ArticlePDF Available

Writing to Read: A Meta-Analysis of the Impact of Writing and Writing Instruction on Reading

December 2011
Harvard Educational Review 81(4):710-744

DOI:10.17763/haer.81.4.t2k0m13756113566

Authors:

Steve Graham

Arizona State University

Michael Hebert

University of Nebraska–Lincoln

Reading is critical to students' success in and out of school. One potential means for improving students' reading is writing. In this meta-analysis of true and quasiexperiments, Graham and Herbert present evidence that writing about material read improves students' comprehension of it; that teaching students how to write improves their reading comprehension, reading fluency, and word reading; and that increasing how much students write enhances their reading comprehension. These findings provide empirical support for long-standing beliefs about the power of writing to facilitate reading.

Studies and effect sizes included in the analysis of each research question

…

Proportion of studies meeting requirements for each quality category by research question

…

Meta-regression for writing about reading (researcher-created comprehension outcomes)

…

Meta-regression for studies examining summary writing

…

Meta-regression for studies examining note taking

…

Figures - uploaded by Steve Graham

Content may be subject to copyright.

Content uploaded by Steve Graham

Content may be subject to copyright.

710

Harvard Educational Review Vol. 81 No. 4 Winter 2011

Writing to Read: A Meta-Analysis of

the Impact of Writing and Writing

Instruction on Reading

STEV E GR AHAM

MICHAEL HEBERT

Vanderbilt University

Reading is critical to students’ success in and out of school. One potential means

for improving students’ reading is writing. In this meta-analysis of true and quasi-

experiments, Graham and Herbert present evidence that writing about material read

improves students’ comprehension of it; that teaching students how to write improves

their reading comprehension, reading ﬂuency, and word reading; and that increas-

ing how much students write enhances their reading comprehension. These ﬁndings

provide empirical support for long-standing beliefs about the power of writing to

facilitate reading.

Reading is one of the most critical skills that students must master to be suc-

cessful educationally, occupationally, and socially. Students’ educational suc-

cess depends on their abilities to read and critically analyze information pre-

sented in textbooks and other classroom materials (Berman, 2009; Klein,

1999). Reading is essential to success in most white-collar and blue-collar jobs

(Greene, 2000), with forecasters predicting an increase in the proportion of

new jobs requiring strong reading skills (Carnevale & Derochers, 2004; Kirsch,

Braun, Yamamoto, & Sum, 2007). Reading is part of the basic fabric of twenty-

ﬁrst-century life, as e-mailing, blogging, texting, Facebook, and other forms of

written text are now common means for social contact and communication.

Written text permeates everyday life, from messages prominently displayed on

billboards and sides of buses to information provided on everyday essentials,

such as cans of food and bottles of medicine.

Despite the importance of reading, many students are not skilled readers by

the end of high school. The 2009 National Assessment of Educational Progress

(NAEP) reported that only 38 percent of twelfth-grade students performed at

711

Writing to Read

steve graham and michael hebert

or above the “proﬁcient” level in reading (deﬁned as solid academic perfor-

mance) (NCES, 2010). In terms of younger students, only 33 percent of fourth

graders and 32 percent of eighth graders performed at these levels (NCES,

2009). In contrast, 34, 43, and 36 percent of fourth-, eighth-, and twelfth-grade

students, respectively, scored at the “basic” level, denoting only partial mastery

of the literacy skills needed at their grade levels. The rest of the tested stu-

dents’ scores were below this basic level.

Furthermore, the reading performance of students who do not speak Eng-

lish as their ﬁrst language, students who have a disability, and students who are

black, Hispanic, or Native American was signiﬁcantly lower than the reading

performance of students who were native English speakers, did not have a dis-

ability, or were white, respectively (NCES, 2009, 2010). The results from the

NAEP clearly demonstrate that large numbers of students need help if they

are to become skilled readers.

Such concerns have spurred large-scale policy actions during this decade

and the last to improve children’s reading, including No Child Left Behind

and Reading First. Both of these approaches relied heavily on the use of sci-

entiﬁcally based practices (involving the teaching of phonological awareness,

phonics, vocabulary, ﬂuency, and comprehension) identiﬁed by the National

Reading Panel (NRP) (NICHD, 2000). Neither of these policy endeavors

resulted in the types of reading gains envisioned by their advocates. While

there are many possible explanations for these outcomes, one contributing

factor may be that the instructional practices identiﬁed by the NRP report

were too narrow and not complete.

An important policy question, then, is what else can schools do to strengthen

students’ reading? There are many potential actions that policy makers and

educators can undertake to improve reading, ranging from enhancing chil-

dren’s overall language skills to providing more engaging and meaningful

reading instruction. In this article, we examine the effectiveness of writing as

a tool for improving students’ reading. We concentrated our efforts on this

often-overlooked tool for two basic reasons. One, several meta-analyses have

found that writing about content classroom material can facilitate learning it

(Bangert-Drowns, Hurley, & Wilkinson, 2004; Graham & Perin, 2007a). It is

also possible that writing about material read enhances comprehension of it.

Two, reading and writing share a close and reciprocal relationship (Fitzger-

ald & Shannahan, 2000), and there is evidence that reading instruction can

improve students’ writing skills (e.g., Graham, 2000; Krashen, 1989). Conse-

quently, it is likely that writing instruction in turn improves students’ read-

ing skills. Three theoretical perspectives are particularly informative in under-

standing the possible impact of writing on reading.

