ArticlePDF Available

Monte Carlo evaluations of methods of grade distribution in group projects: simpler is better

Abstract and Figures

Group projects are widely used in higher education, but they can be problematic if all group members are given the same grade for a project to which they might not have contributed equally. Most scholars recommend addressing these problems by awarding individual grades, computing some kind of individual weighting factor (IWF) from peer and (sometimes) self-assessments, which is then multiplied by the group grade to generate an individual grade. Several variants of the IWF method have been proposed, sometimes with complex algorithms. However, theory suggests they are inaccurate and their accuracy has not been evaluated. This article uses Monte Carlo experiments to assess the accuracy of the original IWF method and variants proposed in the past decade. Findings show that the earlier, simpler methods work best and that self-assessments should definitely be avoided.
Content may be subject to copyright.
Monte Carlo evaluations of grade distribution methods for group projects:
Simpler is better
Sebastián G. Guzmán
West Chester University
Universidad Andrés Bello
This manuscript has been published by Assessment and Evaluation in Higher Education. If you
need the printed version, the paid version is available at and 50 free copies of the full article are
available at
Group projects are widely used in higher education, but they can be problematic if all group
members are given the same grade for a project to which they might not have contributed
equally. Most scholars recommend addressing these problems by awarding individual grades,
computing some kind of Individual Weighting Factor (IWF) from peer- and (sometimes) self-
assessments, which is then multiplied by the group grade to generate an individual grade. Several
variants of the IWF method have been proposed, sometimes with complex algorithms. However,
theory suggests they are inaccurate and their accuracy has not been evaluated. This article uses
Monte Carlo experiments to assess the accuracy of the original IWF method and variants
proposed in the past decade. Findings show that the earlier, simpler methods work best and that
self-assessments should definitely be avoided.
Keywords: peer assessment, self-assessment, group projects, Individual Weighting Factor,
Monte Carlo.
Group projects are a common type of assignment in higher education today. They can help
students develop teamwork skills and confront the challenges of undertaking large and complex
assignments they could not complete individually in a semester, and they require less grading
effort from instructors, among other well-known benefits (for a review, see Jaques and Salmon
2007). However, there are common problems with these assignments, most notably, free-riding
and unfair grades, which can have severe implications on students’ motivation to work and the
group learning processes. To avoid these problems, several scholars have suggested ways of
computing individual grades for individual contributions to group projects (e.g. Conway et al.
1993; Goldfinch 1994; Goldfinch and Raeside 1990; Ko 2014; Lejk and Wyvill 2001a, 2001b; Li
2001; Neus 2011; Sharp 2006; Spatar et al. 2015; Tu and Lu 2005; Zhang and Ohland 2009).
One of the most common techniques is the computation of some variety of Individual Weighting
Factor (IWF), which is multiplied by the group grade to calculate each member’s individual
grade. Recent developments of the technique propose variations of the IWF formula that aim to
avoid the unfairness of biased evaluations and other distortions of the final grade. To this end,
scholars have replaced the original simple average of peer- and self-assessments by more
sophisticated algorithms (e.g. Bushell 2006; Ko 2014; Li 2001; Neus 2011; Sharp 2006; Spatar et
al. 2015).
One of the limitations of these developments is that we do not know how accurate they are
because the bars against which to measure the final grade distortions introduced by each
approach are imprecise. To evaluate each approach, scholars typically focus on special cases in
which distortions are evident, see if there are correlations between individuals’ grades in
individual work and group work, or simply check if students like the system (e.g. Baker 2007;
Jin 2012; Zhang and Ohland 2009). However, these methods are not the best way to test
algorithms. In several scholars’ case analyses it is unclear which assessments are distorted and it
is therefore difficult to claim that one method is better than the other. Additionally, the
correlation between the students’ real and estimated work may be affected by an error that is
correlated with the IWF’s or its variants’ estimation—this is actually the case, my results show.
Furthermore, given the number of assessments involved in computing IWF variants and thus a
large number of variables intervening in the outcome, it is difficult to predict the result in many
Zhang and Ohland (2009) tackle these issues evaluating older variants of the IWF with
Monte Carlo experiments. The experiments consist of simulating a large sample of real
contributions of students and of errors in their peer and self-assessments, and then comparing
IWF variants to each student’s real contribution. In this article, I use the same approach to
evaluate classic and more recent and sophisticated methods, adding two important
methodological improvements as well. First, my analysis focuses not only on the average
distortions each method produces, but also on whether distortions tend to favour students who
work less and those who distort assessments to boost their grades. Second, my experiments are
run in different scenarios of group size, dispersion of contributions, assessment errors, and self-
assessment inflation, making their results more robust.
Findings show that the original IWF and the Normalised IWF (NIWF) introduced by Sharp
(2006), these simple methods, when used excluding self-assessments, are more accurate than the
more complex recent methods. I also find that their distortions are small enough that the methods
can be safely used to estimate individual contributions.
In the following section, I summarise the existing approaches to computing an individual
grade for group projects, highlighting their theoretical contributions and limitations. Next, I
evaluate all approaches using Monte Carlo experiments. Finally, I offer some conclusions about
the CNIWF and previous methods.
Peer assessment of group projects
It is well-known that group projects have several benefits and are common today at all levels of
education and across disciplines. Unfortunately, there are also major challenges associated with
group projects. Several challenges relate to two issues emerging when all members of the group
receive the same grade: free-riding and unfair grades. Free-riding occurs when some students
decide to expend less effort in the group project because they can rely on their peers’ effort to
obtain nearly the same grade. This may translate into some students learning less than they
would otherwise, high workloads for other students, problems in group management and
interpersonal relationships, and students’ dislike for group assignments (Webb 1995). Unfair
grades exist because some group members work and learn more than other, but will be assessed
for the group’s results. This can translate into a disincentive to work more, dislike for group
assignments, and students being passed without actually learning the skills or contents of the
course. Studies show that free-riding and unfair grades are students’ main concerns regarding
group projects (Feichtner and Davis 1984; Macfarlane 2016), and scholars tend to agree these are
serious problem that need to be addressed if we are to use group assignments. The vast majority
of scholars addressing these problems suggest that the best solution is to derive individual grades
for each member, combining instructor assessments of group performance and peer assessments
of individual participation. As I explain below, these scholars offer several alternative methods
to compute such individual grades.
Since the mid-1990s, most approaches to assigning individual grades for group projects
involve variants of Goldfinch and Raeside’s (1990) IWF method (for exceptions, see Dommeyer
2012; Tu and Lu 2005). The core of the method consists of computing an estimator of individual
contributions (IWF) for each member of the group, which is then multiplied by the group’s grade
to obtain an individual grade. The different variants of the IWF for student A are a (sometimes
weighted) index resulting from A’s peers’ assessments of A’s work and sometimes A’s self-
assessment. The IWF’s variants attempt to address several potential problems of the original
method and of peer and self-assessments in general, most notably, those resulting from biased
Bias is a serious concern. Lejk and Wyvill (2001a) find that students tend to favour
themselves, with those who work less inflating their self-assessments more than those who
contribute more. This leads to a pattern of boosting the grades of those who work less to the
detriment of the rest. Additionally, there can also be bias in favour or against some peers and
different bars against which to measure contributions, leading to over- or under-marking by some
group members.
