The Effects of Writing on Learning in Science,Social Studies,and Mathematics: A Meta-Analysis

Cognitive Theories

One of the earliest theories championing the view that writing inherently elicits learning was offered by Britton et al. (1975), who placed special emphasis on the role of language in learning, arguing that language facilitates learning and that writing was a form of language (a derivative of expressive speech). Britton (1982) claimed that writers do not know exactly what they will say when they begin to convert an idea into written text, and that the semantics and syntax of language shape this process, resulting in new learning about an idea at the “point of utterance.” According to Britton, writing about content material most likely influences learning when it involves an authentic and personal act of meaning-making. He placed special emphasis on expressive writing. This form of writing was assumed to be closest to the spoken word, making it a useful tool for learning for students at all developmental levels.

A more recent theoretical model of how writing inherently elicits learning was offered by Galbraith and Baaijen (2018), who described a knowledge-constituting process which evokes cognitive operations and structures that operate below the level of conscious thought. This model draws on connectionist principles of information processing. A writer’s semantic knowledge is stored in long-term memory “as a fixed set of connection strengths between units within a distributed architecture . . .” (pp. 241–242). It is assumed that the writer’s understanding of a concept is implicit and not directly retrievable. It is only through the act of composing text that the writer discovers or has access to this knowledge. More specifically, the writer must synthesize these implicit connections in memory until an initial fusion of understanding is obtained. This initial understanding of knowledge may undergo additional modification and synthesis as the writer evaluates or considers other ideas or additional text is created. The extent to which writing leads to new learning depends on the extent to which the content produced by the writer in this way differs from content already held in the writer’s episodic memory. When this differs, the writer presumably experiences a new development in understanding.

Galbraith and Baaijen (2018) proposed a second and effortful process by which writing facilitates learning. This knowledge-transforming process occurs when writing ideas and content retrieved through episodic memory, external sources, or produced through the knowledge-constituting processes (described above) are subjected to evaluation and manipulation in working memory in order to satisfy the learner’s goals. Galbraith and Baaijen claimed that this process operates best when the ideas under consideration are represented in a fixed form (e.g., notes, a sentence in text) so that working memory is not overtaxed and the writer can concentrate on evaluating how well the target information addresses their learning goals and contributes to how the overall text is structured. This process transforms the writer’s understanding of information by creating a more organized representation in memory.

Somewhat similarly, Silva and Limongi (2019) proposed that writing about content material facilitates learning by consolidating information in long-term memory. This consolidation is achieved through rehearsal of information, elaboration of it, or both. To illustrate, the evaluation of whether information from long-term memory or external sources satisfies a learner’s goals requires active maintenance of these goals and information in working memory. This can be achieved only through rehearsal. Likewise, the evaluation and manipulation of this information in working memory are likely to involve elaboration. Even if the ideas held in working memory are not transformed through elaboration, rehearsal can lead to learning as when the writer acquires new information from external sources or becomes more familiar and facile with ideas drawn from long-term memory.

Other cognitive theories of how writing affects learning emphasize the agency of the learner, indicating that the goals and strategies students apply when writing about information leads to learning (Klein & Boscolo, 2016). When writing about content material, students must make decisions about how to handle new information and execute strategies to satisfy their goals (Graham & Hebert, 2011). Such strategies include but are not limited to the following: identifying which information is most important and needs to be presented in text (fostering explicitness), organizing ideas into a coherent written whole (establishing specific relationships among ideas), deciding how best to present content material in writing (fostering a personal involvement with the information), transforming ideas into written sentences (thinking about what these ideas mean to create text that satisfies the writer’s intentions), and reexamining, evaluating, or revising the text generated (leading to additional reflection and retention of information). The deliberate use of any of these strategies, whether they involve the pursuit of rhetorical goals (i.e., backward search strategies) or the revising of text iteratively to resolve contradictions (i.e., forward search strategies), presumably results in specific learning effects (Klein, 1999).

Specific effects of writing on learning were also championed by Bangert-Drowns et al. (2004). They argued that particular writing tasks promote learners’ use of cognitive learning strategies. For instance, writing can serve as a rehearsal strategy, such as when learners record in a journal the most important things learned, increasing both content exposure and time on task. Similarly, writing a report about a particular subject can evoke the use of organization and elaboration learning strategies, as learners build a structure for the text, link new learning to current understandings, and synthesize and elaborate on their ideas. Specific writing activities may further encourage comprehension monitoring and general cognitive awareness. For example, this can happen when the writing activity prompts students to reconsider old ideas in light of new information. As noted by Langer and Applebee (1987), however, different kinds of writing-to-learn activities promote different kinds of thinking and lead to differences in what is learned.

Klein (1999) identified an additional way in which genre can influence learning, indicating learning can occur when writers “use genre structures to organize relationships among elements of text, and thereby among elements of knowledge” (p. 203). Genres are particular kinds of text, which are commonly distinguished by purpose as well as text structure, and they are embodied though genre elements that form specific relationships with each other. Examples of genres used to promote learning are argumentation, informational writing, and narrative. Writing scholars (Langer & Applebee, 1987) contend that as writers generate and specify relationships among content appropriate to each genre element (e.g., evidence, claims, and warrants in argumentation), they acquire new learning as they construct corresponding relations among their own knowledge.

Social-Cultural Theories

The explanations presented so far on how writing supports learning focused on cognitive mechanisms. Learning effects from writing can also be approached from a sociocultural view. Accordingly, writing is a tool that learners use in a goal-directed fashion to mediate meaning construction, thought, and activity (Smagorinsky, 1995). Both learning and writing are considered a product of a system that includes people, community goals, social practices, typified actions, tools, and a collective history (Graham, 2018). Members of a specific system use writing to accomplish community goals (e.g., writing-to-learn), applying sanctioned social practices and actions to do so. Theoretically, the effects of writing on learning vary depending on the purposes of the writer’s community. Writing does not automatically enable a learner to construct new meaning. Instead, it represents the potential to do so (Wertsch, 1991).

