You are on page 1of 30

RESEARCH ARTICLE

PEER EFFECTS IN COMPETITIVE ENVIRONMENTS:


FIELD EXPERIMENTS ON INFORMATION PROVISION
1
AND INTERVENTIONS
Zhuoxin Li
Carroll School of Management, Boston College, 140 Commonwealth Avenue,
Chestnut Hill, MA 02467 U.S.A. {zhuoxin.li@bc.edu}

Gang Wang
Lerner College of Business and Economics, University of Delaware, 42 Amstel Avenue,
Newark, DE 19716 U.S.A. {gangw@udel.edu}

Harry Jiannan Wang


Lerner College of Business and Economics, University of Delaware, 42 Amstel Avenue,
Newark, DE 19716 U.S.A. {hjwang@udel.edu}

Prior studies of social-norms interventions have focused on nudging behavior in noncompetitive settings. This
research evaluates such interventions in competitive environments, for example, a class with a competitive
grading policy. Field experiments on a learning management system show that providing descriptive infor-
mation about peers’ behavior has mixed effects in reducing procrastination and improving performance
outcomes. Specifically, the effects are moderated by individual characteristics and contextual variables. First,
peer information interventions are more effective for males, and the effects are stronger in a male-majority
environment than in a female-majority environment. These findings differ from prior studies of social-norms
interventions conducted in noncompetitive settings, in which females are found to be more responsive to inter-
ventions. Gender differences in our competitive settings can be explained by males’ and females’ differential
preferences for competition: males are more competitive-oriented and thus are more responsive to peer infor-
mation in competitive environments. Second, we find that individuals who are in great need of interventions,
that is, those with poor past behavior and performance, are also more likely to benefit from peer information
interventions, suggesting that peer information interventions motivate positive change. This study highlights
the heterogeneous effects of peer information interventions and has implications for targeted interventions.

Keywords: Peer information, peer effects, social norms, social comparison, gender differences, nudging,
procrastination, digital interventions, randomized field experiment, learning management systems (LMS)

Introduction1 activities. For instance, health and fitness apps such as Lose
It! and MyFitnessPal help individuals to plan their nutrition
Individuals use a variety of personal applications and enter- and exercise programs, monitor and keep track of progress,
prise systems for work, learning, entertainment, and other and learn about the activities of other members. In an anal-
ogous fashion, online learning management systems (LMS)
such as Canvas and BlackBoard create an IT-mediated envi-
1 ronment for the teaching and learning process, where students
Bin Gu was the accepting senior editor for this paper. Ling Xue served as
the associate editor. get access to course materials, complete assignments online,

DOI: 10.25300/MISQ/2021/16085 MIS Quarterly Vol. 45 No. 1 pp. 163-191/March 2021 163
Li, Wang, & Wang/Peer Effects in Competitive Environments

and receive performance feedback once the teacher posts Previous attempts to reduce procrastination in college classes
assignment grades. These information systems have led to the have used different approaches. For example, Lamwers and
accumulation of detailed data on individuals’ behaviors and Jazwinski (1989) design a number of intermediate deadlines
outcomes. There is an opportunity to analyze these data to and impose these deadlines for the completion of coursework.
learn about individual behavior, understand what behavioral Ariely and Wertenbroch (2002) examine another approach of
patterns are associated with superior or unsatisfactory out- requiring students who delayed their work to meet with the
comes, and, ultimately, design targeted interventions to teacher. Although demonstrated effective in experimental
improve individual performance. For instance, such interven- settings, these interventions are costly to implement because
tions may involve nudging behavior and performance they require significant changes to the current curriculum
improvement by disclosing information about peers (Cialdini design or require extra responsibilities for the teachers.
and Goldstein 2004). Prior studies of social-norms interven-
tions have focused on nudging individual behavior in non- In this research, we designed peer information interventions
competitive settings. However, much is unknown about the through LMS to reduce procrastination in a classroom setting
effectiveness of such interventions in competitive environ- where students’ class ranking affects their final grade. Peer
ments; for example, individuals competing for a higher information reveals the prevalence of certain behaviors in a
ranking on a fitness app’s leaderboard, students in a class- population, such as the number (or the percentage) of students
room with a ranking-based grading policy, employees within in the class who have already submitted an assignment. This
a department competing for performance-based bonuses, etc. type of descriptive information may shape individuals’ beliefs
about what others are doing and thus can motivate their own
This research investigates targeted interventions in the context change to conform to socially appropriate behavior (Cialdini
of online LMS. LMS are an integral part of higher education, and Goldstein 2004) or to achieve a superior position vis-à-vis
with millions of students spending a significant amount of others (Garcia et al. 2013). Compared to existing procras-
their time every week with these systems (Dahlstrom et al. tination reductions through teacher interventions (e.g., Ariely
2014). The wide adoption of LMS in almost every college and Wertenbroch 2002), peer information interventions have
has motivated teachers to consider digital interventions to the appeal of being able to be seamlessly integrated into LMS.
improve the learning process. Such digital inventions take
advantage of rich behavioral data about students’ learning Despite a long tradition of effectiveness in laboratory tests,
activities recorded in system logs (Lam et al. 2012). For peer information interventions have had mixed success in
instance, Canvas keeps track of students’ browsing histories, changing behavior in field contexts, with some studies
clickstream data, and assignment submission records. These showing boomerang effects (Perkins et al. 2005; Schultz et al.
fine-grained data are available at both individual and aggre- 2007). The specific contexts and populations for which these
gate levels, which allow teachers to track each student’s interventions are effective are still not clear (Silva and John
academic performance relative to their peers. Furthermore, 2017). In the context of college classrooms, the critical ques-
such detailed data are also immediately available for targeted tions of interest to educators are whether and under what
experimentation or gamification to improve students’ learning conditions providing peer information will yield positive out-
behaviors and academic performance (Liu et al. 2017; Wang comes, such as improved comprehension and higher academic
et al. 2017). performance. Answers to these questions, together with the
detailed individual-level data from LMS, have the potential to
We focus on interventions of a specific type of individual improve student performance through targeted interventions.
behavior: procrastination. Poor study habits such as procras-
tination are among the major antecedents of unsatisfactory We investigate the effectiveness of peer information through
academic performance for college students (Ariely and a build–evaluate approach (Hevner et al. 2004). Specifically,
Wertenbroch 2002). A recent survey indicates that over 87% we first developed a computer program (an add-on for
of college students procrastinate, particularly when it comes Canvas) that automatically retrieves real-time data about each
to doing their coursework or assigned readings (StudyMode student from the system logs of Canvas, computes descriptive
2014). The absolute amount of procrastination is also signi- summary statistics (descriptive peer information), and iden-
ficant, with students reporting that it typically occupies over tifies individuals to receive treatments. We then conducted a
one-third of their daily activities, often enacted through set of field experiments with undergraduate students in the
sleeping, playing, watching TV, or browsing social media northeastern United States. Our computer program randomly
feeds (Pychyl and Sirois 2016). Students recognize procras- assigned subjects into different treatment groups to receive
tination as being harmful and foolish, and over 95% of interventions. At the end of the experiments, we collected
procrastinators wish to reduce it (O’Brien 2002). data from Canvas to evaluate whether and when peer infor-

164 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

mation interventions are effective in promoting behavioral ments. Peer information interventions can also target
change (i.e., reducing procrastination) and improving individuals with poor past behavior and performance, the
academic outcomes (i.e., achieving a higher grade). individuals most in need of and benefitting more from inter-
ventions. The peer information interventions feature we
Our field experiments reveal that peer information interven- developed can be built into LMS and can be used for classes
tions reduce procrastination behavior and increase individual throughout the entire institution.
performance (i.e., achieving a higher grade). However, the
effects of peer information interventions are moderated by
individual characteristics and contextual variables. Sur-
Literature Review and
prisingly, peer information interventions are more effective
for males than females. In addition, the effects are also
Hypothesis Development
stronger in a male-majority environment than a female-
In this section, we review prior studies on procrastination and
majority environment. The experiments also show that
its measurement. We also discuss the literature related to peer
individuals who are most in need of interventions, that is,
effects, gender differences, and interventions for behavior
those with poor past performance (previously receiving a
change. We then develop the hypotheses on the effects of
lower grade) and bad past behavior (previously procras-
providing peer information on procrastination mitigation.
tinating more), are also more likely to benefit from the inter-
ventions, suggesting such interventions motivate positive
change.
Procrastination and Measurement
This research extends the literature of peer information inter-
Procrastination is defined as the delay in the beginning or
ventions to competitive settings. Prior studies have focused
completion of an intended task (Ferrari 1992). Prior studies
on peer information interventions in noncompetitive settings
have recognized procrastination as self-regulation deficiency
(e.g., reusing towels in a hotel room), in which an individual’s
or failure, for example, lack of motivation and weak impulse
performance does not interfere with others’ performance. control (Arvey et al. 2006; Schouwenburg et al. 2004).
Therefore, peer information interventions motivate behavior
change due to social conformity (Cialdini and Goldstein Often acknowledging procrastination as an irrational delay,
2004). Since females are more socially tuned, they would be individuals expect the outcome to be worse because of the
more likely to respond to such interventions. In competitive delay (Steel 2007). Procrastination in academic contexts, in
settings, providing peer information may induce comparison particular, is a severe problem among students (Kim and Seo
concerns, enhance the perceived competitiveness of the envi- 2015). For example, male students tend to use regulated
ronment, and thus motivate individuals’ own behavior change strategies less effectively than female students, and are more
to achieve a superior position vis-à-vis others (Garcia et al. likely to procrastinate (Karatas 2015; Özer et al. 2009).
2013). In contrast to the prediction from the social con-
formity theory, our results show that males are actually more Prior studies have mostly measured procrastination based on
likely to respond to peer information interventions due to self-assessment questionnaires. However, in these self-
social comparison. Our findings show that peer information reported surveys, students tend to over-report their grades and
interventions in a competitive setting motivates males more exaggerate their procrastination (Kim and Seo 2015). Thanks
strongly than females because males are more competitive- to the adoption of LMS, detailed data on student behavior
oriented. (e.g., clickstream data) have now enabled researchers to
create objective metrics for students’ learning habits. Prior
The findings from this study have implications for the design studies have used the start time for specific tasks as indicators
of data-driven, targeted interventions to promote healthy of procrastination (Cerezo et al. 2017; Rotenstein et al. 2009;
behavior and improve performance outcomes. Providing peer You 2016).
information may help reduce procrastination for college
students, but the effects depend on individual characteristics
and the contextual environment. To achieve effective inter- Procrastination Interventions
ventions, educators may consider gender differences, gender
composition, and the individual’s behavior trajectory and past The prevalence of procrastination and its negative effects are
performance. For instance, peer information interventions well recognized. Educators and researchers have attempted
may target males in a male-majority environment because to help students overcome procrastination. Extant intervent-
these individuals are more likely to benefit from such treat- ions mostly take the form of therapeutic programs (e.g.,

