Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Australasian Journal of Educational Technology, 2023, 39(6).
1
Uncovering socio-temporal dynamics in online discussions: An
event-based approach
Bodong Chen
Graduate School of Education, University of Pennsylvania, Philadelphia, United States of America;
College of Education and Human Development, University of Minnesota, Minneapolis, United States of
America
Oleksandra Poquet
School of Social Sciences and Technology, Technical University Munich, Munich, Germany; Education
Futures, University of South Australia, Adelaide, Australia
Online discussions are widely adopted in higher education to promote student interaction.
However, prior research on online discussions falls short to estimate the effect of multiple
factors collectively shape student interaction in online discussion activities. In this study, we
applied a dynamic network analysis approach named relational event modelling to a data
set from an online course where students participated in weekly discussion activities. In the
relational event models, we incorporated multiple factors including participant
characteristics, network formation mechanisms and immediate participation shifts. Results
indicated that the instructor was more likely to initiate interactions but less likely to receive
responses. Popularity, activity and familiarity established in prior relational events
positively affected future events. Immediate participation shifts such as local popularity,
immediate reciprocation and activity bursts also played a positive role. The study highlights
the importance of considering multiple factors when examining online discussions,
demonstrates the utility of relational event modelling for analysing online interaction and
contributes empirical insights into student interaction in online discussions.
Implications for practice or policy:
• Supporting online discussions in college classrooms requires instructors to consider
multiple actors including pedagogical designs, technological affordances, learner
characteristics and social dynamics.
• Educators could go beyond simply counting student posts to paying attention to how
students interact at a micro level.
• Educators and instructional designers could pay attention to socio-temporal dynamics
in online discussions and evaluate whether emerging dynamics in a particular course
are desirable and conducive to student learning.
Keywords: online learning, online discussions, network analysis, relational event modelling
(REM), temporal analysis
Introduction
Asynchronous online discussions are widely adopted to promote student interaction at all levels of
education. Online discussions can enrich student learning, sustain participation and support under-served
students, as shown by previous studies (Jo et al., 2017; Rakovic et al., 2020; Zheng & Warschauer, 2015).
However, productive online discussions do not occur automatically. To effectively scaffold discussions
through proper and well-timed interventions, one needs to understand the factors that propel discussion
activities.
Unfortunately, explanatory models of online discussions remain rare. It is well understood that
pedagogical designs and technological affordances would shape online learning discussions (Guzdial &
Turns, 2000). Research also shows that other factors such as learner characteristics and social dynamics
may influence online discussions (B. Chen & Huang, 2019; Vaquero & Cebrian, 2013). However, holistically
Australasian Journal of Educational Technology, 2023, 39(6).
2
considering these factors – pedagogical, technological, social, temporal – when modelling online
discussions to explain how they develop, remains a challenge.
This study responds to the research gap by tackling the following question: How do multiple factors –
social, cognitive, temporal – collectively shape social interaction among peers in asynchronous online
discussions? To this end, we seriously considered the temporal dimension of online discussion and
conceptualised social interaction in online discussions as a complex, dynamic phenomenon that is shaped
by these many factors in tandem. To properly capture temporal dynamics, we adopted a network analysis
method named relational event modelling (REM), which is sensitive to both the timing and sequence of
events, and applied it to a data set of online discussion activities. In the following sections, we elaborate
on factors shaping social interaction in online discussions. We then introduce REM and report the findings.
We conclude by discussing the study’s conceptual, methodological and practical implications.
Related work
Factors shaping interaction in asynchronous online discussions
Social participation and interaction in online discussions are shaped by diverse factors including
pedagogical design, technological affordances, learner characteristics and social dynamics.
When it comes to pedagogical factors shaping online discussions, various elements of teacher decisions
and behaviours can come into play, including discussion prompts, task structures, instructor feedback and
instructor presence. In a teacher-centred class, the interactions would naturally centre on the teacher. In
contrast, pedagogical models that emphasise student agency are conducive to patterns in which the
teacher would be in the periphery within class communication (J. Zhang et al., 2009). Besides the activity
structure, the task environment of a discussion activity also shapes how students respond to each other.
An instructor could ask students to share news articles, critique course readings, provide peer feedback
or construct shared knowledge, leading to different discussion behaviours and outcomes (Lee & Recker,
2021). In addition, the instructor’s participation in the discussion also matters as well as students’
perception of the instructor’s presence (Cho & Kim, 2013; Mullen & Tallent-Runnels, 2006).
Technological configurations of online discussion environments also impact learner activity in online
discussions. Technology designs that direct user attention away from unread notes can hasten the death
of discussion threads and throttle inactive threads from being reactivated (Hewitt, 2005). Specialised
discussion environments such as Knowledge Forum may have encoded strong pedagogical viewpoints on
social interaction (Scardamalia & Bereiter, 2008). In contrast, general discussion tools such as social
network sites may offer technological features that serve educational purposes (Ractham et al., 2012)
while also potentially distracting student learning as social network sites prioritise personal profiles and
social ties over subject matter-focused dialogues.
Interaction in online discussions is also shaped by learners’ knowledge, skills and dispositions. For
example, learners’ individual motivation and goal orientation influence both the quantity and quality of
their contributions (Cho & Kim, 2013). Some learners prefer lurking over posting, especially when English
is not their first language (Shafie et al., 2016). Gender differences in online discussions have also been
observed, with female students overall writing more messages than males but less so in mixed groups
(Cho et al., 2022). These personal differences manifest in different levels of participation or post intensity,
often measured by the count of forum posts.
Sociological factors prominent in human society also play significant roles in online discussions. For
example, students would be more likely to respond to someone with a “superior” status, such as being
the course instructor or being “popular” in a class (Vaquero & Cebrian, 2013). Reciprocity, another
common social behaviour, can also come into play as peers tend to reciprocate responses (Cheung et al.,
2008). Peer pressure generated by active participants in online discussions can drive more students to
Australasian Journal of Educational Technology, 2023, 39(6).
