Source-linked AI summary
Class attendance, peer similarity, and academic performance in a large field study
Valentin Kassarnig, Andreas Bjerre-Nielsen, Enys Mones, Sune Lehmann, David Dreyer Lassen
TL;DR
Attendance research often relies on costly, bias-prone surveys and self-reports. This paper measures attendance using smartphone sensor data from nearly 1,000 undergraduates and examines its academic and social patterns. Attendance is weakly correlated with academic success, while attendance homophily is stronger among high performers.
Problem
Most prior attendance research uses surveys or self-reports that are prone to bias and costly to gather at scale.
Source: Introduction · p. 2Method
The study infers class locations and attendance from smartphone GPS and Bluetooth data for nearly 1,000 undergraduates, then examines attendance and peer patterns.
Source: Introduction · p. 2, Class location and attendance · p. 4Results
<0.3 correlation with academic success was observed, while attendance homophily was present across performance groups and stronger among high performers.
Source: Discussion · pp. 12, 14Takeaways & Limitations
In this population, attendance relates to academic success and social interactions, with students of similar performance tending to cluster in the network.
Source: Discussion · p. 12The measurements cover only a subset of enrolled students; about 40% of first-year students participated, and participants achieved higher grades than average.
Source: Discussion · p. 14
Abstract
from arXiv · showhide
Identifying the factors that determine academic performance is an essential part of educational research. Existing research indicates that class attendance is a useful predictor of subsequent course achievements. The majority of the literature is, however, based on surveys and self-reports, methods which have well-known systematic biases that lead to limitations on conclusions and generalizability as well as being costly to implement. Here we propose a novel method for measuring class attendance that overcomes these limitations by using location and bluetooth data collected from smartphone sensors. Based on measured attendance data of nearly 1,000 undergraduate students, we demonstrate that early and consistent class attendance strongly correlates with academic performance. In addition, our novel dataset allows us to determine that attendance among social peers was substantially correlated ($>$0.5), suggesting either an important peer effect or homophily with respect to attendance.
Introduction
The paper addresses how to measure attendance reliably and assess its links to academic performance and peers. It proposes smartphone-based measurement to study these questions in a large student population.
Introduction
- Smartphone data from nearly 1,000 undergraduates is used to study attendance, grades, and social aspects of academic performance.The authors describe the dataset as the first of comparable richness used for attendance analyses.Source: Introduction · p. 2
- Attendance research commonly relies on surveys, sign-in sheets, or self-reports, which are prone to bias and error.Conventional attendance information can also be costly to collect at school or university scale.Source: Introduction · p. 2
- Most prior studies find attendance positively predicts course grades, and a meta-study calls it the most accurate known predictor of academic performance.Initial attendance has also been reported as predictive, while average attendance tends to decline over the semester.Source: Introduction · pp. 2–3
- Prior peer-attendance research is limited to preliminary co-occurrence observations and lacks data on actual communication or interactions.Source: Introduction · p. 3
- The study evaluates smartphone-based attendance inference, its prediction of subsequent exam performance, and similarity of attendance among students’ social peers.A specific inference question is whether large clusters of students enrolled in the same course suffice.Source: Introduction · p. 3
Data
The analysis uses smartphone and academic records from the Copenhagen Networks Study at DTU, covering two academic years and nearly 1,000 undergraduates.
Data
- Copenhagen Networks Study data came from dedicated smartphones used by nearly 1,000 DTU undergraduates over two years.The analysis covers academic years 2013/14 and 2014/15.Source: Data · p. 3
- Recorded channels included GPS location, Bluetooth proximity scans, and mobile-phone communication, while DTU supplied grades and course schedules.Source: Data · pp. 3–4
- Participants represented 24 study lines; about 78% were male and 22% female, and more than 60% enrolled in 2013.More than 25% enrolled in 2012, with the remainder enrolling in 2011.Source: Data · p. 4
- Attendance was not mandatory or officially logged, and participants could leave the study and request deletion of their data.Participation varied over time as students could exit at any point.Source: Data · p. 4
Class location and attendance
The method estimates class locations from nearby course-enrolled students’ Bluetooth proximity and GPS data, then assigns attendance using distance and repeated observations.
Class location and attendance
- For each 15-minute bin, the highest-degree student among nearby same-course peers and that student’s neighbors form the primary cluster used to estimate class location.The class location is the median location of cluster members, using each member’s most accurate location in that bin; only courses with at least eight participants and standard four-hour duration are included.Source: Class location and attendance · p. 4
- A student is assigned to a class when within 200 m of its estimated location in at least three time bins; the distance threshold was robust from 5 to 500 m.The 200 m radius accounts for location noise, especially GPS indoors; main-cluster members are automatically assigned.Source: Class location and attendance · p. 5
- More than 75% of over 26,000 estimated class locations were within 100 m of the scheduled location, and a 200 m threshold covered 90%.Errors were estimated against building centers rather than rooms, and unrecorded relocations meant actual error was typically lower than estimated.Source: Class location and attendance · p. 5
- Figure 1a shows assigned students, an estimated class location, and its range for one day; Figure 1b shows estimated locations for a course across the semester.Source: Figure 1 · p. 5
Social ties
The study uses text-message exchange to infer social ties and identify students’ peers for attendance-pattern analyses.