According to the functional view of reading-writing connections (Fitzgerald

& Shanahan, 2000), writing about text should facilitate comprehension in ﬁve

ways (Applebee, 1984; Emig, 1977; Klein, 1999; Stotsky, 1982):

712

Harvard Educational Review

It fosters explicitness, as the writer must select which information in 1.

text is most important.

It is integrative, as it encourages the writer to organize ideas from text 2.

into a coherent whole, establishing explicit relationships among the

ideas.

It facilitates reﬂection, as the permanence of writing makes it easier to 3.

review, reexamine, connect, critique, and construct new understandings

of text ideas.

It can foster a personal involvement with text, as it requires active deci-4.

sion making about what will be written and how it will be treated.

It involves transforming or manipulating the language of text so that 5.

writers put ideas into their own words, making them think about what

the ideas mean.

In short, writing about text should facilitate comprehending it, as it provides

students with a tool for visibly and permanently recording, connecting, analyz-

ing, personalizing, and manipulating key ideas in text.

The impact of writing on reading likely extends beyond just writing about

text to the possible impact of teaching about the process of writing. Accord-

ing to the shared knowledge view of reading-writing connections, reading and

writing are not identical skills, but both rely on common knowledge and pro-

cesses (Fitzgerald & Shanahan, 2000). Consequently, instruction that improves

writing skills and processes should improve reading skills and processes. We

illustrate this with two examples. One, Ehri (2000) and others (e.g., Moats

2005/2006) hypothesize that teaching students how words are spelled provides

them with schemata about speciﬁc connections between letters and sounds,

and this should make it easier for them to identify and remember words in

text containing these connections. Two, Neville and Searls (1991) hypothesize

that teaching students how to construct more complex sentences by combin-

ing smaller, less complex ones should result in greater skill in understanding

such units in reading.

According to the rhetorical relations view of reading-writing connections

(Tierney & Shanahan, 1991), the process of composing text should enhance

one’s skills at comprehending text. This theoretical perspective views reading

and writing as communication activities, and it assumes that writers gain insights

about reading by creating text for an audience to read, even if the writer is the

intended audience. Theoretically, the process of creating text should prompt

students to be more thoughtful and engaged when reading text produced by

others. Because writers need to make their assumptions and premises explicit

as well as observe the rules of logic when composing text, this presumably

makes them more aware of these same issues in the material they read.

To examine the robustness of each of these theoretical viewpoints about the

impact of writing on reading, we conducted a meta-analysis to answer the fol-

lowing three questions for students in grades 1–12:

713

Writing to Read

steve graham and michael hebert

Does writing about material read enhance students’ comprehension of 1.

text?

Does writing skills instruction strengthen students’ reading skills? 2.

Does increasing how much students write improve how well they read? 3.

Meta-analysis is used to summarize the direction and magnitude of the

effects obtained in a set of empirical studies examining the same basic phe-

nomena (Lipsey & Wilson, 2001). The meta-analysis reported in this article

was funded by the Carnegie Corporation of New York (Graham & Hebert,

2010), and the evidence used to answer each question was derived from either

true or quasi-experiments. In both types of experiments, a writing treatment

(e.g., writing a summary of text read) is compared to a control condition (e.g.,

reading and rereading text) to determine its impact on reading (e.g., reading

comprehension). Meta-analysis is well suited to answering the three questions

posed in this review, as it provides an estimate of the effectiveness of a treat-

ment “under conditions that typify studies in the literature” (Bangert-Drowns

et al., 2004, p. 34). When enough studies are available, and the variability of

individual study effects is greater than the variability due to sampling error

alone, meta-analysis also permits examining the relationship between study

outcomes and features. In other words, conducting a meta-analysis on a large

number of studies with different ﬁndings may enable us to more closely inves-

tigate the relationships between a speciﬁc set of writing procedures (e.g., writ-

ing about text) and students’ grade levels, type of text read, and so forth.

This meta-analysis differs from Writing to Read (Graham & Hebert, 2010)

in four important ways. First, we calculated effects for all quasi-experiments

included in the review and adjusted these effects for possible data cluster-

ing due to hierarchical nesting of data (i.e., researchers assigned classes to

treatment or control conditions but then examined student-level effects). We

included both true experiments and quasi-experimental studies in our review,

as a sensitivity analysis revealed that statistically similar results were obtained

for both types of experiments. Second, we assessed the quality of each study,

allowing us to make better judgments about the conﬁdence that can be placed

in our conclusions. Third, we updated the search for studies to include stud-

ies published after June 2008 (this involved updating the 260 original elec-

tronic searches). And fourth, we applied meta-regression to examine moder-

ating effects of study characteristics. This procedure allowed us to examine

the unique contribution of individual variables (e.g., grade) in accounting for

variability in study effects, after variability due to other variables (e.g., partici-

pant training and study quality) were ﬁrst controlled. While the central results

and recommendations of Writing to Read remained the same, our use of cur-

rent meta-analytic procedures allowed us to draw more nuanced conclusions

and increased our conﬁdence in the ﬁndings.

We anticipated that writing about reading would enhance students’ com-

prehension of text, that writing instruction would improve students’ reading

714

Harvard Educational Review

skills, and that increasing how much students wrote would improve their read-

ing. These predictions were based on the three theoretical views of reading-

writing connections discussed earlier. We further anticipated that the variabil-

ity of study effects would exceed the anticipated variability due to sampling

error, at least for some of our analyses. We thought this was especially likely

for Question 1 (Does writing about material read enhance students’ compre-

hension of text?), as there are many different ways to write about text. Thus, in

addition to conducting an overall analysis of the effects of writing about text,

we examined the effects of speciﬁc types of writing (e.g., summary writing).