3
participate (X. Zhang, 2023). These social factors are not only related to the intensity of participation but
can also be closely interrelated to the quality of contributions (Galikyan et al., 2021).
Taken together, social interaction in online discussions is collectively shaped by pedagogical,
technological, learner and social factors. Of course, what is said within a discussion or how learners
respond to each other as reflected within their discourse is also an important factor of how online
discussions unfold. The cognitive content of discussion posts has been examined in prior studies to
support its importance (G. Chen et al., 2020; Zingaro & Oztok, 2012). However, it can be argued that the
content of the discussion is a function of pedagogical and technological affordances of the online
discussion environment as well as individual characteristics that manifest through social factors during
the process of interaction (B. Chen & Chen, 2023). In other words, what and how learners are talking
about can be actually explained by these factors, and therefore, they can be of primary importance in
explaining how online discussions unfold.
Consideration of time in online discussions
It is of great importance to consider a temporal dimension in modelling online discussions as they are
generated by the interplay of the factors discussed above. Although prior studies on online discussions
have uncovered various factors contributing to online interactions, the temporal dimension often gets
side-lined in these studies (Knight et al., 2017). Traditional research methods rely on the coding and
counting of discussion activities (Jeong et al., 2014), by and large ignoring the pace, order and duration of
learning events. In rare cases when temporal factors are considered, the analysis of time often lacks in
granularity and includes only factors such as the learner’s perception of time, the instructor’s temporal
expectations and high-level summative information about temporal participation patterns. For instance,
B. Chen and Huang (2019) used temporal information about student participation to divide students into
two categories (early-starters and late-arrivers) and found early-starters tend to occupy central positions
in their interaction network. Such analysis extracts high-level temporal information for traditional
methods (e.g., group comparisons) and falls short in uncovering granular temporal patterns in student
interactions. To consider the temporal dimension of online discussions, researchers need to interrogate
the conceptualisation of time and order and seek analytical methods that are both congruent with the
conceptualisation and capable of uncovering the turn-by-turn unfolding of interactional events.
In this paper, we argue for modelling social interactions as discrete relational events that unfold over time
with the occurrence of each event depending on various factors. A relational event can be defined as a
“discrete event generated by a social actor and directed toward one or more targets” (Butts, 2008, p.
159). A relational state between two social actors (e.g., between two students in a course), in contrast,
can be considered as a function of a series of relational events between these actors. Prior studies have
often considered student interaction in online forums as network ties. However, a tie in a network in
essence represents a relational state of “being replied to,” such as Student A has replied to Student B five
times while B has replied to A three times in a course. In this case, relational events such as replies are
aggregated to depict a state. Although investigating social interaction as relational states allows
researchers to discern interaction patterns, conceptual rigour could be hampered if the aggregation of
relational events over time is done superficially. Indeed, every relational event, for example, Student A
replying to a post created by Student B at t1, arises from its unique context. This event is essentially
different from another reply from A to B at t2. Simply aggregating two events together to make a claim
about a relational state between the students treats these events as equal, neglecting the unique contexts
in which these events occur and different forces that may drive these events. As elaborated above, when
one student decides whether to interact with another student, many factors might be at play, including
the student’s personal traits, the other student’s post content, whether they are assigned to the same
group and any earlier interactions between these two students. The accumulated relational events
between students contribute to the formation and changes of their relational state, such as whether two
students have developed a shared interest, which creates new conditions for future relational events. To
understand how online discussions unfold, it is more conceptually sound to focus on the event level,
treating a relational event as a distinct phenomenon that is bounded by its context and contributes to the
dynamic and emergent process of social interaction in online discussions.
Australasian Journal of Educational Technology, 2023, 39(6).
4
The present study
The present study put these various perspectives together to model online learning discussions. Informed
by prior work around the factors shaping online discussions such as pedagogy, technology, individual
characteristics, social dynamics and the need to model them as relational events unfolding over time, we
asked: Considering pedagogical and technological factors, in which ways do learner characteristics and
social dynamics collectively shape social interaction in asynchronous online discussions in higher
education context?
Research context
Participants
The study was situated in an undergraduate, online course at a large public university in the United States
of America. The course covered topics related to technology and ethics and attracted undergraduate
students from a wide range of disciplines. One section of this course (n = 20) participated in this study.
The course design followed social constructivist perspectives of learning. The instructor used a tool named
Yellowdig to support asynchronous online discussions among students. We elaborate on the technological
and pedagogical conditions for online learning discussions in this course to properly explain the type of
influence they could have over the unfolding discussions.
The discussion tool
The discussion environment Yellowdig resembles social network sites like Reddit. As illustrated in Figure
1, a student can share posts – known as pins – that can be commented on or reacted to (e.g., like) by
other students. Students could also mention each other in a post or comment.
Figure 1. The interface of Yellowdig. This figure shows one pin created by one student that attracted
four comments.
Australasian Journal of Educational Technology, 2023, 39(6).
5
Similar to many social network websites, a class on Yellowdig has a news feed featuring posts by its
members. This news feed can be sorted in different ways, and the default sorting algorithm places the
most recent discussion activities on the top. When a pin receives a new comment or reaction, it will be
placed on top of the news feed. However, Yellowdig is different from social network sites in that its user
profile features are less emphasised. Rather than following or making friends with each other, students
visit Yellowdig to participate in content-based discussions similar to those in online discussion forums.
Overall, Yellowdig combines technological features of both discussion forms and social network sites to
promote content-based discussions and social interaction among students.
