An empirical study of observational learning

• Published date: 01 May 2016
• DOI: http://doi.org/10.1111/1756-2171.12132
• Date: 01 May 2016
• Author: Peter W. Newberry

RAND Journal of Economics

Vol.47, No. 2, Summer 2016

pp. 394–432

An empirical study of observational learning

Peter W. Newberry∗

This article provides an empirical examination of observational learning. Using data from an

online market for music, I find that observational learning benefits consumers, producersof high-

quality music, and the online platform. I also study the role of pricing as a friction to the learning

process by comparing outcomes under demand-based pricing to counterfactual pricing schemes.

I find that employing a fixed price (the industry standard) can hamper learning by reducing the

incentive to experiment, resulting in less consumer surplus, but more expected revenue for the

platform.

1. Introduction

The way in which consumers discover new products is an important issue in many of

today’s markets. To help with this discovery,markets often feature opportunities for observational

learning, or learning about the quality of a product through the purchase decisions of peers.

Popularity (i.e., sales) rankings in online markets are an especially prevalent example. The

primary goal of this article is to empirically study the role of observational learning in a real-world

online market. Specifically, I estimate the effect of observational learning on three market-level

outcomes: the probability of success for a high-quality product, the expected consumer welfare,

and the expected revenue. The uniqueness of this exerciselies in the fact that I estimate the impact

of this type of learning, rather than test for its existence. Therefore, I am able to quantify the

benefit of learning to firms, consumers, and producers of high-quality products.

An additional goal is to study how different pricing mechanisms affect the learning process.

Price is an important determinant of consumer experimentation, meaning the way a firm prices

its product affects the efficiency of learning. Toanalyze this, I compare the three outcomes across

different pricing mechanisms. The results of this exercise provides policy makers and/or firms

with information about what mechanism is “best” in a learning environment.

∗The Pennsylvania State University; pwnewberr y@psu.edu.

I gratefully acknowledge the help and support of mydisser tation committee: J.-F.Houde, Ken Hendricks, Alan Sorensen,

and Daniel Quint. I also benefited from comments and suggestions from Chris Adams, Russ Cooper, Amit Gandhi,

Paul Grieco, Mark Roberts, Joel Waldfogel, and JuanJuan Zhang. I thank Chad Syverson and two anonymous referees

for providing comments which substantially improved the article. I am also grateful for comments from audiences at

Colby College, the Universityof Pennsylvania, the Wharton School, the Midwest Economics Association 2012 meetings

in Chicago, IL, the IIOC 2012 meetings in Washington, DC, the Econometric Society 2013 summer meetings in Los

Angeles, CA, the NBER Summer Institute 2013, Economics of Information Technology and Digitization Workshop in

Cambridge, MA, and the Quantitative Marketing and Economics Conference 2013 in Chicago, IL.

394 C2016, The RAND Corporation.

NEWBERRY / 395

I focus on an online market for music named Amie Street Music. Amie Street provides an

ideal setting in which to study observational learning. The reason is that the learning process

is quite transparent. Specifically, Amie Street employs a “demand-based pricing” scheme. With

this, a song is free when it is first posted, and then the price increases with every download up to

98¢. This scheme not only encourages consumers to experiment with newmusic, it also ser ves as a

signal of how many times a song has been purchased. Additionally, like most online music stores,

users can listen to a sample of a song before they buy it (hereafter, “listen to”) at no monetary

cost. Unique to Amie Street, however, is the fact that a consumer observes how many of her

peers have listened to the sample. Therefore, a consumer has access to two pieces of information:

(i) the price, which is a signal of the number of purchases, and (ii) the number of listens.

The transparency of these signals allows me to formulate a tractable structural model of

individual observational learning in this market. Specifically, I assume a consumer arrives at a

song and observes the price and the number of listens. Because preferences include a common

quality component (either high or low), the consumer can use this public information to form a

belief about her own utility.Based on this belief, she decides to either ignore the song or listen to

it. If she listens, all uncertainty disappears, and the purchase decision is straightforward. If she

ignores the song, she exits the market.

Hendricks, Sorensen, and Wiseman (2012) showthat in this type of environment it is possible

for a high-quality product to fail. In the model described above, this would happen if the first few

individuals who arrive at the song have unusual tastes for it, provide “inaccurate” information to

the market, resulting in others ignoring it. The more often this occurs, the higher the probability

the high-quality song fails.1

To quantify the effect of observational learning on this outcome, along with surplus and

revenue,I first estimate the parameters of the lear ning model. Todo this, I iterate over the individual

consumer-decision process until all songs have either failed or succeeded. This iteration produces

the long-run distribution of the number of listens and the price. With this joint distribution and

the distribution of the outcomes observed in the data, I form a likelihood function. I then find the

parameters which maximize the likelihood of the data.

To assess the impact of observational learning, I use the estimated parameters to compare

the outcomes on Amie Street to a counterfactual environment in which consumers have no

information. Under this assumption, consumers are unaware that prices are tied to demand and

base their listening decisions only on the prior (i.e., the overall quality of music on Amie Street).