MIS Quarterly Vol. 45 No. 1/March 2021 165


Li, Wang, & Wang/Peer Effects in Competitive Environments

helping procrastinators build self-regulatory skills and self- Neighbors 2004; Perkins 2002), and promoting sustainable
efficacy) to establish better habits and overcome procras- environmental behavior (Schultz et al. 2007).
tination (van Eerde and Klingsieck 2018). A few applied
studies have attempted to reduce procrastination. For Prior studies have focused on peer information interventions
example, Lamwers and Jazwinski (1989) design different in noncompetitive settings (e.g., reusing towels in a hotel
arrangements of deadlines and find that frequent deadlines room), in which an individual’s performance does not inter-
reduce procrastination and improve student grades compared fere with others’ performance. In noncompetitive settings,
to a fully self-paced arrangement. Ariely and Wertenbroch peer information interventions motivate behavior change due
(2002) discuss another approach to decrease procrastination to social conformity (Cialdini and Goldstein 2004). In com-
by requiring students who delayed their work to meet with the petitive settings, providing peer information may induce
teacher. Although proved effective to some extent, the inter- social comparison—the tendency to self-evaluate by com-
ventions to reduce procrastination in the literature, are not paring oneself to others (Garcia et al. 2013). The social cues
without limitations. Typically, a deadline imposed by the from descriptive information may shape individuals’ beliefs
teacher involves the threat of another punishment (e.g., failing about what others are doing, enhancing the perceived com-
an assignment or being forced to meet with the teacher) if the petitiveness of the environment, and thus can motivate an
desired academic behavior is not performed by the specified individual’s own behavior change to achieve a superior
date. Perrin et al. (2010) point out that “improved studying position vis-à-vis others (Garcia et al. 2013).
due to an intervention of this kind would appear to be main-
tained by negative reinforcement” (p. 465) in the form of Evidence for the success of peer information interventions has
avoiding some additional teacher-imposed punishment. been mixed. Although some studies appear to confirm the
Furthermore, these interventions can be costly to implement effectiveness of the peer information interventions, other
because they either require significant changes to curriculum studies have failed to produce substantial changes in behavior
design or demand extra responsibilities from the teachers (Clapp et al. 2003; Granfield 2005). It is still not clear for
(e.g., additional time to meet with students who procras- which contexts and populations these interventions are effec-
tinated). tive (Silva and John 2017). In fact, some studies indicate that
peer information interventions may backfire; they have
Our study investigates an alternative intervention that relies actually increased undesirable behaviors and misperceptions
on the revelation of descriptive peer information to students, they set out to decrease (Perkins et al. 2005; Wechsler et al.
which does not interfere with a teacher’s course schedule. 2003; Werch et al. 2000). For instance, providing peer
The feature of peer information notification we developed can information (for example, the fraction of other employees in
be built into LMS and used for classes throughout the entire the company who participate in the 401(k) plan) may decrease
institution. Several field experimental studies have inves- the enrollment of nonparticipants due to discouragement from
tigated student behavior on MOOC platforms, but their focus upward social comparisons (Beshears et al. 2015). Also,
is on content contribution (Baek and Shore 2020) or assign- Hashim et al. (2017) find that random provision of informa-
ment completion (Huang et al. 2018) rather than learning tion about others’ actions is no different than not providing
outcomes in a classroom setting (the completion rate is around information, suggesting the need for targeted interventions.
97%, and thus less of a concern in our setting). The objective
of our research is to identify when peer information interven-
tions work (identifying moderators that determine hetero- Hypotheses Development
geneous treatment effects) and how they work (identifying the
mediator that leads to grade improvement). Effects of Peer Information

Individuals use information about their peers to determine the


Effects of Peer Information: Social most appropriate course of action in a given situation
Conformity and Social Comparison (Cialdini and Goldstein 2004; Cialdini and Trost 1998).
Providing peer information may create a sense of urgency
Policymakers are increasingly interested in the potential of (Michael 1991) and produce a salience effect (Leland et al.
leveraging peer effects to nudge individual behavior (Peng et 2019) that focuses the attention of participants on the targeted
al. 2017; Qiu and Kumar 2017). While much is unknown action. In competitive settings, disclosing descriptive infor-
about peer information interventions through LMS, several mation about other students’ progress may also encourage
studies have looked into the effect of peer information in social comparisons, making the competitive environment
other contexts, such as stimulating prosocial behavior (Huang more salient to those who received the information (Garcia et
et al. 2019), reducing alcohol consumption (Lewis and al. 2013). Competition effects may motivate students to start

166 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

earlier (and spend more time on the assignment), hoping to above may drive the gender effects in opposite directions. On
achieve a higher grade compared to their peers. one hand, females may be more likely to respond to the social
cues because they are more socially attuned (Gordon et al.
Peer information about other students’ progress may motivate 2000). On the other hand, males are more likely to be
procrastinators to reduce procrastination and start working on motivated by a competitive environment induced by peer
an assignment earlier than they typically would. Such information (Ruble et al. 2006; Weber et al. 2004). Given
positive behavior change may be associated with performance prior findings in the literature and the specific context we
improvement for several reasons. First, starting the assign- study, we lean toward the prediction that, in our setting, male
ment earlier grants students more time to understand students are more likely to opt into a competitive setting
expectations and to acquire the knowledge and skills needed induced by the peer information they received (Gneezy et al.
to solve the problems in the assignment. Second, early 2003). As a result, providing descriptive peer information can
starters may also seek help from teaching assistants or the motivate them to work on the assignment early to achieve a
instructor—they can go to office hours when they are stuck better grade compared to others. We, therefore, propose the
with difficult questions. In contrast, procrastinators may find following hypothesis:
themselves short of time and resources to complete a quality
submission before it is due. Therefore, reducing student pro- Hypothesis 2a (H2a). The effect of peer information
crastination may lead to better performance outcomes (Ariely interventions is stronger for males than females.
and Wertenbroch 2002). We thus hypothesize:
Gender Composition. The effects of peer information are
Hypothesis 1a (H1a). Peer information interventions reduce very likely to be context-specific. In particular, this section
student procrastination. discusses how the gender composition of the reference group
moderate the effectiveness of peer information interventions.
Hypothesis 1b (H1b). Peer information interventions
increase student performance. Individuals’ behavior and performance depend not only on
their own gender but also on the gender of people with whom
they interact (Gupta et al. 2013). Groups of different gender
The Moderating Role of Gender compositions may reveal distinct idiosyncrasies emerging
and Gender Composition from the interactions among group members. For example,
groups with a higher female-to-male ratio may reveal more
The effects of peer information interventions are very likely “feminine” personality and dynamics. In a study of a business
to be individual- and context-specific (Lu et al. 2016). This competition game, Apesteguia et al. (2012) observe that all-
section discusses how gender and gender composition moder- female teams are less aggressive than mixed-gender or all-
ate the effectiveness of peer information interventions. male teams in running their businesses. All-female teams
compete less aggressively in pricing and invest more in social
Gender: Gender differences and their impacts on economic sustainability initiatives. In a dictator game setting, Dufwen-
and social outcomes have attracted increasing attention berg and Muren (2006) find that female-majority groups are
(Ahuja and Thatcher 2005). The literature suggests that more generous and egalitarian.
females are more socially attuned and thus more cooperative
(Stocks et al. 2009). For example, Frank et al. (1993) find Gender composition may influence the behavior of group
that females are significantly more cooperative than males in members across gender. Lavy and Schlosser (2011) find that
prisoner’s dilemma games. In contrast, males are prone to an increase in the proportion of female students in the class-
exhibit stronger individualism (Kashima et al. 1995). Males room improves both male and female students’ cognitive
were found to exhibit significantly more aggressiveness than outcomes. Such academic gains are mediated through
females (Brooks 1982). Furthermore, males prefer competi- decreased classroom disruptions and improved inter-student
tion more than females. Gneezy et al. (2003) let participants and student–teacher relationships. Lee et al. (2014) investi-
solve computerized mazes in different competitive settings. gate the impact of school and classroom gender composition
They find that males are more strongly motivated by com- on student achievement. They find that male students
petitive incentives. attending single-sex classes within coed schools score below
male students in mixed-gender classes. Interestingly, class-
While prior studies suggest females are more likely to respond room or school gender composition does not affect the out-
to social cues in noncompetitive settings, it is not clear comes of female students (Lee et al. 2014), which suggests
whether males or females are more responsive to peer infor- that gender composition has disproportional effects on male
mation in competitive settings. The two forces discussed students’ behavior.

MIS Quarterly Vol. 45 No. 1/March 2021 167


Li, Wang, & Wang/Peer Effects in Competitive Environments

Following the findings in the literature, we expect an environ- Research Method


ment with more females to reveal more feminine personality
and dynamics (Apesteguia et al. 2012; Dufwenberg and Experimental Setting
Muren 2006). Students, including male students, may per-
ceive an environment of a higher female-to-male ratio less We designed and implemented a set of randomized field
competitive and more cooperative. Thus, compared to a experiments with Canvas, a widely adopted LMS in the
male-majority environment, male students in a female- United States. The subjects are students of undergraduate
majority class may respond less aggressively to the peer infor- courses in business analytics using Excel and other tools.
mation they receive. (Dewan et al. 2017; Lee et al. 2015; Each course has multiple sections taught by different teachers.
Zhang et al. 2017). We, therefore, hypothesize as follows: The reasons for choosing such courses are threefold. First,
the size of each section of the class is relatively large so that
Hypothesis 2b (H2b). The effect of peer information inter- we can assign different treatment conditions to students
ventions is stronger in a male-majority environment than in within the same section. Second, different sections of the
a female-majority context. course are highly comparable: they used the same core
syllabus and curriculum, and all used the Canvas site. Third,
assignments in these courses require considerable efforts from
The Moderating Role of Past the students. More details related to the courses, the assign-
Behavior and Performance ments, and sample questions are presented in Appendix E.

