Content uploaded by Joshua McGrane
Author content
All content in this area was uploaded by Joshua McGrane on Apr 03, 2018
Content may be subject to copyright.
1
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Abstract
National standardized assessment programs have increasingly included extended written
performances, amplifying the need for reliable, valid and efficient methods of assessment.
This article examines a two-stage method using comparative judgments and calibrated
exemplars as a complement and alternative to existing methods of assessing writing. Written
performances were taken from Australia’s NAPLAN assessment, which included both
narrative and persuasive performances from students aged 8 to 15. In Stage 1, assessors
performed comparative judgments on 160 performances to form a scale of 36 calibrated
exemplars. These comparative judgments showed a very high level of reliability and
concurrent validity. In Stage 2, assessors scored 2380 new performances by matching them to
the most similar calibrated exemplar. These matching judgments showed a generally high
level of reliability and concurrent validity and were reasonably efficient after a
familiarization period. Further research is suggested to enhance Stage 2 by simplifying the
exemplar scale and scaffolding it with detailed descriptors. Overall, the findings support the
use of the method in standardized writing assessment and its application to various learning
areas.
2
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Applying a Thurstonian, two-stage method in the standardized assessment of writing
This article investigates the reliability, validity and efficiency of a two-stage method
of writing assessment, which adapts the pioneering methodology of Thurstone (1928) and
Thurstone and Chave (1929), with respect to written performances from Australia’s large-
scale standardized assessment program; the National Assessment Program - Literacy and
Numeracy (NAPLAN). The two-stage process involves (a) the application of comparative
judgments to a large set of written performances to establish a scale of calibrated exemplars,
and (b) the subsequent use of these calibrated exemplars to assess new written performances
through matching judgments. This two-stage method combines the benefits of comparative
judgments in creating a reliable and valid ‘ruler’ of writing quality, where each calibrated
exemplar represents a different level of quality to match with new performances, with the
greater efficiency of matching judgments since comparative judgment designs are often
criticized as highly time consuming. The two-stage method has been successfully applied to
the assessment of writing in early childhood by classroom teachers with very high levels of
reliability and concurrent validity observed (Heldsinger & Humphry, 2013). The current
research extends the application of the method to the standardized assessment context,
including both narrative and persuasive performances written by students across grades 3, 5,
7, and 9 (age range 8 to 15), and implements two criteria for the judgments in both stages of
the process. The aim is to establish the reliability and validity of the two-stage method to
explore its viability as a complement or alternative to current scoring methods in the
standardized assessment of student writing.
Background
As governments across the world increasingly invest in a range of large-scale
assessment programs with the aim of enhancing accountability in their educational systems,
the robustness of the standardized assessment of writing is becoming an increasingly
3
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
pertinent issue. Moreover, there has been a push for such assessments to develop beyond
traditional multiple-choice formats to performance-based assessments that more closely align
with the construct being investigated (Shermis, 2014). For example, under the Common Core
State Standards initiative, which is presently transforming the educational assessment
landscape in the United States, repeated, extended written performances are a central part of
the formative and summative assessment of various content areas (Shermis, 2014). In the
United Kingdom, extended written performances are a part of the various Key Stage
assessments, and existing writing assessment practices have attracted a large amount of
scrutiny and criticism (House of Commons Education Committee, 2017). Moreover, in the
Australian context, extended written performances are already a core aspect of the annual
National Assessment Program—Literacy and Numeracy (NAPLAN) assessments (Wyatt-
Smith & Jackson, 2016).
The complexities associated with the standardized assessment of writing are
intertwined with the complexities of the skill itself. Writing is a foundational aspect of
students’ literacy involving a complex set of cognitive processes and physical skills, which
enables students to express, communicate, make connections and construct meaning in print
(Hayes & Berninger, 2014; Mackenzie, Scull, & Munsie, 2013). Writing is also an integral
part of schooling, as students beyond the age of eight may spend up to half of their classroom
hours engaged in writing tasks across a range of learning areas, despite spending very
minimal classroom time on learning and engaging in writing in the earlier grades of schooling
(Brindle, Graham, Harris, & Hebert, 2016; McHale & Cermak, 1992). The minimal attention
paid to the learning of writing in early schooling is noteworthy given that education systems
across the world have increasingly expressed concern regarding declines in students’ writing
achievement (Mackenzie et al., 2013; Wyatt-Smith & Jackson, 2016), which has also
motivated the broadening of standardized assessments beyond the traditional tests of
4
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
mathematics and reading (Mackenzie et al., 2013). With the increase in the volume of written
performances in high-stakes standardized assessments, there is a growing need to develop
methods that optimize the validity and reliability of writing assessment, as well as the
efficiency with which they are implemented by assessors and teachers.
Standardized writing assessment, to date, has been dominated by analytic, rubric-
based methods where written performances are rated on multiple criteria intended to cover
the scope of the writing construct (Spandel, 2005). These rubric-based methods are also used
across a range of learning areas throughout the school years and across a range of disciplines
in higher education (Heldsinger & Humphry, 2013; Rezaei & Lovorn, 2010; Sadler, 2009;
Tierney & Simon, 2004). It has been argued that rubric marking carries a number of
advantages such as providing criterion-level diagnostic information, bridging formative and
summative forms of assessment, and reducing subjectivity with high levels of reliability
between raters (Brookhart & Chen, 2014; Jonsson & Svingby, 2007; Reddy & Andrade,
2010). However, a number of authors have argued that the validity and reliability of the
rubric approach have not been adequately empirically studied (Heldsinger & Humphry, 2013;
Humphry & Heldsinger, 2014; Rezaei & Lovorn, 2010).
Specifically, rubric-based assessment has been shown to be affected by numerous
rating tendencies, including rater leniency, central tendency, restriction of range and the halo
effect (Myford & Wolfe, 2003). Humphry and Heldsinger (2014) provided evidence that the
common, grid-like structure of rubrics, where each criterion has a common number of
categories, induces violations of local independence across ratings, which undermines
validity by limiting construct-relevant variation in scores. Rezaei and Lovorn (2010) showed
that if raters were not well trained in employing a rubric effectively, their assessments were
biased toward the more mechanical features of writing. This latter point is particularly
pertinent to large-scale, high-stakes standardized writing assessments, as effective assessor
5
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
training requires time and resources. In response to some of these limitations, a number of
authors have proposed the application of comparative judgments as an alternative approach to
writing assessment, and to performance assessments more generally (Bramley, Bell, & Pollitt,
1998; Heldsinger & Humphry, 2010, 2013; Pollitt, 2012; Steedle & Ferrara, 2016; Tarricone
& Newhouse, 2016).