When excessive variability was evident in effects for Question 1, and enough

studies were available to conduct a meta-regression (for all studies in general

or a particular type of writing), we were particularly interested in the unique

and mediating inﬂuence of four variables: grade (as text becomes increas-

ingly difﬁcult with grade; Baker, Dreher, & Guthrie, 2000), subject area (as

the impact of writing about text may differ by content; Bangert-Drowns et al.,

2004), participant training (as writing about text may be more effective when

students are taught how to do this; Graham & Perin, 2007a), and study quality

(Lipsey & Wilson, 2001).

It should be noted that participant training in Question 1 involved teach-

ing students how to use a speciﬁc writing activity as a tool for understanding

material read, whereas writing instruction in Question 2 (Does writing skills

instruction strengthen students’ reading skills?) was not directly connected to

reading or reading instruction (instruction here was focused on learning how

to write). Both Questions 2 and 3 examined if the impact of writing instruc-

tion or increased writing generalized to reading.

Method

Study Inclusion and Exclusion Criteria

Studies had to meet the following six criteria to be included in this review:

Was a true experiment (assignment to conditions is random) or a quasi-1.

experiment (assignment to conditions is not random)

Involved a treatment group that wrote about what they read, were taught 2.

to write, or increased how much they wrote

Included at least one reading measure that assessed the impact of the 3.

writing treatment or condition; quasi-experimental studies had to include

a comparable pretest reading measure since students were not randomly

assigned to conditions

Involved students in grades 1–12 4.

Was published in English5.

Contained the statistics necessary to compute a weighted effect size (or 6.

statistics were obtained from the authors)

715

Writing to Read

steve graham and michael hebert

We excluded studies for the following four reasons. One was if the writing

treatment or condition did not involve the creation of meaningful text. Con-

sequently, studies where students copied text verbatim, practiced typing, wrote

single words, and added a missing word to a sentence were excluded. The

only exception involved studies where the writing treatment involved teaching

spelling. Numerous literacy theorists claim that spelling instruction enhances

word reading skills (e.g., Adams, 1990; Ehri, 1987), and such instruction com-

monly involves copying or writing individual words (Graham et al., 2008).

We also excluded studies if the control condition wrote or received writing

instruction. There were two exceptions to this rule. We did include a study if

treatment and control students received the same amount of writing or writing

instruction as part of their typical language arts program, but the experimen-

tal manipulation for the writing condition involved additional writing or writ-

ing instruction. We also included a study if students in the control condition

copied text (this never occurred) or completed written cloze activities (this

occurred once).

We also excluded studies if the writing treatment was tested in a special

school for students with disabilities (e.g., school for the deaf), as the moder-

ating factors to be controlled in such settings warrant special care/attention

beyond the scope of this meta-analysis. And, ﬁnally, we excluded studies if the

only reading outcome measure was identical to the writing treatment, as the

treatment and assessment of it could not be separated.

Search Strategies Used to Locate Studies

To identify possible studies for this review, we conducted 260 electronic

searches (ending in January 2010) in four databases: ERIC, PsychINFO, Edu-

cation Abstracts, and ProQuest (including Dissertation Abstracts Interna-

tional). We both read each item identiﬁed in these searches (more than four-

teen thousand). If the item looked promising based on its abstract or title,

we obtained it (agreement between us was 99.2 percent, with disagreements

resolved by Graham). We hand-searched the following peer-reviewed journals:

Assessing Writing, Journal of Literacy, Reading and Writing: An International Jour-

nal, Reading Research and Instruction, Reading Research Quarterly, Research in the

Teaching of Writing, Scientiﬁc Studies of Reading, and Written Communication. Other

sources for locating studies included the report from the National Reading

Panel, Teaching Children to Read (NICHD, 2000), as well as chapters examining

the relationship between writing and reading in inﬂuential books, such as the

Handbook of Reading Research (Kamil, Mosenthall, Pearson, & Barr, 2000). Once

we obtained a document, we searched the reference list to identify other stud-

ies. Of 752 documents collected, 95 experiments met the inclusion criteria.

Through our independent reading, we established reliability for selected doc-

uments, with only three disagreements (reliability = 99.6%), which we resolved

through discussion.

716

Harvard Educational Review

Categorizing Studies According to Questions and Methods

We read each study and placed it into a category based on the question it

answered. Next, we further examined studies assigned to a speciﬁc question

and placed them into pre-identiﬁed instructional subcategories. For Question

1 (impact of writing about text read), these categories were:

Answering questions in writing (writing short answers to questions about 1.

text before, during, or after reading it) or generating questions in writ-

ing to ask about text

Taking written notes about text during or after reading it; notes could 2.

be unstructured or organized via an outline, graphic organizer, column

method, and so forth

Summarizing, or self-generating written synthesis of text read as well as 3.

summaries written using a speciﬁc format or with a set of speciﬁc rules

Extended writing, or written responses that extended beyond single 4.

statements in response to questions, notes, or summaries and focused

on students’ personal reactions to material read; analysis, interpretation,

or application of the material presented in text; or explaining the text

material to others

Writing short responses about text read (brief analogy, metaphor, and 5.

compare/contrast statement)

All Question 1 studies had a reading comprehension outcome measure.

We initially placed studies assigned to Question 2 (impact of writing instruc-

tion) in two categories: process writing (an approach that involves creating

a supportive writing environment where instruction is typically personalized

and students engage in cycles of planning, translating, and revising text; see

Graham & Perin, 2007b, for a more complete deﬁnition) and skills instruction

(writing skills taught systematically). We modiﬁed the categorical structure of

Question 2, however, as few studies (k = 3; 14%) actually assessed the impact

of process writing instruction on reading. Based on a subsequent reanalysis of

the studies, we decided to categorize investigations by their impact on read-

ing outcomes. This resulted in three categories that assessed the impact of:

(1) process writing, text structure, and paragraph/sentence skills instruction

on reading comprehension; (2) sentence/spelling instruction on reading ﬂu-

ency; and (3) spelling instruction on word reading skills.