We developed a computer program that automatically


The effect of peer information interventions is also likely to
retrieves real-time information about each student from the
be moderated by an individual student’s past behavior and
system log of Canvas, computes descriptive statistics of
performance. Specifically, we expect the effect of peer infor-
student progress in the reference group (descriptive peer
mation interventions will be stronger for students with
information), and identifies individuals to receive interven-
relatively poor past behavior and low past performance.
tions. Specifically, as an assignment deadline approached
Descriptive information about desirable behavior by peers (e.g., due in a few days), the computer program compiled a
provides a reference point, with which people are likely to list of students who had not started working on the assignment
conform (Cialdini and Goldstein 2004). With such a cue in (i.e., not yet downloaded the data file for the assignment).
mind, an individual can measure the appropriateness of their The computer program then randomly assigned subjects into
own behavior by how far away they are from the descriptive different treatment groups. Students received these messages
information of their reference group (Schultz et al. 2007). via email and were also notified when they logged into
The discrepancy between peers’ and his own behavior may Canvas. We designed and executed three complementary
motivate an individual to determine the most appropriate studies, summarized in Table 1.
course of action in a given situation (Cialdini and Goldstein
2004; Cialdini and Trost 1998). Studies I and II were conducted in classes with a competitive
grading policy where a student’s final letter grade depended
Providing peer information induces social comparison con- on her/his class ranking (see grading policies in Appendix E).
cerns (Garcia et al. 2013) and the desire to achieve or main- Standard experiments leveraging peer information consider all
tain a better relative position among others, which increases individuals in a class section as the reference group and pro-
perceived competitiveness especially for those with lower vide descriptive statistics based on the observed number in the
prior performance. While high-performing individuals can group. We adopted this design in Study I. However, this
also be responsive to interventions because they are more experimental design is subject to one identification issue
capable of improving themselves, the effects from peer infor- when exploring the role of the contextual environment: there
mation interventions are possibly weaker for these individuals is no variation in gender composition within a section.
because they are already more self-motivated, even without Increasing the number of classes/sections could not fully ad-
interventions. In sum, individuals who realize they are below dress this issue because doing so would introduce an
peers are more likely to exert extra effort to catch up, whereas additional degree of heterogeneity at the section level. There-
those who find they are above are less likely to respond posi- fore, we explored an alternative design. In Study II, students
tively. We, therefore, propose the following hypothesis: were randomly assigned to small study groups before working
on the individual assignment. In the experiment, students still
Hypothesis 3 (H3). The effect of peer information inter- received descriptive peer information (as in Study I), but the
ventions is stronger for individuals who have poor past reference group is the study group they were in rather than the
behavior and past performance. entire class. All experimented assignments in this research

168 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 1. Experimental Design


Study I Study II Study III
Experiment A course that had a non-ranking final
Courses that had a ranking-based final grade policy
Context grade policy
Experiment Group-Level Peer Info + Manipulation of
Section-Level Peer Info Group-Level Peer Info
Design Competition
Add on top of Study II: create variations
Baseline design from the Create variations in gender in competitiveness (with vs. without
Objective
literature composition ranking-based grading policy) for the
experimented assignment.
“Assignment 3 is due in 4 “Assignment 3 is due in 4
Control Condition No Message
days.” days.”
“Assignment 3 is due in 4 “Assignment 3 is due in 4
T1: “Assignment 3 is due in 4 days.”
days. X% of students in days. X% of students in
Treatment T2: “Assignment 3 is due in 4 days. X%
your class have started your group have started
Condition(s) of students in your group have started
working on this working on this
working on this assignment.”
assignment.” assignment.”
Assignment
Assignment with a
Manipulation of without a ranking-
— — ranking-based
Competition based grading
grading policy
policy
Total # Courses 2 2 1 1
Individual Assignments Individual Assignments
(Course I-1: Excel (Course II-1: Excel
Assignment Individual Assignments
modeling questions; modeling questions;
Format (Excel modeling questions)
Course I-2: Analytics Course II-2: Analytics
questions) questions)
Total # Sections 6 4 5 6
Total # Groups — 39 67 80
Total Class Size 446 199 213 255
Total # Subjects
136 71 49 57
in Control
Total # Subjects
120 63 102 116
in Treatment
Notes: More details related to the courses, the assignments and sample questions are in the Appendix E.

are individual assignments. Students were required to com- Study I (Section-Level Peer Information). This standard
plete them independently. peer information intervention considers all students in the
class section as the reference group. As shown in Table 2,
In Study III, we used a different course without a ranking- students within a class section were randomly assigned to a
based grading policy for the final letter grades (a student’s control group to receive messages without peer information
final grades depend on which bucket her/his final score falls or the treatment group to receive messages with peer
into). However, we manipulated the grading policy for the information computed at the section level (i.e., “X% of
experimented assignment so that students in the competitive students in your class have started working on this
conditions receive a final assignment score that is determined assignment”).
by their ranking in the class (see Table E1 and grading
policies in Appendix E), whereas students in the benchmark Study II (Group-Level Peer Information). At the
conditions do not follow a ranking-based policy. We beginning of the semester, students were randomly assigned
elaborate on the three studies below. to a four- to six-person group. As shown in Table 3, students
within a class section were randomly assigned to a control

MIS Quarterly Vol. 45 No. 1/March 2021 169


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 2. Treatment Conditions with Section-Level Peer Information (Study I)


Condition Message
Control T0 (Reminder only): “Assignment 3 is due in 4 days.”
Treatment T1 (Reminder with Peer Info): “Assignment 3 is due in 4 days. X% of students in
(Class-Level Peer Info) your class have started working on this assignment.”

Table 3. Treatment Conditions with Group-Level Peer Information (Study II)


Condition Message
Control T0 (Reminder only): “Assignment 3 is due in 4 days.”
Treatment T1 (Reminder with Peer Info): “Assignment 3 is due in 4 days. X% of students in
(Group-Level Peer Info) your group have started working on this assignment.”

Table 4. Treatment Conditions with Manipulation of Competition (Study III)


Condition Message
TNR0 (No Message)
TNR1 (Reminder only): “Assignment 3 is due in 4 days.”
Non-ranking based grading
TNR2 (Reminder with Peer Info): “Assignment 3 is due in 4 days. X% of students in
your group have started working on this assignment.”
TR0 (No Message)
TR1 (Reminder only): “Assignment 3 is due in 4 days.”
Ranking-based grading
TR2 (Reminder with Peer Info): “Assignment 3 is due in 4 days. X% of students in
your group have started working on this assignment.”
Notes: Right before sending out the messages, our computer program retrieved and analyzed system log data to compute the variable X, the
percent of students who had started working on the assignment.

group to receive messages without peer information or the StartTime, a continuous variable recording the length of time
treatment group to receive messages with peer information, (hours) between two events: the time a student started
where the descriptive information comes from the group (i.e., working on the assignment (indicated by downloading the
the percent of students in their group who have started necessary data file for the assignment) and the time the
working on the assignment). assignment was due (Cerezo et al. 2017; Rotenstein et al.
2009; You 2016).2 A smaller value of StartTime indicates
Study III (Manipulation of Competition). Study III builds that the student started late and had experienced more
on Study II with additional manipulation of the competitive procrastination, whereas a larger value suggests that the
environment (Table 4). Students in the competitive condi- student started early. The second dependent variable is
tions received a final assignment score that is determined by Grade, which is the points a student received from the assign-
their ranking in the class, whereas students in the benchmark ment. We normalized this grade by the total credits for the
conditions did not follow a ranking-based policy. assignment such that Grade is a percentage ranging from 0 to
100. Both StartTime and performance Grade are of interest
to educators: StartTime captures students’ behavior change
Variables and Measurement (the extent of procrastination) due to the treatments, and such
behavior improvement may lead to a higher performance
Dependent Variables. We constructed two dependent outcome as measured by Grade.
variables to measure the outcomes of experimental conditions
(Table 5). The first dependent variable is related to a 2
Appendix F provides more evidence on the validity of using StartTime as an
student’s procrastination behavior, which is measured by indicator of procrastination.

170 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 5. Notation and Variable Definition


Variables Definitions
Dependent Variables
Hours that a student started working on the assignment (indicated by downloading
StartTime
the necessary data file) before the due time.
Percentage of points that a student achieved over the total points of the
Grade
assignment.
Other Variables (Student Characteristics and Environmental Variables)
Male The sex of a student: 1 for male, and 0 for female.
The StartTime from the previous assignment. A measure of past procrastination
StartTime_Prev
tendency.
Grade_Prev The Grade from the previous assignment. A measure of past performance.
Section_Male_Composition Percentage of male students in a section.
Group_Male_Composition Percentage of male students in a group (Study II and III).

Other Variables. To investigate the heterogeneous treatment as the reference group (i.e., “X% of students in your class
effects for students, we obtained gender information for each have started working on this assignment”). Randomization
student coded by the binary variable, Male. We also con- tests in Table A1 of Appendix A show that there are no
structed two important variables related to students’ significant differences between the treatment group and the
procrastination tendency and past performance as observed in control group with respect to preexperiment factors and
the previous assignment: (1) StartTime_Prev: hours that a individual traits.
student started working on the previous assignment before the
corresponding due time, and (2) Grade_Prev: a student’s Before statistically testing the hypotheses, we first visualize
grade from the previous assignment. In addition, we calcu- some of the patterns observed in Study I. As shown in Figure
lated the percentage of male students, Section_Male_ 1a, peer information interventions overall have positive ef-
Composition, to measure gender composition in a section. fects on procrastination behavior. However, from Figure 1b,
We also calculated the percentage of male students in a group peer information interventions have a positive effect on male
Group_Male_Composition for Study II and III. students’ procrastination behavior, but there is a seemingly
negative effect for female students.

Effects of Providing Peer Information Gender Differences


on Behavior Change
In a set of regression analyses with data for Study I, we esti-
This section presents the effects of peer information interven- mated the treatment effect of peer information interventions
tions on reducing student procrastination. We discuss the and the moderating role of gender (Table 6). The results of
intervention effects on performance outcomes in the next Model 1 show that peer information interventions do not have
section. a significant main effect on students’ procrastination
behavior. That is, H1a is not supported for male and female
combined. Further, the results of Model 2 show that the
Study I (Section-Level Peer Information) treatment effect for female students is negative (-9.43, not
statistically significant, p > 0.10), and that the estimation of
As discussed earlier (Table 1), Study I follows the standard the interaction term of Treatment*Male is 18.41 and statis-
design of peer information interventions and treats all students tically significant. Therefore, the marginal treatment effect
in the class as the reference group. Students were randomly for male students is 8.98, indicating that peer information
assigned to the control group to receive a reminder message interventions have a stronger positive effect on male students’
or the treatment group to receive a message with reminder and behavior than female students as reflected by an earlier
descriptive peer information computed using the whole class starting time of 9 hours.

MIS Quarterly Vol. 45 No. 1/March 2021 171


Li, Wang, & Wang/Peer Effects in Competitive Environments

(a) Impact on Starting Time (b) Impact on Starting Time by Gender

Figure 1. Effects of Peer Information Interventions on Student Behavior

Table 6. Moderating Effects of Gender


All
DV: StartTime Model 1 Model 2 Males Female
42.88*** 56.52*** 30.43* 20.37
Intercept
(15.56) (16.49) (17.70) (45.06)
1.37 -9.43 9.57* -8.52
Treatment
(4.43) (6.78) (5.33) (7.65)
-15.01**
Male – – –
(6.38)
18.41**
Treatment*Male – – –
(8.92)
0.04* 0.04* 0.02 0.09*
StartTime_Prev
(0.02) (0.02) (0.02) (0.05)
0.09 0.08 0.23 -0.04
Grade_Prev
(0.15) (0.15) (0.18) (0.27)
Section Fixed Effect Yes Yes Yes Yes
N 256 256 150 106
R-squared 0.07 0.09 0.14 0.11
Notes: Standard errors are reported in parentheses; ***p < 0.01, **p < 0.05, and *p < 0.1.