The Comparative Judgment and Two-Stage Approaches to Writing Assessment
The theoretical background and practical applications of the comparative judgment
approach is overviewed in detail elsewhere (Andrich, 1978; Bond & Fox, 2012; Bramley et
al., 1998; Heldsinger & Humphry, 2010; Thurstone, 1927). The approach was proposed by
Thurstone (1927) as a method to scale the perceived property of a set of stimuli (e.g., their
perceived mass) by statistically modelling comparisons of the property of the stimuli in pairs.
In the present context, a set of writing performances may be scaled in terms of their perceived
quality by comparing them in pairs and inferring the scale locations from the proportion of
judgments in favor of each performance versus each other performance. While there have
been few educational applications of the comparative judgment approach compared to other
methods of writing assessment, Heldsinger and Humphry (2010, 2013) demonstrated that it
could be used to order performances in a highly reliable and valid manner across a range of
learning areas.
A distinctive advantage of comparative judgments over rubric ratings is that teachers
directly compare one performance against another so that there is reduced possibility for
assessor harshness or lenience to affect judgments (Andrich, 1978; Heldsinger & Humphry,
2010). Similarly, because there is no rating required, comparative judgments are not subject
to the biasing effects of different rater tendencies, or the response dependencies that are
induced by structural features of the design of rubrics (Humphry & Heldsinger, 2014).
6
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Heldsinger and Humphry (2010) also showed that the comparative judgment approach
could be efficiently implemented with a very high level of reliability across different content
areas in a classroom context in the absence of the extensive training processes often
employed in large-scale testing programs. The primary disadvantage of comparative
judgments, however, is that they can be time-consuming and therefore cognitively intensive
for the judges (Bramley et al., 1998; Pollitt, 2012; Steedle & Ferrara, 2016).
To overcome this disadvantage, Heldsinger and Humphry (2013), drawing upon the
work of Thurstone (1928) and Thurstone and Chave (1929), implemented a two-stage process
designed to exploit the above advantages of comparative judgments while attaining greater
efficiency. As briefly described above, the first stage of the process is to create a scale of
exemplars by pairwise comparisons of written performances. Subsequently, teachers assess
new performances using this scale of exemplars by matching the quality of the novel written
performance with the most similar calibrated exemplar on the pre-existing scale. Heldsinger
and Humphry (2013) applied this two-stage method in the assessment of narrative writing for
early childhood and, despite very minimal training, found very high levels of inter-judge
reliability (the average was .923). The authors also cross-referenced assessments using the
two-stage method with rubric ratings and found a high level of concurrent validity (r = .895).
Finally, Heldsinger and Humphry (2013) observed that the method was a reasonably efficient
way of obtaining highly reliable judgments with minimal training, as teachers on average
were able to assess approximately 20 written performances per hour.
The present study aims to extend the findings of Heldsinger and Humphry (2013) in
several important respects. First, Heldsinger and Humphry (2013) applied the method in a
classroom context, with an early childhood student cohort and only a single writing genre.
The current research applies the method to performances taken from NAPLAN over two
calendar years, including performances by students aged 8 to 15, and both narrative and
7
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
persuasive genres. Moreover, Heldsinger and Humphry (2013) implemented the method
using a single holistic judgment. The present study extends this research by implementing
two criteria, writing conventions and authorial choices, in both stages of the two-stage
process. The inclusion of separate judgments for these two criteria was based on feedback
from teachers that, in some instances, it was difficult to reconcile these two broad aspects of
the written performances to make a holistic comparison. The study therefore also extends the
existing research on the use of comparative and matching judgments for writing assessment,
which have generally relied upon holistic judgments, by examining the two-stage method
using two criterion judgments. These two judgments, in turn, provide more information about
the quality of the written performance. The method is investigated in terms of reliability,
concurrent validity and efficiency, to ascertain whether it constitutes a viable complement or
alternative to existing methods in the standardized assessment of student writing.
The study
Procedure – Overview
In the first stage of the method, both narrative and persuasive writing performances
were compared by experienced assessors using the method of pairwise comparison
(Thurstone, 1927). Comparative judgments were made with respect to two broad criteria: (i)
authorial choices, which includes features of writing where the writer is free to make choices
including subject matter, language choices, development of tone, style, voice and reader-
writer relationship; and (ii) conventions, where the writer is expected to largely follow rules,
including spelling, punctuation, correct sentence formation, and clarity of referencing.
Assessors were instructed to make an on-balance comparison for each broad criterion
judgment, rather than favoring any particular aspect of the criterion, which is consistent with
previous research using comparative judgments for writing assessment (Heldsinger &
Humphry, 2010, 2013). Only these concise criteria were provided to orient assessors to the
8
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
key features of performances that they need to focus on when judging which in a given pair is
better because, unlike rubric-style approaches, this process does not necessitate the
description of gradations or levels for each criterion. Nonetheless, while concise criteria were
used in the present context, all assessors involved in the study were able to draw from their
familiarity with the writing marking guides for NAPLAN (Australian Curriculum,
Assessment and Reporting Authority (ACARA), 2010, 2013).
The NAPLAN Narrative and Persuasive writing marking guides include ten marking
criteria and only differ with respect to a single criterion (ACARA, 2010, 2013). The first six
criteria of each guide are subsumed within the broad ‘authorial choices’ criterion of the
present study, including ‘Audience’, ‘Text structure’, ‘Ideas’, ‘Character and setting’ (in the
Narrative guide only), ‘Persuasive devices’ (in the Persuasive guide only), ‘Vocabulary’, and
‘Cohesion’. The last four criteria of each guide are subsumed within the broad ‘conventions’
criterion of the present study, including ‘Paragraphing’, ‘Sentence structure’, ‘Punctuation’,
and ‘Spelling’. For further information see ACARA (2010, 2013).