Discussion activiity design
Online discussion on Yellowdig was an integral part of this online class. Students participated weekly on
Yellowdig to discuss readings and share ideas. Each week, they were asked to either make a post in
response to a prompt or post a commentary on a news article related to the course topic. They were
required to minimally contribute one post and comment on two posts each week. A labour-based grading
practice was adopted. Students received points when they demonstrated enough effort to participate by
posting substantive content (e.g., minimally 100 words in a post) and engaging with their peers.
Data sources
We were granted approval from the Institutional Review Board to analyse system logs from these classes.
The scope of approval was limited to timestamped event logs; therefore, the post content had to be
excluded from the analysis.
Interaction data from Yellowdig, including 274 posts, 514 comments, 36 mentions and 74 reactions, were
the primary data source. Course materials, including the syllabus and weekly announcements, were
gathered to inform the interpretation of results.
Research hypotheses
Drawing on the research literature (see the Related work section) and considering the research context,
we narrowed down to a list of research hypotheses about participant characteristics and social dynamics.
Our analysis was limited to system logs that did not include students’ discussion content, which limited
the scope of our hypotheses. The hypotheses were formulated in the order of how patterns may form in
a social setting: sometimes it is an individual-level characteristic that shapes if someone replies or receives
replies (the first set of hypothesis), sometimes it is the dynamics between a pair of people or an emergent
triad (the second set of hypothesis), and sometimes it is a micro-level activity between the types of
behaviours that usually follow each other such as responding to a set of comments from multiple peers
(the third set of hypothesis). In each of these sets of hypotheses, we added either individual
characteristics or social dynamics that may be influencing the emergence of online learning discussions,
as per our proposed conceptualisation of social interaction in online discussions.
Our first set of hypotheses explains patterns of sending and receiving replies by considering individual
characteristics and individual social status of discussion participants. Regarding participant characteristics,
the literature shows that the timing of participation matters. Students’ posts written earlier in the week
tend to receive more replies (Zingaro & Oztok, 2012), and students who join online discussions earlier are
more likely to occupy central positions in their interaction networks (B. Chen & Huang, 2019). Being time-
consistent, meaning not participating in online discussions close to the deadline but spreading out effort
across a week, is likely to be associated with social interaction. Also, sometimes being the instructor may
or may not attract student interaction (Zingaro & Oztok, 2012). Therefore, we proposed the following
hypotheses about participant characteristics that explain if online discussion participants with a particular
characteristic are more likely to reply to others or receive replies:
• H1a: Learners with consistent temporal patterns of participation are more likely to reply to
others.
Australasian Journal of Educational Technology, 2023, 39(6).
6
• H1b: Learners with consistent temporal patterns of participation are more likely to receive
replies from others.
• H1c: The instructor is more likely to reply to others.
• H1d: The instructor is more likely to receive replies.
Our second set of hypotheses pertains to social mechanisms in the interaction process itself. First, factors
associated with familiarity established in prior relational events are predictive of future events. There are
different types of familiarity in the context of online discussion including popularity, persistence,
reciprocity and triadic closure. Participants of online discussion also tend to reciprocate each other
(Cheung et al., 2008), meaning that prior replies are likely to be reciprocated. Finally, because the
discussion environment Yellowdig resembles the features of social network sites that aim to facilitate peer
interaction in discussion threads, it is reasonable to hypothesise that when two students interact with a
common peer within or across discussion threads, they are likely to interact with each other by forming a
triadic closure (Bianconi et al., 2014). Based on these ideas, we proposed the following hypotheses
pertinent to different types of familiarity formed in prior relational events – to model social dynamics at
the level of dyads and triads.
• H2a: The number of a learner’s prior interactions affects future incoming ties.
• H2b: Past interactions from A to B are to be reciprocated by new interactions from B to A.
• H2c: Past interactions from A to B are to be repeated, leading to new interactions from A to B.
• H2d: Two learners sharing an outbounding discussion partner (i.e., a peer being replied to by
both learners) also tend to interact with each other.
The third set of hypotheses pertains to the micro patterns of activity that may always happen together.
Besides factors related to familiarity reflected in prior events, the literature suggests that the patterns of
such local, turn-by-turn dynamics could also contribute to global interaction patterns. Recognising the
importance of the details of interaction and sequential constraints in group conversations, Gibson (2005)
proposed a number of participation shifts patterns. His inventory of participation shifts includes turn
claiming, turn receiving, turn usurping and turn continuing (see also Butts & Marcum, 2017). Prior work
in contexts such as email communication has demonstrated that these micro-level social dynamics
patterns have explanatory power as they are building-blocks of socio-cognitive phenomena such as
transactive discussion in learning settings where reasoning and uptakes of prior reasoning statements are
essential for meaningful collaboration (Gweon et al., 2013). Technologically, we recognised Yellowdig’s
unique features and their potential impact on participation shifts. In particular, because Yellowdig places
the more recent discussion activities (either the newest post or the post with the newest reply) on the
top of the news feed, a reply from Student A to Student B may not only trigger a response from Student
B (AB → BA) but also a third student, X, who happens to be viewing the news feed (AB → XB). Meanwhile,
given Yellowdig is an asynchronous discussion environment which students may visit at any time, it could
be possible that Student B would first respond to another peer, Y, before responding to Student A (AB →
BY). Each student may interact with peers in a batch during each Yellowdig session (AB → AY). Based on
the literature and technological features of Yellowdig used in the study, the following hypotheses about
immediate, turn-by-turn effects were proposed:
• H3a: When B receives a message from A, B tends to immediately respond to A (AB → BA).
• H3b: When B receives a message from A, B tends to receive another message next from a peer
other than A (AB → XB).
• H3c: When B receives a message from A, B would immediately participate but not necessarily
respond to A (AB → BY).
• H3d: A student tend to interact with their peers in a batch, replying to multiple peers at a time
(AB → AY).
