Distinction between learning and decision making and social influence mechanisms
To understand why it is crucial to keep a proper distinction between learning and decision making, let us focus on the southeast part of the map. This quadrant concerns settings characterized by a high level of social influence and a low level of clarity of the values of the alternatives. Such settings are characterized by a high level of unpredictability (Salganik et al. Reference Salganik, Dodds and Watts2006) and highly skewed popularity distributions. Why would an alternative become much more popular than other available choice alternatives? The target article explains this by invoking a high social-influence parameter (J
t
). But what the article does not fully acknowledge is that the dynamics of the system will strongly depend on how learning (i.e., the dynamic updating of quality estimates by the agents) and decision making (i.e., the choice between alternatives) combine in driving social influence.
Suppose, for example, that agents are motivated to select the same alternative as others (due to network externalities, for example), and coordinating a switch to a different, but potentially superior alternative is difficult (maybe because there are communication issues, or high switching costs). In this situation, learning does not matter, because the driving social influence mechanism is a coordination failure. Consequently, there will be a low turnover among the most popular alternatives, independently of the transparency of the qualities of the alternatives (Arthur Reference Arthur1989).
Contrast this situation to what happens in settings where the main social influence mechanism is social learning. In such settings, agents make inferences about the relative qualities of the alternatives on the basis of relative popularities because they do not have enough information to know which alternative is the best. They thus select what is popular. Prior research has shown that convergence toward one potentially suboptimal alternative relies on a very small amount of information. This implies that the release of a small amount of contradictory public information can shatter the “cascade” (Bikhchandani et al. Reference Bikhchandani, Hirshleifer and Welch1992). More generally, when the dominant social influence mechanism is social learning, this suggests that there will likely be large swings in popularities and thus a high turnover rate among popular alternatives.
The prediction regarding turnover again differs if social influence operates through sampling (Denrell & Le Mens Reference Denrell and Le Mens2007; Reference Denrell, Le Mens, Knauff, Pauen, Sebanz and Wachsmuth2013). Such social influence operates when agents are uncertain about the qualities of the alternatives: They are motivated to select what is popular (maybe because there are network externalities), but they learn about the qualities of the alternatives from their own experiences only, maybe because they do not trust others, or agents have different skills or tastes. In such settings, choices and beliefs will tend to coalesce around one alternative. But because beliefs will tend to justify differences in popularities as a result of an information bias in favor of popular alternatives, what has become popular will tend to remain popular. There will thus be a low turnover among the most popular alternatives.
What can be done by a policy maker who wants to affect the dynamics of popularities depends on the precise nature of the social influence process, which, in turn, depends on the articulation of beliefs and choices. Suppose our policy maker intends to increase the popularity of an alternative that is currently unpopular, but of higher quality than the most popular alternative. If the population is subject to a lock-in due to a coordination failure, the solution is to create a coordination mechanism (possibly a new institution, or a vote) to facilitate the switch. If a suboptimal alternative is the most popular as a result of an information cascade, making available a small amount of information in favor of the unpopular but superior alternative should be enough to lead to a switch. Such information release would not lead to a switch if the problem were a coordination failure. And if the population selects a suboptimal alternative as a result of social influence through sampling, the remedy is to make people sample the unpopular alternative. This will work even if persuasive campaigns are ineffective.
Decisions about information feeds and the technologies at the core of the “big data” revolution
The technologies that support the “big data” discussed by Bentley et al. do not only facilitate access to more information more quickly, but they are also characterized by a high level of customizability: Agents are now free to configure information sources in order to get selectively exposed to the information they want to be exposed to. They can do so by tweaking the settings of RSS (Rich Site Summary) readers, selecting the tweets they attend to, or customizing the news feed on their Facebook account. This is in sharp contrast with older forms of information broadcasting such as television, radio, or newspapers where agents were receivers with less latitude in deciding what to get exposed to. The novel customizability of information sources has the potential to reduce rather than increase the diversity of information agents get exposed to. We researchers can only hope to understand the implications of this increased customization for the dynamics of opinions, tastes, and popularity if we decouple learning, beliefs, and decision making in our analyses and aim to understand the dynamic process by which agents adjust their exposure to information.