One way of addressing validity concerns is through methods that encourage members to
make more valid peer assessments, that is, improving the raw data before an index is computed.
There are four techniques along this line. First, assessments should be made confidential so that
students can freely express criticism of peers that they would not be willing to express publicly
(Lejk and Wyvill 2001a). This recommendation is broadly accepted and uncontroversial today.
Second, students may be asked to offer qualitative comments justifying their scores, so that they
have to reflect on them and they are less arbitrary, although it is time-consuming for instructors
to consider the comments (Loddington et al. 2009). Third, students may be asked to assess each
peer in several categories that reflect actual behaviour or to assess each peer holistically with
only one assessment or score. Unfortunately, it is so not clear which method is best (Lejk and
Wyvill 2001b; Ohland et al. 2012; Sharp 2006). Fourth, students can receive an incentive or
penalty for providing valid or invalid assessments (Tu and Lu 2005); yet we do now know
whether instructors and students will find such a policy acceptable.
Another important aspect, although neglected by the literature, is that scales that start from
0—studies often use use scales starting from 1. This inflates the grades of all those who worked
less. If A contributed about half of what B contributed and B’s contribution is the maximum of 5,
the middle point in the 1–5 scale is 3, which is 60% of B’s contribution, not half. If the average
contribution is 4, A would receive 1.25 and B would receive .75. By adapting the scale to go
from 0–4, the middle point is 2, exactly 50% of B’s contribution. The average contribution would
be 3, giving A an IWF of 1.33 and B only .67. In other words, with the wrong scale, B’s grade
was artificially increased in .08 times the group’s grade and 12% of what should be her real
grade, to the detriment of A.
A different but complementary approach to issues of validity involves improving the final
computation of a member’s mark with algorithms that weight and scale raw assessments in
different ways or may even exclude self-assessments, which tend to be biased. This approach
typically leads to variations of the IWF method, which are discussed below.
The IWF method and its variants
Conway and colleagues’ (1993) basic IWF for each student is computed as follows: First,
add the assessments of member A’s contribution assigned for that student by her peers and
herself. This can be done with whatever scale used. We call this the Individual Effort Rating
(IER) for each student. Next, compute the Average Effort Rating (AER) of the group, which is
the average of all the IERs. Finally, for each student, divide the IER by the AER to obtain the
IWF (see example in Table 1). We may express the IWF of student j assessed by assessors i’s
Raw Assessments (RAij) as:
== ====
Table 1 here
Each student’s IWF may then be multiplied by the group grade to compute an individual
grade. Or the final individual grade may be a weighted average of the IWF-based grade and the
group grade, with the weights set arbitrarily, usually 50% each.
One of the problems of this method is that some assessors are more generous than others, so
they do not all use the same rating scale. This has two implications. If we exclude self-
assessments, under-raters would artificially decrease their peers’ IWF-esa (Neus 2011, -esa
indicating ‘excluding self assessments’ hereafter). For instance, if everyone in a group of four
did the same amount of work and all students mark their peers with 25 except for a student who
rates them with 20, her IWF-esa would be 1.05, while everyone else would receive .98. And if
self-assessments are included, over-raters have a higher influence in the final grade than their
peers (Neus 2011). The best solution for this problem is to normalise each assessors’ ratings so
that all her assessments average (or add) one (or 100 or N) before computing the IER. This
makes all assessors’ evaluations comparable as indicators of relative contributions of each
member (Sharp 2006). This is achieved simply by dividing each individual assessment by the
sum of all the assessments made by the same assessor (see Table 1) or by asking students to split
a set number of points, say 100. Spatar and colleagues (2015) call the IWF computed from
normalised assessments the normalised IWF or NIWF (see Table 1). Thus, student i’s normalised
assessment of j’s work can be computed as:
Normalized assessments could also be amplified simply by multiplying the result in the
formula by 100, n, 100/n, or another constant. If this formula is used—as opposed, for instance,
from asking students to split a given number of points—the (normalised) AER will be equal to n,
the number of assessors. Replacing the (normalised) AER by n, the NIWF becomes simply the
average of all NAs of a given student’s work:
Following Lejk and Wyvill’s (2001a) general recommendation of excluding self-
assessments due to the fact that they tend to be inflated, Sharp (2006) recommends excluding
self-assessments in the normalization process—note that in this case n, the number of
assessments for a given student, is the group size minus 1. It should be noted, however, that there
is no strong agreement about whether to include self-assessments, and multiple scholars continue
to use self-assessments (Carvalho 2013; Ko 2014; Ohland et al. 2012; Zhang and Ohland 2009).
Ko argues that SA is better because excluding it involves the assumption that the assessor
should be self-assessed as an average, favouring low-contribution students to the detriment of
high-contribution students.
However, Bushell (2006) notes a serious problem when the normalization process excludes
self-assessments: it involves the assumption that the assessor should be self-assessed as an
average contributor, favouring low-contribution students to the detriment of high-contribution
students. This distortion produced by the normalization process is especially noticeable in
smaller groups. Bushell proposes a solution that requires the instructor to make decisions on a
case-by-case basis. Unfortunately, this solution is impractical because it cannot be easily
automated as an algorithm.
Li (2001) had an idea similar to Sharp’s, but Li’s method leads to producing the same grade
for all students if they make correct assessments. It is based on correcting the assessments if an
assessor gives to others fewer points than what everyone receives in average—without
considering self-assessments. This is avoided if the instructor manually corrects the model when
she notices an assessor is right in giving fewer points to others. Yet even after such case-by-case
analysis, which is often difficult, everyone’s grade comes closer to the group grade, favoring
those who worked less. Unfortunately, it is difficult to notice this in the examples discussed by
Li because opinions of students are inconsistent, with no agreed-upon ranking of contributors.
Several authors have offered alternatives to deal with the problem of inconsistent
assessments. Sharp (2006) proposed a statistical test to decide whether the difference in
estimated contributions is large enough and agreed-upon sufficiently to warrant any difference in
final grades. Simplifying this proposal, Neus (2011) suggests using an ‘agreement-corrected
IWF’ (ac-IWF) based on scaling each assessee’s NIWF according to the level of agreement
among assessments for that assessee. Only if there is significant agreement in the assessments of
one assessee, does that assessee’s grade change. The ac-IWF for a given student can be
computed as IAF×(IWF–1)+1, where IAF is an Individual Agreement Factor for that student.