Accordingly, social context is pivotal in writing-to-learn (Langer & Applebee, 1987). In social-cultural theories, the effects of writing on learning are attributed to social, institutional, cultural, historical, and political factors (Graham, 2018). For example, social context can exert its effects by supporting or not supporting the use of specific instructional procedures, like writing-to-learn (Klein, 1999). Writing may be a useful tool for promoting learning in a class where writing and writing for learning are sanctioned and valued, but it is not effective in a class where this is not the case. As a result, the effects of writing on learning vary according to the purposes, values, and history of the institutions in which writing is applied as a tool for learning. Its effects are also expected to vary depending on learners’ goals, values, identity, expectations, and beliefs.

Social-cultural theories further emphasize that writing becomes embedded and particular to the context in which it is learned (Smagorinsky, 1995). Consequently, a writing-to-learn activity that becomes a typical and routinized tool for accomplishing goal-directed learning in one classroom may not transfer readily to another classroom. The success of a writing-to-learn activity then depends on the instructional environment in which it is implemented, as a complex set of values, social interactions, and typified instructional practices moderate its impact.

Theoretical Implications

Both cognitive and social-cultural theories provide multiple rationales for how writing can facilitate learning. Cognitively, writing elicits implicit and explicit use of cognitive operations and structures that can facilitate learning. Socially, writing provides members of a community with a tool for learning that can be used in a goal-directed fashion to foster thinking, meaning-making, and action. These theories provide support for the proposition, tested in this meta-analysis, that writing about content facilitates learning for school-aged students.

These theories also raise questions about the conditions under which writing-to-learn is effective. This led to the second purpose of this meta-analysis: the identification of moderator variables that are reliably related to the magnitude of effects obtained with writing-to-learn activities. Such an analysis provides a more nuanced view of the effects of writing on learning. Theoretically derived moderators in this review included content area, grade level, features of the writing-to-learn activities, features of assessment, and features of instruction. We indicate why such moderators are theoretically important below. Unfortunately, we were unable to examine if the social, cultural, institutional, political, and historical factors so important to social-cultural theory also accounted for variability in study effects. The available writing-to-learn investigations provided little, if any, documentation about these factors (Klein, 1999).

Purpose

Because of the theoretical promise of writing-to-learn, the issues surrounding previous meta-analyses, and the accumulation of new studies (two thirds of the studies in this meta-analysis were new), we conducted a meta-analysis examining the effects of writing on learning. We limited this analysis to writing about science, social studies, and mathematics in Grades 1 to 12, as most of the writing-to-learn research at these grades involves these disciplines (Miller et al., 2018). We did not include studies with college students, as they are fundamentally different from elementary and secondary students and are also a more select group of learners (Rivard, 1994). Similar to previous meta-analyses, only true and quasi-experiments were examined.

We further addressed multiple issues evident in the two prior meta-analyses regarding methodological design. Quasi-experiments had to include pretest and posttest measures of learning. Treatment and control conditions had to devote equal instructional time to learning and teach the same basic content material. Learning measures were excluded from any analyses if pretest scores approached ceiling levels (it is more difficult to demonstrate learning effects when this occurs). Grades as a measure of learning were also excluded, as they can reflect more than just academic growth. All variables were operationalized and reliability of coding was established. ESs were adjusted for small sample size, outliers were winzorized, clustering effects for quasi-experiments were corrected, and possible publication bias was examined.

Like the two previous meta-analyses, we examined both published and unpublished studies. As described above, a number of criteria were instituted to ensure that all of the included investigations were methodologically sound. We further evaluated each study on the following nine quality indicators: research design, reliability of measures, pretest equivalence, posttest ceiling effects, posttest floor effects, teacher effects controlled, multiple classes in each condition, problems with attrition, and treatment fidelity. This allowed us to determine if study quality was related to variability in obtained effects, and to draw methodological recommendations for future research (this was not done in the previous meta-analyses). Moderator analyses also involved characteristics of participants, writing activities, assessments, and treatment/control comparisons, which included a broader array of variables than were included in Bangert-Drowns et al., 2004).

Research Questions and Predictions

We asked the following research questions for students in Grades 1 to 12:

We anticipated an affirmative answer to each research question. We expected that writing would enhance learning in the studies reviewed (RQ1), basing this prediction on the multiple theoretical explanations for writing-to-learn effects presented earlier. This prediction was further consistent with findings from Bangert-Drowns et al. (2004) and Graham and Perin (2007).

It was further expected that writing-to-learn effects would be moderated by content area (RQ2). We predicted writing would result in statistically significant different effects in science, social studies, and mathematics, and that effects would be larger in disciplines where writing is more valued and sanctioned (Smagorinsky, 1995). This was based on findings that teachers in different disciplines place more or less emphasis on writing (Gillespie et al., 2014).

We anticipated that the learning gains of students who were in earlier grades would be smaller than those made by students in later grades (RQ3). Students in earlier grades may not have acquired needed content knowledge or writing skills, making it less likely they could effectively use writing to support learning. We further expected that features of the writing-to-learn activities would be related to variability in ESs (RQ3). Different types of writing (e.g., argumentation, informational) contain different structures for organizing relationships among knowledge (Klein, 1999); the depth of processing promoted by a writing activity through analysis and interpretation should create more opportunities than merely recording information; metacognitive prompting is more likely to encourage learners to evaluate and modify their understandings (Hacker, 2018); and the highest level of learning a writing activity was designed to promote (e.g., knowledge, comprehension, application, analysis, synthesis, or evaluation; i.e., Bloom et al., 1956) should influence what is learned.