You have
Access
The two-dimensional “map” proposed by Bentley et al. relies on a set of four simplifying assumptions – three of which I find unproblematic. But I would like to discuss the third assumption: the simplification that consists in blurring “the distinction between learning and decision making” (target article, sect. 2, para. 7). Although the authors recognize that these are distinct actions, they also claim this distinction is somewhat too subtle to be of serious consideration in the target article. This simplification unfortunately (1) obscures the social influence processes that operate in the various empirical settings the map aims to categorize, thus limiting the usefulness of the potential policy implications of the studies being mapped; and (2) it limits our potential understanding of the implications of the new technologies underlying the “big data” revolution for the dynamics of popularities of choices and opinions.
Distinction between learning and decision making and social influence mechanisms
To understand why it is crucial to keep a proper distinction between learning and decision making, let us focus on the southeast part of the map. This quadrant concerns settings characterized by a high level of social influence and a low level of clarity of the values of the alternatives. Such settings are characterized by a high level of unpredictability (Salganik et al. Reference Salganik, Dodds and Watts2006) and highly skewed popularity distributions. Why would an alternative become much more popular than other available choice alternatives? The target article explains this by invoking a high social-influence parameter (J t ). But what the article does not fully acknowledge is that the dynamics of the system will strongly depend on how learning (i.e., the dynamic updating of quality estimates by the agents) and decision making (i.e., the choice between alternatives) combine in driving social influence.
Suppose, for example, that agents are motivated to select the same alternative as others (due to network externalities, for example), and coordinating a switch to a different, but potentially superior alternative is difficult (maybe because there are communication issues, or high switching costs). In this situation, learning does not matter, because the driving social influence mechanism is a coordination failure. Consequently, there will be a low turnover among the most popular alternatives, independently of the transparency of the qualities of the alternatives (Arthur Reference Arthur1989).
Contrast this situation to what happens in settings where the main social influence mechanism is social learning. In such settings, agents make inferences about the relative qualities of the alternatives on the basis of relative popularities because they do not have enough information to know which alternative is the best. They thus select what is popular. Prior research has shown that convergence toward one potentially suboptimal alternative relies on a very small amount of information. This implies that the release of a small amount of contradictory public information can shatter the “cascade” (Bikhchandani et al. Reference Bikhchandani, Hirshleifer and Welch1992). More generally, when the dominant social influence mechanism is social learning, this suggests that there will likely be large swings in popularities and thus a high turnover rate among popular alternatives.
The prediction regarding turnover again differs if social influence operates through sampling (Denrell & Le Mens Reference Denrell and Le Mens2007; Reference Denrell, Le Mens, Knauff, Pauen, Sebanz and Wachsmuth2013). Such social influence operates when agents are uncertain about the qualities of the alternatives: They are motivated to select what is popular (maybe because there are network externalities), but they learn about the qualities of the alternatives from their own experiences only, maybe because they do not trust others, or agents have different skills or tastes. In such settings, choices and beliefs will tend to coalesce around one alternative. But because beliefs will tend to justify differences in popularities as a result of an information bias in favor of popular alternatives, what has become popular will tend to remain popular. There will thus be a low turnover among the most popular alternatives.
What can be done by a policy maker who wants to affect the dynamics of popularities depends on the precise nature of the social influence process, which, in turn, depends on the articulation of beliefs and choices. Suppose our policy maker intends to increase the popularity of an alternative that is currently unpopular, but of higher quality than the most popular alternative. If the population is subject to a lock-in due to a coordination failure, the solution is to create a coordination mechanism (possibly a new institution, or a vote) to facilitate the switch. If a suboptimal alternative is the most popular as a result of an information cascade, making available a small amount of information in favor of the unpopular but superior alternative should be enough to lead to a switch. Such information release would not lead to a switch if the problem were a coordination failure. And if the population selects a suboptimal alternative as a result of social influence through sampling, the remedy is to make people sample the unpopular alternative. This will work even if persuasive campaigns are ineffective.
Decisions about information feeds and the technologies at the core of the “big data” revolution
The technologies that support the “big data” discussed by Bentley et al. do not only facilitate access to more information more quickly, but they are also characterized by a high level of customizability: Agents are now free to configure information sources in order to get selectively exposed to the information they want to be exposed to. They can do so by tweaking the settings of RSS (Rich Site Summary) readers, selecting the tweets they attend to, or customizing the news feed on their Facebook account. This is in sharp contrast with older forms of information broadcasting such as television, radio, or newspapers where agents were receivers with less latitude in deciding what to get exposed to. The novel customizability of information sources has the potential to reduce rather than increase the diversity of information agents get exposed to. We researchers can only hope to understand the implications of this increased customization for the dynamics of opinions, tastes, and popularity if we decouple learning, beliefs, and decision making in our analyses and aim to understand the dynamic process by which agents adjust their exposure to information.