The IAF for a student k is calculated as 1–sj(k)/max(sj), where sj is the (sample) standard
deviations of assessments of each assessee, max(sj) is the highest sj, and sj(k) is the standard
deviation of assessments of the assessee k’s work (see Table 1).
While it is reasonable to consider assessments as more valid if there is agreement among
assessors, the idea that without agreement the grade should be the average is problematic: why
are we to assume that in such a case the group average is the best estimate of that assessee’s
contribution? The peer assessments of that assessee’s work may still be a better approximation
than the average. Two of the main problems with Neus’s logic are evidenced by a paradox: if
everyone agrees that A did more work than the rest, A receives a smaller grade than the rest if
they disagree about how much more work A did (Ko 2014, 304-5). Additionally, if all but one
member agree on how much an assessee worked, the single divergent opinion, likely to be
biased, can unfairly bring the assessee’s grade close to the average (see Table 2 of Ko 2014, ,
also Table 1 below).
Spatar et al. (2015) attempt to moderate the ac-IWF’s sensitivity to biased assessments of
free riders by damping the effect of the IAF, replacing it by a scaled IAF (SIAF) computed as 1–
sj(k)/[2×max(sj)]. Yet while the problems of the ac-IWF will be smaller in this case, they are
likely to still occur because its logic is faulty. (Beyond the ASNIWF, Spatar and colleagues also
adopt a pre-existing method to deal with grades above the permitted maximum, but this is
unrelated to the question addressed here about estimating the proportion of the group work done
by each student.)
Ko (2014) addresses Neus’s problems with a formula that weights each assessor’s
assessments by the assessor’s reliability. The logic is that the existence of disagreements about
an assessee does not imply all assessments of that assessee are invalid; some are more valid than
others, and we should automate the process of identifying their validity as much as possible to
avoid time-consuming case-by-case checks proposed by other methods.
Unfortunately, Ko’s proposal still has some limitations, stemming from the formula for his
‘iterative IWF’ (it-IWF), which must be explained to grasp the problem. Ko’s reliability weight
factor depends on the distance between the assessor’s assessments for each assessee and the
mean of all assessments for each corresponding assessee. However, Ko notes, the mean is
artificially affected by unreliable assessments. Thus, Ko proposes an iterative method: first, we
compute a weighted average of assessments for each assessee assuming all assessors are equally
reliable; next, we define the reliability of each assessor; then we repeat the process, but using an
updated weighted average of assessments for each assessee, where the reliability factor defines
the weight. We can then repeat the process, computing new reliability factors and again updated
weighted averages, until convergence.
One problem of Ko’s it-IWF is that it produces different results depending on the size of the
scale used. For instance, using the data in Table 1, in which each assessment could range
between 0 and 20, A’s it-IWF is .33. But A receives .48 if we use normalised the assessments or
asked students to indicate the proportion of the group’s work done. The problem of inconsistency
could be avoided by simply starting from normalised assessments and setting what a normal
assessment would be beforehand—whether 1, 1/n, 100 or 100/n. But how are we to determine
what is the proper scale, if all scales return different results? The more serious problem lies in
how to set parameter b of Ko’s algorithm, which defines the level of discrimination of outliers.
Ko sets it as the mean of standard deviations of each assessor’s assessments, x
̅(σi). Since the rest
of the formula is based on variance (σi2) rather than standard deviation, there is a problem of
scales, as one is the square of the other. One solution to this issue is to set b as the mean of
assessors’ variances instead—i.e. x
̅(σi2). This it-IWF with corrected b and based on normalised
assessments produces consistent results regardless of the scale of normalised assessments.
However, Ko provides no rationale to identify an optimum b, as the it-IWF formula is not
grounded in statistical theory. In fact, to prevent inflation of the it-IWF through self-assessment
inflation, it is better to set b at a proportion of the mean of variances, such as a tenth of it—a
number suggested by some Monte Carlo tests I ran. I here call this variant it-IWF2.
Yet the it-IWF has another problem, which the it-IWF2 does not solve. Although the it-IWF
gives less weight to inflated self-assessments, it still allows inflated self-assessments to inflate
the final grade. There are two reasons for this. First, because each assessment is weighted by the
average reliability of the assessor, a student who inflates only her own assessment but is
consistent with peers’ opinion in the rest of her assessments can have a high reliability. This is
especially true when the rest of the peers are not completely consistent in their opinions and
when self-assessment inflation is not extremely exaggerated. The high reliability of the student
inflating her self-assessment will increase the weight of the biased self-assessment throughout all
iterations. In Table 1, for instance, if student A had been a bit more honest and assessed her peers
similar to what they did and inflated her self-assessment less, she would have inflated her it-IWF
considerably more. If she gave ratings of 9, 15, 17, and 15, respectively, her it-IWF would be .37
while her peers agree she deserved .31. The effect would be even higher if the peers disagreed
more about B, C, and D’s contributions.
There is second reason for why the it-IWF and it-IWF2 allow for grade self-inflation. When
peers are not consistent in assessing a student, such student’s mean contribution will be boosted
by an inflated self-assessment and the boosted mean contribution will sometimes remain through
iterations. For instance, assume all three members worked the same and agree on everything but
on A’s work, with B saying A contributed 110 and C saying A contributed 90. If A said she
contributed 120 instead of 100, she would increase her it-IWF to 1.07 and her it-IWF2 to 1.06.
Thus, iterations are not effective at dealing with the fact that biased assessments distort the mean
contribution from which we calculate assessments’ reliability when there is assessor
disagreement. Furthermore, the it-IWF method’s treatment of all disagreements as equal is
problematic given the pattern of students inflating their self-assessments (Lejk and Wyvill
The intuitive solution to the problem of self-assessment inflation is to eliminate self-
assessments. However, this is still problematic. If assessments are not normalised, students have
the incentive of deflating their peers’ assessment to effectively inflate their own grade. Since we
know students tend to inflate their self-assessments when allowed, we would expect them to
deflate their peers’ if self-assessments are excluded. Thus, the it-IWF-esa would still be
computed on the basis of distorted raw assessments. Those by low-contribution students would
tend to be more biased, but the it-IWF-esa cannot identify this pattern and will tend to treat those
moderately biased as more accurate than the unbiased assessments of high contribution students.
If, on the other hand, assessments are normalised, we have the abovementioned problem noted
by Bushell. Thus, there is no clear solution to this issue.
Beyond the mathematical problems of the it-IWF approach, its iterative method makes it
somewhat unpractical. Although the iteration is automated with an Excel macro, it requires some
advanced knowledge of spreadsheets and some adjustments in the spreadsheet or even the macro
program code if dealing with multiple groups of different sizes. This fact, along with the
complexity of the algorithm, which students may not understand, makes it unlikely that the
method will be used as much as simpler methods, especially in the less mathematically oriented
disciplines. To be worthwhile, the method would need to be substantially more accurate than its
best alternative.