Observing patterns of differential gender effects in response effect for male students than female students, that is, H2a is
to peer information, we further conducted subsample analyses supported.
by gender. The treatment effect for male students is 9.56 and
statistically significant (p = 8), whereas -8.52 and not
statistically significant (p = 0.27). Figure 2 illustrates the Evidence of the Moderating Role
heterogeneous treatment effects between male and female of Gender Composition
students. The interventions have a positive effect in moti-
vating male students to start earlier, whereas a negative effect, In this section, we report the preliminary observations of
although not statistically significant, was observed for female heterogeneous treatment effects with respect to the role of
students. Thus, peer information interventions had a stronger gender composition in a class. Specifically, we divided the

172 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Figure 2. Moderating Effects of Gender (Based on Model 2 in Table 6)

samples receiving treatments into three groups of roughly The detailed experimental procedure works as follows. At the
equal size based on the male composition in the class (i.e., beginning of the semester, students were randomly assigned
Section_Male_Composition): Low (<42%, female-dominant), to a group of four to six members.3 This randomization
medium (45%–60%, balance), and high (>60%, male- strategy generated 39 groups across four sections, with a
dominant). As shown in Figure 3a, in a female-dominant group male composition in the range between 0% and 100%
class, the treatment effect is negative. When the gender (see the details in Table B1 of Appendix B). Before parti-
composition environment becomes more male-dominant, the cipating in the experiment (an individual assignment),
treatment effect becomes more salient. Figure 3b further students had worked together and completed several group
shows that for male students, the treatment effect increases as assignments. As shown in Table 3, students were randomly
male composition increases. These results provide pre- assigned to the control group to receive reminder messages
liminary support for H2b. without peer information or the treatment group to receive
messages with the descriptive peer information calculated
from their study group (i.e., “X% of students in your group
A potential concern with the preliminary results in Figure 3 is
have started working on the assignment”). Randomization
that confounding factors other than gender composition may
tests in Table A2 of Appendix A shows that there are no
also contribute to the observed effects, which weakens the
significant differences between the treatment group and the
causal interpretation of gender composition. In Study II control group with respect to preexperiment factors and
below, we implemented an alternative experimental design to individual traits.
create exogenous variations in gender composition and for-
mally tested the moderation effect of gender composition. Experiment results in Table 7 show that male students in
male-majority groups are more responsive to peer information
interventions and reduce procrastination (by 12.03 hours),
Study II (Group-Level Peer Information) whereas other individuals are not. This finding is consistent
with the prediction from Hypothesis H2b. Consistent with
Causal estimation of the effect of gender composition requires previous results in Study I, female students did not show a
randomization in gender composition, which is challenging, treatment effect from the interventions. We formally test the
if not impossible, at the class/section level because students H2b by pooling data from Study II and III, which allows us to
self-select to enroll into a class/section. Randomization at the test the interaction effect with a larger sample size. Empirical
group-level, however, is possible because instructors often results (Table B2 in Appendix B) show that the treatment
assign students to work on group assignments. In Study II, effect of peer information for male students in a male-
we explored an alternative design, in which students received majority group is stronger than that for male students not in a
descriptive peer information with their randomly assigned male-majority group (19.70, p = 0.06). Hence, H2b is
study group as the reference group. This experimental design supported.
creates random variations in gender composition within a
class/section, which helps avoid having to conduct the experi-
3
ments with a large number of classes, where we run into the For each class/section, we use the following procedure/constraints to assign
identification challenge caused by cross-class heterogeneity. students into groups: (1) create five-member groups, as many as possible; (2)
no group should have less than four students or more than six students.

MIS Quarterly Vol. 45 No. 1/March 2021 173


Li, Wang, & Wang/Peer Effects in Competitive Environments

(a) Effects with Respect to Male Composition (b) Effects with Respect to Male Composition by
Gender

Figure 3. Moderating Effect of Gender Composition

Table 7. Effects of Group-Level Gender Composition


Males Females
Male Majority Non-Male Majority Female Majority Non-Female
DV: StartTime Group Group Group Majority Group
3.77 38.15 59.24** 26.74
Intercept
(18.22) (20.43) (23.03) (28.67)
12.03* 14.72 0.09 -10.79
Treatment
(6.10) (13.15) (7.06) (12.62)
0.12* -0.05 0.05 0.15
StartTime_Prev
(0.06) (0.10) (0.08) (0.09)
0.17 0.02 0.01 0.14
Grade_Prev
(0.16) (0.32) (0.16) (0.23)
Section Fixed Effect Yes Yes Yes Yes
N 56 20 41 17
R-squared 0.25 0.18 0.21 0.58
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

Study III (Manipulation of Competition) Experiment results show that gender differences exist for
students in ranking-based conditions (Table 8a) but are absent
As described earlier, Study III was conducted in a different for students in non-ranking-based conditions (Table 8b).4
course without a ranking-based grading policy for the final Male students are more responsive to peer information and
letter grades. This course, with 11 sections, allowed us to start working on the assignment earlier in the ranking-based
extend the design in Study II by manipulating the competitive setting. These findings provide direct evidences that males
environment for the experimented assignment. Specifically, prefer competition more than females. Therefore, males are
students in a section were first randomly assigned into the more likely to opt into a competitive setting induced by the
competition condition (with a ranking-based grading policy) descriptive peer information they received (Gneezy et al.
or the benchmark condition (without a ranking-based grading 2003). Females are not responsive to interventions, especially
policy). Within the competition/benchmark condition, the
experimental design is identical to that of Study II, that is,
students in treatment conditions received reminder message 4
In Study III, we do not have data on students’ prior behavior and perfor-
with peer information, whereas students in control conditions mance because the experiment was conducted at the beginning of the
received reminder messages without peer information. semester.

174 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 8. Peer Information in Ranking Versus Non-Ranking Based Setting


(a) Ranking Based Setting
All
DV: StartTime Model 1 Model 2 Male Female
32.18*** 44.47*** 34.03*** 24.57
Intercept
(8.46) (10.06) (10.18) (15.00)
11.67** -0.45 17.73** -1.47
Treatment
(5.69) (9.07) (7.39) (8.71)
-18.31**
Male – – –
(8.26)
18.99*
Treatment*Male – – –
(10.59)
Section Fixed Effect Yes Yes Yes Yes
N 116 116 71 45
R-squared 0.26 0.29 0.33 0.27
(b) Non-Ranking Based Setting
All
DV: StartTime Model 1 Model 2 Male Female
42.86*** 51.95*** 37.05*** 49.55***
Intercept
(7.38) (8.65) (10.72) (9.93)
5.70 -1.82 10.26 -1.17
Treatment
(5.75) (9.37) (7.35) (9.01)
-17.33**
Male – – –
(8.36)
12.20
Treatment*Male – – –
(11.75)
Section Fixed Effect Yes Yes Yes Yes
N 102 102 65 37
R-squared 0.08 0.12 0.10 0.16
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

when the percentage value X is not sufficiently large because showing the preliminary evidence supporting H1b. Moreover,
social conformity occurs only at high prevalence level (we consistent with the results of student behavior, the inter-
further investigate this issue in the next section). ventions had differential effects on male and female students.
Peer information interventions have a stronger effect for male
students than female students (Figure 4b).
Effects of Peer Information Interven-
tions on Performance Outcome To test our hypotheses regarding the average treatment effect
of the interventions on performance outcome, we first con-
ducted regression analyses using all data combined from the
Effects of Peer Information Interventions
three studies (Model 1 in Table 9). We see that peer
on Student Grades
information interventions have a positive main effect on
In this section, we further investigate the impact of peer students learning performance (4.13, p < 0.01), which sup-
information interventions on students’ performance, that is, ports our H1b. Further, consistent with H2a, the estimation
Grade. Similarly, we first visualize some of the interesting of the interaction term of Treatment*Male is 6.33 and statis-
patterns observed using pooled data from the three studies. tically significant (Model 2 in Table 9). Therefore, the
As shown in Figure 4a, peer information interventions overall interventions have a stronger treatment effect on Grade for
have a positive average effect on performance outcomes, males than for females.

MIS Quarterly Vol. 45 No. 1/March 2021 175


Li, Wang, & Wang/Peer Effects in Competitive Environments

(a) Impact on Grade (b) Impact on Grade by Gender

Figure 4. Effects of Peer Information Interventions on Performance Outcome

Table 9. Effects on Performance Outcome


Pooled Data (Study I&II&III)
All
DV: Grade Model 1 Model 2 Male Female
83.23*** 87.86*** 79.87*** 91.64***
Intercept
(4.35) (4.63) (5.49) (7.04)
4.13*** 0.31 6.54*** 0.27
Treatment
(1.57) (2.45) (2.18) (2.18)
-6.42***
Male – – –
(2.27)
6.33**
Treatment*Male – – –
(3.19)
Section Fixed Effect Yes Yes Yes Yes
N 608 608 362 246
R-squared 0.08 0.09 0.16 0.06
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

Having observed heterogeneous treatment effects of gender, student behavior and performance outcomes. In this section,
we then conducted the analyses for male students and female we aim to establish the potential underlying mechanisms of
students separately (Table 9). Results show that the inter- the effectiveness of the interventions by conducting a media-
ventions have a positive effect on students learning perfor- tion analysis. Since there are heterogeneous treatment effects
mance only for male students (6.54), and not for female of gender, we therefore conducted the analyses for male
students (0.27), indicating a stronger treatment effect for male students and female students separately.
students.
Following the two-step approach proposed by Zhao et al.
(2010), we estimated two regression equations with StartTime
Mediating Role of Behavior Change and Grade as dependent variables for each (shown in Table
10). Specifically, male students receiving peer information
From the results of earlier sections, we have observed the started to work on the assignment 10.23 hours earlier than
positive effects of peer information interventions on both those in the control group did. The estimated coefficient of

176 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 10. The Mediating Role of Behavior Change


Males Females
Step 1 Step 2 Step 1 Step 2
StartTime Grade StartTime Grade
35.93*** 76.77*** 66.77*** 88.13***
Intercept
(8.20) (5.61) (13.24) (7.41)
10.23*** 5.66** -3.91 0.47
Treatment
(3.26) (2.20) (4.10) (2.18)
0.09** 0.05
StartTime – –
(0.04) (0.04)
Section Fixed Effect Yes Yes Yes Yes
N 362 362 246 246
R-squared 0.20 0.17 0.20 0.07
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