The comparison of these environments allows me to quantify the valueof obser vationallear ning.

Results suggest that observational learning leads to an increase in total welfare by1.35%. However,

it only reduces the number of failed high-quality songs by about 1% (one song).

I also investigate how the choice of pricing scheme can impact outcomes in this market.

Different pricing schemes act as frictions to learning if they discourage experimentation, limiting

the amount of information that is passed to the market. First, I compare demand-based pricing to

a fixed price. Although the results of this exercise serve as a general analysis of pricing in these

types of markets, they are of particular interest in the music industry because of the tradition of a

fixed price of 99¢ per song (e.g., iTunes).2Ifind that the counterfactual environment with a fixed

price of 98¢ has a higher percentage of failed high-quality songs (3.6%), lowerexpected consumer

surplus (25%), and higher expected revenue (142%), compared to the Amie Street environment.3

The transfer from consumers to the platform in this environment leads to an overall welfare loss

of about 3%. These results indicate that the firm can increase revenues by fixing the price, but

this comes at the expense of consumers and producers of high-quality music. Next, I compare

1This is analogous to a “bad herd” discussed in Hendricks, Sorensen, and Wiseman (2012). However, due to

assumptions made to fit the data, the interpretation of bad herds and how and why they happen differs from Hendricks,

Sorensen, and Wiseman (2012).

2In 2009, iTunes movedto a three-tiered pricing system with prices of $0.69, $0.99, and $1.29.

3I use 98 instead of 99 because this is the maximum price reached on Amie Street. I believe this is a good

approximation of the 99¢ price point.

C

The RAND Corporation 2016.

396 / THE RAND JOURNAL OF ECONOMICS

the observed structure of the demand-based pricing scheme on Amie Street to two alternative

demand-based schemes and find that the mechanism on Amie Street generates at least as much

welfare as the others. Finally, I compare two revenueenhancing policies: (i) increasing the amount

of information and (ii) fixing the price. I find that having a fixed price is the most efficient way to

increase revenue, in that it hurts the consumers the least.

This article contributes to several different strands of the literature. Classic theoretical

observational learning articles (e.g., Banerjee, 1992; Bhikchandani, Hirshleifer, and Welch,1992;

Smith and Sørensen, 2000) show that a decision maker,when observing the decisions of her peers,

may ignore any private information she has and base her decision purely on what others have

done. This can lead to herd behavior: individuals making a decision purely because others made

that decision. Hendricks, Sorensen, and Wiseman (2012) show that in a model for a search good

with observational learning, it is possible that beliefs converge to a point where a high-quality

product gets ignored in the long run. I contribute to the above articles by empirically studying

these long-run effects of observational learning.

There are many articles in the empirical literature which study how consumers learn about

new products. Erdem and Keane (1996), Ackerberg(2003), and Crawford and Shum (2005) study

consumer learning through experimentation, whereas Chevalier and Mayzlin (2006), Sorensen

(2006), and Luca (2011) analyze the impact of peer recommendations on the decisions of con-

sumers. In this article, I focus on observational learning as the avenue for product discovery.

Empirical articles studying this type of learning are sparse. In two examples, Cai, Chen, and Fang

(2009) and Zhang (2010) test for the presence of this type of learning in the restaurant industry

and donated kidney market, respectively. Knight and Schiff (2010) tests for observational learn-

ing but does so using voting data from presidential primaries. I contribute to this literature by

estimating the effect of observational learning, rather than testing whether it is occurring.

In a project closely related to the current study, Salganik, Dodds, and Watts (2006) show

descriptive evidence of the long-run effect of social learning in an experimental music market.

I quantify these effects by estimating a structural model. Additionally, I measure the effect of

different pricing mechanisms in a market with observational learning. Although Bose et al. (2006,

2008) examine this topic theoretically,I am not aware of any empirical articles which do the same.

Articles which examine different aspects of the music industry include Hendricks and

Sorensen (2009), which studies skewness due to information problems, Olberholzer-Gee

and Strumpf (2007), which looks at the effect of piracy, and Shiller and Waldfogel (2011) and

Waldfogel (2012), whichexamine pricing and the effect of digitization, respectively. I contribute

to this literature by studying observational learning and pricing in a unique real-world market for

digital music.

Finally, I estimate the learning model using only the long-run outcomes of products. To my

knowledge, this is new to the empirical learning literature.

The remainder of the article is organized as follows. Section 2 discusses some of the insti-

tutional details of Amie Street Music. Section 3 describes the structural model of observational

learning, whereas Section 4 provides a more detailed description of the data. In Sections 5 and 6,

I present the estimation procedure and the results. Section 7 concludes.

2. Amie Street Music

To motivate the model, I introduce some of the institution details of Amie Street Music.

On Amie Street, users can download individual songs from many different artists and genres.4

Although a few recognizable artists appear on Amie Street, it is primarily focused on “indie”

music, implying that most of the artists who appear on the site are unknown, or known to very

4The site was active from 2006 to September 2010, when it was bought by Amazon.com, and subsequently shut

down.

C

The RAND Corporation 2016.