In sum, no method to date deals with self-assessment inflation and assessor disagreement
properly, but it is unclear which produces the smallest distortions. In the next two sections I
discuss the methodology to evaluate these IWF variants’ accuracy using Monte Carlo
experiments and their results. I confirm the expectation that some variants produce large errors
and that the rest are still systematically biased against students contributing more. Nevertheless,
but some of these distortions are smaller than expected by scholars proposing newer alternatives.
In real cases in which assessors disagree, it is impossible to know what the real contribution
of each student was. This makes the evaluation of IWF variants’ accuracy difficult, as there is no
good bar against which to measure the IWF variants’ estimate of the students’ real contribution.
Monte Carlo experiments allow circumventing this situation, providing a more robust assessment
of each IWF variant. The method consists of estimating results from a large sample of machine-
generated random data that follows a patterned distribution. We can do this starting from
randomly generated ‘real contributions’ from which assessments deviate, which is what happens
in real life if we assume there is something like a real contribution. This data allows calculating
each IWF variant’s ‘estimate error’, that is, how much it deviates from the real contribution.
From this sample we can calculate statistics for each IWF variant’s estimate error. Specifically, I
look at the median and the 5th and 95th percentiles for each level of contribution between 0 and
1.55 rounded to .1, where a contribution of 1 (100%) indicates a fare share of the group’s work
and average contribution—Lejk and Wyvill (2001a, 2001b) show there are very few cases of
more extremely high or low contribution, with 99.3% of the cases have a contribution between
.55 and 1.45 in my main sample (see Supplementary Appendix 2).
Additionally, I look at how the error is distributed across levels of self-assessment inflation,
to identify which methods favour students who inflate their self-assessments and how much.
Finally, I also evaluate the Root Mean Square Error (RMSE), computed as NΕ/
, where Ε
is the error and N the number of cases. While this statistic does not specify which students
receive a larger error, it is a good summary of the overall error for the sample.
Because results can vary depending on the dispersion of the contributions and the
assessments, I first generated data with a distribution that would closely match the dispersion of
the real data reported by Lejk and Wyvill (2001a, b, hereafter, L&W distribution) for groups of
five: a standard deviation of contributions of 11.64% and a standard deviation of peer
assessments of the same assessee of 2.88%, with a tendency to inflate self-assessments or deflate
peer-assessments. This deviation of contributions produced an μ(sGrp.IWF-esa) of 9.56%. This is
1.77% higher than the category-based μ(sGrp.IWF-esa) of 7.79% reported by Lejk and Wyvill
(2001a, 558), because holistic assessments produce a standard deviation higher than category-
based assessments in about 1.76%—based on estimations from Lejk and Wyvill’ (2001b)
reported IWF data. The standard deviation of peer-assessment agreement I estimated from Lejk
and Wyvill’ (2001b) holistic assessments data is 2.93%.
Considering an average contribution as 1 or 100%—with assessments varying in 5%
increments—contributions are randomly assigned with a normal distribution capped at 0% and
200%, and later scaled so that the average for each group is 100%. The between-group
heterogeneity of contributions within a group, the within-group contributions, and peer
assessments of a single peer have a normal distribution.
The inflation of self-assessments tends to be greater among those contributing less, as
reported by Lejk and Wyvill (2001a, 558), producing a difference between μ(sIWF-esa[Grp]) and
μ(sIWF[Grp]) of about 1.28% in groups of five—μ(sIWF-esa[Grp]) and μ(sIWF[Grp]) indicate the
between-group mean of within-group standard deviation of IWF-esa scores and of IWF scores,
respectively. In theory, several distributions could match this requirement. With the formula I
use (see Supplementary Appendix 1), the average self-inflation is 21.45% and the standard
deviation of self-assessment inflation is 9.14% in the L&W distribution.Graphs and tables with
more details about the distributions can be found in the online Supplementary Appendixes 2-4
(all supplementary appendixes are available through the link
I do most of the analysis with groups of four students with a similar distribution for three
reasons: distortions are more evident in smaller groups, the literature commonly illustrates
analyses in groups of four, and Monson (2017) finds that group of four or more are more
effective than groups of three. Nonetheless, I repeat the analysis for smaller and larger group
Because the dispersion of the data may differ in each class, I repeat the test for three levels
of agreement in the assessments, three levels of average dispersion of contributions, and three
levels of self-assessment inflation. This should provide more robust findings. I vary the levels of
agreement, dispersion of contributions, and self-assessment inflation to about half and to about
double than the L&W distribution—but in the case of self-assessment inflation, to 0 and about
1.5 times the mean of the L&W distribution, to show more relevant and realistic cases—these
factors are approximate, as the resulting distribution also depends on the combinations of other
distribution variables and group sizes. The results for each IWF variant are rather similar across
scenarios, so I only present some variations here with the most important differences. The rest
can be replicated automatically by running the Stata code available in the online Supplementary
Appendix 5.
Even if most classes were to approximate Lejk and Wyvill’s data, it is possible that some of
these other distributions are more realistic for another reason. The dispersion of contributions
and the inflation of self-assessments were computed in relation to a ‘real contribution’ in Lejk
and Wyvill’s data but on their IWF and IWF-esa results respectively. So if, for example, my
results show that the IWF diminishes the dispersion of estimated contributions, then a sample
with a higher dispersion of contributions would better represent the real dispersion of
contributions in Lejk and Wyvill’s data.
One limitation of the sample is that it does not distinguish disagreement among assessors
about who contributed more from disagreement based on someone overrating or underrating
everyone. Therefore, if in real courses much of the disagreement comes from some individuals
generally underrating or overrating everyone, the distribution of disagreements may not be
normal, as the L&W distribution assumes. Consequently, my results may underestimate the
distortions produced by the methods that do not normalise assessments. With these methods, one
case of overrating in a group can generate large distortions.
I included the IWF variants from the last decade (the ac-IWF, ASNIWF, and it-IWF), plus
the original IWF and its simplest and possibly most influential variant, the NIWF. I considered
each variant in a version with self-assessments and one excluding them (indicated with the suffix
-esa). In the case of the it-IWF, I also analyzed the possible corrections suggested above, namely,
using normalised assessments and a corrected b in the computation (it-IWF2).
To compute the IWF-esa and the it-IWF-esa, which use non-normalised peer-assessments,
excluding self-assessments, I used ‘deflated peer-assessments’ to simulate the effect analogous to
inflating self-assessments allowed by the lack of normalization. Deflated peer-assessments are
the same as the normalised peer-assessments when self-assessments are considered in the
normalization, rounded to the nearest 5%.