Question 3 (impact of extra writing) contained a single category: studies

that increase the amount of student writing. Reading comprehension was the

outcome measure in all of these studies.

We used a variety of outcome measures to assess reading outcomes across

the three questions, including researcher-devised and standardized norm-

referenced tests. For example, researcher-devised measures of reading com-

prehension included answering questions about text (multiple choice and

short answers), retelling what was read (orally or in writing), summarizing

text read in one sentence, and identifying words systematically omitted from

717

Writing to Read

steve graham and michael hebert

text (cloze procedure). No single reading measure was used by a majority of

investigators for the three reading constructs studied (comprehension, ﬂu-

ency, and word reading), and multiple measures for the same construct were

applied in many studies.

Study Feature Coding

We coded each study for study descriptors, quality indicators, and variables

necessary to calculate effect sizes. Study descriptors included: grade, type of

student (e.g., struggling writers, English language learners, etc.), number of

participants, locale, treatment length, participant training, description of the

treatment, description of the control condition, subject, genre, outcome mea-

sures, publication type, and research design.

There were eleven quality indicators:

Random assignment of participants to conditions1.

Total attrition of less than 10 percent of total sample 2.

Total attrition of less than 10 percent and equal attrition across condi-3.

tions (i.e., conditions did not differ by more than 5 percent)

Control condition speciﬁcally described4.

Treatment ﬁdelity greater than .805.

Teacher effects controlled (e.g., random assignment of teachers)6.

More than a single teacher carrying out each condition7.

Reliability of reading measure established and greater than .608.

Posttest ceiling/ﬂoor effects for reading measure not evident (more 9.

than one standard deviation [SD] from ﬂoor and ceiling)

Pretest ceiling/ﬂoor effects not evident for reading posttest measure 10.

in quasi-experiments (more than one SD from ﬂoor and ceiling)

Pretest equivalence evident for pretest reading measure in quasi- 11.

experiments (i.e., conditions did not differ by more than 0.5 SD)

Each quality indicator was scored as 1 (met) or 0 (not met), except for

quality indicators related to measures (8–11 were scored for the proportion

of measures that met the criteria). A total score was calculated for each study

(nine possible points for true experiments and eleven possible points for quasi-

experiments). This was converted to a proportion by dividing obtained score

by total possible points.

Independently, we both completed coding for study descriptors and qual-

ity indicators, as well as calculation of effect sizes. We resolved disagreements

by reexamining the study. Our initial agreement was 94.8 percent for all

variables.

Calculation of Effect Sizes

For true experiments, an effect size (ES) was calculated by subtracting the

mean score of the treatment group at posttest (XT) from the mean score of

the control group at posttest (XC) and dividing by the pooled standard devia-

718

Harvard Educational Review

tion of the two groups (sp). For quasi-experiments, we used the same proce-

dure, except we ﬁrst adjusted pretest differences between treatment and con-

trol conditions by subtracting mean pretest score for each group from their

mean posttest score. We divided the difference between the adjusted means

for treatment and control by the pooled standard deviation for posttest, as rec-

ommended by Lipsey and Wilson (2001). For some quasi-experiments, it was

necessary to estimate the posttest ES from covariate-adjusted posttest means.

In a few instances, we had to compute an ES separately for pretest and post-

test as the pretest and posttest used different measures to assess the same con-

struct (e.g., reading comprehension); we obtained an adjusted ES by subtract-

ing pretest ES from posttest ES.

As a prelude to calculating some ESs, it was sometimes necessary to average

the performance of two or more groups in each condition (e.g., statistics were

reported separately by grade). We did this by using procedures recommended

by Nouri and Greenberg (as cited in Cortina & Nouri, 2000). It was also neces-

sary in some cases to estimate missing means and standard deviations from the

statistics reported by the authors. For example, we calculated missing standard

deviations by estimating residual sums of squares to compute a root mean

squared error (RMSE) (e.g., Shadish, Robinson, & Congxiao, 1999). For cova-

riate or complex factorial designs, we estimated pooled standard deviations by

calculating and restoring the variance explained by covariates and other “off-

factors” to the study’s error term and recalculating the pooled standard devia-

tion from the composite variance.

The quasi-experiments in this review assigned whole classes to treatment

or control conditions and then examined student-level effects. Adjusting stan-

dard errors (SE) for these studies was necessary, as a portion of the total vari-

ance in such quasi-experiments was likely due to grouping or clustering within

treatments, with the total variance representing a sum of group and student

variances. To correct effect variance, we imputed an estimate of the ratio of

between-group variance to total variance (i.e., intraclass correlation [ICC]; see

Hedges, 2007) from ICC estimates provided by Hedges and Hedberg (2007).