The comparative judgment data were analyzed using the Bradley-Terry-Luce (BTL)
model (Bradley & Terry, 1952; Luce, 1959) to develop authorial choices, writing conventions
and combined criteria scales on which the ordering of location estimates represent the judged
ordering of the quality of the written performances on the respective criteria and overall. The
combined criteria scale estimates were obtained by combining all comparative judgments in
the BTL model analysis, irrespective of which criterion was used for the judgment. Moreover,
performances across all grades and across both genres were simultaneously calibrated, which
is consistent with the operational NAPLAN procedures where students across the grades in a
calendar year respond to a common prompt and performances are scored on a common
NAPLAN writing scale, irrespective of the writing genre. Humphry and McGrane (2015) also
show how comparative judgments are operationally used to longitudinally equate the
9
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
NAPLAN writing scale in the absence of a common-person or common-item equating
design. After calibration was complete, a subset of performances representing the full range
of performances from the combined criteria scale was then selected for use as calibrated
exemplars.
In Stage 2, ACARA engaged experienced assessors to score a new pool of narrative
and persuasive written performances from two years of the NAPLAN writing assessment.
The assessors scored each of the new written performances by judging which calibrated
exemplar from Stage 1 was most alike the new performance, and the latter was assigned the
scale estimate of the matched calibrated exemplar. Each assessor made a matching judgment
of each performance separately for the two criteria and they were instructed to make an on-
balance judgment for each broad criterion.
The Stage 2 data were examined to establish the inter-judge reliability of the matching
judgments in terms of the two criteria and a combined scale. In addition, the concurrent
validity of the matching data was assessed by comparing the estimates with a traditional
rubric score assigned to the same performances by a different set of highly experienced
assessors using the above described NAPLAN writing marking guides (ACARA 2010, 2013).
Finally, the efficiency of the methodology was examined by inspecting the duration of the
matching judgments over the full set of judgments by the assessors.
Stage 1 – Calibration of the Exemplars
Stage 1 method. The results of Stage 1 are summarized as the Stage 2 results are the
focus of this article.
Participants. The student performances were sampled from two years (2011 and
2012) of NAPLAN. The NAPLAN writing tests are administered under standardized
conditions according to protocols. Time is allowed to each student in each grade level for
10
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
planning (five minutes), writing (30 minutes) and editing (five minutes). In a given calendar
year, every student from each grade level received the same writing prompt/topic.
A total of 3000 performances were obtained from a clustered random sample across
grades 3, 5, 7 and 9. From this larger sample, a subsample of 160 performances, including 80
narrative performances from the 2011 NAPLAN writing assessment and 80 persuasive
performances from the 2012 NAPLAN writing assessment, was obtained for the comparative
judgment stage of the study. This subsample was selected to have an approximately uniform
distribution with between 1 and 8 performances for each score in the range from 0-48 (min-
max possible score) on the NAPLAN rubric. The subsample was also drawn to have
approximately equal numbers of students from each grade level (grade 3 n = 43; grade 5 n =
38; grade 7 n = 38; grade 9 n = 41).
Participating assessors. Twenty-one expert assessors drawn from a pool of
experienced NAPLAN assessors in one Australian State were involved as judges in Stage 1.
Assessors made independent judgments either in a central venue or at other sites.
Comparative judgment procedure. The number of possible paired comparisons in
which each performance is compared with every other, is 160(160-1)/2=12720. Given it is
not possible to efficiently make this number of comparisons, a design was constructed so that
across the 21 assessors, each performance was compared with an average of 40 other
performances, giving an average of 80 judgments across the two criteria. This number of
comparisons was used to attain sufficiently small standard errors of measurement for
selecting ordered, calibrated exemplars to be used in Stage 2. The specific pairs compared by
each individual assessor were randomly selected from the list of all pairs.
Data analysis. The authorial choices, conventions and combined scales were
produced by analyzing the criterion-level and combined comparative judgments using the
BTL model. This model is essentially the same as Thurstone’s (1927) Case V of the Law of
11
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Comparative Judgment, but the logistic function is substituted for the cumulative normal
function. The BTL is also mathematically equivalent to the conditional form of the Rasch
model when the item parameter has been eliminated (Andrich, 1978). The BTL model is
pij=Pr
(
i>j
)
=exp
(
θi−θj
)
1+exp
(
θi−θj
)
(1)
where, in the case of written performance judgments,
θi
and
θj
in Equation 1 represent
the level of writing quality instantiated by the performances
i
and
j
, and
pij
is the
probability that performance
i
is judged better than performance
j
. In general terms,
written performances with a low proportion of favorable judgments obtain more negative
estimates, performances with a high proportion of favorable judgments obtain more positive
estimates, and performances that have a similar proportion of favorable judgments receive
similar estimates. The BTL model parameters were estimated using the PairWise software
(Holme & Humphry, 2008), which implements maximum likelihood estimation and
constrains the mean of the estimates to zero. The PairWise software also provides a Person
Separation Index (PSI), which is analogous to Cronbach’s (1951) alpha coefficient (Andrich,
1988) and provides an index of the internal consistency of the judgments as a whole. The PSI
is defined as the ratio of true variance to observed variance in the writing quality estimates,
PSI=var
[
^
θ
]
−MSE
var
[
^
θ
]
(2)
where
MSE=∑
n
σn
2
is the mean square error of the writing quality estimates from the BTL model. The PSI ranges
in value from 0 to 1, with higher values indicating greater reliability in the comparative
judgments.
12
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Stage 1 results.
Comparative judgment reliability. The PSI was .971 for the authorial choices scale, .
967 for the conventions scale, and .982 for the combined scale. These results indicated that
all judgments had a very high level of reliability. The PSIs for the specific criterion and
performance genre combinations are provided in Table 1. These similarly showed a very high
level of reliability, ranging from .971 to .983 for the narrative performances across the
criteria, and .962 to .980 for the persuasive performances across the criteria. The PSIs
calculated by grade level were similarly very high, ranging from .939 to .968 for grade 3, .