It was expected that features of instruction (RQ3) would be associated with writing-to-learn effects. Longer treatments or ones that involve more writing should result in stronger effects than shorter treatments or ones that involve less writing, as learners have more opportunities to become proficient in using the writing activities. We also anticipated that professional development as well as teaching students to use writing activities would produce larger effects, as teachers receiving such preparation would be better prepared to implement these activities and students better prepared to apply them, respectively. We further assumed that larger effects would be obtained when teachers implemented instruction when compared to researchers. Teachers know their students better than researchers, and teacher-led instruction signals that writing-to-learn activities are sanctioned and valued.

We predicted that features of assessments would also be correlated with obtained ESs (RQ3). Different types of measures (e.g., multiple-choice or essay tests) as well as measures that assess different levels of knowledge (as specified by Bloom et al., 1956) focus on distinct aspects of learning and should yield different effects. Measures created by a researcher should result in higher effects than measures designed by developers of standardized assessments, as researchers often align their assessments closely to the goals of the experiment and standardized measures by design do not. Alignment between assessments and content material should also create variable effects, as measures more closely aligned to content are likely to produce larger effects. Last, the match between the type of knowledge assessed and the type of knowledge that the writing activities is designed to promote should result in variability in effects. For instance, a study in which the assessment is exactly matched to the learning goals of the writing activity should produce a larger effect than a study in which the assessment measures a higher level of knowledge than the writing activity promotes.

We further expected that variability in effects would be related to features of the experimental methods used to assess the effects of writing-to-learn. We anticipated that type of treatment and control comparisons (RQ3) would account for variability in effects, as writing effects are likely to be stronger in studies where writing is compared to no writing versus studies in which more and less writing are compared. We also anticipated that measures of study quality would account for variability in effects (RQ3). Indicators of study quality such as study design, reliability of measures, and ceiling/floor effects are likely to contribute to variability in the effects of writing-to-learn. To illustrate, quasi-experiments may produce larger effects than true experiments as they are less tightly controlled. Variability of effects may be greater in studies where measures are not reliable versus studies that employ reliable measures, as unreliable measures increase the likelihood of error. Studies where measures exhibit ceiling/floor effects may exhibit less variability in effects than studies without such problems, as such measurement problems can restrict the amount of growth observed.

Inclusion/Exclusion Criteria

Each study had to meet nine criteria to be included in this meta-analysis: (1) involved students in Grades 1 to 12; (2) tested if writing enhanced students’ learning in science, social studies, or mathematics; (3) applied a true- or quasi-experimental design to test the effects of writing on learning (quasi-experiment had to include a comparable pretest assessment for each outcome of interest); (4) included at least one measure assessing student learning (letter grades were not considered a viable assessment); (5) incorporated the writing-to-learn activity as part of classroom instruction; (6) controlled for the amount of time students in the treatment and control conditions spent learning content material; (7) focused on teaching the same content knowledge to both treatment and control students; (8) reported the statistics necessary to compute a weighted ES (or data were obtained from the authors); and (9) were written in English.

Studies were excluded for four reasons: (1) the control condition used writing to support learning in science, social studies, and mathematics (one exception was made: Treatment students did considerably more writing to support learning than controls); (2) the intervention was conducted in special schools for students with disabilities; (3) the attrition rate at the end of the experiment was 20% or greater (attrition greater than 20% can bias study findings; Dumville et al., 2006); and (4) pretest scores had a high ceiling effect (i.e., a mean sore a full SD below the highest possible score). We did not include studies with kindergarten students, as they do little writing at school, and they are still learning the most rudimentary writing skills (Puranik et al., 2014).

The goal of this meta-analysis was to isolate the effects of writing on learning in science, social studies, and mathematics and not to compare different writing-to-learn activities. The inclusion and exclusion criteria were designed to achieve this objective, as included experiments involved students in the treatment condition using writing-to-learn activities, with the students in the control condition learning similar content and doing either no writing or engaging in considerably less writing-to-learn activities than treatment students. These criteria addressed possible confounds due to a lack of a control condition, pretest differences in quasi-experiments, differences in the content and the amount of time spent learning by treatment and control students, ceiling effects on learning measures at pretest, and high attrition rates.

Search Strategies

Selection of Studies for Review

After eliminating duplicate listings from the five search strategies, 31,456 separate listings were identified. Ninety-eight of the 206 publications identified by examining prior reviews, searching relevant journals, and contacting the study authors were duplicate listings.

A three-step process was applied to determine which studies to include in this meta-analysis. One, the second author with the assistance of the last author and two graduate research assistants read the title and abstract of the 31,348 listings identified through the electronic searches and the 108 separate listings located through prior reviews, journal searches, and study author contacts. This involved looking for evidence in the title and abstract that the paper presented an empirical study involving students in Grades 1 to 12 (Criterion 1) who engaged in writing activities to learn about science, social studies, or mathematics (Criterion 2). Five hundred and four publications met these criteria. Interrater agreement for this phase of the selection process was 99%, with disagreements on 315 listings. The first author, who had 40 years of experience conducting literacy research, resolved these differences after discussion with the second author. When disagreements arose, the full paper was obtained and examined to help the two authors rectify these disagreements. Sixty percent of these disagreements involved whether content learning was assessed, whereas 30% of the disagreements centered on the degree to which writing-to-learn was part of both treatment and control conditions.

The second phase of the selection process examined the full text of the remaining 504 publications (84% of these texts were from the electronic searches). The distribution among these studies was 43% in science, 21% in social studies, and 36% in mathematics. The second and last author read each document to determine if inclusion criteria were met and all exclusion criteria were not met (agreement was 95%). Two inclusion criteria were not included in this analysis (6 and 7). They were completed in Phase 3. As with Phase 1, all disagreements were resolved via discussion between first and second authors. Our analysis at this point yielded 56 publications that included 59 experiments (three publications yielded two experiments each; Boscolo & Mason, 2001; Leopold & Leutner, 2012; Willey, 1988). The reasons for excluding the 448 publications were the following: (1) 57% of studies did not meet the criterion for design (k = 252, Criterion 3), 34% of studies did not include a learning measure (k = 151; Criterion 4), and (3) the remaining 10% of the studies (k = 45) did not involve instruction (Criterion 5), include a writing-to-learn activity (Criterion 1), or need statistics (Criterion 8).