To sum up, the three sets of research hypotheses were generated based on the proposed framework of
factors shaped online discussions, with the consideration of temporal dynamics and constraints imposed
on the patterns of discussion from the study’s unique pedagogical and technology context.
Australasian Journal of Educational Technology, 2023, 39(6).
7
REM
To test these research hypotheses, the REM approach (Butts, 2008) was applied to the data set.
As a method of social network analysis, REM provides a systematic approach to model the emergence of
network structures based on the collective influence of historical relational events, actor characteristics,
and contextual factors (Butts, 2008). The purpose of REM is to “understand how past interactions affect
the emergence of future interactions” (Quintane et al., 2013, p. 533). For instance, an event of Student s
replying to Student r at a time point, t1, is a relational event, a1, and Student r responding to s at a later
moment, t2, is a subsequent event, a2, which could be influenced by a1. REM derives a set of statistics
based on historical relational events (such as the count of prior s → r events) and combines them with
actor attributes (such as gender) and contextual factors (such as group affiliation and friendship) to
predict the next relational event. Mathematically, a relational event model could be expressed as follows
(Butts & Marcum, 2017, p. 55):
if )
otherwise
where λ represents the hazard of potential event a at time t given history At; whereas θ is a vector of real-
valued parameters; and u is a vector of REM statistics about the event a iniated by its sender s(a) to its
receiver r(a) at time τ(a). For example, reciprocity as a statistic measures the odds of Student i replying to
j provided that i has previously responded to j. If reciprocity is of interest in a study, a statistic of reciprocity
will be incorporated in the u vector of the relational event model, together with other variables, including
potential covariates, Xa. Statistics such as reciprocity capture both social and temporal information in
relational events and provide rich measures to describe social interactions (Leenders et al., 2016). More
detailed explanation of REM is beyond the scope of this paper and can be found elsewhere (Butts &
Marcum, 2017; Leenders et al., 2016).
To test our hypotheses, we computed measures related to participant characteristics and sequential
structural signatures reflected in their interaction logs (see Table 1). The relevent R package was used to
calculate these measures and estimate model parameters. Below we explicate these measures derived
for REM.
Actor or participant characteristics
In terms of actor or participant characteristics, two variables were included to allow the testing of
Hypotheses H1a–H1d:
• Being the instructor. This variable characterised whether a participant was the instructor of a
course.
• Profiles of the timing of participation. Given the importance of timing in online discussions (see
the Consideration of time in online discussions section), we derived profiles of participants’
timing behaviours. To characterise the timing of student participation, we used finite mixture
models (Deb & Trivedi, 2013) to cluster learners in each course based on the distribution of a
student’s logged activities per weekday, similar to the approach adopted in Park et al. (2018).
Hence, students with similar activity on the same days of the week in a course would be
clustered. For network modelling purposes, learners were grouped based on their cluster
membership, which indicated the timing of their participation.
Australasian Journal of Educational Technology, 2023, 39(6).
8
Table 1
Independent variables in the relational event models
Effects
Visualisation
Description
Actor attributes:
Time-consistent
Being time-consistent (or not)
Role
Instructor (vs. student)
Past relational events:
Nodal:
Total degree (NTDegRec)
Well-connected students to receive
more interactions
Dyadic:
Reciprocity (RRecSnd)
Past interactions to be reciprocated
Persistence (RSndSnd)
Past interactions to happen again
Triadic:
Outbounding shared partner (OSPSnd)
Shared outbounding partner leads
to direct interactions
Immediate effects:
Participation shifts:
AB → BA
Immediate reciprocation
AB → XB
Popularity, persistence of target
AB → BY
Pay-it-forward or hand-off
AB → AY
Activity bursts, persistence of
source
Here we provide additional details about the finite mixture models used to identify timing profiles of
learners. System logs used in the clustering included a rich set of behavioural indicators including creating
or revising a post, reading a post, creating or revising a comment, reacting to a post or comment (liking
or loving), and removing a post. Learners were then grouped based on the number of logs in Yellowdig
they had each day of the week. The flexmix and mixtools R packages were used for the analysis. During
finite mixture modelling, learners with consistent levels of activity on the same days were classified as
belonging to the same latent group. In this course, a cluster of learners with significantly higher number
of logs in general and on certain weekdays (Thursdays) than the rest of the class was labelled as time-
consistent participants. These learner profiles – time-consistent and other learners – characterised the
learners’ temporal participation patterns and were incorporated in the models.
Social dynamics
Measures of social dynamics on Yellowdig, as reflected by the sequential structural signatures, were
calculated using the relevent R package. First, the following measures reflecting different types of
familiarity in prior relational events were computed to examine H2a–H2d.
• Total degree provides information about the impact of a participant’s past events (as sender or
receiver) on the probability of that same participant to be replied in the future. A positive
coefficient indicates an actor’s higher prior relational events predicts a higher propensity for
this actor to be involved in the next event.
• Social persistence, also called inertia, captures the tendency for an actor, i, to reply to another
actor, j, if i has previously replied to j. This variable captures how much a relational event
reoccur. A positive coefficient indicates the occurrence of a relational event increases the
propensity for this event to happen again.
• Reciprocity describes the tendency of an actor, j, to reciprocate i if i commented on j’s posts
before.
Australasian Journal of Educational Technology, 2023, 39(6).
9
• Outbounding shared partner is a triadic effect that indicates the tendency for two actors, i and
j, with a shared outbounding partner, k, to interact with each other. In the context of Yellowdig
discussions, this effect could be understood as when two students comment on a post made by
a peer, they have a tendency to interact with each other as well.
Statistics of turn-by-turn social dynamics, also referred to as “immediate effects” in REM (Butts, 2008),
were also calculated for the testing of H3a–H3d. In particular, the following participation shifts (Gibson,
2005) were specified as part of REM:
• AB → BA indicates a tendency for an interaction from A to B to be immediately reciprocated.