952 to .977 for grade 5, .960 to .978 for grade 7, and .960 to .976 for grade 9, across the
criteria. These were not further differentiated into performance genre, as the sample sizes
were not sufficient to obtain robust estimates.
Correlations between judgment criteria. The correlation between the estimates
obtained from the authorial choices and conventions criteria comparative judgments for each
of the performances was .964. This correlation was very similar for both the narrative
performances (.958) and the persuasive performances (.973). Moreover, it was also very
similar for grade 3 (.948), grade 5 (.969), grade 7 (.960), and grade 9 (.936). These findings
provide justification for combining criteria in an analogous manner to combining criteria in
rubrics with mutually correlated scores. Moreover, the magnitude of these correlations
suggests that the judgments for these two criteria may not be discriminating between different
aspects of writing quality. Nonetheless, given that assessors from a previous application of
the two-stage method indicated that, for select performances, it was difficult to reconcile
these aspects to make a holistic comparison, the two criteria were carried over to Stage
2.Concurrent validity with rubric scores. To establish the concurrent validity of the
comparative judgment procedure, the three comparative judgment criteria scales were
correlated with the rubric scores for the 160 performances. This correlation was .944 for the
13
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
authorial choices scale, .944 for the conventions scale, and .951 for the combined scale.
Further, these correlations were essentially identical when examined for just the narrative
performances (.946, .943, & .953), and just the persuasive performances (.942, .946, & .948).
These results provide strong evidence that the concurrent validity was not moderated by the
type of writing performance.
Stage 2 – Using the Matching Procedure with Calibrated Exemplars to Assess Writing
Stage 2 method. In Stage 2, experienced assessors scored a new set of narrative and
persuasive writing performances using 36 calibrated exemplars from Stage 1. This number of
exemplars was selected to cover the full range of NAPLAN writing performance quality
across the four grades and to provide a similar level of granularity in scores to the NAPLAN
analytic rubrics. Sufficient comparative judgments were performed in Stage 1 to ensure with
reasonable certainty that the 36 exemplars were correctly ordered. Moreover, the exemplars
were carefully selected to ensure that they were consistent in their ordering for both the
authorial choices and conventions criteria, as determined by their proximity to the regression
line between the criteria estimates in Stage 1. Finally, a content expert involved in the
NAPLAN operations reviewed the exemplars to ensure that they were free of any anomalous
content that may affect the assessors’ perception of their relative ordering.
A linear transformation was carried out on the scale of exemplars displayed to
assessors so that their scores ranged from 50 to 95, which was a more familiar score range to
these assessors than the logit scale estimates from Stage 11. The performance scale was
displayed graphically as shown in Figure 1 with reference exemplars to the left of a scale and
the performances to be assessed on the right-hand side of the display area. Assessors were
asked to match each performance assessed to a reference exemplar on the scale, as described
in more detail in the procedure section to follow.
1 The lowest score of 50 was not intended to indicate a ‘pass’ performance, but rather as a
more intuitive lower bound for assessors than the negative values of the logit scale.
14
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
As described above, a key difference from Heldsinger and Humphry (2013) in the
current study was that matching judgments were made using both the authorial choices and
conventions criteria from Stage 1, rather than a holistic judgment of performance quality.
Thus, each of the new narrative and persuasive written performances in Stage 2 was assessed
by matching the performance qualities to the most similar calibrated exemplar from Stage 1
with respect to the specific broad assessment criterion. Each performance was assigned the
logit scale estimate of the calibrated exemplar with which it was matched for that criterion. In
addition, a combined scale estimate was calculated for each performance in Stage 2 by
averaging their matched scale estimates across the authorial choices and conventions criteria.
These matched logit scale estimates were used in the calculation of all inter-judge reliability
correlations, which are explained further in the Data Analysis subsection below.
Collection of the pool of writing performances for assessment at Stage 2. The new
pool of narrative and persuasive writing performances was drawn from the larger sample of
approximately 3,000 writing performances from which the 160 Stage 1 performances were
also drawn. In total, 2380 new performances were selected, including 586 from grade 3, 532
from grade 5, 679 from grade 7, and 583 from grade 9.
Participating assessors. Twenty-one experienced assessors participated in the Stage 2
assessment. They were drawn from a pool of experienced NAPLAN assessors in a different
Australian State from Stage 1. Assessors made independent judgments either in a central
venue or at other sites.
Procedures. Assessors were provided with brief instructions regarding the assessment
task to match new written performances to their most similar exemplars, and were instructed
to make an on-balance judgment for the two broad criteria. No further training was provided.
To complete the task, assessors were digitally presented with 36 exemplars displayed
adjacent to a vertical scale in customized software, as shown in simplified form in Figure 1.
15
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Assessors were also provided with a hardcopy booklet containing the 36 exemplars and a
response sheet. Each of the written performances to be matched was presented on the right-
hand side of the screen. Performances were presented in a random sequence, with each
assessor making a total of 346 matching judgments (173 each for the authorial choices and
conventions criteria). Once assessors had determined the reference exemplar X that was
closest in quality to the assessed performance Y for that criterion, they assigned that scale
score of X to the assessed performance by dragging an icon into a box at the appropriate scale
location adjacent to that exemplar. Assessors first matched their allocated writing
performances based on the authorial choices criterion and then for the conventions criterion.
The duration of the judgment was also recorded for each matching judgment on each
criterion.
Data analysis. Similar to Heldsinger and Humphry (2013) and Gyagenda and
Engelhard (2009), the inter-judge reliability for all of the 210 possible pairs of assessors was
computed as the mean of the inter-judge correlations in their matching judgment estimates for
the commonly judged performances. These were calculated for the authorial choices,
conventions and combined criteria, as well as for the individual grade levels. As an additional
indicator of inter-judge reliability, the correlations between the matched estimates of each
assessor and the mean estimates for all assessors (excluding the assessor being compared)
were calculated for the authorial choices, conventions and combined criteria. Moreover, the
correlation between the authorial choices and conventions judgment criteria estimates was
calculated and the median durations of matching judgment time were examined to explore the
method’s efficiency.
Stage 2 results.