The third phase of the selection process involved the first and second authors independently examining the 59 identified experiments (in 56 documents) to determine if any study violated inclusion Criterion 6 (instructional time was equivalent for treatment and control conditions) or Criterion 7 (instruction for treatment and control focused on the same content). There was one disagreement that was resolved through discussion. One study was eliminated as students in the treatment condition received more instruction than the control condition before the pretest (Criterion 6). Two other studies were eliminated as there was uncertainty concerning whether treatment and control students were taught the same content material (Criterion 7). A total of 56 studies in 53 documents meet all criteria.

It is important to make clear the relationship between this meta-analysis and the two prior ones examining writing-to-learn. From the Bangert-Drowns et al. (2004) meta-analysis, we identified 20 investigations from the 48 included in their review that met all of our inclusion/exclusion criteria. The Graham and Perin (2007) meta-analysis (Grades 4–12) yielded two additional experiments that were suitable for our meta-analysis (i.e., Boscolo & Mason, 2001; Hand et al., 2004).

Coding

Analyses

Statistical Analysis of Effect Sizes

Characteristics of Studies

Content Areas

Characteristics and findings for the experiments in this meta-analysis are presented in Table 1. Twenty-six experiments (46% of all studies) tested the effects of writing on science; 21 experiments (38% of studies) assessed the impact of writing on mathematics; eight experiments (14% of studies) examined the influence of writing on social studies; and one experiment (2% of studies) investigated the effects of writing in both science and social studies.

OPEN IN VIEWER

Table 1

Characteristics and overall effect sizes for writing-to-learn treatments

Study	Grade	Content area	N	Student type	Tx (Wr) days	Type of writing	Metacog prompting	Writing comparison	Overall quality	Effect size
Wells (1986)	1	M	77	FR	15 (15)	J/A	Yes	Some vs. No	5	−–0.15
V. M. Johnson (1998)	2	M	40	AV	12 (12)	I/A	No	Some vs. No	6	0.09
Hyser (1992)	3	SS	57	FR	120 (120)	J/R	No	Some vs. No	3	0.25
Jitendra et al. (2013) a	3	M	109	LA	60 (60)	G/R	Yes	Some vs. No	8	0.35
Brodney (1993)	3 – 5	M	542	FR	80 (80)	J/A	Yes	Some vs. No	5	0.06
Davis (1990)	4	SS	77	FR	45 (18)	J/R	No	Some vs. No	6	0.35
DiSimoni (2002)	4	SC	24	FR	40 (40)	J/A	Yes	Some vs. No	5	1.06
Gillespie Rouse et al. (2017) a	4	SC	46	FR	1 (1)	AR/A	Yes	More vs. Less	7	0.21
Mason & Boscolo (2000) a	4	SC	36	FR	50 (50)	I/A	Yes	Some vs. No	4	0.48
Jang (2010) a	4	SC	114	FR	24 (12)	MIX/A	Yes	Some vs. No	5	0.50
Millican (1994)	4	M	243	FR	50 (25)	J/A	Yes	Some vs. No	5	0.54
Lodholtz (1980)	4 - 6	M	456	FR	14 (14)	N/A	No	Some vs. No	7	0.18
Peterson (2003)	4 - 6	M	30	NS	25 (25)	J/A	Yes	Some vs. No	3	0.14
Chen et al. (2013) a	4	SC	838	FR	40 (3)	AR/A	Yes	More vs. Less	5	0.23
Bauman (1992)	5	M	182	FR	90 (54)	J/A	Yes	Some vs. No	5	0.23
Boscolo & Mason (2001) a	5	SS	32	FR	10 (10)	I/A	Yes	More vs. Less	4	1.06
Bosolo & Mason (2001) a	5	SC	32	FR	6 (6)	I/A	Yes	More vs. Less	3	0.88
DiPillo (1994)	5	M	54	FR	40 (24)	J/R	Yes	Some vs. No	4	0.04
Moynihan (1994)	5	M	121	FR	15 (15)	J/A	Yes	Some vs. No	4	0.51
Merchie & Van Keer (2016) a	5, 6	SC, SS	405	FR	10 (8)	G/R	Yes	Some vs. No	3	0.06
Jurdak & Zein (1998) a	5 - 7	M	104	FR	36 (36)	J/A	Yes	Some vs. No	3	0.26
Greer (2010)	6	M	192	FR	50 (50)	I/CD	Yes	Some vs. No	4	0.56
Keown (2008)	6	SC	34	FR	11 (11)	G/R	No	Some vs. No	4	0.42
Konopak et al. (1990) a	6	SS	48	FR	4 (3)	I/A	No	More vs. Less	3	0.59
Parson (2013)	6	SC	47	FR	48 (24)	I/R	No	Some vs. No	5	0.24
Pillsbury (2008)	6	SC	38	FR	CD (CD)	N/A	No	More vs. Less	3	−0.1
Vidal-Abarca & Gilabert (1995) a	6	SC	94	FR	25 (19)	G/R	No	More vs. Less	5	0.28
Keselman et al. (2007) a	7	SC	46	FR	12 (4)	AR/A	No	Some vs. No	1	0.15
Rosa (1999)	7	M	30	EL	30 (30)	N/A	No	Some vs. No	4	0.59
Waschle et al. (2015) a	7	SC	46	NS	6 (3)	J/A	Yes	Some vs. No	5	1.05
Bigelow (1992)	7, 8	SC	50	AV	2 (1)	G/R	No	Some vs. No	4	0.84
Akkus et al. (2007) a	7 – 10	SC	99	FR	CD (CD)	AR/A	Yes	More vs. Less	6	0.17
Ayers (1993)	8	SC	35	AV	40 (20)	MIX/A	No	Some vs. No	2	−0.74
Kramarski & Mevarech (2003) a	8	M	192	FR	10 (10)	I/A	Yes	Some vs. No	7	−0.41
Morphy (2013)	8	SS	37	LA, AV	8 (5)	I/A	No	More vs. Less	8	0.50
Rivard (1996)	8	SC	22	FR	30 (5)	I/A	No	Some vs. No	4	−0.26
Schmitz & Perels (2011) a	8	M	195	FR	49 (49)	J/A	Yes	Some vs. No	3	0.32
Willey (1988.2)	8	SC	46	FR	90 (77)	J/A	Yes	More vs. Less	3	1.43
Bell & Bell (1985) a	9	M	38	FR	20 (20)	I/A	Yes	Some vs. No	3	0.24
Cross (2008) a	9	M	112	FR	50 (50)	AR/A	Yes	Some vs. No	6	0.48
Faber et al. (2000) a	9	SS	58	LA, AV	40 (12)	G/R	No	Some vs. No	5	0.14
Feather (1998)	9	SC	52	FR	75 (75)	J/A	No	Some vs. No	5	−0.09
Kingir et al. (2012) a	9	SC	122	FR	50 (CD)	AR/A	No	More vs. Less	7	0.74
Klugh (2008)	9	SS	40	FR	30 (20)	G/R	No	More vs. Less	4	−0.12
Reaves (1991)	9	SC	177	FR	60 (60)	I/R	No	Some vs. No	1	−0.09
L. Johnson (1991)	9, 10	M	68	FR	91 (57)	I/A	No	Some vs. No	7	0.51
Stewart (1992)	9, 10	M	62	FR	84 (48)	J/R	Yes	Some vs. No	3	0.58
H& et al. (2004) a	10	SC	73	FR	30 (8)	I/A	Yes	More vs. Less	5	0.63
Idris (2009) a	10	M	169	FR	25 (25)	I/A	Yes	Some vs. No	3	1.26
Leopold & Leutner (2012; Experiment 1) a	10	SC	45	FR	1 (1)	I/R	No	Some vs. No	7	−0.48
Leopold & Leutner (2012; Experiment 2) a	10	SC	36	FR	1 (1)	G/R	No	Some vs. No	7	−0.68
Tsai (1995)	10	SC	44	AV	3 (3)	I/R	No	Some vs. No	5	0.14
Willey (1988.1)	10	SS	51	FR	90 (77)	J/A	Yes	More vs. Less	3	0.04
Kasparek (1993)	10, 11	M	68	AV	60 (60)	I/A	Yes	Some vs. No	4	0.37
Caukin (2010)	11	SC	31	HA	25 (CD)	AR/A	No	More vs. Less	2	2.01
O’Sullivan (2013)	11	SC	141	FR	25 (CD)	I/A	No	Some vs. No	3	0.18