StartTime in Step 2 is statistically significant (0.09, p < 0.05), show that the treatment effect is stronger for males with
indicating a positive effect of reducing procrastination on poorer past behavior than those with better past behavior
performance outcomes. Hence, the mediation effect of (9.06, p < 0.10). After receiving peer information, a male
behavior improvement (StartTime) on male students’ perfor- student with poorer past behavior started to work on the
mance outcomes (Grade) is 0.92. We conducted a bootstrap assignment 9.06 hours earlier than a male with better past
analysis (10,000 samples) and found that the mediation effect behavior. Parameter estimate of the three-way interaction
is statistically significant (p < 0.05). Hence, the mediating (Model 3) is 32.54 (statistically significant, p < 0.05). That is,
role of behavior change is found for male students. That is, the effect of peer information is stronger for students with
peer information interventions have a positive effect on poorer past behavior and past performance. Hence, H3 is
reducing male students’ procrastination behavior, which leads supported.
to better performance outcomes (if a male student started 24
hours earlier to work on the assignment, he would achieve a
2.16% higher grade). However, we did not find such a Effect of the Prevalence of Peer Behavior
significant mediating effect for female students.
The effect of peer information interventions may depend on
The mediation analysis above highlights that behavior how prevalent certain behavior is observed among peers (Sun
improvement serves as the underlying mechanism for perfor- et al. 2019).5 To investigate the potential heterogeneous
mance improvement. effects with respect to different percentage X, we created two
dummy variables: StartPercent_Low (1 if start percent < 50%
and 0 otherwise) and StartPercent_High (1 if start percent $
Additional Analyses and Robustness Checks 50% and 0 otherwise). Empirical results (Table C2 in Appen-
dix C) show that for females, the effect of peer information is
Past Behavior and Performance not significant (p = 0.36 and p = 0.18 for low and high per-
centages, respectively), although we observe some trend that
We further explore how students’ past procrastination disclosing a low percentage X may have a negative effect (not
behavior and performance affect male students’ responses to statistically significant, p > 0.10). For males, the effect of
peer information interventions. We created two dummy peer information interventions is positive and statistically
variables: (1) StartLate_Prev, equal to 1 if a student started significant (for either low or high percentages), and the effect
working on the previous assignment later than at least half of is stronger when the percentage X is high.
the other students, and equal to 0 otherwise, and
(2) GradeLow_Prev, equal to 1 if a student’s grade from the However, the prevalence of peer behavior above (Study I, II,
previous assignment was lower than at least half of the other and III) is not (strictly) exogenous. To ensure strictly exo-
students, and equal to 0 otherwise. genous prevalence levels, in Study IV we utilized an alterna-

To test Hypothesis H3, we analyzed the combined data from 5


Table C1 in Appendix C shows the distribution of the start percentages (i.e.,
three studies (results in Table 11 below). Results in Model 1 X used in our treatment messages).

MIS Quarterly Vol. 45 No. 1/March 2021 177


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table 11. Regression Results by Past Behavior and Performance (Males)


DV: StartTime Model 1 Model 2 Model 3
43.03*** 40.47*** 35.00***
Intercept
(8.42) (8.38) (8.88)
3.71 9.94** 12.49*
Treatment
(5.08) (4.51) (6.66)
-10.02** -5.97* 1.66
StartLate_Prev
(4.75) (3.45) (6.34)
-2.70 -1.62 13.11*
GradeLow_Prev
(3.56) (4.86) (7.10)
9.06* -5.10
Treatment * StartLate_Prev –
(4.98) (9.00)
-3.12 -20.29**
Treatment * GradeLow_Prev –
(7.19) (10.23)
-26.87***
StartLate_Prev * GradeLow_Prev – –
(9.59)
32.54**
Treatment* StartLate_Prev * GradeLow_Prev – –
(14.14)
Section Fixed Effect Yes Yes Yes
N 303 303 303
R-squared 0.17 0.17 0.20
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

tive design, in which we manipulated the prevalence level of Cross-talk with Other Subjects
peer behavior revealed to students.6 We randomly assigned
students from one single class to a control group to receive One common issue in uncontrolled field experiments is the
reminder messages without peer information or to one of the possibility of contamination, in which a subject happens to be
three treatment groups to receive messages with peer infor- exposed to multiple treatment conditions. In our setting, it is
mation at different levels (low = 21%, medium = 41%, or high likely that such contamination occurs through cross-talk with
= 61%) (Table C3 in Appendix C). other students in the class. For instance, a student in the con-
trol group happens to know the message intended for students
Figure 5 visualizes the treatment effects by different preva- in the treatment group.
lence level of peer information. We also conducted formal
analyses and summarized results in Table C5 (Appendix C). To provide additional evidence that contamination is not a
We observe differential responses by males and females. For major concern, we conducted a post-experiment survey for
females, providing peer information does not help them both Study II and Study III, in which we asked subjects if
reduce procrastination when prevalence level is low, but we they discussed the message they received with classmates.
see a positive trend for medium or high level prevalence Survey results show that the likelihood of cross-talk and
(although not statistically significant, p > 0.10). For males, receiving different messages was only 6%~8% (Table F2 in
however, providing peer information helps reduce procras- Appendix F). This result suggests that cross-talk does exist,
tination even when the prevalence level is only 21% (statis- but it does not seem to be a major concern in our setting. As
tically significant, p = 0.09). Males’ response to low preva- a robustness check, we excluded those subjects who might
lence can be because peer information induces a competitive have discussed their message with friends. The empirical
environment: males want to play ahead of the game to results remain consistent (Table F3 in Appendix F).
achieve a higher grade than their peers. These findings
suggest that females are more affected by conformity,
whereas males can also be influenced by competition. Are Reminder Messages Alone Effective?

6 In Study III, we also look into if reminder messages without


This type of manipulation may be subject to ethical concern and should not
be the standard design in large-scale interventions. peer information would help reduce student procrastination.

178 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

(a) Impact of Reminder on Start Time (b) Impact of Reminder by Gender

Figure 6. No Message Versus Reminder Messages

We compared the StartTime for students received reminder competitive settings, males are more likely to respond to peer
messages without peer information and those received no information interventions than females. Such differential
messages (Figure 6). Results in Table D1 (Appendix D) show effects are even more salient in a male-dominated environ-
that reminder messages without peer information do not have ment. These results are consistent with the theory of gender
a significant effect in reducing procrastination for either male differences in competitive environments (Gneezy et al. 2003),
or female students. This is not surprising because many LMS which predicts a significant increase in performance for males
already have some sort of reminder features built into the as the competitiveness of the environment increases, but not
system. For instance, when students log into their Canvas for females. Males have been found to exhibit significantly
account, Canvas shows students a “To Do” list along with the more aggressiveness than females (Brooks 1982). Further-
deadline for each item. more, males prefer competition more than females. There-
fore, males are more likely to opt into a competitive setting
induced by the descriptive peer information they received
(Gneezy et al. 2003).
Discussion and Conclusion
To further dig into the competition effects induced by peer
Social Conformity or Social Comparison? information, we pool the findings across multiple studies for
more in-depth discussions. The empirical results suggest that
The literature has suggested that peer information inter- peer information motivates behavior change by inducing
ventions influence individual behavior because people may comparison concerns, which enhances the perceived compe-
choose to comply with peers (Schultz et al. 2007). However, titiveness of the environment (Gneezy et al. 2003). Gender
social conformity alone would predict that females are more differences observed in our experiments can be explained by
responsive to interventions, which contradicts the empirical the differential responses to competition between males and
findings from our study. We demonstrate that such incon- females. The direct support for the competition argument
sistency is due to the different contexts under which previous comes from Study III, in which we manipulated the compe-
studies were conducted. Social conformity is more salient in tition conditions. Compared with females, males are more
noncompetitive settings, whereas social comparison can responsive to interventions in ranking based conditions.
dominate in competitive environments. Prior studies have However, we do not observe such gender differences in non-
focused on noncompetitive settings, in which an individual’s ranking based conditions. Another direct evidence comes
performance does not interfere with others’ performance. from Study IV, in which we manipulated the prevalence of
Therefore, the primary effects of peer information interven- peer behaviors. For males, however, providing peer informa-
tions come from social conformity. In our competitive class- tion helps reduce procrastination even when the prevalence
room setting, however, students may perceive a certain level level is low (21%), which cannot be explained by social con-
of competition with their classmates and providing peer formity. Males’ response to low prevalence can be because
information may induce comparison concerns. Therefore, the peer information introduces comparison concerns and compe-
primary effects of peer information interventions may come tition effects: males want to play ahead of the game to
from social comparison, rather than social conformity. In achieve a higher grade than their peers do.

MIS Quarterly Vol. 45 No. 1/March 2021 179


Li, Wang, & Wang/Peer Effects in Competitive Environments

Theoretical Implications require significant changes to the current curriculum design


or impose extra responsibilities on the teachers. This research
Prior studies have focused on peer information interventions evaluates a data-driven approach to reduce student procras-
in noncompetitive environments, in which an individual’s tination via digital interventions. Information systems have
performance does not interfere with others’ performance. We led to an aggregation of detailed data on individuals’ behavior
demonstrate that social conformity is more salient in non- trajectory and outcomes. Our research demonstrates how to
competitive settings, whereas social comparison can dominate analyze these data to study individual behavior, understand
in competitive environments. This research complements to what behavioral patterns are associated with superior or
the literature by investigating peer information interventions unsatisfactory outcomes, and ultimately, to improve
in environments with some degree of competition. In contrast individual performance by targeted interventions.
to the prediction from the social conformity theory, our results
show that males are actually more likely to respond to peer The findings from this study have implications for the optimal
information interventions. Our results suggest males are more design of data-driven, targeted interventions to promote
strongly motivated because they are more competitive- healthy behaviors and outcomes. Providing peer information
oriented (Gneezy et al. 2003). Our research highlights the may help reduce procrastination for college students, but the
role of environmental contexts in peer information interven- effect depends on individual characteristics and the contextual
tions. environment. To achieve effective interventions, educators
may consider individuals’ gender, behavior trajectory, past
This research provides insights into the mixed results in the performance, and gender composition in the class. For
literature about peer information interventions (Schultz et al. instance, peer information interventions may target males in
2007). Our research demonstrates that peer information inter- a male-majority environment because these individuals
ventions are moderated by individual characteristics and benefit more from such treatments. The peer information
contextual variables. Individual differences such as gender interventions artifact we developed for LMS can auto-
and behavior trajectory play a significant role in an indi- matically extract real-time information from Canvas’ system
vidual’s responses to peer information interventions. We also log to implement targeted interventions.
find that peer information interventions motivate behavior
change, which in turn leads to a better performance outcome.
This finding adds to the literature by establishing the media- Limitations and Future Research
tion link between digital interventions and performance
improvement. Our research, therefore, adds to the broader One important issue in peer information interventions is that
literature of digital interventions through social influence and the prevalence of peer behavior (“X% of students have
peer effects by highlighting the possibility of targeted inter- started…”) depends on the timing of interventions and other
ventions based on individual characteristics and the factors. It is likely that the percentage X is small at the time
contextual environment (Lu et al. 2016). of interventions. For males, our research shows that
providing peer information is beneficial even when the
This research also points to the dark side of peer information percentage X is relatively low. However, disclosing a low
interventions: widening the gender gap. The literature has percentage to females may backfire. To address the potential
documented the gender and performance gap in competitive negative effect of low percentage value, future research may
settings. Our research suggests that peer information inter- experiment with alternative framing of the message, for
ventions can widen such gender gap because providing peer example, by hiding the percentage information in the message
information enhances the competitiveness of the environment. to females: “Students in your class/group have already
In competitive environments, females are at a greater disad- started….” Future research may also leverage cross-group
vantage due to peer information interventions. tournament, in which instructors disclose information about
other groups in the class (“X % of other study groups have
started…”). Such a design may cultivate a more cooperative
Practical Implications environment within a group but promote competition between
groups. It is likely that gender and gender composition may
This research has implications for the design of digital inter- play a role in such settings (Delfgaauw et al. 2013). In
ventions. Prior research has discussed how alternative addition, it is likely that some students delay on purpose.
pedagogical designs such as imposing multiple intermediate Future research may investigate whether peer information
deadlines may reduce student procrastination (Ariely and interventions are beneficial to these students. Finally, we
Wertenbroch 2002; Lamwers and Jazwinski 1989). Such have conducted field experiments on multiple business
interventions, however, are costly to implement because they courses in the classroom setting, but future research can