Inter-judge reliability. The average inter-judge correlation between each of the 210
unique pairs of assessors was high for each of the judgment tasks, indicating good inter-judge
16
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
reliability for the matching judgments. The mean correlation was .799 (range .585 to .919) for
the authorial choices criterion, .781 (range .606 to .928) for the conventions criterion, and .
858 (range .702 to .932) for the combined criteria. These mean inter-judge correlations were
similar for just narrative and just persuasive performances, although larger for the latter (see
Table 1). Moreover, they were similar for three of the individual grade level performances,
ranging from .746 to .838 for grade 3, .719 to .822 for grade 5, and .689 to .820 for grade 9,
across the criteria. The mean inter-judge correlations were lower for grade 7, ranging from .
569 to .672 across the criteria, although the variability in the estimates was much smaller for
this grade (SD range from 1.442 to 1.495) compared to the other grades in the sample (SD
range from 1.742 to 2.279). Similar to Stage 1, the grade level correlations were not further
differentiated into performance genre due to insufficient sample sizes.
The correlations between each assessor’s matching estimate and the average estimate
of all other assessors similarly indicated a high level of inter-judge reliability for each of the
matching judgment tasks. The mean correlation was .893 (range from .798 to .955) for the
authorial choices criterion, .888 (range from .805 to .956) for the conventions criterion, and .
928 (range from .847 to .953) for the combined scale. These mean correlations were similar
when examining the judgments for just the narrative performances (.857, .842, & .901) and
the persuasive performances (.928, .919, & .950).
Correlations between judgment criteria. The correlation between the average
estimates obtained from the authorial choices and conventions criteria judgments was .969,
where the average was taken across all assessors. This correlation was similarly very high for
the subsets of narrative performances (.949) and the persuasive performances (.980). This
finding provided evidence that the strong correspondence between judgment criteria was not
substantially moderated by performance genre.
17
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Concurrent validity with rubric scores. To establish the concurrent validity of this
method using calibrated exemplars, the estimates obtained by the matching judgments were
then correlated with scores obtained from the NAPLAN analytic rubric (ACARA, 2010,
2013) for the same performance scripts.
Procedure. Six experienced NAPLAN assessors double-marked the full sample of
2,380 performances using these rubrics. This marking exercise was undertaken separately
from Stage 1 and 2 by a separate group of assessors. The assessments were made according to
the procedures for marking NAPLAN writing performances set out in the above described
NAPLAN writing marking guides (ACARA 2010, 2013). This rubric assessment provided a
total score, authorial-choices sub-score for the six criteria consistent with this broad criterion,
and conventions sub-score for the four criteria consistent with this broad criterion. Similar to
the findings for the comparative judgment procedure and matching procedure estimates, these
NAPLAN rubric sub-scores were highly inter-correlated, with an overall correlation of .929
across all performances, .947 for the persuasive performances, and .907 for the narrative
performances.
The NAPLAN rubric has consistently had a PSI of between 0.95 and 0.96 over the
course of the testing program, which shows a very high, historical level of reliability. The
inter-rater reliability for this set of written performances was very high for the total rubric
scores with an inter-rater correlation of .894 for the narrative performances and .922 for the
persuasive performances. However, these correlations may be somewhat inflated relative to
the general level of reliability in NAPLAN rubric marking across all States, as these six raters
were some of the most experienced raters from a single marking center, and had therefore
undergone training on the rubric together over a number of years. Given that all raters were
highly experienced and showed a high level of inter-rater reliability, the final rubric score and
18
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
sub-scores used for correlations with the matched estimates were the average scores of the
two raters for each performance.
Concurrent validity results. To first examine the concurrent validity of the matched
estimates, these average NAPLAN rubric scores and sub-scores were correlated with the
average matched estimates (taken across both criteria) for the 37 performances that were
matched by all 21 assessors, as these estimates are based on multiple-observations and so
should be more reliable. These correlations showed a very high level of concurrent validity
between the matched estimates and NAPLAN rubric scores. The correlation with the total
rubric score was .925 for the average authorial choices matched estimate, .926 for the average
conventions matched estimate, and .933 for the average combined matched estimate. These
correlations were similar for just the narrative performances (.917, .908, & .924) and just the
persuasive performances (.942, .951, & .952), although they were somewhat larger for the
latter.
Next, the individual matched estimates of all of the performances were correlated
with their rubric score and sub-scores, and showed a similarly high level of concurrent
validity with correlations ranging from .819 to .900 for the combined and individual
performance genres (see Table 1).
Efficiency of matching judgments. The computerised administration of Stage 2
enabled the recording of judgment duration for each assessor across their 346 judgments. To
overcome the influence of outliers, the median duration for all assessors was examined. The
judgments showed a consistent downward trend in average duration from the beginning to the
end of task. The median duration commenced at 16.36 minutes for the first matching
judgment. By a third of the way through the task, the median duration fell to 6.68 minutes per
judgment. By the final judgment, the median duration was 3.77 minutes compared to an
19
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
overall median duration of 6.03 minutes for all of the matching judgments across the entire
task.
Discussion
Discussion of the Stage 1 Results
Similar to the previous findings of Heldsinger and Humphry (2010, 2013), the
development of the scale of exemplars by comparative judgments in the Stage 1 calibration
showed a very high level of reliability. The separation indices were all very high and similar
for the different genres, criteria and grades. These high separation indices may be attributable
in part to the number of judges, as explained by Heldsinger and Humphry (2013), and so they
are not directly comparable to an index of a rubric marked by a single judge. Nonetheless, the
indices show that the assessors were very consistent in their interpretation and assessment of
the written performances, and this consistency was not moderated by genre, criterion or
grade. The very high correlation between the two criterion judgments in Stage 1 also
provided strong evidence that it was appropriate to form a single scale of calibrated
exemplars for the matching judgments in the second stage of the study.
Discussion of the Stage 2 Results
The aim of stage two was to ascertain the reliability and validity of the matching
judgments using calibrated exemplars for the standardized assessment of writing, applied to
performances from a broad age range across multiple genres and criteria.
Reliability and validity. The present findings provide further evidence that assessors
are able to use calibrated exemplars to make quite reliable judgments at both the criterion and
overall levels without extensive training in the procedure, and without any group moderation
processes to establish consistency across assessors. Unlike the Stage 1 results, the level of
reliability was somewhat moderated by genre type, criterion and grade level, although the
20
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
majority of reliability indices were at or above the common .8 correlation threshold for
acceptable reliability in rubric scores (Brookhart & Chen, 2014).