• AB → XB indicates a local level popularity, meaning that when B receives a reply from A, B
tends to receive another message next from a peer other than A.
• AB → BY indicates a case where the target of the second event is another actor Y other than A
who initiated the first event.
• AB → AY captures the tendency of an actor A to initiate a series of events in a batch.
Model fitting and selection
For this course, we trained a series of relational event models using these variables following a forward
selection strategy so that we evaluate the influence of each set of variables on relational events and how
these variables were collectively shaping the emergence of relational events. Following Butts (2008), the
models were evaluated based on the Akaike information criterion (AIC) and Bayesian information criterion
(BIC) scores that are widely used to evaluate how well a model fits the data it was trained on.
Findings
Descriptive analysis of the interaction networks
Table 2 reports descriptive statistics of discussion activities in the class. On average, each participant
authored 13.7 posts and 25.7 comments and connected with 16.3 class members. Based on descriptive
statistics of the network structure, the class was densely connected and lowly centralised, with its nodes
reachable to each other (see Table 2). Figure 2 presents a network visualisation of social interaction in the
class. In this visualisation, the instructor and time-consistent students are colour coded. As shown in the
figure, eight learners were identified as being time-consistent and the other 12 learners were not.
Table 2
Descriptive measures of the interaction network (undirected)
Variable
Value
Total posts
274
Total comments
514
Mean degree
16.28
Network density
0.81
Diameter
3
Mean distance
1.18
Degree centralisation
0.18
Australasian Journal of Educational Technology, 2023, 39(6).
10
Figure 2. Network visualisation of social interaction in the course. Each node represents one participant,
indexed by the number, with time-consistent students coloured in blue and the instructor coloured in
orange. Each edge is directed, with its density representing the count of its occurrences throughout the
whole semester.
Results of REM
To examine the ways in which participant attributes and social dynamics collectively shape social
interaction in the study, we trained a series of relational event models for the class.
Goodness of fit
The relational event models were fitted in a step-wise manner, with new factors gradually added to the
model to test all the hypotheses (Butts & Marcum, 2017). Overall, the AIC and BIC scores decreased when
variables were added to the null model showing improvement of the models. Based on the AIC and BIC
scores, the full model provided the best fit. Below we report results of hypothesis testing based on the
full model (see Table 3).
1
23
4
5
6
7
8
9
10
11
12
13 14
15
16
17
18
19
20
21
time_consistent
instructor
time_consistent
instructor
Australasian Journal of Educational Technology, 2023, 39(6).
11
Table 3
Coefficient estimates in REM
Coefficient
Standard error
p value
Sending time-consistent
0.10
0.13
.425
Receiving time-consistent
0.20
0.12
.085
Sending instructor
0.85
0.19
< .001***
Receiving instructor
-2.05
0.38
< .001***
Degree (NTDegRec)
12.33
1.45
< .001***
Reciprocity (RRecSnd)
0.53
0.16
< .001***
Persistence (RSndSnd)
1.30
0.15
< .001***
Closure (OSPSnd)
-0.01
0.01
.200
Reciprocation (AB-BA)
1.30
0.48
.007**
Pay-it-forward (AB-BY)
-0.61
0.38
.114
Activity bursts (AB-AY)
1.25
0.14
< .001***
Popularity (AB-XB)
2.96
0.09
< .001***
Note. p < .001; p <. 01; p <. 05
Participant attributes
First, we hypothesised that peer interaction in the setting could be predicted by participant attributes. Based
on REM, we found the first two hypotheses could not be supported, meaning that time-consistent students
did not show higher propensity of initiating interactions with their peers (H1a) or receiving replies from
peers (H1b), suggesting that the timing profile of participant was not a significant factor for the occurrence
of student interactions in the discussion environment.
We also hypothesised that the instructor role would positively impact social interaction. As revealed by
the relational event models, being the instructor was associated with higher odds of initiating interactions
in four classes (H1c). However, the instructor was significantly less likely to be responded by students
(H1d). These findings suggested the instructor in the course was actively interacting with students but did
not necessarily have stronger propensity of attracting student replies. Such dynamics reflected
Yellowdig’s strength in fostering peer-to-peer rather than student–instructor interactions.
Social dynamics
In terms of statistics related to prior relational events, total degree was confirmed to be a significantly
positive predictor with the largest coefficient in the model (H2a). This indicated that students with more
connections at a time point had higher propensity to involve in new interactions. Reciprocity was also
found to be a positive contributor to peer interaction in the class (H2b), even though the coefficient was
small. This means prior replies from one student to another was predictive of future replies in the opposite
direction. This finding is consistent with prior work that reports reciprocity as a definitive mechanism of
network formation in online environment. Social persistence or inertia was found positive and significant
as well (H2c), indicating prior replies from one student to another had positive propensity to reoccur. This
suggested that social dynamics in the discussion environment had a memory of interactions across the
dyads that moved from serendipitous encounters to familiarity and eventually relational ties. Triadic
closure, in particular the outbounding shared partner mechanism, was not a significant factor in the class
(H2d). This finding revealed that two students were not more likely to interact with each other if they had
previously replied to another peer, suggesting that there was not much of group formation processes
based on triadic closure mechanism in the Yellowdig environment.
Finally, in terms of the turn-by-turn participation shifts, four hypotheses were tested. The pay-it-forward
pattern (AB → BY) was non-significant (H3c), while the other three factors were significantly positive
predictors, including local popularity (AB → XB) (H3b), immediate reciprocation (AB → BA) (H3a), and
activity bursts (AB → AY) (H3d). These results suggested at the local turn-by-turn level, a student replied
by a peer has high propensity to respond to the peer (immediate reciprocation) and be immediately
replied by another peer (local popularity). Students also tended to send replies to multiple peers in a
batch (activity burst), indicating a dominant feature of posting behaviour in this context was that learners
Australasian Journal of Educational Technology, 2023, 39(6).