RMSE analysis reported in Table 2 show that the IWF-esa and the it-IWF-esa produce the
smallest error in all scenarios and on average among all scenarios (average RMSE = .010), with
the former performing slightly better in smaller groups. They are followed closely by the it-
IWF2-esa and the NIWF-esa (average RMSEs of .012 and .013, respectively). The it-IWF2 it-
IWF, IWF and NIWF follow closely. The rest of the variants have substantially larger RMSEs,
ranging between .032 and .064. As we would expect, all tend to have higher RMSE in cases of
high assessor disagreement. They also tend to have a larger error in groups of three and smaller
in groups of seven, although the IWF-esa performs well in small groups. More importantly, the
IWF, NIWF, and it-IWF have large errors in cases of high dispersion of contributions (> .031),
indicating that they are less stable in certain contexts. Since it is especially important to
differentiate between contributions when they are large, these high errors make these estimators
Table 2 here
Note that the it-IWF2-esa’s already complex formula had to be adjusted. This was because
in cases of low differences in contribution and low disagreement within a group, the iteration
formula works with extremely small numbers and by rounding it can generate a division by 0
that generates distorted it-IWF2-esa. Without this correction, the it-IWF2-esa generated an
RMSE of .050 in the L&W distribution and an RMSE of .115 in the sample with low
disagreement. Such cases would be easily identified by instructors, but would entail additional
work. This should warn us about the potential disadvantages of complex algorithms.
Figure 1 shows the estimate error of several IWF variants by level of contribution for the
main sample, that is, groups of four students and L&W distribution. It shows that most variants
that include self-assessments, as well as the ac-IWF and ASNIWF, produce a broadly spread
error or systematically favours those who contribute less. The IWF-esa, it-IWF-esa, and it-IWF2-
esa perform almost identically, favouring low contribution students about .04 more than high
contribution students, with a 5th-95th percentile error range of ±.02 at each level of contribution.
The NIWF-esa error range is slightly narrower than that of the IWF-esa, it-IWF-esa, and it-
IWF2-esa, but it penalises high contribution students substantially more.
Figure 1 here
While their error’s range is narrow, the IWF-esa and the it-IWF-esa allow students to inflate
their grade by deflating the assessments of their peers’ work. Figure 2 reports the error of the
estimate for the IWF-esa and it-IWF-esa, which are nearly identical. With both estimators,
students tend to increase their weighting factor in slightly more than .01 if they deflate their peer
assessments in .04—total adding all peer assessment deflations, equivalent to inflating their self-
assessment in .04. Note that only about 1.25% of the students would do more than that in the
L&W sample. This is significantly less than their similar variants that include self-assessments.
Nonetheless, in this regard the NIWF-esa and it-IWF2-esa are better estimates, as they are
immune to inflation of self-assessments.
Figure 2 here
As the RMSE analysis suggests, the patterns are rather consistent across scenarios,
especially for the best estimators (graphs can be replicated with the code in Supplementary
Appendix 5). Thus, one of the oldest and simplest methods, the IWF-esa has the best
performance. The sophistication of the it-IWF-esa, supposedly aimed at increasing accuracy,
produces no better results, indicating that the simpler IWF-esa should be preferred. However,
both are slightly sensitive to efforts biased deflation of peer assessments that could increase the
assessor’s grade. The it-IWF2-esa performs almost as well, but is immune to deflation of peer
assessments. However, it performs nearly as well in best case scenario distributions, while its
error is 56% and 60% larger in the L&W distribution and the high dispersion of contribution
samples, respectively. Additionally, another old method, the NIWF-esa performs almost as well
as the IWF-esa but, while it prevents self-assessment inflation, it favours low-contribution
students to the detriment of those who contribute more at a more substantial rate. The minimal
loss in accuracy might be compensated by the benefit of not favouring efforts to boost one’s
grade by deflating peer assessments, especially for larger groups, where it performs just as well
as the IWF-esa. Other it-IWF, ac-IWF, and ASNIWF variants actually increase the distortions
produced, which should caution against their use.
While the error of the IWF-esa, NIWF-esa, and it-IWF-esa are fairly small, graphs in
Figures 3-5 comparing them can better illuminate which one should be preferred. In groups of
four that approximate the L&W distribution, a student who deflates her peers’ assessments in .4
(total) to inflate her grade will likely increase her grade in about .01. Fair assessors are likely to
be penalised about .01 (Figure 2). In groups of five, fair assessors are penalised .01 but other
students tend to boost their grade less than .01. The effects are negligible for larger groups
(Figure 3). Low contribution students’ grade will likely be inflated—in the median case—but the
inflation will be near 0 for near-average contributors and for complete free-riders, and about .02
for students contributing half of their share. High contribution students are likely to be penalised
less than .02 (Figure 2). Both of these figures are reduced to about .01 in groups of five students,
and are smaller for larger groups (Figure 4). Additionally, in 90% of the cases, the error
introduced by assessor disagreement is less than ±.02 in groups of four (Figure 2) or slightly
wider than ±.01 in groups of six (Figure 4).
Figures 3-5 here
Given that the it-IWF2-esa’s advantage over the IWF-esa is that it is immune to the effect of
deflating peer assessments but this effect is so small, the difference is unlikely to be worthwhile
for instructors. Additionally, much of that advantage is counteracted by the it-IWF2-esa’s
slightly higher overestimation of the work of low-contribution students to the detriment of grade
of high-contribution students (Figure 5, left).
The main differences between the IWF-esa and the NIWF-esa are three. First, the NIWF-esa
does not allow students to boost their grades by deflating their peers’ assessments. Second, the
IWF-esa is highly sensitive to the distortions created by over- and under-raters. Third, the NIWF-
esa is likely to penalise a student who contributes 1.4 times his share in about .04 in groups of
four, 0.3 in groups of 5, and .02 in groups of 6 (Figures 2 and 4). Fourth, as a consequence of the
third difference, the NIWF-esa’s error is larger in cases of high dispersion of contributions,
which is when grade distribution becomes more important. The NIWF-esa’s RMSE in this
context more than doubles the IWF-esa’s RMSE, at .021 and .010 respectively. Students
contributing about 50% of their share tend to be favoured more, with a median .04 extra points,
whereas students contributing 1.4 of their share are penalised about .05 (Figure 5, right). Since
the IWF-esa slightly underestimates the dispersion of contributions because it tends to favour
low contribution students and penalise high contribution students, a dispersion slightly higher
than that of the L&W distribution may be more realistic. This makes the behaviour of the
estimates in samples with larger dispersion of contributions a critical condition. The fourth
difference may seem to make the IWF-esa a better choice. However, this depends on whether
disagreement is normally distributed or skewed by a few over- or under-raters. If the latter is
true, the IWF-esa’s distortions from under- and overrating may be less acceptable than the still
moderate distortions that the NIWF-esa produces in cases of high dispersion of contributions.
Concluding remarks
Several scholars have proposed ways of computing individual grades in group projects to prevent
free-riding and unfair grades. However, all methods have their drawbacks and the methods’
accuracy has not been adequately tested. Some problems are common to all methods. There are
some agreements in the literature about how to improve the validity of peer assessments, but
there are also disagreements about some alternatives. A new policy identified here to address the
validity of peer assessment instruments is to use scales that start from 0. Starting from 1, as is
often done, unjustifiably benefits students who contribute less to the detriment of those who
contribute more. Other problems, such as that of collusion, cannot be simply solved with a
formal instruction or algorithm, and have to be addressed on a case-by-case basis by instructors.