The inter-judge correlations for the matching estimates were noticeably lower for the
narrative performances, and particularly for the conventions criterion on these performances.
This finding was corroborated by the lower correlations between each assessor’s performance
estimate and the average estimate of all other assessors for the 19 narrative performances
with common marking. The reliability indices were also consistently lower than the levels of
reliability observed for the calibrated exemplar method in Heldsinger and Humphry (2013)
where it was applied to early-childhood writing. Similarly, the Stage 2 inter-judge
correlations were noticeably lower for the grade 7 performances relative to the other grades.
This finding is likely attributable, in part, to the lower observed variance for grade 7
estimates across the criteria relative to the other grades, which constrains the correlations.
Further potential explanations for the reduced reliability of the matching judgments in the
present study, and suggestions to enhance the reliability of the method are discussed in the
Limitations and Future Research sections below.
A high level of concurrent validity was observed between the matching judgment
estimates and the scores and sub-scores given by expert raters using the NAPLAN rubric.
Again though, these correlations were somewhat lower for the narrative genre, the
conventions criterion, and for grade 7, which may be attributable to the lower reliability of
the matching estimates for these subsets of performances. Nevertheless, the majority of
correlations between the matching estimates and rubric scores/ sub-scores exceeded .8,
providing evidence that the criterion-level and overall matching estimates were valid and
comparable to rubric scores. The individual grade level correlations were below the .8
criterion, which is likely due to the smaller variance and range of abilities within a grade
level, as, for a given association, observed correlations are higher where the total variance is
21
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
greater. Moreover, this finding likely reflects the cognitive demands of matching each
individual's performance against a relatively large number of exemplars, as discussed in the
Limitations and Future Research sections below.
Efficiency. The median duration for all matching judgments in the entire task in Stage
2 was approximately six minutes. It was evident in the results that the assessors became, on
average, substantially quicker at the matching judgments throughout the overall duration of
the task. Allowing for an initial period during which assessors familiarized themselves with
the calibrated exemplars, it was possible to attain an average judgment rate between 10 and
16 performances per hour across assessors. Moreover, the method alleviates some of the
need for the extensive training and moderation that effective rubric scoring typically entails,
although it is noted that all assessors involved in the matching task had separately undertaken
NAPLAN rubric training and so they likely drew upon some of this experience when making
their matching judgments, which may have enhanced the efficiency of these judgments
(Heldsinger & Humphry, 2013; Jonsson & Svingby, 2007; Shermis, 2014). Nonetheless, a
downward trend in median judgment duration across the matching exercise suggests that the
increased efficiency was at least partly attributable to the assessors familiarizing themselves
with the calibrated exemplars, and not based on their prior experience.
Limitations. Verbal feedback from assessors in Stage 2 clearly indicated that the
matching process placed a relatively high cognitive load on them, particularly when more
complex performances by older students were assessed. This issue was not as evident in the
original application by Heldsinger and Humphry (2013) because they focused on early
childhood writing with less sophisticated and complex features than the texts produced by
older children in the present research. In addition, feedback and evaluation of the process by
assessors indicated that too many calibrated exemplars were used in Stage 2. A total of 36
exemplars were selected so that the number of scores that could be assigned was similar to
22
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
the possible level of granularity from the NAPLAN analytic rubric. However, assessors did
not always agree with the ordering of the calibrated exemplars that were presented, despite
the very positive results from Stage 1 of the research.
In hindsight, this lack of consensus on the ordering of the exemplars was bound to be
a limitation because the mean distance between the scale locations of adjacent exemplars was
somewhat smaller than the mean standard error of the estimates on the scale established via
comparative judgments. Consequently, in some cases there was insufficient certainty around
whether adjacent exemplars were correctly ordered. Because the assessors were asked to
match performances against calibrated exemplars, and they may not have perceived the
exemplars to be ordered in a manner that was consistent with the overall solution from Stage
1, this discrepancy would have added to the difficulty of the decisions. This additional
difficulty is likely to have had an adverse effect upon the reliability of the matching
judgments.
In addition, one assessor annotated hard copies of performances to assist with the
judgments and indicated to the researchers that brief descriptors of ranges of performance
would have assisted the matching assessment process. As described in the Introduction, the
research aimed to investigate whether calibrated exemplars could be used to capitalize on the
advantages of comparative judgments while attaining greater efficiency using a two-stage
process directly analogous to that adopted by Thurstone (1928). However, Thurstone’s
applications involved attitude and value statements, each of which were one sentence and
could be rapidly read and ascertained. In contrast, because it is necessary to become familiar
with a number of exemplars of extended writing and to hold information about these in mind
while assessing a new performance, the use of calibrated exemplars to assess writing
performances is more cognitively challenging. Therefore, any aids that can be provided to
23
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
assessors may make the process more efficient, as well as potentially increase the reliability
and validity of the method.
The results indicated somewhat lower inter-judge correlations for narrative
performances than persuasive performances. It is possible that the cognitive demands on the
assessors were somewhat higher for narrative than persuasive performances. The persuasive
texts generally have consistent macrostructural components because the instructions indicate
students should include: (a) an introduction, (b) a body (with reasons in support); and (c) a
conclusion. This greater consistency in macrostructure makes it a somewhat easier task for
assessors to focus on key features of a persuasive performance when performing the
matching. This macrostructural consistency may also explain the somewhat greater
correlation between the authorial choices and conventions criteria in both the matching
judgment estimates and rubric scores for the persuasive performances. In contrast, narrative
texts have no standard macrostructural features, other than the use of paragraphs by
sufficiently capable students. The assessment of narrative performances might also be more
influenced by the subjective level of engagement of the assessor through creative elements
such as ideas and voice (Grainger, Goouch, & Lambirth, 2005). For these reasons, there may
be a greater need for descriptors and annotations for narrative than persuasive performances,
but feedback from assessors indicated that such aids will be helpful for both types of writing.