12
tended to respond to several learners at a time when they logged into the environment. These findings
together point to a localised nature of online communication that occurs within that time when learners
logged online and within the space of what is visible to them and still available to contribute.
General discussion
In this study, we conceptualised online discussions in education as a dynamic phenomenon influenced by
various forces and sophisticated interactions among them and posit that the research of online
discussions needs to consider the temporal unfolding of discussion activities. Based on this understanding,
we argue it would be more conceptually rigorous to investigation student interaction in these discussion
activities as relational events. This event-based approach differs from the traditional state-based
approach which makes claims about interpersonal relational states based on temporal aggregates of
relational events between actors. The event-based approach advanced in this paper recognises the
probable impact of prior events on later ones, whereas state-based approach fails to account for these
nuanced patterns enacted by diverse factors. We suggest that understanding online learning discussions
is more appropriate through an event-level examination that considers each relational event and the
context where it occurs. To demonstrate this approach, we applied REM to a data set of student online
discussions and tested three sets of hypotheses grounded in the literature and the research context.
The first set of hypotheses were related to participant characteristics and their average affect to make
contributions or receive them. Results did not find time-consistent participants (i.e., those who logged
into the system on the same weekdays across several days) more likely to send and receive replies. This
finding contradicts earlier studies that recognised different timing of learner participation and its
connection with discussion performance (B. Chen & Huang, 2019; Riel et al., 2018). We tested whether
the instructor attracted or extended more replies and found them more likely to send replies but actually
less likely to be replied. The instructor was actively participating but was less likely than the students to
be replied to. This finding agrees with that reported in Zingaro and Oztok (2012). It is worth further
research to find out whether the social feature of Yellowdig played a role in this study since Yellowdig
centres attention on student posts, whereas traditional threaded discussion forums are often structurally
centred on instructor prompts.
The second set of hypotheses were generally about different types of familiarity reflected in previous
relational events, modelled at the level of a dyad and a triad to account for various types of social
dynamics that describe these two different social structures. Results identified total degree (i.e.,
popularity), reciprocity and social persistence as three significant contributors to the occurrence of
relational events. That is, more active students were more likely to receive replies; more replies from one
student to another predicted future replies between these two students in both directions. However, the
triadic closure hypothesis was rejected, suggesting having a shared interaction partner did not lead to two
students interacting with each other. The finding on popularity agrees with earlier studies, such as
Vaquero and Cebrian (2013), that popular participants are more likely to receive replies. Social persistence
found in earlier studies (B. Chen & Huang, 2019) was also significant in the study. The positive role played
by reciprocity also agrees with earlier findings that participants of online discussion tend to reciprocate
each other (Cheung et al., 2008). However, triadic closure found important in other contexts (Bianconi et
al., 2014) was not a significant factor in this study.
Finally, the third set of hypothesis examined micro-level patterns of behaviour such as turn-by-turn
participation shifts. We found participation shifts to explain online discussion patterns. Immediate
reciprocation, local popularity and activity bursts were positive predictors of relational events. These local
patterns were rarely examined in prior studies, even in studies of online discussions that applied REM
(e.g., Vu et al., 2015). A learner who just replied to a peer had a high propensity to be replied to by another
peer. The local popularity could be attributed to the technological design of Yellowdig that promotes the
most recent discussion activity to the top of the news feed. Also, learners tended to respond to multiple
peers at a time, that is, having activity bursts, in the discussion environment. The strong presence of
Australasian Journal of Educational Technology, 2023, 39(6).
13
activity bursts is understandable given the discussion was asynchronous, so students were rarely present
at the same time in the discussion environment.
This study’s contribution to the literature is two-fold. First, we surveyed multiple sets of factors
influencing social interaction in asynchronous online discussions. These factors include student
characteristics, pedagogical practices, technological features and social dynamics. Building on the
literature, this study recognises the nuances of online discussions and calls for methods that can cope
with the complex dynamics. Second, by applying REM, this study contributes fresh empirical findings
about socio-temporal dynamics in online discussions. Departing from prior work, REM allowed us to
model the effects of these factors simultaneously (Butts, 2008; Butts & Marcum, 2017). This event-based
approach enabled us to drill down to the event level and examine how specific temporal and social factors
facilitate or hinder the occurrence of relational events. While earlier studies based on other methods have
examined some of these factors, it is important to point out that REM allowed us to examine how these
factors synergisticly contributed to network formation in this study.
Empirical findings from this study have practical implications. First and foremost, sophisticated socio-
temporal dynamics uncovered in this study suggest fresh angles to scaffold student participation and
interaction in online discussions. Although prior work has documented various ways to scaffold online
discussions, such as adopting questioning and participation roles strategies (K.-Z. Chen & Yeh, 2021;
Wang, 2005; Wise et al., 2012; Zhu et al., 2023), this study suggests it might be fruitful to also help learners
attend to not only posts at the top of the feed (Hewitt, 2005). Insights revealed by REM suggest various
ways to facilitate peer interaction. Instructors and instructional designers who aspire to promote student
interaction could consider those network effects as potential levers when designing discussion activities.
For example, given the significance of familiarity for the occurrence of interactional events, instructors
could support students to interact with less familiar classmates to boost familiarity in the class as a whole;
similarly, as students were found to binge-post in this study, future iterations of the course could engage
students to join the online conversation multiple times during a week to enhance the frequency and
quality of their online encounters. Second, technology developers also need to carefully consider how a
particular design decision may shape the nature of online discussions (Guzdial & Turns, 2000). In this case,
Yellowdig has social features that decentralise discussions (to be less focused on the instructor) while also
giving rise to local popularity by placing the posts with most recent replies on the top. We suggest that
developers of educational technologies consider such implications when designing software features and
pay close attention to educational meanings of these features. Finally, given the rise of learning analytics,
analytic tools could be developed to provide information for the instructor and learners to take actions
on information derived from system log data. Findings from this study suggest new ways to expand on
current analytics that mostly describe relational states by incorporating a suite of sequential structure
signatures that describe deeper socio-temporal dynamics.