Yet the topic that has generated most debate is how to compute an individual grade after
obtaining the best peer assessments we can. This article evaluated the most common and recent
computation methods through Monte Carlo experiments.
Results show that newer methods produce large median estimate errors or broadly spread
estimate errors, whereas the older IWF-esa produces fairly small distortions and this is rather
consistent across multiple scenarios. Paradoxically, most methods designed to address issues of
bias usually produced the largest errors—including Ko’s version of the iterative method, the it-
IWF, with self-assessments. The best methods were the simplest ones, the IWF-esa and NIWF-
esa, along with the about equally accurate but substantially more complex—and thus not worth
the trouble—it-IWF-esa and it-IWF2-esa.
However, no method is perfect. The IWF-esa produces low distortions overall, but is
sensitive to distortions produced by over-raters and under-raters. Until studies identify the real
distribution of disagreements, a more conservative approach would be to prefer the NIWF-esa, at
least when there seems to be disagreement due to overrating or underrating—although this is
sometimes difficult to identify. My personal experience is that in every large class, at least one
student clearly overrates everyone else, which would create a large distortion for that group if I
used the IWF-esa. Even when the average level of distortions in the course is smaller with the
IWF-esa, if they are large within one group, instructors may prefer the more moderate distortions
of the NIWF-esa.
Either way, all methods will tend to generate some distortions, even if they are usually small
ones. Students and instructors need to acknowledge that this is part of the process of team work,
where it is nearly impossible to exactly quantify how much each member contributed to the final
result. In fact, outside of academia individual contributions to team work are also imperfectly
assessed, if they are at all, and students should be prepared for this.
To summarise, Monte Carlo simulations have shown that sometimes the complex algorithms
designed to increase accuracy actually produce large distortions. The size of these distortions
should serve as a cautionary tale against testing methods only by their correlations with grades, a
few ideal cases, and students’ opinions. Future variants of the IWF method should always
compare the results they would produce against the ‘real contribution’ of students in simulated
data and should pay attention to where the distortions concentrate. Scholars and instructors
should also consider this article’s results as a reminder that sometimes simpler is better—and
usually no method is perfect.
Baker, D. F. 2007. "Peer assessment in small groups: a comparison of methods." Journal of
Management Education 32 (2):183-209.
Bushell, G. 2006. "Moderation of peer assessment in group projects." Assessment & Evaluation
in Higher Education 31 (1):91-108.
Carvalho, A. 2013. "Students' perceptions of fairness in peer assessment: Evidence from a
problem-based learning course." Teaching in Higher Education 18 (5):491-505.
Conway, R., D. Kember, A. Sivan, and M. Wu. 1993. "Peer assessment of an individual‘s
contribution to a group project." Assessment & Evaluation in Higher Education 18
Dommeyer, C. J. 2012. "A new strategy for dealing with social loafers on the group project: The
segment manager method." Journal of Marketing Education 34 (2):113-27.
Feichtner, S. B., and E. A. Davis. 1984. "Why some groups fail: A survey of students'
experiences with learning groups." Organizational Behavior Teaching Review 9 (4):58-
Goldfinch, J. 1994. "Further developments in peer assessment of group projects." Assessment &
Evaluation in Higher Education 19 (1):29-35.
Goldfinch, J., and R. Raeside. 1990. "Development of a peer assessment technique for obtaining
individual marks on a group project." Assessment & Evaluation in Higher Education 15
Jaques, D., and G. Salmon. 2007. Learning in Groups: A Handbook for Face-to-Face and Online
Environments. Abingdon, UK: Routledge.
Jin, X.-H. 2012. "A comparative study of effectiveness of peer assessment of individuals’
contributions to group projects in undergraduate construction management core units."
Assessment & Evaluation in Higher Education 37 (5):577-89.
Ko, S.-S. 2014. "Peer assessment in group projects accounting for assessor reliability by an
iterative method." Teaching in Higher Education 19 (3):301-14.
Lejk, M., and M. Wyvill. 2001a. "The effect of the inclusion of selfassessment with peer
assessment of contributions to a group project: A quantitative study of secret and agreed
assessments." Assessment & Evaluation in Higher Education 26 (6):551-61.
———. 2001b. "Peer assessment of contributions to a group project: A comparison of holistic
and category-based approaches." Assessment & Evaluation in Higher Education 26
Li, L. K. Y. 2001. "Some refinements on peer assessment of group projects." Assessment &
Evaluation in Higher Education 26 (1):5-18.
Loddington, S., K. Pond, N. Wilkinson, and P. Willmot. 2009. "A case study of the development
of WebPA: An online peer‐moderated marking tool." British Journal of Educational
Technology 40 (2):329-41.
Macfarlane, B. 2016. "The performative turn in the assessment of student learning: A rights
perspective." Teaching in Higher Education 21 (7):839-53.
Monson, R. 2017. "Groups that work: Student achievement in group research projects and effects
on individual learning." Teaching Sociology 45 (3):240-51.
Neus, J. L. 2011. "Peer assessment accounting for student agreement." Assessment & Evaluation
in Higher Education 36 (3):301-14.
Ohland, M. W., M. L. Loughry, D. J. Woehr, L. G. Bullard, R. M. Felder, C. J. Finelli, R. A.
Layton, H. R. Pomeranz, and D. G. Schmucker. 2012. "The comprehensive assessment of
team member effectiveness: development of a behaviorally anchored rating scale for self-
and peer evaluation." Academy of Management Learning & Education 11 (4):609-30.
Sharp, S. 2006. "Deriving individual student marks from a tutor’s assessment of group work."
Assessment & Evaluation in Higher Education 31 (3):329-43.
Spatar, C., N. Penna, H. Mills, V. Kutija, and M. Cooke. 2015. "A robust approach for mapping
group marks to individual marks using peer assessment." Assessment & Evaluation in
Higher Education 40 (3):371-89.
Tu, Y., and M. Lu. 2005. "Peer-and-self assessment to reveal the ranking of each individual's
contribution to a group project." Journal of Information Systems Education 16 (2):197-
Webb, N. M. 1995. "Group collaboration in assessment: Multiple objectives, processes, and
outcomes." Educational Evaluation and Policy Analysis 17 (2):239-61.
Zhang, B., and M. W. Ohland. 2009. "How to assign individualized scores on a group project:
An empirical evaluation." Applied Measurement in Education 22 (3):290-308.