180 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

benefit from expanding the scale and scope of the experiments Dahlstrom, E., Brooks, D. C., and Bichsel, J. 2014. “The Current
to test the generalizability of our findings to various contexts. Ecosystem of Learning Management Systems in Higher
For example, we may expand the study to different disciplines Education: Student, Faculty, and IT Perspectives,” EDUCAUSE
and different levels of courses and MOOC platforms. Future Research, September 17.
Delfgaauw, J., Dur, R., Sol, J., and Verbeke, W. 2013. “Tourna-
research may also look at the effects of peer information in
ment Incentives in the Field: Gender Differences in the
other business settings, such as workplaces.
Workplace,” Journal of Labor Economics (31:2), pp. 305-326.
Dewan, S., Ho, Y.-J., and Ramaprasad, J. 2017. “Popularity or
Proximity: Characterizing the Nature of Social Influence in an
Acknowledgments Online Music Community,” Information Systems Research
(28:1), pp. 117-136.
The authors thank the senior editor, the associate editor, and the Dufwenberg, M., and Muren, A. 2006. “Gender Composition in
reviewers for a highly constructive and developmental review Teams,” Journal of Economic Behavior and Organization (61:1),
process, which has substantially improved the manuscript. The pp. 50-54.
authors also thank all instructors at Boston College and the Univer- Ferrari, J. R. 1992. “Procrastinators and Perfect Behavior: An
sity of Delaware who helped set up the experiments in their classes. Exploratory Factor Analysis of Self-Presentation, Self-
Awareness, and Self-Handicapping Components,” Journal of
Research in Personality (26:1), pp. 75-84.
Frank, R. H., Gilovich, T., and Regan, D. T. 1993. “Does Studying
References
Economics Inhibit Cooperation?,” The Journal of Economic
Perspectives (7:2), pp. 159-171.
Ahuja, M. K., and Thatcher, J. B. 2005. “Moving beyond
Garcia, S. M., Tor, A., and Schiff, T. M. 2013. “The Psychology
Intentions and toward the Theory of Trying: Effects of Work
of Competition: A Social Comparison Perspective,” Perspectives
Environment and Gender on Post-Adoption Information Tech-
on Psychological Science (8:6), pp. 634-650.
nology Use,” MIS Quarterly (29:3), pp. 427-459.
Gneezy, U., Niederle, M., and Rustichini, A. 2003. “Performance
Apesteguia, J., Azmat, G., and Iriberri, N. 2012. “The Impact of
in Competitive Environments: Gender Differences,” Quarterly
Gender Composition on Team Performance and Decision
Journal of Economics (118:3), pp. 1049-1074.
Making: Evidence from the Field,” Management Science (58:1),
Gordon, F. M., Welch, K. R., Offringa, G., and Katz, N. 2000.
pp. 78-93.
“The Complexity of Social Outcomes from Cooperative, Com-
Ariely, D., and Wertenbroch, K. 2002. “Procrastination, Deadlines,
petitive, and Individualistic Reward Systems,” Social Justice
and Performance: Self-Control by Precommitment,” Psycho-
Research (13:3), pp. 237-269.
logical Science (13:3), pp. 219-224.
Granfield, R. 2005. “Alcohol Use in College: Limitations on the
Arvey, R. D., Rotundo, M., Johnson, W., Zhang, Z., and McGue, M.
Transformation of Social Norms,” Addiction Research & Theory
2006. “The Determinants of Leadership Role Occupancy:
(13:3), pp. 281-292.
Genetic and Personality Factors,” The Leadership Quarterly
Gupta, N. D., Poulsen, A., and Villeval, M. C. 2013. “Gender
(17:1), pp. 1-20.
Matching and Competitiveness: Experimental Evidence,” Econ-
Baek, J., and Shore, J. 2020. “Forum Size and Content Contribu-
omic Inquiry (51:1), pp. 816-835.
tion per person: A Field Experiment,” Management Science
Hashim, M. J., Kannan, K. N., and Maximiano, S. 2017. “Informa-
(66:12), pp. 5906-5924.
tion Feedback, Targeting, and Coordination: An Experimental
Beshears, J., Choi, J. J., Laibson, D., Madrian, B. C., and Milkman,
Study,” Information Systems Research (28:2), pp. 289-308.
K. L. 2015. “The Effect of Providing Peer Information on
Hevner, A. R., March, S. T., Park, J., and Ram, S. 2004. “Design
Retirement Savings Decisions,” The Journal of Finance (70:3),
Science in Information Systems Research,” MIS Quarterly
pp. 1161-1201.
(28:1), pp. 75-105.
Brooks, V. R. 1982. “Sex Differences in Student Dominance
Huang, N., Burtch, G., Gu, B., Hong, Y., Liang, C., Wang, K.,
Behavior in Female and Male Professors’ Classrooms,” Sex
Fu, D., and Yang, B. 2019. “Motivating User-Generated Con-
Roles: A Journal of Research (8:7), pp. 683-690.
tent with Performance Feedback: Evidence from Randomized
Cerezo, R., Esteban, M., Sánchez-Santillán, M., and Núñez, J. C.
Field Experiments,” Management Science (65:1), pp. 327-345.
2017. “Procrastinating Behavior in Computer-Based Learning Huang, N., Zhang, J., Burtch, G., Li, X., and Chen, P. 2018.
Environments to Predict Performance: A Case Study in “Combating Procrastination on MOOCs via Optimal Calls-to-
Moodle,” Frontiers in Psychology (8), Article 1403. Action: Evidence from a Field Experiment,” in Academy of
Cialdini, R. B., and Goldstein, N. J. 2004. “Social Influence: Management Annual Meeting Proceedings, p. 14171.
Compliance and Conformity,” Annual Review of Psychology Karatas, H. 2015. “Correlation among Academic Procrastination,
(55), pp. 591-621. Personality Traits, and Academic Achievement,” Anthropologist
Cialdini, R. B., and Trost, M. R. 1998. Social Influence: Social (20:1,2), pp. 243-255.
Norms, Conformity and Compliance, New York: McGraw-Hill. Kashima, Y., Yamaguchi, S., Kim, U., Gelfand, M. J., and Yuki, M.
Clapp, J. D., Lange, J. E., Russell, C., Shillington, A., and Voas, 1995. “Culture, Gender, and Self: A Perspective from
R. B. 2003. “A Failed Norms Social Marketing Campaign,” Individualism–Collectivism Research,” Journal of Personality
Journal of Studies on Alcohol (64:3), pp. 409-414. and Social Psychology (69:5), pp. 925-937.