Note. N = number of students in study; Tx = treatment; Wr = writing; metacog = metacognitive; SS = social studies; SC = science; M = mathematics; FR = full range; SD = students with disabilities; EL = English learner; LA = low achieving; AV = average students; NS = not specified; CD = cannot determine; J = journal writing; A = analysis and interpretation; I = informative writing; R = mostly recording information; G = graphical representations; AR = argumentative writing; MIX = multiple types of writing; N = narrative.

a

Journal article.

Quality of Studies

Table 1 includes a quality score for each experiment. This score represents the sum of all criteria met, except for treatment fidelity. We did not include treatment fidelity in this sum, as it occurred in just three experiments (5%). In 31 of the experiments (55%), a true-experimental design was used to assess the effects of writing on learning. Attrition greater than 10% was not an issue in 47 experiments (84%), and differential attrition did not occur in 42 experiments (75%). In 45 experiments (80%) there was no meaningful difference at pretest between treatment and control condition. Ceiling and floor problems were not evident in 30 experiments (54%). Treatment and control conditions each included more than two instructional groups in 30 experiments (54%). Reliability of measures was established in 29 experiments (52%). Teacher effects were controlled in 29 experiments (52%). Also, 43% of studies were journal articles that were peer reviewed, and the remaining studies were dissertations/thesis.

The Impact of Writing on Learning

Writing about content improved students’ learning, as we obtained a statistically significant average weighted ES of 0.30 (see Table 2). While a positive ES was obtained in 82% of the investigations (see Table 1), there was considerable variability in effects. The I² statistic showed that 71% of variance resulted from between-study factors. Publication bias did not appear to be an issue in this omnibus analysis. The trim and fill method imputed no ESs to the left or the right of the funnel plot. Furthermore, the correlation between study effects and variance in effects was not statistically significant (p for the Egger’s test = .15).

