model allowed only a single learning process instead of the bivariate learning processes in the proposed model.
Model fit measures are reported in Table 8. The proposed model outperformed the two benchmark models according to three measures: AIC (in-sample), AIC (holdout-sample), and a fitting ratio (true positive). Note that the baseline model with no learning fits better than the single learning model in terms of in-sample AIC and fitting ratio.
Table 8 Model comparisons |
| | Model 1: No learning | Model 2: Single learning | Proposed model |
| AIC (in-sample) | 3442000 | 3632578 | 2884920 |
| AIC (holdout-sample) | 1781125 | 1730389 | 1700486 |
| Fitting ratio (true positive) | 0.496 | 0.358 | 0.768 |
| We randomly set aside 100 subjects to be used as a holdout sample. |
5.5 Counterfactual Simulations: Advertising Schedule Plans
Planning advertising schedules is an important part of real marketing practices. Our results showed that advertising stimulated consumer engagement through social learning via community-based website browsing. Therefore, changing advertising scheduling is likely to increase the long-term effects on consumers' purchase decisions.
To assess such effects, two different simulated advertising plans for Wii were examined. We assumed that the total amount of TV GRPs, which was equivalent to the total advertising spending, remained equal to that of the original plan during the analysis period. For the first simulated plan, we distributed the total TV GRPs evenly across the period. The second advertising plan decreasingly allocated the TV GRPs. Then, we calculated the estimated Wii purchase probabilities in our sample using the above estimated parameters and the actual datasets for the different advertising schedule policies.
Figures 2 and 3 report the simulated updated paths of average quality beliefs for the Wii for the flat advertising plan, the decreasing advertising plan, and the actual plan. While the simulated updates of the average personal beliefs about the three plans converged to almost the same value at the end, the terminal values of the average social beliefs differed among the three advertising schedules. The decreasing and flat advertising plans exhibited more rapid growth in both quality beliefs but also exhibited more rapid decay than that for the actual plan. The actual plan allocated increasing TV GRPs across the period and the increasing amount of advertising prevented quality beliefs from decaying during the final periods. However, the decreasing advertising plan resulted in higher values for social learning at the beginning of the updates. Even though those values decayed later, they remained substantially larger at the terminal points than those of the actual adverting plan. Consequently, the purchase probability of Wii for the decreasing advertising plan is estimated to be , which is greater than the estimated buying probability for the actual plan.
Figure 2 Simulated updates of mean personal beliefs for Wii |
Full size|PPT slide
Figure 3 Simulated updates of mean social beliefs for Wii |
Full size|PPT slide
6 Conclusion
Many marketers have struggled to manage consumer engagement with their products and have missed opportunities to enhance social learning. We evaluated the following empirical questions. (a) Do traditional advertisings and PRs campaigns stimulate consumers' social and personal learning about products? (b) What is the relative importance of the two types of learning on consumer purchase choices?
Empirical analyses yielded substantive insights for marketing managers into the implementation of engagement marketing and enhancing consumers' social learning. (ⅰ) When consumers consider their purchase of new videogame consoles, they weigh social beliefs about product quality at least three times more heavily than personal beliefs. (ⅱ) Consumers are found to be risk-taking in terms of product quality; they are likely to buy videogame consoles when their social belief and personal belief are at either high or low levels, but not at moderate levels. (ⅲ) While consumers update their quality beliefs, the information signal from social learning is perceived to be more informative and becomes greater over time than the personal learning signals. (ⅳ) Two types of TV commercials (console ads and software ads) and PRs have positive effects on social learning. In contrast, only Wii's software ads are marginally significant for personal learning. Finally, (ⅴ) the counterfactual decreasing advertising plan results in a slightly higher estimated purchase probability in our simulation. Marketers are able to optimize their advertising schedule by accounting for the dual learning processes.
Further studies should consider the influence of other factors in the analysis. First, this model does not directly consider the roles of social network structure and opinion leadership (e.g., [
33]). Second, while we classified the videogame-related websites with respect to the community feature, we did not consider the detailed contents of the websites. It may be useful to classify the websites further and use text-mining approaches to incorporate websites' contents and word-of-mouth messages.
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
PDF(393 KB)
Figure 1 Outline of the research and the proposed model
Table 1 Selected studies and the position of this study