MIS Quarterly Vol. 45 No. 1/March 2021 181


Li, Wang, & Wang/Peer Effects in Competitive Environments

Kim, K. R., and Seo, E. H. 2015. “The Relationship between Perrin, C. J., Miller, N., Haberlin, A. T., Ivy, J. W., Meindl, J. N.,
Procrastination and Academic Performance: A Meta-Analysis,” and Neef, N. A. 2010. “Measuring and Reducing College
Personality and Individual Differences (82), pp. 26-33. Students’ Procrastination,” Journal of Applied Behavior Analysis
Lam, P., Lo, J., Lee, J., and McNaught, C. 2012. “Evaluations of (44:3), pp. 463-474.
Online Learning Activities Based on LMS Logs,” in Virtual Pychyl, T. A., and Sirois, F. M. 2016. “Procrastination, Emotion
Learning Environments: Concepts, Methodologies, Tools and Regulation, and Well-Being,” Chapter 8 in Procrastination,
Applications, Hershey, PA: IGI Global, pp. 1767-1784. Health, and Well-Being, Cambridge, MA: Academic Press, pp.
Lamwers, L. L., and Jazwinski, C. H. 1989. “A Comparison of 163-188.
Three Strategies to Reduce Student Procrastination in PSI,” Qiu, L., and Kumar, S. 2017. “Understanding Voluntary Knowl-
Teaching of Psychology (16:1), pp. 8-12. edge Provision and Content Contribution through a Social-
Lavy, V., and Schlosser, A. 2011. “Mechanisms and Impacts of Media-Based Prediction Market: A Field Experiment,”
Gender Peer Effects at School,” American Economic Journal: Information Systems Research (28:3), pp. 529-546.
Applied Economics (3:2), pp. 1-33. Rotenstein, A., Davis, H. Z., and Tatum, L. 2009. “Early Birds
Lee, S., Turner, L. J., Woo, S., and Kim, K. 2014. “All or Nothing? Versus Just-in-Timers: The Effect of Procrastination on
The Impact of School and Classroom Gender Composition on Academic Performance of Accounting Students,” Journal of
Effort and Academic Achievement,” Working Paper, National Accounting Education (27:4), pp. 223-232.
Bureau of Economic Research. Ruble, D. N., Martin, C. L., and Berenbaum, S. A. 2006. “Gender
Lee, Y.-J., Hosanagar, K., and Tan, Y. 2015. “Do I Follow My Development,” in Handbook of Child Psychology, New York:
Friends or the Crowd? Information Cascades in Online Movie John Wiley & Sons.
Ratings,” Management Science (61:9), pp. 2241-2258.
Schouwenburg, H. C., Lay, C. H., Pychyl, T. A., and Ferrari, J. R.
Leland, J., Schneider, M., and Wilcox, N. 2019. “Minimal Frames
2004. Counseling the Procrastinator in Academic Settings,
and Transparent Frames for Risk, Time, and Uncertainty,”
Washington, DC: American Psychological Association.
Management Science (65:9), pp. 4318-4335.
Schultz, P. W., Nolan, J. M., Cialdini, R. B., Goldstein, N. J., and
Lewis, M. A., and Neighbors, C. 2004. “Gender-Specific Misper-
Griskevicius, V. 2007. “The Constructive, Destructive, and
ceptions of College Student Drinking Norms,” Psychology of
Reconstructive Power of Social Norms,” Psychological Science
Addictive Behaviors (18:4), pp. 334-339.
(18:5), pp. 429-434.
Liu, D., Santhanam, R., and Webster, J. 2017. “Toward Mean-
ingful Engagement: A Framework for Design and Research of Silva, A., and John, P. 2017. “Social Norms Don’t Always Work:
Gamified Information Systems,” MIS Quarterly (41:4), pp. An Experiment to Encourage More Efficient Fees Collection for
1011-1034. Students,” PloS One (12:5), p. e0177354.
Lu, X., Zhao, X., and Xue, L. 2016. “Is Combining Contextual and Steel, P. 2007. “The Nature of Procrastination: A Meta-Analytic
Behavioral Targeting Strategies Effective in Online Adver- and Theoretical Review of Quintessential Self-Regulatory
tising?,” ACM Transactions on Management Information Systems Failure,” Psychological Bulletin (133:1), pp. 65-94.
(7:1), pp. 1-20. Steel, P., Brothen, T., and Wambach, C. 2001. “Procrastination and
Michael, J. 1991. “A Behavioral Perspective on College Personality, Performance, and Mood,” Personality and Individual
Teaching,” The Behavior Analyst (14:2), pp. 229-239. Differences (30:1), pp. 95-106.
Michinov, N., Brunot, S., Le Bohec, O., Juhel, J., and Delaval, M. Stocks, E. L., Lishner, D. A., and Decker, S. K. 2009. “Altruism or
2011. “Procrastination, Participation, and Performance in Online Psychological Escape: Why Does Empathy Promote Prosocial
Learning Environments,” Computers & Education (56:1), pp. Behavior?,” European Journal of Social Psychology (39:5), pp.
243-252. 649-665.
O’Brien, W. K. 2002. “Applying the Transtheoretical Model to StudyMode. 2014. “Eighty-Seven Percent of High School and
Academic Procrastination,” Unpublished Doctoral Dissertation, College Students Are Self-Proclaimed Procrastinators,”
University of Houston. PRNewswire, May 27.
Özer, B. U., Demir, A., and Ferrari, J. R. 2009. “Exploring Sun, M., Zhang, X., and Zhu, F. 2019. “U-Shaped Conformity in
Academic Procrastination among Turkish Students: Possible Online Social Networks,” Marketing Science (38:3), pp.
Gender Differences in Prevalence and Reasons,” The Journal of 461-480.
Social Psychology (149:2), pp. 241-257. van Eerde, W., and Klingsieck, K. B. 2018. “Overcoming Procras-
Peng, C.-H., Yin, D., Wei, C.-P., and Zhang, H. 2017. “Impact of tination? A Meta-Analysis of Intervention Studies,” Educational
Perspective Taking on Reviewer Behavior: A Multi-Method Research Review (25), pp. 73-85.
Exploration,” in Proceedings of the 38th International Conference Wang, L., Gunasti, K., Gopal, R., Shankar, R., and Pancras, J.
on Information Systems, Seoul. 2017. “The Impact of Gamification on Word-of-Mouth Effec-
Perkins, H. W. 2002. “Social Norms and the Prevention of Alcohol tiveness: Evidence from Foursquare,” in Proceedings of the 50th
Misuse in Collegiate Contexts,” Journal of Studies on Alcohol, Hawaii International Conference on System Sciences, Los
Supplement (14), pp. 164-172. Alamitos: IEEE Computer Society Press.
Perkins, H. W., Haines, M. P., and Rice, R. 2005. “Misperceiving Weber, J. M., Kopelman, S., and Messick, D. M. 2004. “A
the College Drinking Norm and Related Problems: A Nation- Conceptual Review of Decision Making in Social Dilemmas:
wide Study of Exposure to Prevention Information, Perceived Applying a Logic of Appropriateness,” Personality and Social
Norms and Student Alcohol Misuse,” Journal of Studies on Psychology Review (8:3), pp. 281-307.
Alcohol (66:4), pp. 470-478.

182 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Wechsler, H., Nelson, T. E., Lee, J. E., Seibring, M., Lewis, C., and Ph.D. in Information, Risk, and Operations Management from the
Keeling, R. P. 2003. “Perception and Reality: A National University of Texas at Austin. His research interests focus on
Evaluation of Social Norms Marketing Interventions to Reduce information and coordination issues in digital platforms and value
College Students’ Heavy Alcohol Use,” Journal of Studies on chains. His research has been published or is forthcoming in MIS
Alcohol (64:4), pp. 484-494. Quarterly, Management Science and Production and Operations
Werch, C. E., Pappas, D. M., Carlson, J. M., DiClemente, C. C., Management.
Chally, P. S., and Sinder, J. A. 2000. “Results of a Social Norm
Intervention to Prevent Binge Drinking among First-Year Resi- Gang Wang is an assistant professor of Management Information
dential College Students,” Journal of American College Health Systems at the Lerner College of Business and Economics at the
(49:2), pp. 85-92. University of Delaware. He received his Ph.D. in Operations and
You, J. W. 2016. “Identifying Significant Indicators Using LMS Information Management from the University of Connecticut. His
Data to Predict Course Achievement in Online Learning,” The research interests include social media, e-Business platforms, and
Internet and Higher Education (29), pp. 23-30. firm strategies in e-Markets. His research has been published in
Zhang, D. J., Allon, G., and Van Mieghem, J. A. 2017. “Does ACM Transactions on Management Information Systems and
Social Interaction Improve Learning Outcomes? Evidence from Decision Support Systems.
Field Experiments on Massive Open Online Courses,” Manu-
Harry Jiannan Wang is a professor of Management Information
facturing & Service Operations Management (19:3), pp. 347-367.
Systems at the Lerner College of Business and Economics at the
Zhao, X., Lynch Jr., J. G., and Chen, Q. 2010. “Reconsidering
University of Delaware. He received his Ph.D. in MIS from the
Baron and Kenny: Myths and Truths about Mediation Analysis,”
University of Arizona. He was the founding director of OneConnect
Journal of Consumer Research (37:2), pp. 197-206.
(NYSE: OCFT) US Research Institute and served as Vice President
of Technology for the Association for Information Systems. His
research interests involve artificial intelligence, business analytics,
About the Authors
business process management, and enterprise systems. He has
published research articles in journals such as Information Systems
Zhuoxin Li is an assistant professor of Information Systems in the
Research, Decision Support Systems, and ACM Transactions on
Carroll School of Management at Boston College. He received his
Management Information Systems.

Appendix A
Randomization Tests

Table A1. Randomization Test for Study I


Control Treatment p-value
N 136 120 –
Male 0.60 0.58 0.74
StartTime_Prev 75.78 77.50 0.46
Grade_Prev 93.13 92.48 0.43

Table A2. Randomization Test for Study II


Control Treatment p-value
N 71 63 –
Male 0.59 0.54 0.55
StartTime_Prev 82.56 74.86 0.55
Grade_Prev 85.81 88.42 0.40

MIS Quarterly Vol. 45 No. 1/March 2021 183


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table A3. Randomization Test for Study III


Non-Ranking
Ranking Conditions Conditions p-value
# Sections 6 5
# Survey Responses 144 136 –
FirstFactor 0.12 -0.13 0.41
Number of Courses 5.20 5.18 0.65
Level_PrerequisiteGrade 4.74 4.79 0.83
Interest_level 4.57 4.68 0.19
Level_PrevGPA 4.34 4.41 0.50
Note: FirstFactor measures self-reported procrastination tendency (see Appendix F for details).

Appendix B
Group-Level Summary and Results

Table B1. Summary of Descriptive Statistics of Group-level Male Proportion


Group_Male_Composition # Groups Mean SD Min Median Max
Overall 39 54.34 34.01 0 60.00 100
Section 1 7 64.29 26.92 0 60.00 83.33
Study II Section 2 7 62.29 21.21 16.67 66.67 83.33
Section 3 11 43.79 38.44 0 50.00 100
Section 4 14 53.69 38.99 0 58.33 100
Overall 147 61.27 24.68 0 66.67 100
Section 1 13 50.00 16.67 33 50.00 100
Section 2 13 57.69 24.17 33 66.67 100
Section 3 14 64.74 25.49 33 66.67 100
Section 4 13 65.38 24.02 33 66.67 100
Section 5 14 66.67 22.96 33 66.67 100
Study III
Section 6 13 66.67 24.02 33 66.67 100
Section 7 13 57.69 24.17 33 66.67 100
Section 8 14 64.74 16.72 33 66.67 100
Section 9 13 58.33 25.00 33 66.67 100
Section 10 13 70.21 28.76 33 71.05 100
Section 11 14 58.97 36.40 0 66.67 100

184 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table B2. Effects of Group-Level Gender Composition (Pooling Data)


Males Only
DV: StartTime Male Majority Group Non-Male Majority Group Combined
33.02*** 36.09** 24.48***
Intercept
(7.87) (16.43) (9.04)
19.66*** 4.06 -0.71
Treatment
(4.97) (10.29) (8.77)
12.38*
Male_majority – –
(7.08)
19.70*
Treatment*Male_majority – –
(10.26)
Section Fixed Effect Yes Yes Yes
N 109 38 147
R-squared 0.46 0.36 0.43
Notes: (1) Pooling data are from Study II and the competitive setting from Study III.
(2) Male_majority is a dummy variable, equal to 1 if the percentage of male students (Group_Male_Composition) in a group is above
60%, 0 otherwise.
(3) ***p < 0.01, **p < 0.05, and *p < 0.1.