OPEN IN VIEWER

Table 2

Overall analysis of weighted effect sizes and confidence levels

Analysis	K	g	95% CI	p	I 2	τ2
Overall	56	0.30	[0.20, 0.41]	<.001	71.16	0.10
Content area
Science	26	0.31	[0.16, 0.48]	.002	74.67	0.17
Social studies	8	0.31	[0.10, 0.52]	.003	12.93	0.01
Mathematics	21	0.32	[0.16, 0.48]	<.001	74.79	0.09
Grade level
Elementary	19	0.29	[0.18, 0.40]	<.001	33.21	0.02
Middle school	18	0.30	[0.07, 0.53]	.011	78.14	0.20
High school	18	0.30	[0.06, 0.54]	.013	79.12	0.18
Features of writing activities
Informational text	18	0.34	[0.09, 0.60]	.001	80.64	0.23
Argumentative text	7	0.42	[0.17, 0.67]	.001	60.11	0.06
Journal writing	18	0.33	[0.17, 0.49]	<.001	58.51	0.06
Graphical representation	8	0.19	[−0.03, 0.41]	.094	—	—
Analysis and interpretation	38	0.36	[0.22, 0.50]	<.001	76.72	0.13
Recording of information	17	0.18	[0.03, 0.33]	.017	47.39	0.04
Metacognitive prompting	30	0.40	[0.26, 0.55]	<.001	75.13	0.10
No metacognitive prompting	25	0.15	[−0.01, 0.31]	.063	—	—
Promote knowledge/comprehension	18	0.16	[0.02, 0.31]	.030	43.16	0.04
Promote analysis/synthesis	15	0.32	[0.04, 0.59]	.023	82.40	0.22
Promote evaluation	17	0.44	[0.21, 0.68]	<.001	74.09	0.16
Features of instruction
Writing taught to students	31	0.29	[0.15, 0.43]	<.001	69.42	0.09
Writing not taught to students	25	0.32	[0.15, 0.49]	<.001	73.89	0.13
PD provided to teachers	21	0.30	[0.15, 0.45]	<.001	77.03	0.09
PD not provided to teachers	35	0.31	[0.16, 0.49]	<.001	67.80	0.14
Teacher Implementation	44	0.28	[0.16, 0.40]	<.001	73.56	0.11
Researcher Implementation	12	0.40	[0.15, 0.65]	<.001	60.56	0.11
Features of assessment
Assessed knowledge	7	0.24	[−.02, 0.50]	.070	—	—
Assessed comprehension	24	0.30	[0.14, 0.47]	<.001	66.65	0.10
Assessed application	17	0.45	[0.29, 0.61]	<.001	65.39	0.06
Norm-referenced test	9	0.31	[0.17, 0.45]	<.001	35.91	0.02
Criterion-referenced test	8	0.39	[0.07, 0.70]	.002	69.55	0.14
Open-ended response test	13	0.38	[0.16, 0.59]	.001	50.72	0.07
Multiple-choice test	15	0.10	[−0.16, 0.35]	.461	—	—
Mixed-assessment test	10	0.18	[−0.07, 0.44]	.162	—	—
Researcher-designed assessment	39	0.30	[0.15, 0.46]	<.001	78.03	0.17
Commercial assessment	12	0.24	[0.02, 0.47]	.030	55.75	0.08
Standardized assessment	11	0.31	[0.19, 0.43]	<.001	30.47	0.01
Proximal alignment	41	0.29	[0.16, 0.42]	<.001	74.74	0.15
Distal alignment	11	0.31	[0.19, 0.43]	<.001	30.47	0.01
Knowledge match assessment/writing activity	17	0.32	[0.21, 0.44]	<.001	00.00	0.00
Knowledge match assessment lower than writing activity	32	0.34	[0.18, 0.49]	<.001	79.17	0.14
Treatment vs. control comparison
Writing vs. no writing	41	0.25	[0.12, 0.37]	<.001	72.57	0.10
More writing vs. less writing	15	0.49	[0.27, 0.71]	<.001	63.10	0.10
Quality of study
Peer reviewed	24	0.34	[0.17, 0.52]	<.001	77.21	0.13
Not peer reviewed	32	0.27	[0.14, 0.41]	<.001	88.43	0.09
True-experimental design	31	0.21	[0.08, 0.33]	.001	60.49	0.06
Quasi-experimental design	25	0.42	[0.24, 0.59]	<.001	74.70	0.13
Test reliability adequate (>.70)	29	0.26	[0.11, 0.40]	.001	74.08	0.11
Test reliability inadequate/not reliable	27	0.36	[0.20, 0.52]	<.001	67.85	0.11
No ceiling/floor effects (<1 SD)	30	0.23	[0.08, 0.38]	.002	70.65	0.10
Ceiling/floor effects (>1 SD)	26	0.39	[0.23, 0.54]	<.001	70.78	0.11
Pretest equivalence	45	0.27	[0.16, 0.39]	<.001	66.23	0.09
No pretest equivalence	11	0.42	[0.15, 0.70]	.002	82.90	0.16
Not N-of-1 study	30	0.41	[0.22, 0.60]	<.001	75.43	0.21
N-of-1 study	26	0.22	[0.10, 0.34]	<.001	61.73	0.05
Teacher effects controlled	29	0.25	[0.13, 0.36]	<.001	61.50	0.05
Teacher effects not controlled	27	0.40	[0.19, 0.60]	<.001	77.07	0.22
Attrition (<10%)	47	0.36	[0.24, 0.48]	<.001	69.83	0.11
Attrition (>10%)	9	0.05	[−0.14, 0.25]	.602	—	—
No differential attrition	42	0.38	[0.25, 0.51]	<.001	70.46	0.01
Differential attrition	14	0.11	[−0.05, 0.27	.180	—	—

Note. K = number of effect sizes; g = Hedges’s g; PD = professional development.

Content Area Differences

Writing about content improved students’ learning in science, social studies, and mathematics. Average weighted ESs for these three areas were 0.31, 0.31, and 0.32, respectively. Each of these effects was reliably greater than no effect, and there was considerable variability in ESs in each subject area, except social studies (see Table 2). A metaregression, however, did not find that discipline moderated writing-to-learn effects, as this predictor did not account for a statistically significant proportion of variance, Q = .04, df = 2, p = .98. Merchie and Van Keer (2016) was not included in this analysis as it involved a science/social studies treatment.