Table 1. Example of IWF variants’ results using Spatar et al’s (2015) Table 5 data
Assessee (j)
Row sum
Original Raw Assessments (RA)
Assessor (i)
A 20 20 20 20 80
B 4 16 17 15 52
C 4 15 18 14 51
D 4 15 17 15 51
IERs [=∑RA] 32 66 72 64 58.5
IWF [=IER/AER] .55 1.13 1.23 1.09
IERs-esa (excl. self-assmnt.) [=∑RA-esa] 12 50 54 49 41.25
IWF-esa [=IER-esa/AER-esa] .29 1.21 1.31 1.19
Normalised assessments (NAs) [RA/∑RAi] σi
A .25 .25 .25 .25 1.00 .00
B .08 .31 .33 .29 1.00 .10
C .08 .29 .35 .27 1.00 .10
D .08 .29 .33 .29 1.00 .10
NIWF [=∑NAj/∑NAi] .48 1.15 1.26 1.11
ac-IWF computation using NAs max(sj)
Assessee standard deviation [sj] .09 .03 .05 .02 .07
IAF [=1–sj(k)/ max(sj)] .00 .71 .47 .77
ac-IWF [=IAF×(NIWF–1)+1] 1.00 1.10 1.12 1.08 4.31
ASNIWF computations using NAs
SAIF [=1–sj(k)/(2×max(sj))] .50 .85 .74 .89
ASNIWF [=SIAF×(NIWF–1)+1] .74 1.12 1.19 1.09 4.16
Normalised assmnts., excl. self-assmnt. (NA-esa) σi
A 0.33 0.33 0.33 1.00 .00
B 0.11 0.47 0.42 1.00 .16
C 0.12 0.45 0.42 1.00 .15
D 0.11 0.42 0.47 1.00 .16
NIWF-esa [=∑NA-esa/∑NAi] .34 1.20 1.28 1.17 4.00
ac-IWF-esa computation using NAs-esa max(sj)
Assessee standard deviation [sj] .00 .06 .08 .05 .08
IAF-esa [=1–sj(k)/ max(sj)] .93 .23 .00 .37
ac-IWF-esa [=IAF-esa×(NIWF-esa–1)+1] .39 1.05 1.00 1.06 3.50
ASNIWF-esa computations using NAs-esa
SIAF-esa [=1–sj(k)/(2×max(sj))] .96 .61 .50 .69
ASNIWF-esa [=SIAF-esa×(NIWF-esa–1)+1] .37 1.13 1.14 1.12 3.75
it-IWF* .33 1.19 1.34 1.14 4.00
it-IWF using NAs* .48 1.15 1.27 1.11 4.00
it-IWF using b=μ(σi
2)* .39 1.17 1.31 1.13 4.00
it-IWF using NAs and b=μ(σi
2)* .42 1.17 1.30 1.12 4.00
it-IWF using NAs and b=μ(σi
2)/10 (i.e. it-IWF2)* .34 1.19 1.34 1.13 4.00
it-IWF-esa* .31 1.19 1.33 1.16 4.00
it-IWF2-esa* .32 1.20 1.31 1.17 4.00
* See Ko (2014) for computation details.
Table 2. RMSE of twelve IWF variants in nine scenarios
Scenario IWF IWF-
it-IWF2 it-IWF2-
ac-IWF ac-IWF-
L&W distribution 0.017 0.009 0.017 0.011 0.016 0.009 0.014 0.014 0.061 0.050 .038 .030
Dispersion of contrib.
Low 0.011 0.008 0.011 0.008 0.010 0.008 0.015 0.008 0.031 0.024 .020 .015
High 0.033 0.010 0.033 0.021 0.031 0.010 0.016 0.016 0.124 0.104 .078 .062
Low 0.017 0.009 0.017 0.011 0.016 0.008 0.013 0.009 0.061 0.050 .038 .030
High 0.022 0.020 0.022 0.021 0.021 0.020 0.030 0.024 0.061 0.049 .038 .030
Self-asmnt. inflation
Low 0.006 0.007 0.006 0.011 0.006 0.007 0.007 0.009 0.055 0.049 .028 .030
High 0.023 0.010 0.023 0.011 0.021 0.010 0.017 0.009 0.062 0.050 .041 .030
Group size
3 0.020 0.011 0.020 0.020 0.020 0.014 0.023 0.015 0.062 0.041 .040 .032
7 0.011 0.006 0.011 0.006 0.011 0.007 0.007 0.006 0.058 0.062 .0.34 .032
Mean 0.018 0.010 0.018 0.013 0.017 0.010 0.016 0.012 0.064 0.053 .040 .032
Figure 1. Estimate error’s median and 5th and 95th percentiles by level of contribution (rounded to .1) in groups of four members, for twelve IWF
variants (continues on next page).
Figure 1. (continued)
Figure 2. Estimate error’s median and 5th and 95th percentiles by level
of self-assessment inflation in groups of four members for IWF-esa
and it-IWF-esa
Figure 3. Estimate error’s median and 5th and 95th percentiles by level
of self-assessment inflation (rounded to .1) for the IWF-esa in groups
of five (grey) and six (black)
Group size: 5 Group size: 6
Figure 4. Estimate error’s median and 5th and 95th percentiles by level of contribution (rounded to .1) for the IWF-esa and NIWF-esa
in groups of five (left) and six (right)
Figure 5. Estimate error’s median and 5th and 95th percentiles by level of contribution (rounded to .1) for the it-IWF2-esa, IWF-esa,
and NIWF-esa in sample with high dispersion of contribution and groups of four
... This research's findings indicated that a single mark should not be allocated to all students in a group if free-riding issues are present [65,89]. Giving the same mark for each individual is perceived as unfair, making hardworking students unhappy and resentful, while it also sends the wrong message to slackers [90]. ...
Full-text available
Group-based assessments have been recognized as one of the ways of developing work-ready attributes in project management graduates. This paper is aimed at exploring the best practices for improving the administration process of group-based assessments to make it fit for purpose. The study adopted a systematic literature review and 22 semi-structured interviews with project management academics in higher education. The study found that teams should be small in size and formed by the faculty. Tasks should be based on a real-world context and be sufficiently challenging for students. The group should create a team charter to establish expectations and ground rules for group members. Group assessments should include team member peer evaluations and oral presentations to curb social loafing. Academics should be actively involved in giving regular feedback, training students about teamwork, and communicating their expectations to students. The study findings are useful to inform project management academics about the design and administration characteristics that need to be considered to achieve the ultimate purpose of group-based assessments in polishing the job-readiness attributes of project management graduates.
Full-text available
There is a soaring demand for work-ready graduates who can quickly adapt to an ever-challenging work environment. Group-based assessments have been widely recommended as a means to develop the skills required for the world of work. However, group-based assessments are perceived as challenging for both students and educators. This systematic literature review (SLR), based on the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA), focuses on analyzing and synthesizing the existing literature on group-based assessments. A four-step approach was undertaken in order to conduct this research. The SLR identified 71 relevant articles, analyzed using thematic analysis with the aid of NVivo software. An open coding approach was adopted to generate codes. The validity of the SLR process and the reliability of the research tool were maintained through the use of trustworthiness. The findings identified dominant themes such as self- and peer evaluations, training students for group work assessments, group formation, group size, and the role of academics and technology in facilitating group processes. The outcomes of this review contribute significantly to the design and administration of group-based assessments in higher education by providing academics with practical guidelines to effectively facilitate group-based assessments which fit the purpose.