Social ties
- The authors treat message exchange as evidence of a strong social tie when inferring students’ social connections.Source: Social ties · p. 5
- The peer-network method relies on message exchange rather than class-location error measurements.Source: Social ties · pp. 5–6
- Text messages sent or received between participants define a student’s peer list, with one message sufficient to establish a network link.Source: Social ties · p. 6
Statistical methods
The paper uses rank-based and nonparametric tests to compare attendance and performance, and a robust estimator to assess temporal trends. Its results examine grades, attendance changes over a semester, and peer similarity.
Statistical methods
- Kruskal–Wallis tests compare group distributions, with Dunn’s multiple comparisons and Bonferroni correction identifying which groups differ; Spearman correlation tests rank association without assuming linearity.The correlation coefficient is applied to ranks rather than raw values.Source: Statistical methods · p. 6
- The Theil–Sen estimator tracks temporal trends and is more robust to outliers and skewed data than simple linear regression.Source: Statistical methods · pp. 6–7
Results
- The results examine attendance associations with course and term grades, semester-long attendance trends, and similarity between students and social peers.Source: Results · p. 7
Academic achievement
Attendance is positively associated with grades, and higher-attendance groups show fewer exam failures. The association distinguishes failure risk more clearly than grades among students who pass, but does not establish causality.
Academic achievement
- Exam failure fell from 23% in the lowest attendance quintile to less than 4% in the highest.The two highest attendance groups showed no observable difference in grade distributions.Source: Academic achievement · p. 7
- Attendance better distinguishes whether students fail than predicts the grade earned by those who pass; the observed association does not establish causality.Source: Academic achievement · p. 8
- .255 (Spearman correlation, p < .001) quantifies the weak positive association between attendance and final grade.Figure 3 presents attendance conditional on grade, with box plots and observation counts.Source: Academic achievement · p. 7, Figure 3 · p. 8
Temporal effects
Attendance declined over the semester in all performance groups, but low performers started lower and their attendance fell faster than that of moderate and high performers.
Temporal effects
- Low performers began the first week 10% below high performers, and their attendance drop exceeded 20 percentage points versus 10 and 8 for moderate and high performers.Source: Temporal effects · pp. 9–10
- −1.4, −.6, and −.4 percentage points/week were the Theil–Sen slopes for low, moderate, and high performers, respectively.Slope differences were significant for low versus moderate and low versus high performers, but not moderate versus high.Source: Temporal effects · p. 10
- By semester’s end, the attendance gap between low and high performers had widened to 24 percentage points.Source: Temporal effects · p. 10
Peer similarity
Students’ attendance was moderately correlated with the attendance of their text-message contacts. Peer attendance also varied across performance groups, and similarity persisted after class-level effects were removed.
Peer similarity
- Peers of low performers also had lower attendance, though the differences were less pronounced than in students’ own attendance trends.Strong social contacts were inferred from exchanged text messages.Source: Peer similarity · p. 10
- The attendance similarity pattern remained after class-level effects were removed, consistent with homophily or peer effects.Source: Peer similarity · p. 11
- .48 (p < .001) was the moderate correlation between students’ attendance and their contacts’ mean attendance.Figure 6 plots own attendance against contacts’ mean attendance at course level, with color indicating relative data density.Source: Peer similarity · p. 10, Figure 6 · p. 12
Discussion
The discussion reports weak attendance–grade correlation alongside temporal and social-pattern findings, then qualifies their interpretation with measurement, confounding, sampling, participation, and outcome limitations.
Discussion
- Attendance was weakly correlated (< 0.3) with academic success, while students with similar performance clustered socially; sorting or peer effects appeared early and persisted across the term.The authors describe the social pattern as a mixture of either sorting or peer effects, not as evidence of a causal peer effect.Source: Discussion · p. 12
- 80%: the failing rate drops relative to the low-attendance quintile once attendance reaches the 75% threshold; early attendance distinguishes low performers from others, with weaker effects at high attendance and performance.Attendance declines overall during the semester, but its pace and level differ between low performers and the rest; moderate and high performers do not show the same distinction.Source: Discussion · pp. 13–14
- Homophily occurs across performance groups but is stronger among high performers; Figure 7 plots students’ course-averaged self and peer deviations from class-average attendance.Each dot represents a student's own and peers' median attendance deviation from the class average, averaged across a course's classes.Source: Figure 7 · p. 13, Discussion · p. 14
- GPS and Bluetooth signals are noisy; after correction, less than 10% of estimated class locations fall outside the 200 m radius around official locations.The method estimates class locations from clusters of physical meetings and uses a connected student's median location to reduce location-data noise.Source: Discussion · p. 14
- The estimates are not causal: personal characteristics were uncontrolled, and unobserved factors may drive both attendance and performance.The authors caution that attending more classes is not likely to change achievement by the amount observed.Source: Discussion · p. 14
- Measurements cover only a subset of enrolled students: around 40% of first-year students accepted phones, and participants achieved higher grades than the average student.The authors note that attendance and performance still vary substantially within the study dataset.Source: Discussion · p. 14
- Attendance does not establish active participation, and other contributors to academic performance remain unknown.Students who attend may or may not participate actively in class activities.Source: Discussion · p. 14
- Raw grades oversimplify multidimensional academic performance; subject interests differ, and distinguishing higher performers requires more individual-level information.The authors suggest that further research is needed to understand factors beyond attendance that influence academic success.Source: Discussion · p. 15
Supporting information
The supporting information provides dataset and data-collection details, describes class-level attendance correction, and reports distinct attendance–performance correlations across subject areas.
Supporting information
- The supporting information details data collection and recorded data types, explains class-level attendance correction, and shows distinct attendance–performance correlations across subject areas.Source: Supporting information · p. 15