Grade-Level Differences

We conducted a metaregression to determine if writing-to-learn effects were moderated by grade level (i.e., elementary, middle school, and high school). Writing about content did improve students’ learning at all three grade levels. Average weighted ESs for elementary, middle school, and high school students were 0.29, 0.30, and 0.30, respectively. Effects at each of these grade levels were statistically significant, and there was considerable variability in ESs at the middle and high school levels but not at the elementary grades (see Table 2). The overall effects of writing-to-learn were not moderated by grade level. This predictor did not account for a statistically significant proportion of variance in the metaregression, Q = .08, df = 2, p = .96. Akkus et al. (2007) was not included in this analysis as it did not contain separate data for middle and high school students.

Differences Related to Features of the Writing Activities

Three metaregressions were conducted to determine if features of writing-to-learn treatments were related to variability of effects. Two metaregressions included one predictor. This included an analysis examining type of writing (informational, journal, argument, or graphical writing), and another examining the highest level of the Bloom et al. (1956) taxonomy that promoted (knowledge/comprehension, analysis/synthesis, and evaluation). The other metaregression included two predictors: (1) writing promoted analysis and interpretation versus recording information and (2) metacognitive prompt included or not included. We were unable to apply all four predictors in a single analysis or even two analyses as we would have had to drop too many studies. For type of writing, there were not enough studies examining narrative writing (k = 3) or a mixture of writing activities (k = 2) to include them in the analyses. For highest level of the Bloom et al. taxonomy, we combined several levels to increase our power (knowledge [k = 3] and comprehension [k = 15]; analysis and synthesis [k = 5 and k = 10, respectively]), and there were not enough studies with writing activities categorized as application (k = 5) to include them in an analysis. We could include only Greer (2010) in the analysis on type of writing, as the writing activities were not well enough described to include it in the other two metaregressions.

The average weighted ESs for types of writing were argumentative = 0.42, informational = 0.34, journal = 0.33, and graphical = 0.19. Each of these effects were statistically significant (and evidenced high to moderate heterogeneity of ESs), except for graphical writing, which did not yield an effect greater than zero (see Table 2). However, the overall effects of writing-to-learn were not moderated by type of writing. This predictor did not account for a statistically significant proportion of variance, Q = 1.42, df = 3, p = .70.

The average weighted ES for writing-to-learn treatments that involved analysis and interpretation of content material was 0.36, and 0.18 for writing that involved recording information. The ES for analysis and interpretation as well as recording information was reliably greater than no effect (see Table 2). When the writing-to-learn treatment involved metacognitive prompting, the average weighted ES was 0.40 and 0.15 when it did not include such prompting. The ES for metacognitive prompting was reliably greater than no effect; this was not the case when such prompting was not included. These two predictors evidenced high to moderate variability. They did not moderate the effects of writing-to-learn, Q = 5.58, df = 2, p = .06.

The average weighted ES for writing activities that promoted knowledge/comprehension, analysis/synthesis, and evaluation was 0.16, 0.32, and 0.44, respectively. The ES for each level of this predictor was reliably greater than no effect (see Table 2), but there was high to moderate variability for each variable. The overall effect of writing-to-learn was not moderated by the highest level of knowledge prompted by the writing activity/activities, Q = 3.31, df = 2, p = .19.

Differences Related to Features of Instruction

We conducted two metaregressions to determine if features of instruction moderated writing-to-learn effects. For the first metaregression, we examined the collective impact number of treatment days (M = 36.26) and number of days treatment students wrote (M = 28.55). We were not able to include five studies in this analysis as it was not possible to determine number of days for treatment, writing, or both (see Table 1). Effects of writing-to-learn were not moderated by number of treatment and writing days, Q = 0.64, df = 2, p = .73.

The second metaregression examined if writing-to-learn effects were moderated by three predictors: treatment students taught or not taught how to apply writing activities, treatment teachers did or did not receive professional development, and teachers or researchers delivered the treatment. The average weighted ESs for writing activities taught (ES = 0.29) and not taught (ES = 0.32), professional development provided (ES = 0.30) or not provided (ES = 0.31), and teachers (ES = 0.28) or researchers (ES = 0.44) delivered instruction were all statistically significant (see Table 2). Considerable variability was evident for all predictors. The effects of writing-to-learn were not moderated by these three predictors, Q = 0.17, df = 3, p = .98.

Differences Related to Features of the Assessments

Five metaregressions were conducted to determine if features of the assessments used to gauge the effects of learning in the studies reviewed were related to variability of effects. Each metaregression included one predictor. It was necessary to conduct a separate analyses for the first metaregression as we had to eliminate eight studies—type of knowledge assessed in Bell and Bell (1985) could not be determined, and studies involving assessing synthesis (k = 4) and evaluation (k = 3) of knowledge were too few to include in the analysis. For the other three metaregressions, collinearity problems occurred when more than one predictor was included. This mostly happened because of the overlap in categories across predictors (e.g., norm-referenced, standardized, and less proximally aligned assessments).

For the first metaregression, average weighted ESs for assessing knowledge, comprehension, and application were 0.24, 0.30, and 0.45, respectively. The ESs for assessing comprehension and application of knowledge were reliably greater than no effect (see Table 2), but the ES for assessing knowledge was not. There was considerable variability in ESs for comprehension and application. The overall effect of writing-to-learn was not moderated by the highest level of knowledge assessed, Q = 3.31, df = 2, p = .19.

For the second metaregression, the average weighted ESs for types of assessment were the following: criterion-related measure = 0.39; open-ended response format = 0.38; norm-referenced = 0.31; multiple-choice = 0.10, and mixed-assessment format = 0.18. Each of these effects were statistically significant, except for multiple-choice and mixed-assessment formats (see Table 2). Norm-referenced measures were the only type of assessment that did not demonstrate a high or moderate degree of variability in ESs. The overall effects of writing-to-learn, however, were not moderated by type of assessment, Q = 4.70, df = 4, p = .32. We did not include four studies in this analysis as they did not clearly indicate the type of tests (Bell & Bell, 1985; Greer, 2010; Idris, 2009) or used an assessment other than the ones tested (Gillespie Rouse et al., 2017). We could not use all assessments in Parson (2013) and Pillsbury (2008) for the same reasons.