The extraction of individual marks from a group submission is one approach to address the problem of students who don't contribute adequately to group work. This is a highly active research area. An alternative is proposed in this paper: alter the method by which groups are allocated for assessed coursework in tertiary education. The proposed two-step method involves (i) soliciting feedback from the students on the effort invested by their classmates in preliminary assignments and (ii) using these data to allocate groups for a subsequent larger assignment. Students who invest a lot of effort in the preliminary assignments are grouped with similarly motivated students, reducing their dissatisfaction, while students who formerly acted as ‘freeloaders’ are now obliged to engage with the larger assignment and thereby achieve the desired learning outcomes. This paper describes a trial of this group allocation approach for an undergraduate engineering design course of 111 students. © 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
Full-text available
Group research projects frequently are used to teach undergraduate research methods. This study uses multivariate analyses to examine the characteristics of higher-achieving groups (those that earn higher grades on group research projects) and to estimate the effects of participating in higher-achieving groups on subsequent individual learning (grade on final paper). The sample includes 257 students who completed a sociology research methods course at a small liberal arts institution between 2004 and 2015. Group achievement (grade on group research project) is predicted by group size, gender composition of the group, and group’s average grade on midterm exams. Group achievement on the research project contributes to subsequent individual learning (grade on final paper) even after controlling for individual characteristics (student’s gender, race, and average grade on midterm exams) and group characteristics. The findings support claims that group research projects are an effective pedagogy for undergraduate sociology research methods courses and point to some guidelines for instructors’ composition of groups for research projects.
Full-text available
This study proposes an advanced method to factor in the contributions of individual group members engaged in an integrated group project using peer assessment procedures. Conway et al. proposed the Individual Weight Factor (IWF) method for peer assessment which has been extensively developed over the years. However, most methods associated with IWF use a simple average algorithm which still has limitations in effectively reflecting abnormal situations. Therefore, this study proposes an iterative method for measuring a fair IWF by considering each assessor's reliability. The proposed algorithm considers each assessor's reliability which is evaluated by an analysis of variance of the scores given by the assessor. This relative reliability factor is used to improve the fairness of the IWF and reduce the effect of unfair or unreliable assessors' scores. A comparison with previous methods demonstrates that the proposed method resulted in substantially more fair grades being given to individual students for group projects.
Full-text available
This paper describes the experience of peer assessment in a problem-based learning course in a Portuguese university, and presents the results of a survey of 120 students who attended the course over seven semesters. Cluster analysis revealed that students group into two distinct clusters with significantly different experiences of peer assessment. Although the larger cluster (70%) expressed a positive experience, the other group perceived peer assessment results and final marks to be unfair, and reported incidents of friendship-marking and conflict arising from peer assessment. Results show that perceptions of fairness (both regarding peer assessment and final marks) are very much associated with those problems. We reflect on these findings to explore implications for practice as well as future research.
Full-text available
Instructors often incorporate self- and peer evaluations when they use teamwork in their classes, which is common in management education. However, the process is often time consuming and frequently does not match well with guidance provided by the literature. We describe the development of a web-based instrument that efficiently collects and analyzes self- and peer-evaluation data. The instrument uses a behaviorally anchored rating scale to measure team-member contributions in five areas based on the team effectiveness literature. Three studies provide evidence for the validity of the new instrument. Implications for management education and areas for future research are discussed.
Full-text available
Professors often give members of a group project the same grade or they use peer evaluations to provide individualized grades. Unfortunately, both these methods have shortcomings. This article describes the segment manager method (SMM), a method for assigning individualized grades on the group project that does not rely on peer evaluations. A quasi-experiment is described where marketing research students are graded in groups either with peer evaluations (the control group) or with the SMM (the treatment group). Students participating in the quasi-experiment were more likely to prefer the SMM over the approach that uses peer evaluations. Moreover, when the SMM is compared with the peer evaluations approach, the SMM appears to be more likely to generate individualized grades on the group project that reflect each student's academic ability, to deter social loafing, to increase scores on the group project, and to increase scores on an exam that is heavily loaded with project material. Recommendations for future users of the SMM are offered.
Full-text available
Large-scale assessment programs are increasingly requiring students to work in collaborative small groups instead of, or in addition to, requiring them to work individually. This article considers theoretical and practical issues that need to be taken into account in the design, use, and interpretation of the results of such assessments. Theoretical issues include the compatibility of group work with different purposes of assessment and the way emerging group processes may work toward or against different purposes of assessment. Practical issues include designing and administering assessments to produce scores that are consistent with the goals of the assessment, composing groups so that the results are fair, and preparing students for collaborative group assessments.
Active learning and group-based processes in higher education are central to student engagement strategies. Forms of assessment regarded as evidencing student engagement, including attendance, class participation grading and group-based projects, have become commonplace in the university curriculum on an international basis. Whilst the literature has focused on evaluating such forms of assessment in terms of learning gain, analysis of their impact from a student rights perspective has been largely overlooked. This paper will analyse student perspectives of three forms of assessment entailing the measurement of observable student attitudes and behaviour: attendance, class participation and group work grading. The evidence from a survey of undergraduates based in a Hong Kong university suggests that the majority of students are concerned about whether such practices are appropriate and fair, potentially undermining their freedom of choice to learn as adults.
Group work can form a substantial component of degree programme assessments. To satisfy institutional and student expectations, students must often be assigned individual marks for their contributions to the group project, typically by mapping a single holistic group mark to individual marks using peer assessment scores. Since the early 1990s, various mapping methods that use self- and peer ratings have been developed. They are based on (normalised) individual weighting factors, partial scaling of the group mark, inter-rater agreement corrections or parabolic functions. We show that no single existing method can be successfully applied to most practical peer assessment scenarios such as different marking scale interpretations, intra-group ranking errors, biased free-riders and marks exceeding 100%. We present a combined analytical mapping method that incorporates the benefits of existing mechanisms, but alleviates their weaknesses with minimum computational effort and tutor input. The robustness of the method is illustrated through problematic assessment examples and empirically evaluated in multiple group work environments involving a total of 243 students and five different disciplines.
The use of peer assessment as a way of differentiating between individual students on a group project is discussed. A new style of peer appraisal questionnaire for the students to complete is introduced, together with the detailed description of a method of calculating a ‘peer assessment factor’ from these questionnaires. This factor allocates to an individual group member a percentage of the mark awarded to the group's project submission. The results obtained when this scheme was used on a large first year course are discussed, together with some possible modifications and their effects. Examples of the appraisal questionnaire and the calculations are included