Despite these contributions and implications, it is important to point out a few limitations with the present
study. First, data available for analysis in the study were limited to behavioural logs in the discussion
environment. As a result, we were unable to incorporate features of discussion content in the relational
event models. Also, even though the main goal of this paper was to demonstrate potential of analysing
relational events in online discussions, the interpretation of some findings would be complemented by
qualitative interviews with students and/or student self-reports about their discussion practice.
Further work is needed to address these limitations. Future studies could incorporate other factors
mentioned in the literature review but not examined in the study. For example, one direction could be to
incorporate natural language processing techniques to capture features of the discussion content, so that
the relational event models can be further enhanced. Another direction could be to create clear mapping
between design patterns of online discussions and the REM statistics so that an instructor could
purposefully adopt relevant statistics to evaluate discussion activities. Overall, this paper makes a strong
case for examining online discussions as relational events and encourages future research and design to
consider learner interactions as relational events that unfold over time in context.
Australasian Journal of Educational Technology, 2023, 39(6).
14
Author contributions
Author 1: Conceptualisation, Investigation, Formal analysis, Methodology, Writing – original draft, Writing
– review and editing; Author 2: Data curation, Investigation, Formal analysis, Methodology, Writing –
review and editing.
References
Bianconi, G., Darst, R. K., Iacovacci, J., & Fortunato, S. (2014). Triadic closure as a basic generating
mechanism of communities in complex networks. Physical Review E, 90(4), Article 042806.
https://doi.org/10.1103/PhysRevE.90.042806
Butts, C. T. (2008). A relational event framework for social action. Sociological Methodology, 38(1), 155–
200. https://doi.org/10.1111/j.1467-9531.2008.00203.x
Butts, C. T., & Marcum, C. S. (2017). A relational event approach to modeling behavioral dynamics. In A.
Pilny & M. Poole (Eds.), Group processes: Data-driven computational approaches (pp. 51–92).
Springer. https://doi.org/10.1007/978-3-319-48941-4_4
Chen, B., & Chen, Z. (2023). Why do students reply? Uncovering the socio-semantic entanglement in
web annotation activities. Education and Information Technologies. https://doi.org/10.1007/s10639-
023-12187-5
Chen, B., & Huang, T. (2019). It is about timing: Network prestige in asynchronous online discussions.
Journal of Computer Assisted Learning, 35, 503–515. https://doi.org/10.1111/jcal.12355
Chen, G., Lo, C. K., & Hu, L. (2020). Sustaining online academic discussions: Identifying the characteristics
of messages that receive responses. Computers & Education, 156, Article 103938.
https://doi.org/10.1016/j.compedu.2020.103938
Chen, K.-Z., & Yeh, H.-H. (2021). Acting in secret: Interaction, knowledge construction and sequential
discussion patterns of partial role-assignment in a MOOC. Australasian Journal of Educational
Technology, 37(6), 41–60. https://doi.org/10.14742/ajet.6979
Cheung, W. S., Hew, K. F., & Ng, C. S. L. (2008). Toward an understanding of why students contribute in
asynchronous online discussions. Journal of Educational Computing Research, 38(1), 29–50.
https://doi.org/10.1016/j.compedu.2020.103938
Cho, M.-H., & Kim, B. J. (2013). Students’ self-regulation for interaction with others in online learning
environments. The Internet and Higher Education, 17, 69–75.
https://doi.org/10.1016/j.iheduc.2012.11.001
Cho, M.-H., Lim, S., Lim, J., & Kim, O. (2022). Does gender matter in online courses? A view through the
lens of the community of inquiry. Australasian Journal of Educational Technology, 38(6), 169–184.
https://doi.org/10.14742/ajet.7194
Deb, P., & Trivedi, P. K. (2013). Finite mixture for panels with fixed effects. Journal of Econometric
Methods, 2(1), 35–51. https://doi.org/10.1515/jem-2012-0018
Galikyan, I., Admiraal, W., & Kester, L. (2021). MOOC discussion forums: The interplay of the cognitive
and the social. Computers & Education, 165, Article 104133.
https://doi.org/10.1016/j.compedu.2021.104133
Gibson, D. R. (2005). Taking turns and talking ties: Networks and conversational interaction. American
Journal of Sociology, 110(6), 1561–1597. https://doi.org/10.1086/428689
Guzdial, M., & Turns, J. (2000). Effective discussion through a computer-mediated anchored forum.
Journal of the Learning Sciences, 9(4), 437–469. https://doi.org/10.1207/S15327809JLS0904_3
Gweon, G., Jain, M., McDonough, J., Raj, B., & Ros´e, C. P. (2013). Measuring prevalence of other-
oriented transactive contributions using an automated measure of speech style accommodation.
International Journal of Computer-Supported Collaborative Learning, 8(2), 245–265.
https://doi.org/10.1007/s11412-013-9172-5
Hewitt, J. (2005). Toward an understanding of how threads die in asynchronous computer conferences.
Journal of the Learning Sciences, 14(4), 567–589. https://doi.org/10.1207/s15327809jls1404_4
Jeong, H., Hmelo-Silver, C. E., & Yu, Y. (2014). An examination of CSCL methodological practices and the
influence of theoretical frameworks 2005–2009. International Journal of Computer-Supported
Collaborative Learning, 9(3), 305–334. https://doi.org/10.1007/s11412-014-9198-3
Australasian Journal of Educational Technology, 2023, 39(6).