We based these estimates on the ﬁndings from national intervention studies

where the reading comprehension outcomes were adjusted for pretest covari-

ates. If an adjusted ICC was not available (this was the case for several grades),

we applied an unadjusted ICC for that grade. If an unadjusted ICC for a grade

was not available, we based ICC on an average of the adjusted ICCs for the grade

levels adjacent to it. If a quasi-experiment in this review included more than

one grade level, we averaged the ICCs for those grades. Most quasi-experiments

had equal sample sizes across clusters, and we assumed equal cluster sizes when

this information was not provided. We adjusted all computed effects for true

and quasi-experiments for small-sample-size bias using the formula

dadj = d * γ

where γ = 1 – 3/4 (ntx + nctrl ) – 9 (Hedges, 1982)

719

Writing to Read

steve graham and michael hebert

Because using multiple ESs from the same study violates the assumption

of independent data points fundamental to most statistical procedures (Wolf,

1986), we used three procedures to preserve the statistical independence of

ESs in this review. First, if a study contained more than one ES for a single

reading construct (e.g., reading comprehension), we combined these effect

sizes as an unweighted average effect. Such aggregation is preferable when

intercorrelations among measures are unknown (which was the case in this

review), as standard error estimation is complicated when this information is

missing (Gleser & Olkin, 1994).

Second, we separated ESs for different constructs (e.g., reading compre-

hension and word reading) for stratiﬁed analyses (i.e., they were not aggre-

gated together in any statistical analysis). We did this also for researcher-

designed and standardized norm-referenced measures of the same construct,

except for analyses involving the effects of writing instruction on reading ﬂu-

ency and word reading (Question 2). Both of these analyses involved a small

number of studies, and the researcher-devised and norm-referenced measures

for each construct were similar.

Third, if a study had multiple treatment or control conditions, then we

selected only one treatment comparison for inclusion in any analysis (i.e., if

the study included a full version and a partial version of the treatment, then

we compared only the full version to the control condition). There were some

exceptions to this rule. One, in papers reporting multiple studies (Barton,

1930; Doctorow, Wittrock, & Marks, 1978; Vidal-Abarca & Gilabert, 1995), we

computed an ES for each study. Two, in studies where one version of a treat-

ment was compared to a control condition and another version of a treatment

was compared to a separate control condition (Denner, 1987; Slater, 1982), we

computed an ES for each treatment-control comparison. Three, in investiga-

tions where two different treatments (e.g., process writing and skills instruc-

tion; Kelley, 1984) or two different dosages of a treatment were compared to

a single control (Weber & Henderson, 1989), we computed a separate ES for

each comparison.

Statistical Analysis of Effect Sizes

Our meta-analysis employed a weighted random-effects model. For each of

the three research questions, we calculated an average weighted ES (weighted

to take into account sample size by multiplying each ES by its inverse vari-

ance). We did not calculate an average weighted ES unless there were at least

four studies addressing the question or subquestion. We also calculated a

conﬁdence interval and statistical signiﬁcance of the obtained weighted ES,

as well as two measures of homogeneity (Q and I2). The homogeneity mea-

sures allowed us to determine if variability in the ESs for an average weighted

effect was larger than expected based on sampling error alone. When

this was the case, and there were at least sixteen ESs, we conducted meta-

regression moderator analysis to determine if this excess variability could be

720

Harvard Educational Review

accounted for by identiﬁable differences between studies (e.g., training ver-

sus no training).

The meta-regression involved a mixed-effects model with maximum likeli-

hood estimates using macros developed for SPSS. Meta-regression is similar to

traditional regression analysis, where the contribution of speciﬁc variables to

the prediction of the target outcome measure (in this case, ESs for individual

studies) is estimated (see Konstantopoulous & Hedges, 2009). We assumed

that in addition to a random effect due to sampling error, there was a sys-

tematic component to the variance among studies. The macros added a ran-

dom effect variance component and recalculated the inverse variance weight

before reﬁtting the model (Lipsey & Wilson, 2001). We entered the predictor

variables as a single block.

Results

Table 1 contains information about each study for Questions 1–3. Within each

question, studies are arranged from early to later grades. The following infor-

mation is reported: grade(s), type of student, treatment and control compari-

son, training for treatment, subject area, genre of writing, total quality of study

score, and ES (an asterisk denotes an ES from a norm-referenced test). For

studies that contain multiple treatments, we give the overall ES as well as the

ES for each speciﬁc treatment (bracketed).

Quality of Research

For Questions 1–3, the proportion of studies meeting each of the eleven quality

indicators is presented in table 2. While random assignment occurred in just 48

percent of all studies, most quasi-experiments evidenced pretest equivalence.

We decided to retain quasi-experiments in this meta-analysis even though there

was a relatively large number of true experiments, because there was no statisti-

cally signiﬁcant difference in average weighted effect sizes for true and quasi-

experiments (p = .97; ESs for true and quasi-experiments were each 0.50).

Pretest ceiling/ﬂoor effects were evident for 42 and 50 percent of measures

for Questions 2 and 3 studies, respectively. Most posttest measures (73%) did

not evidence ceiling/ﬂoor effects, and we established that acceptable reliabil-

ity was evident for close to two-thirds of posttest measures. We were able to

deﬁne the control conditions in most studies (82%), but rarely did we ﬁnd

that researchers established treatment ﬁdelity (22%), and there was only one

teacher per condition in 49 percent of studies (this was especially problematic

for Question 1). For studies in Questions 1 and 2, attrition and teacher effects

were not major issues, as 70 percent of studies met both attrition criteria and

teacher effects were controlled in 69 percent of studies (attrition and teacher

effects were major concerns for Question 3). The mean proportion of quality

indicators met for studies in Questions 1 and 2 were .63 each (SDs = 0.13 and

0.19, respectively) and .55 for Question 3 (SD = 0.22).