Appendix C
Prevalence of Peer Behavior

Table C1. Summary of Descriptive Statistics of Start Percent (X) in a Reference Group
# of
Reference Group Subjects Mean SD Min Median Max
Overall 608 38.26 11.22 17 35 67
Study I Section level 256 42.32 6.26 35 48 48
Study II Group level 134 31.51 13.30 17 25 67
Study III Group level 218 37.84 0.12 25 33 67

Table C2. Effect of Start Percent of Peers (Study I&II&III)


DV: StartTime All Males Female
49.51*** 36.53*** 62.44***
Intercept
(7.27) (8.13) (13.67)
2.39 8.54*** -5.67
Treatment*StartPercent_Low
(2.58) (3.29) (4.18)
24.37*** 31.86*** 10.82
Treatment*StartPercent_High
(7.38) (9.86) (11.87)
-8.32***
Male – –
(2.66)
Section Fixed Effect Yes Yes Yes
N 608 362 246
R-squared 0.17 0.21 0.21
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

MIS Quarterly Vol. 45 No. 1/March 2021 185


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table C3. Treatment Conditions with Section-Level, Randomized Norms (Study IV)
Condition Message
Control
Assignment 3 is due in 4 days.
(No Social Info.)
Treatment 1 Assignment 3 is due in 4 days. 21% of students in your class have started working on this
(Low Prevalence) assignment.
Treatment 2 Assignment 3 is due in 4 days. 41% of students in your class have started working on this
(Medium Prevalence) assignment.
Treatment 3 Assignment 3 is due in 4 days. 61% of students in your class have started working on this
(High Prevalence) assignment.

Table C4. Randomization Test for Study IV


Control T1-Low T2-Medium T3-High p-value
N 58 60 57 53 –
Male 0.61 0.65 0.52 0.60 0.55
StartTime_Prev 70.47 75.27 72.38 84.63 0.53
Grade_Prev 96.94 93.85 95.86 96.31 0.23
Note: There were 270 students in this single class. At the time when we sent messages, 42 (16%) students had started to work on the assignment.
The remaining 228 students were randomly assigned to one of the conditions.

Table C5. Results from Randomization of Peer Behavior Prevalence (Study IV)
DV: StartTime All Males Females
24.45 14.15*** 30.01
Intercept
(15.84) (21.09) (24.18)
4.65 9.64* -3.60
Treatment1 (Low-21%)
(4.74) (5.60) (8.48)
10.44** 14.36** 4.30
Treatment2 (Medium-41%)
(4.75) (5.93) (7.84)
15.42*** 21.73*** 5.18
Treatment3 (High-61%)
(4.89) (5.91) (8.42)
-6.07*
Male – –
(3.50)
0.12*** 0.15*** 0.07
StartTime_Prev
(0.03) (0.03) (0.05)
0.002 -0.01 0.05
Grade_Prev
(0.16) (0.22) (0.25)
N 217 129 88
R-squared 0.17 0.24 0.05
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

186 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Appendix D
Effects of Interventions

Table D1. Effects of Interventions (Study III)


All
DV: StartTime Model 1 Model 2 Males Females
35.92*** 42.02*** 36.32*** 35.39***
Intercept
(6.34) (7.36) (7.65) (11.37)
0.47 6.85 -3.15 6.21
T1 (Reminder-only)
(4.11) (6.60) (5.26) (6.51)
8.98** 5.34 11.01** 4.41
T2 (Reminder and peer info)
(4.08) (6.50) (5.27) (6.40)
-8.55
Male – – –
(5.93)
-9.89
T1*Male – – –
(8.36)
5.89
T2*Male – – –
(8.30)
Section Fixed Effect Yes Yes Yes Yes
N 324 324 201 123
Notes: Standard errors are reported in parentheses; ***p < 0.01, **p < 0.05, and *p < 0.1.

Appendix E
Additional Information about the Experiment Context
Course Information

The courses experimented in this research are business core courses (undergraduate level). These courses cover various topics related to data
analytics methods using Excel, SAS, and other tools.

Grading Policies

Starting from 2017, the business school has enforced grading policies for core courses, which have stated “targets of 15% A grades at most,
35% A- or above at most, and 65% B+ or above at most.” These standards applied to students in all core courses, including those in Study I,
Study II, and Study IV.

For Study III, we use a different course without a ranking-based grading policy for the final letter grades. Instead, we create competitive
conditions by applying a ranking-based grading policy for the experimented assignment only. When the assignment was released, students
from a section assigned into competitive conditions were notified by instructors about the specific policy. That is, a student’ score for the
experimented assignment will be adjusted according to his/her ranking of the raw score within a class (Table E1).

MIS Quarterly Vol. 45 No. 1/March 2021 187


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table E1. Grading Policy in Competitive Conditions


Ranking of Raw Scores (High to Low)
Within a Class Assignment Grade Adjusted Assignment Score
Top 15% A 100
Top 16~35% A- 92
Top 36%~60% B+ 82
Top 61%~75% B 76
Top 76%~83% B- 68
Top 84%~89% C+ 62
Top 90%~93% C 56
Top 94%~97% C- 48
>98% D+ 42

Example Questions
Instructions

A telecom firm was concerned about the amount of churn (customers’ discontinuation of subscription) it was facing. The information about
the customers on the observation date and their decision to churn between 31 and 60 days after observation date are recorded in a SAS dataset
called “cc.Cell”.

The first 10 observations of the data are

The dataset has the following variables:

(1) Customer_id: Unique customer identification number


(2) Churn (dependent variable): A dummy variable indicating the instance of churn between 31-60 days after observation date
(3) Avgmou: Average monthly minutes of use over the life of the customer
(4) Months: Total number of months in service
(5) Eqpdays: Number of days (age) of current equipment
(6) Rent: A dummy variable indicating whether the customer rent (Rent = 1) or own the equipment (Rent = 0)
(7) Retcall: A dummy variable indicating whether the customer has received a retention call in the past 365 days (Retcall = 1) or not (Retcall
= 0).
(8) Creditcard: A dummy variable indicating whether the customer owns a credit card (Creditcard = 1) or not (Creditcard = 0)

188 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Questions

Run a logistic regression, which models the probability that a customer would churn (Churn = 1) using the explanatory variables: (3) Avgmou;
(4) Months; (5) Eqpdays, (6) Rent; (7) Retcall; and (8) Creditcard.

Please answer the following questions:

a. What is the -2 Log-likelihood and AIC value of the model?


b. How much better is this model than the null model with no covariates (that is a model with only intercepts) in terms of -2LogL and AIC?
c. Which variables are significant?
d. Quantify the impact of “rent” on the odds of churning using percentage change. (Answer is supposed to be structured as: Compared to those
who own the equipment (rent = 0), odds of buying for those who rent the equipment (rent = 1) is XX% lower/higher.)

Appendix F
Survey
Validation of StartTime as a Measure of Procrastination

In this appendix, we provide more support on the validity of using StartTime as a measure of procrastination. Specifically, we followed a
procedure used in prior studies and evaluated the correlation between self-reported procrastination tendency and observed behavior (Michinov
et al. 2011; Steel et al. 2001).

At the beginning of the semester for Study III, all students received a questionnaire regarding their basic information (e.g., cumulative GPA,
grade of the prerequisite course, and degree of interest in the course) and procrastination tendency. A six-item procrastination scale was adopted
and adjusted according to our classroom context (Michinov et al. 2011; Steel et al. 2001). Items include “I postpone starting in on things I don’t
like to do,” “When I have a deadline, I wait till the last minute,” “I promise myself I’ll do something then drag my feet,” “Whenever I make
a plan of action, I follow it,” “Even though I hate myself if I don’t get started, it doesn’t get my going,” “I always finish important jobs with
time to spare.” Students respond on a five-point scale ranging from 1 (strongly disagree) to 5 (strongly agree). We conducted a principal
component analysis on the six items, resulting the first component explaining 48.12% of the total variance, which is comparable to the literature
(Michinov et al. 2011). We used the value of the first component (FirstFactor) as a measure of self-reported procrastination tendency.

We calculated the correlation between self-reported procrastination tendency (FirstFactor) and observed StartTime, leading to -0.35, which
is statistically significant (p < 0.001), which is comparable to that in the literature (Steel et al. 2001).

We further regressed the self-reported procrastination tendency (FirstFactor) on StartTime and other control variables. The results (Table F1)
show that observed behavior (StartTime) is a valid indicator for procrastination.

MIS Quarterly Vol. 45 No. 1/March 2021 189


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table F1. StartTime and Self-Reported Procrastination Tendency


DV: FirstFactor Model 1 Model 2 Model 3
0.68* 3.76*** 4.07***
Intercept
(0.32) (0.94) (0.89)
-0.01*** -0.01***
StartTime –
(0.002) (0.002)
-0.05 -0.14
Male –
(0.21) (0.20)
-0.17* -0.12
CumulativeGPA –
(0.09) (0.09)
-0.08 -0.04
Grade_ Prerequisite –
(0.08) (0.07)
-0.60*** -0.06***
Interest_in_Course –
(0.17) (0.16)
Section Fixed Effect Yes Yes Yes
N 280 280 280
R-squared 0.15 0.11 0.21
Notes: (1) About 60% of students filled out the questionnaire.
(2) ***p < 0.01, **p < 0.05, and *p < 0.1.

In sum, to establish the validity of our behavioral measure of procrastination (StartTime), we conducted a questionnaire, which yielded the self-
reported procrastination tendency. Our analyses and results validated that our behavioral measure (StartTime) captures student procrastination.

Survey on Potential Cross-Talk

In study II, we conducted a post-experiment survey with students (four sections across two courses). The survey questions related to cross-talk
are as follows:

• Please list the name(s) of your classmate(s) who you interacted with the most
• Have you discussed the message you got with your classmate? (Yes or No)

In Study III, we asked two questions related to cross-talk as follows:

• Did you discuss the message you received with anyone?


• If yes, please list their full names.

From Table F2, the response rates for both studies are around 60%, which is higher than that of a typical survey of 30~40%. We looked into
the responses and found that those non-responders are less active students in the class (e.g., absent more frequently). In other words, non-
responders are less likely to be cross-talkers because they are less likely to meet other students in the class. The imperfect response rate would
not be very likely to affect the experimental results.

Table F2. Survey Summary of Cross-Talk


Study II Study III
Number(Percent) Number(Percent)
Number of responses (response rate) 147 (59%) 135 (62%)
Number (Percentage) of students who discussed with others 26 (17%) 19 (14%)
Number (Percentage) of students who discussed and their messages
9 (6%) 11 (8%)
different from the classmates who they interacted with
Note: For the response rate of Study III, we only consider students who received messages.

190 MIS Quarterly Vol. 45 No. 1/March 2021


Li, Wang, & Wang/Peer Effects in Competitive Environments

Table F3. Regression Results after Excluding Cross-Talk (Study II&III)


Excluding Students Who Discussed Excluding Students Who Discussed and
DV: StartTime with Others Messages Different
52.51*** 53.44***
Intercept
(7.39) (7.33)
1.45 1.16
Treatment
(4.99) (4.85)
-15.48*** -16.62***
Male
(4.54) (4.45)
11.69* 11.59*
Treatment*Male
(6.38) (6.21)
Section Fixed Effect Yes Yes
N 307 332
R-squared 0.30 0.29
Notes: ***p < 0.01, **p < 0.05, and *p < 0.1.

MIS Quarterly Vol. 45 No. 1/March 2021 191


Copyright of MIS Quarterly is the property of MIS Quarterly and its content may not be
copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

You might also like