15
Jo, I., Park, Y., & Lee, H. (2017). Three interaction patterns on asynchronous online discussion
behaviours: A methodological comparison. Journal of Computer Assisted Learning, 33(2), 106–122.
https://doi.org/10.1111/jcal.12168
Knight, S., Wise, A. F., & Chen, B. (2017). Time for change: Why learning analytics needs temporal
analysis. Journal of Learning Analytics, 4(3), 7–17. https://doi.org/10.18608/jla.2017.43.2
Lee, J.-E., & Recker, M. (2021). The effects of instructors’ use of online discussions strategies on student
participation and performance in university online introductory mathematics courses. Computers &
Education, 162, Article 104084. https://doi.org/10.1016/j.compedu.2020.104084
Leenders, R. T. A. J., Contractor, N. S., & DeChurch, L. A. (2016). Once upon a time. Organizational
Psychology Review, 6(1), 92–115. https://doi.org/10.1177/2041386615578312
Mullen, G. E., & Tallent-Runnels, M. K. (2006). Student outcomes and perceptions of instructors’
demands and support in online and traditional classrooms. The Internet and Higher Education, 9(4),
257–266. https://doi.org/10.1016/j.iheduc.2006.08.005
Park, J., Yu, R., Rodriguez, F., Baker, R., Smyth, P., & Warschauer, M. (2018, July 16–20). Understanding
student procrastination via mixture models [Conference presentation]. International Conference on
Educational Data Mining, Raleigh, NC, United States of Ameria.
Quintane, E., Pattison, P. E., Robins, G. L., & Mol, J. M. (2013). Short- and long-term stability in
organizational networks: Temporal structures of project teams. Social Networks, 35(4), 528–540.
https://doi.org/10.1016/j.socnet.2013.07.001
Ractham, P., Kaewkitipong, L., & Firpo, D. (2012). The use of Facebook in an introductory MIS course:
Social constructivist learning environment. Decision Sciences Journal of Innovative Education, 10(2),
165–188. https://doi.org/10.1111/j.1540-4609.2011.00337.x
Raković, M., Marzouk, Z., Liaqat, A., Winne, P. H., & Nesbit, J. C. (2020). Fine grained analysis of
students’ online discussion posts. Computers & Education, 157, Article 103982.
https://doi.org/10.1016/j.compedu.2020.103982
Riel, J., Lawless, K. A., & Brown, S. W. (2018). Timing matters: Approaches for measuring and visualizing
behaviours of timing and spacing of work in self-paced online teacher professional development
courses. Journal of Learning Analytics, 5(1), 25–40. https://doi.org/10.18608/jla.2018.51.3
Scardamalia, M., & Bereiter, C. (2008). Pedagogical biases in educational technologies. Educational
Technology, 48(3), 3–11. https://www.jstor.org/stable/44429572
Shafie, L. A., Yaacob, A., & Singh, P. K. K. (2016). Lurking and L2 Learners on a Facebook group: The
voices of the invisibles. English Language Teaching, 9(2), 1–12. https://doi.org/10.5539/elt.v9n2p1
Vaquero, L., & Cebrian, M. (2013). The rich club phenomenon in the classroom. Scientific Reports, 3,
Article 1174. https://doi.org/10.1038/srep01174
Vu, D., Pattison, P., & Robins, G. (2015). Relational event models for social learning in MOOCs. Social
Networks, 43, 121–135. https://doi.org/10.1016/j.socnet.2015.05.001
Wang, C.-H. (2005). Questioning skills facilitate online synchronous discussions. Journal of Computer
Assisted Learning, 21(4), 303–313. https://doi.org/10.1111/j.1365-2729.2005.00138.x
Wise, A. F., Saghafian, M., & Padmanabhan, P. (2012). Towards more precise design guidance: Specifying
and testing the functions of assigned student roles in online discussions. Educational Technology
Research and Development, 60(1), 55–82. https://doi.org/10.1007/s11423-011-9212-7
Zhang, J., Scardamalia, M., Reeve, R., & Messina, R. (2009). Designs for collective cognitive responsibility
in knowledge-building communities. Journal of the Learning Sciences, 18(1), 7–44.
https://doi.org/10.1080/10508400802581676
Zhang, X. (2023). Peer pressure and web-based peer learning: An exploratory case study. Education and
Information Technologies, 28(8), 10405–10420. https://doi.org/10.1007/s10639-023-11609-8
Zheng, B., & Warschauer, M. (2015). Participation, interaction, and academic achievement in an online
discussion environment. Computers & Education, 84, 78–89.
https://doi.org/10.1016/j.compedu.2015.01.008
Zhu, X., Shui, H., & Chen, B. (2023). Beyond reading together: Facilitating knowledge construction
through participation roles and social annotation in college classrooms. The Internet and Higher
Education, 59, Article 100919. https://doi.org/10.1016/j.iheduc.2023.100919
Australasian Journal of Educational Technology, 2023, 39(6).
16
Zingaro, D., & Oztok, M. (2012). Interaction in an asynchronous online course: A synthesis of
quantitative predictors. Journal of Asynchronous Learning Networks, 16(4), 71–82.
https://doi.org/10.24059/olj.v16i4.265
Corresponding author: Bodong Chen, cbd@upenn.edu
Copyright: Articles published in the Australasian Journal of Educational Technology (AJET) are available
under Creative Commons Attribution Non-Commercial No Derivatives Licence (CC BY-NC-ND 4.0).
Authors retain copyright in their work and grant AJET right of first publication under CC BY-NC-ND 4.0.
Please cite as: Chen, B., & Poquet, O. (2023). Uncovering socio-temporal dynamics in online discussions:
An event-based approach. Australasian Journal of Educational Technology, 39(6), 1-16.
https://doi.org/10.14742/ajet.8618