721

Writing to Read

steve graham and michael hebert

TABLE 1 Studies and effect sizes included in the analysis of each research question

Study Grade Students Treatment Training Content Genre Quality Measure ES

Question 1: Effect of writing about reading

Adams-Boating, 2001 2 FR EW vs. RI NT LA N 0.55 C 1.02

Denner et al., 1989 2 FR SW vs. NW NT LA N 0.89 C 0.61

Cohen, 1983 3 A & BA QG vs. BAU-RI T LA N 0.78 C *0.75

MacGregor,1988 3 A & AA QGA vs. R NT LA E 0.44 C *0.34

Bayne, 1984 3 & 4 FR NT-S vs. RI T LA N 0.64 C *0.14

Jenkins et al., 1987 3–6 LD SUM vs. BAU-RI T LA E 0.73 C 0.68

Jaekyung et al., 2008 4 & 5 FR EW vs. R T LA N & E 0.67 C *0.36

Saunders & Goldenberg, 1999 4 & 5 FR & ELL EW vs. R+ST NT LA E 0.67 C 0.08

Chang & Sung, 2002 5 FR NT-S vs. R T/NT SCI E 0.70 C 0.49

Jennings, 1990 5 FR

MULT

{SUM vs. BAU-RI}

{EW vs. BAU}

SS E 0.55 C

0.88

{0.34}

{1.83}

Newlun, 1930 5 FR SUM vs. R+ST T SS E 0.64 C *0.36

Linden & Wittrock, 1981 5 FR & ELL SS vs. R, D, & RI NT LA N 0.67 C 0.92

Amuchie, 1983 5 & 6 ELL SUM vs. BAU-RI T LA N & E 0.45 C 1.37

Leshin, 1989 5 & 6 FR NT-U vs. U NT SCI E 0.89 C 0.43

Berkowitz, 1986 6 GR & PR

MULT

{QA vs. R+RR}

{NT-S vs. R+RR}

T SS E 0.64 C

0.63

{0.35}

{0.87}

Coffman, 1992 6 FR QA vs. R NT NR N 0.83 C 0.32

Taylor & Berkowitz, 1980 6 A & AA

MULT

{QA vs. R+DT}

{SUM vs. R+DT}

SS E 0.61 C

0.32

{0.26}

{0.43}

Copeland, 1987 6 GW & PW EW vs. R+RR NT LA E 0.67 C 1.27

Doctorow et al., 1978 6 GW SUM vs. R NT NR N 0.89 C 1.56

Doctorow et al., 1978 6 PW SUM vs. R NT NR N 0.89 C 0.98

(continued)

722

Harvard Educational Review

TABLE 1 Studies and effect sizes included in the analysis of each research question

Study Grade Students Treatment Training Content Genre Quality Measure ES

Keown, 2008 6 FR NT-S vs. R+D T SCI E 0.73 C 0.44

Rinehart et al., 1986 6 FR SUM vs. R+WS T SS E 0.73 C 0.49

Vidal-Abarca & Gilabert, 1995 6 FR NT-S vs. BAU-RI T SCI E 0.45 C 0.21

Olsen, 1991 6 & 8 FR EW vs. BAU-RI NT LA N 0.73 C *0.53

Ryan, 1981 6–8 FR

MULT

{SUM vs. R}

{NT-U vs. R}

NT LA N 0.55 C

0.50

{0.42}

{0.52}

Barton, 1930 7 FR NT-S vs. R T SS E 0.27 C 0.63

Denner, 1987 7 FR NT-U vs. R+RR T LA N 0.73 C 0.45

Denner, 1987 7 FR NT-S vs. R+RR T LA N 0.73 C 0.70

Taylor & Beach, 1984 7 FR

MULT

{QA vs. BAU-RI}

{SUM vs. BAU-RI}

T SS E 0.55 C

0.43

{0.26}

{0.75}

Bigelow, 1992 7 & 8 FR NT-S vs. R NT SCI E 0.67 C 0.78

Denner & McGinley, 1992 7 & 8 FR SW & ShS vs. NW NT LA N 0.78 C 0.60

Salisbury, 1934 7, 9, & 12 FR SUM vs. BAU-RI T LA E 0.36 C *0.57

Armbruster & Anderson, 1980 8 FR NT-S vs. BAU T SCI E 0.36 C 0.34

Ballard, 1988 8 FR NT-S vs. DRTA T LA E 0.45 C 0.28

Chang, 1987 8 FR NT-S vs. R T SCI E 0.63 C 0.69

Denner, 1992 8 GR & PR

MULT

{NT-S vs. R+RR}

{NT-U vs. R+RR}

T SS E 0.89 C

0.54

{0.64}

{0.45}

Vidal-Abarca & Gilabert, 1995 8 FR NT-S vs. BAU-RI T SCI E 0.64 C 0.22

Bates, 1981 9 GR & PR SUM vs. R+RR NT LA N 0.89 C –0.17

Faber et al., 2000 9 GR & PR NT-S vs. R T SS E 0.78 C 0.03

Klugh, 2008 9 FR (SWD) NT-S vs. BAU T SS E 0.36 C 0.10

Slater, 1982 9 FR NT-S vs. R+S NT SS E 0.67 C 0.76

Slater, 1982 9 FR NT-U vs. R NT SS E 0.67 C 0.87

Trasborg, 2005 9 & 10 PR SUM vs. BAU-RI T SS E 0.73 C *0.38

Langer & Applebee, 1987 9 & 11 FR

MULT

{QA vs. R+S}

{SUM vs. R+ST}

{EW vs. R+ST}

NT SS E 0.63 C

0.37

{0.06}

{0.51}

{0.62}

Peverly & Wood, 2001 9–11 RD QA vs. R NT LA N 0.68 C *0.44

Barton, 1930 9–12 FR NT-S vs. R+RI T SS E 0.36 C 0.37

Matthews, 1938 9–12 FR NT-U vs. R NT SS E 0.89 C –0.15

Placke, 1987 9–12 LD SUM vs. TT T SS E 0.73 C 0.53;