Another issue raised by some assessors was that it seemed more difficult to make
matching judgments for conventions, which may explain the generally lower levels of
reliability and concurrent validity for this criterion. The reason indicated by these assessors
was that it was possible to focus specifically on features such as spelling and sentence
structure, as they are distinctive and separate, making it necessary to reconcile more than one
distinctive judgment into an overall conventions judgment. In contrast, authorial choices
comprise components that are more connected, such as ideas, audience and text cohesion. In
24
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
addition, it is possible that some of the components classified as conventions are more useful
for distinguishing performances by the less capable students than the more capable students.
The more capable students have generally mastered skills such as correct paragraphing and
sentence structure, so assessors may have had somewhat greater clarity when assessing
highly capable students on the authorial choices criterion than the conventions criterion, as
differences in the former would be easier to discern in higher quality performances than
differences in the latter.
Future research. Based on feedback and the current results, future research is
planned in which: (a) fewer exemplars will be used, which are more definitively ordered for
the assessors by taking explicit account of the standard error estimate for each of the
exemplar performance estimates; and (b) theoretically and empirically informed descriptors
will be provided in conjunction with exemplars to assessors to support matching judgments.
The planned research will also provide the facility for annotations of calibrated exemplars to
assist the assessors to recall key features and details of each performance, therefore reducing
the cognitive load of the task. Finally, a larger sample that is representative of each grade, and
for each genre at each grade in particular, will be recruited to obtain sufficiently robust
reliability estimates within grades, as the current research was primarily focused upon
obtaining reliability estimates across a range of grade levels.
These additions to the methodological design should make it possible to attain higher
levels of inter-judge reliability with similar or improved efficiency to that found in Stage 2 of
the present research. A drawback of these additions is that they require time and resources to
implement and therefore somewhat detract from the greater efficiency of the method relative
to rubric-based assessment. However, once the calibrated exemplars are scaffolded with
descriptors and annotations, in particular, it is anticipated that new assessors will be able to
25
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
efficiently apply the method with an even higher level of reliability and validity than found in
the present study in the absence of extensive training or moderation processes.
With respect to the inclusion of more specific judgment criteria in both the
comparative judgment and matching stages of the research design, the present study went
beyond the previous application by including two broad criterion for judgments based on
feedback that aspects of these broad criteria were difficult to reconcile for certain writing
performances. While the present research did not find evidence to motivate the inclusion of
separate calibrated exemplar scales for each criterion, inclusion of more specific criteria in
Stage 1 may motivate the inclusion of more specific calibrated exemplar scales for matching
judgments.
The calibrated exemplar method is also applicable beyond writing assessments, as it is
suitable for any sort of extended performance such as essays and media productions for
which holistic judgments can be made (Heldsinger & Humphry, 2013). This broader
applicability than either rubric-based or automated essay scoring is particularly pertinent
given the current global trend toward more construct-relevant, performance-based
assessments (Shermis, 2014). Consequently, future research is either planned or underway to
demonstrate the robustness of the method for performance assessments more broadly. In
particular, assessment tasks, including calibrated exemplars, have been developed to assess
science investigations, informational reports, history, and mathematics. Software tools have
also been concurrently developed to integrate the abovementioned descriptors and
annotations into the matching process for these learning areas.
In addition, while the current findings demonstrate the potential application of the
calibrated exemplar method to large-scale programs, particularly with the abovementioned
improvements, this two-stage method was developed first and foremost to be used by
classroom teachers and Heldsinger and Humphry (2013) provided evidence that teachers can
26
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
reliably and validly use the method with minimal training. Given the various criticisms of
standardized assessments, including the narrowing of curricula, future research will examine
whether the calibrated exemplar method can provide a valid and reliable basis for comparable
performance assessments of classroom-based tasks across schools without the need for large-
scale programs. This research requires providing a common bank of exemplars across
classroom teachers and schools, which is already possible using the online software
employed by the present study.
Summary and Conclusion
To summarize, the aim of this research was to explore the reliability, validity and
efficiency of the two-stage method of writing assessment in a standardized writing
assessment context including multiple age groups, both narrative and persuasive genres, and
two judgment criteria. The results showed good to very high levels of reliability in both
stages of the method. The concurrent validity of the estimates from the second stage was high
across criteria and genres, and the second stage was found to be reasonably efficient after an
initial familiarization period. Nonetheless, the reliability and validity statistics were
noticeably lower than a previous application of the two-stage method, particularly for the
narrative performances on the conventions criterion.
These lower levels of reliability and concurrent validity are perhaps unsurprising
given the more complex application in the present research. Future research will examine the
reliability, validity and efficiency of the method with refinements to the method to help
reduce the cognitive load of the task for assessors, and ultimately across a broader range of
extended performances. Overall, the present research provides evidence for the viability of
this method as a complement or alternative to existing methods of standardized writing
assessment, and supports its potential for application in assessing extended performances
across a range of learning areas.
27
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
References
Australian Curriculum, Assessment and Reporting Authority. (2010). NAPLAN 2010 Writing
Narrative Marking Guide. Retrieved from the National Assessment Program website:
http://www.nap.edu.au/verve/_resources/2010_marking_guide.pdf
Australian Curriculum, Assessment and Reporting Authority. (2013). NAPLAN 2013
Persuasive Writing Marking Guide. Retrieved from the National Assessment Program
website:
http://www.nap.edu.au/_resources/amended_2013_persuasive_writing_marking_guid
e_-with_cover.pdf
Andrich, D. (1978). Relationships between the Thurstone and Rasch approaches to item
scaling. Applied Psychological Measurement, 2(3), 451-462.
Andrich, D. (1988). Rasch models for measurement. Beverly Hills: Sage Publications, Inc.
Bond, T. G., & Fox, C. M. (2012). Applying the Rasch model: Fundamental measurement in
the human sciences. New York: Routledge.
Bradley, R. A., & Terry, M. E. (1952). Rank analysis of incomplete block designs: I. The
method of paired comparisons. Biometrika, 39(3/4), 324-345.
Bramley, T., Bell, J. F., & Pollitt, A. (1998). Assessing Changes in Standards over Time Using
Thurstone's Paired Comparisons. Education Research and Perspectives, 25(2), 1-24.