* –0.59

Tsai, 1995 9–12 FR SUM vs. R+ST T SCI E 0.50 C 0.28

Bean et al., 1983 10 AA QG vs. R+D T SS E 0.45 C 0.26

Graner, 2007 10 NLD & LD SUM vs. TT T LA E 0.77 C 0.20

Hayes, 1987 10 A & AA

MULT

{SUM vs. R+M}

{QG vs. M}

{ShS vs. R+M}

NT SCI E 0.68 C

0.04

{–0.01}

{0.14}

{0.11}

Weisberg & Balajthy, 1989 10–12 PR SUM vs. R+D T SS E C 0.43

(continued)

723

Writing to Read

steve graham and michael hebert

TABLE 1 Studies and effect sizes included in the analysis of each research question

Study Grade Students Treatment Training Content Genre Quality Measure ES

Keown, 2008 6 FR NT-S vs. R+D T SCI E 0.73 C 0.44

Rinehart et al., 1986 6 FR SUM vs. R+WS T SS E 0.73 C 0.49

Vidal-Abarca & Gilabert, 1995 6 FR NT-S vs. BAU-RI T SCI E 0.45 C 0.21

Olsen, 1991 6 & 8 FR EW vs. BAU-RI NT LA N 0.73 C *0.53

Ryan, 1981 6–8 FR

MULT

{SUM vs. R}

{NT-U vs. R}

NT LA N 0.55 C

0.50

{0.42}

{0.52}

Barton, 1930 7 FR NT-S vs. R T SS E 0.27 C 0.63

Denner, 1987 7 FR NT-U vs. R+RR T LA N 0.73 C 0.45

Denner, 1987 7 FR NT-S vs. R+RR T LA N 0.73 C 0.70

Taylor & Beach, 1984 7 FR

MULT

{QA vs. BAU-RI}

{SUM vs. BAU-RI}

T SS E 0.55 C

0.43

{0.26}

{0.75}

Bigelow, 1992 7 & 8 FR NT-S vs. R NT SCI E 0.67 C 0.78

Denner & McGinley, 1992 7 & 8 FR SW & ShS vs. NW NT LA N 0.78 C 0.60

Salisbury, 1934 7, 9, & 12 FR SUM vs. BAU-RI T LA E 0.36 C *0.57

Armbruster & Anderson, 1980 8 FR NT-S vs. BAU T SCI E 0.36 C 0.34

Ballard, 1988 8 FR NT-S vs. DRTA T LA E 0.45 C 0.28

Chang, 1987 8 FR NT-S vs. R T SCI E 0.63 C 0.69

Denner, 1992 8 GR & PR

MULT

{NT-S vs. R+RR}

{NT-U vs. R+RR}

T SS E 0.89 C

0.54

{0.64}

{0.45}

Vidal-Abarca & Gilabert, 1995 8 FR NT-S vs. BAU-RI T SCI E 0.64 C 0.22

Bates, 1981 9 GR & PR SUM vs. R+RR NT LA N 0.89 C –0.17

Faber et al., 2000 9 GR & PR NT-S vs. R T SS E 0.78 C 0.03

Klugh, 2008 9 FR (SWD) NT-S vs. BAU T SS E 0.36 C 0.10

Slater, 1982 9 FR NT-S vs. R+S NT SS E 0.67 C 0.76

Slater, 1982 9 FR NT-U vs. R NT SS E 0.67 C 0.87

Trasborg, 2005 9 & 10 PR SUM vs. BAU-RI T SS E 0.73 C *0.38

Langer & Applebee, 1987 9 & 11 FR

MULT

{QA vs. R+S}

{SUM vs. R+ST}

{EW vs. R+ST}

NT SS E 0.63 C

0.37

{0.06}

{0.51}

{0.62}

Peverly & Wood, 2001 9–11 RD QA vs. R NT LA N 0.68 C *0.44

Barton, 1930 9–12 FR NT-S vs. R+RI T SS E 0.36 C 0.37

Matthews, 1938 9–12 FR NT-U vs. R NT SS E 0.89 C –0.15

Placke, 1987 9–12 LD SUM vs. TT T SS E 0.73 C 0.53;

* –0.59

Tsai, 1995 9–12 FR SUM vs. R+ST T SCI E 0.50 C 0.28

Bean et al., 1983 10 AA QG vs. R+D T SS E 0.45 C 0.26

Graner, 2007 10 NLD & LD SUM vs. TT T LA E 0.77 C 0.20

Hayes, 1987 10 A & AA

MULT

{SUM vs. R+M}

{QG vs. M}

{ShS vs. R+M}

NT SCI E 0.68 C

0.04

{–0.01}

{0.14}

{0.11}

Weisberg & Balajthy, 1989 10–12 PR SUM vs. R+D T SS E C 0.43

(continued)

Hare & Borchardt, 1984 11 AA SUM vs. NW T SCI E 0.45 C 0.44

Wetzel, 1990 11 FR EW vs. R NT LA N 0.55 C *0.23

Andre & Anderson, 1979 11-12 A & AA QG vs. R T & NT PSY E 0.44 C 0.51

Bretzing & Kulhavey, 1979 11-12 FR SUM vs. R NT SS E 0.67 C 0.56

Kulhavey et al., 1975 11-12 FR NT-U vs. R+ST NT LA N 0.67 C 0.37

Schultz & Di Vesta, 1972 11-12 AA NT-U vs. R NT SS E 0.67 C 0.15

Bowman, 1989 12 FR EW vs. R NT LA N 0.55 C 0.47

Walko, 1989 12 FR NT-U & NT-S vs.