Brindle, M., Graham, S., Harris, K. R., & Hebert, M. (2016). Third and fourth grade teacher’s
classroom practices in writing: A national survey. Reading and Writing, 29, 929-954.
Brookhart, S. M., & Chen, F. (2014). The quality and effectiveness of descriptive rubrics.
Educational Review, 1-26.
Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika,
16(3), 297-334.
28
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Grainger, T., Goouch, K., & Lambirth, A. (2005). Creativity and writing: Developing voice
and verve in the classroom. Abingdon, UK: Routledge.
Gyagenda, I. S., & Engelhard, G. (2009). Using classical and modern measurement theories
to explore rater, domain, and gender influences on student writing ability. Journal of
Applied Measurement, 10(3), 225-246.
Hayes, J. R., & Berninger, V. (2014). Cognitive processes in writing: A framework. In B.
Arfé, J. Dockrell, & V. Berninger (Eds.), Writing development and instruction in
children with hearing loss, dyslexia, or oral language problems: Implications for
assessment and instruction (pp. 3-15). Oxford, UK: Oxford University Press.
Heldsinger, S. A., & Humphry, S. M. (2010). Using the method of pairwise comparison to
obtain reliable teacher assessments. The Australian Educational Researcher, 37(2), 1-
19.
Heldsinger, S. A., & Humphry, S. M. (2013). Using calibrated exemplars in the teacher-
assessment of writing: an empirical study. Educational Research, 55(3), 219-235.
Holme, B., & Humphry, S. M. (2008). PairWise software. Perth: University of Western
Australia.
House of Commons Education Committee. (2017). Primary Assessment. Retrieved from the
Parliament UK website:
https://publications.parliament.uk/pa/cm201617/cmselect/cmeduc/682/682.pdf
Humphry, S. M., & Heldsinger, S. A. (2014). Common Structural Design Features of Rubrics
May Represent a Threat to Validity. Educational Researcher, 43(5), 253-263.
Humphry, S. M., & McGrane, J. A. (2015). Equating a large-scale writing assessment using
pairwise comparisons of performances. The Australian Educational Researcher,
42(4), 443-460.
29
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Jonsson, A., & Svingby, G. (2007). The use of scoring rubrics: Reliability, validity and
educational consequences. Educational research review, 2(2), 130-144.
Luce, R. D. (1959). Individual choice behavior: a theoretical analysis. New York: Wiley.
Mackenzie, N. M., Scull, J., & Munsie, L. (2013). Analysing writing: The development of a
tool for use in the early years of schooling. Issues in Educational Research, 23(3),
375-393.
McHale, K., & Cermak, S. A. (1992). Fine motor activities in elementary school: preliminary
findings and provisional implications for children with fine motor problems.
American Journal of Occupational Therapy, 46(10), 898-903.
Myford, C. M., & Wolfe, E. W. (2003). Detecting and measuring rater effects using many-
facet Rasch measurement: part I. Journal of Applied Measurement, 4(4), 386-422.
Pollitt, A. (2012). The method of adaptive comparative judgement. Assessment in Education:
Principles, Policy & Practice, 19(3), 281-300.
Reddy, Y. M., & Andrade, H. (2010). A review of rubric use in higher education. Assessment
& Evaluation in Higher Education, 35(4), 435-448.
Rezaei, A. R., & Lovorn, M. (2010). Reliability and validity of rubrics for assessment
through writing. Assessing Writing, 15(1), 18-39.
Sadler, D. R. (2009). Indeterminacy in the use of preset criteria for assessment and grading.
Assessment & Evaluation in Higher Education, 34(2), 159-179.
Shermis, M. D. (2014). State-of-the-art automated essay scoring: Competition, results, and
future directions from a United States demonstration. Assessing Writing, 20, 53-76.
Spandel, V. (2005). Creating Writers Through 6-trait Writing: Assessment and Instruction
(4th ed.). Boston: Pearson and Allyn & Bacon.
Steedle, J. T., & Ferrara, S. (2016). Evaluating Comparative Judgment as an Approach to
Essay Scoring. Applied Measurement in Education, 29(3), 211-223.
30
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Tarricone, P., & Newhouse, C. P. (2016). Using comparative judgement and online
technologies in the assessment and measurement of creative performance and
capability. International Journal of Educational Technology in Higher Education,
13(1), 1-16.
Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review, 34(4), 273-
286.
Thurstone, L. L. (1928). Attitudes can be measured. American Journal of Sociology, 33, 529-
554.
Thurstone, L. L., & Chave, E. J. (1929). The Measurement of Attitude: A Psychophysical
Method and Some Experiments with a Scale for Measuring Attitudes Toward the
Church. Chicago, IL: University of Chicago Press.
Tierney, R., & Simon, M. (2004). What's still wrong with rubrics: focusing on the consistency
of performance criteria across scale levels. Practical Assessment, Research &
Evaluation, 9(2), 1-10.
Wyatt-Smith, C., & Jackson, C. (2016). NAPLAN data on writing: A picture of accelerating
negative change. Australian Journal of Language and Literacy, 39(3), 233-244.
31
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
Table 1
Summary of reliability and validity statistics across the two stages broken down
by narrative (Nar.), persuasive (Pers.) and pooled (All) performances, as well as
the combined, authorial choices and conventions criteria.
No. of
performances Statistic Index
All Nar.
Pers
. All
Nar
.
Per
s.
Stage 1:
Reliability
160 80 80
Person
Separation Index
Combined
.
98
2
.
98
3
.
980
Authorial
.
97
1
.
97
2
.
968
Conventions
.
96
7
.
97
1
.
962
Stage 2:
Reliability 37 19 18
Combined
.
85
8
.
77
9
.
907
Mean inter-judge
correlation
Authorial
.
79
9
.
71
0
.
868
Conventions
.
78
1
.
64
1
.
852
Stage 2:
Validity
238
0
113
7
124
3
Combined
.
87
2
.
85
7
.
900
Correlation with
rubric score
Authorial
.
85
1
.
83
0
.
882
Conventions . . .
32
Running head: TWO-STAGE METHOD OF WRITING ASSESSMENT
83
4
81
9860
Figure 1. Schematic of the digital display seen by assessors while making each of
their matching judgments.