R+ST NT SCI E 0.89 C -0.11

Wong et al., 2002 12 FR EW vs. R+D NT LA N 0.55 C 0.85

Barton, 1930 HS FR NT-S vs. R T SS E 0.36 C 0.83

Licata, 1993 HS FR

MULT

{EW- vs. R+ST}

NT SCI E 0.64 C

0.45

{0.56}

{0.33}

Weisberg & Balajthy, 1990 NR PR SUM vs. R T SS E 0.33 C 0.81

724

Harvard Educational Review

(continued)TABLE 1 Studies and effect sizes included in the analysis of each research question

Study Grade Students Treatment Training Content Genre Quality Measure ES

Question 2: Effects of writing instruction on reading comprehension

Frey, 1993 1 FR PrW vs. RI T LA NA 0.73 C *0.12

Fuchs et al., 2006 1 PR & PM Sp vs. MI T LA NA 0.83 W

0.39

0.32

Uhry & Shepherd, 1993 1 FR Sp vs. RI T LA NA 0.67

*0.43

*1.78

*0.70

Conrad, 2008 2 A Sp vs. R T LA NA 0.67 W 0.62

Graham et al., 2002 2 WSP Sp vs. MI T LA NA 0.89 W *0.51

Weber & Henderson, 1989 3–5 FR Sp vs. RI T LA NA 0.67 W

0.54

1.17

Weber & Henderson, 1989 3–5 FR Sp vs. RI T LA NA 0.67 W

0.34

0.79

Hunt & O’Donnell, 1970 4 FR SC vs. RI T LA NA 0.27 C *0.26

Licari, 1990 4 FR PrW vs. RI T LA NA 0.55 C *0.41

Crowhurst, 1991 6 FR PE vs. R+D T SS NA 0.88 C 0.36

Hamby, 2004 6 PR WI vs. RI T LA NA 0.36 C *0.25

Kelley, 1984 6 FR

(no SWD) PrW vs. R T LA NA 0.82 C *0.40

Kelley, 1984 6 FR

(no SWD) WS vs. R T LA NA 0.82 C *0.31

Neville & Searls, 1985 6 FR SC vs.

R+Cloze T SS NA

0.36

*0.06

0.32

Hughes, 1975 7 FR SC vs. OL T LA NA 0.64 F *0.57

Shockley, 1975 7 PR WI vs. R+RI T LA NA 0.78 C *0.30

Jones, 1966 8 A & AA Sp vs. NoTr T SCI NA 0.73 C *0.22

NA C 0.22

Phelps, 1978 8 A SC vs. SC-NW T LA NA 0.55 C 0.32

Callaghan, 1977 9 & 11 FR SC vs. BAU T LA NA 0.36 C *0.02

Baker, 1984 10 FR EssW vs. R T LA NA 0.39 C 0.23

Gonsoulin, 1993 10–12 PR SC vs. RI T LA NA 0.55 C *0.17

Question 3: Effects of extra writing on norm-referenced and researcher-created reading comprehension outcome measures

Bode, 1988 1 FR DJ vs. RI NA LA NS 0.55 C *0.36

Healy, 1991 1 FR W-FY vs.

W-HY+RI NA LA NS 0.73 C *0.56

Ramey, 1989 1 FR W vs. R NA LA SELF 0.45 C *0.16

Sussman, 1998 1 WSP XJW vs. RRJ NA LA NS 0.67 C 0.22

Peters, 1991 2 FR W vs. R NA LA SELF 0.64 C *0.31

Reutzel, 1985 3 FR W vs. RI T LA N 0.7 C 0.87

Soundy, 1978 3–6 FR ExpW vs. SSR NA LA SELF 0.78 C *0.42

Roy, 1991 4 & 5 FR RJ vs. R NA LA NS 0.36 C *0.01

Dana et al., 1991 6 FR W vs. RI NA LA PPL 0.09 C 0.24

Notes: A = average, AA = above average, BA = below average, BAU = business as usual, C = comprehension, Cloze = cloze activities, D = discussion, DJ = dialogue journals,

DRTA = directed reading thinking activity, E = expository, ELL = English language learners, EW = extended writing, EssW = essay writing, ExpW = expressive writing, F

= reading ﬂuency, FR = full range, GR = good readers, GW = good writers, HS = high school, J = journals, LA = language arts, LD = learning disabled, M = math, MS =

middle school, N = narrative, NA = not applicable, NLD = non–learning disabled, NoTr = no treatment, NR = not reported, NS = not speciﬁed, NT = no training, NT-S = note

taking–structured, NT-U = note taking–unstructured, NW = no writing, PPL = pen pal letters, PR = poor readers, PrW = process writing, PSY = psychology, PW = poor writers,

QA = question answering, QG = question generation, QGA = question generation and answering, R = reading, RD = reading disabled, RI = reading instruction, RJ = reading

journals, RRJ = rereading journals, SC = sentence combining, SCI = science, SELF = self-selected topics, ShS = short statements, SI = spelling instruction, SS = social

studies, SSR = sustained silent reading, ST = studying, SW = story writing, SUM = summary writing, SWD = students with disabilities, T = training, TT = test taking, U =

underlining, W = word reading, W-FY = writing–full year, W-HY = writing–half year, WS = writing skills, WSP = weak spellers, XJW = extra journal writing

*Norm Referenced Measures

{ } = Speciﬁc treatment

725