For the third metaregression, the average weighted ESs for writing-to-learn treatments that involved researcher-designed tests, tests from textbook, and standardized assessments designed by test developers were 0.30, 0.24, and 0.31, respectively. Each of these effects was reliably greater than zero (see Table 2). High to moderate variability of ESs was evident for researcher-designed tests and standardized tests. The overall effects of writing-to-learn were not moderated by who designed the assessments, Q = 0.48, df = 2, p = .78. Jitendra et al. (2013) included researcher-designed and standardized assessment from test developers.

In the fourth metaregression, the average weighted ES of 0.29 for assessments that were more proximally aligned to content taught was statistically significant, as was the effect of 0.31 for assessments that were less proximally aligned to taught content (see Table 2). There was considerably variability in ESs only for more proximally aligned measures. The effects of writing-to-learn were not moderated by test alignment, Q = 0.09, df = 1, p = .77.

For the fifth metaregression, the average weighted ES of 0.32 for assessments and writing activities that were matched in terms of knowledge level measured and promoted was statistically significant. This was also the case for the average weighted ES of 0.34 for studies where assessment measured a lower level of knowledge than the highest level of knowledge the writing activity was designed to promote. There was considerably variability in ESs only for studies where a lower level of knowledge was assessed (see Table 2). The effects of writing-to-learn were not moderated by the match between the highest level of knowledge assessed and highest level of knowledge promoted by the writing activity, Q = 0.14, df = 1, p = .71. We did not include studies in this analysis where the assessment measured a higher level of knowledge than the writing activity promoted, as there were only five of these investigations.

Differences Related to Treatment/Control Comparisons

Writing about content was effective when writing-to-learn treatments were compared to control conditions where students did not write or wrote significantly less than treatment students. The average weighted ES of 0.25 for the writing versus no writing comparison was reliably greater than no effect as was the ES of 0.49 for the more writing versus less writing comparison (see Table 2). There was considerable variability in the ESs for both of these treatment/control comparisons. The metaregression did not find that writing-to-learn effects were moderated by type of treatment/control comparison, Q = 3.69, df = 1, p = .055.

Differences Related to Study Quality

We conducted two metaregressions to determine if study quality moderated writing-to-learn effects. For the first metaregression, we examined the collective impact of the following quality indicators: study design (true experiment vs. quasi-experiment), reliability of measures (reliable vs. not reliable), ceiling/floor concerns for measures (no concerns vs. concerns), pretest equivalence (equivalent vs. not equivalent), N-of-1 studies (multiple groups per condition vs. only one group for one condition or more), teacher effects (effects controlled vs. effects not controlled), and overall attrition (overall attrition was >10% vs. <10%). We were able to include differential attrition in this analysis due to collinearity problems with overall attrition. As can be seen in Table 2, the average weighted ES for each level of the seven predictors and differential attrition were positive and reliably greater than no effect, except when attrition was >10% (ES = .05) or differential attrition was a problem (ES = 0.11). There was considerable variability in ESs. Collectively, study quality as indicated by these seven predictors identified above did not moderate writing-to-learn effects, 11.38, df = 7, p = .12.

For the second metaregression, we examined peer reviews’ relation to magnitude of effects. Less than one half of the investigations (k = 24) included in this review underwent peer review. As a result, it was possible that the obtained writing-to-learn effects were moderated by differences in investigations that were and were not peer reviewed. The average weighted ES of 0.34 for peer-reviewed studies was statistically significant, as was the effect of 0.27 for studies that did not undergo peer review (see Table 2). Both types of studies evidenced considerably variability in ESs. The effects of writing-to-learn were not moderated by whether a study underwent peer review or not, Q = 0.38, df = 1, p = .54. Furthermore, we found that there was no statistical difference (F = 1.47, df = 1, 54, p = .23) in the overall quality score for studies that were peer reviewed (M = 4.79, SD = 1.81) or not peer reviewed (M = 4.25; SD = 1.52).

Writing Improves Learning in Science, Social Studies, and Mathematics

As predicted, writing about content in science, social studies, and mathematics enhanced learning. In 56 investigations across the three content areas, a statistically significant average weighted ES of 0.30 was obtained. The average weighted ES for science (0.30), social studies (0.33), and mathematics (0.32) were virtually identical. While additional writing-to-learn research is needed in each of these content areas (science contained the largest number of studies at 26), this is especially the case for social studies. Only eight studies isolated the effects of writing-to-learn in this content area. The lack of studies in social studies in relation to science and mathematics is in direct contrast to classroom realities, where social studies teachers are more likely to use writing as a tool for learning than teachers in the other two content areas (Gillespie et al., 2014; Ray et al., 2016). It is possible that this imbalance is a result of how social studies researchers conduct instructional research. As Hicks et al. (2012) indicated, there is very little focus in social studies research on testing the effectiveness of specific strategies. Whatever the reasons for the lack of writing-to-learn investigations in social studies, this is an area that must be a priority in the future.

It is also important to note that the overall ES for writing-to-learn in this meta-analysis exceeded the average weighted ES of 0.17 obtained by Bangert-Drowns et al. (2004) with school-aged and college students and the average weighted ES of 0.23 reported by Graham and Perin (2007) with students in Grades 4 to 12. The effects for science were relatively equivalent in this meta-analysis and the one by Bangert-Drowns et al. (0.30 vs. 0.32). However, our analysis obtained larger effects for social studies (0.33 vs. 0.13) and mathematics (0.32 vs. 0.24). Effects for the three content areas were not provided by Graham and Perin.

The current meta-analysis provided a more reliable and valid estimate of the effects of writing-to-learn in these three content areas than previous meta-analyses, as it included more studies and used more stringent controls than the two previous reviews. In contrast to the prior meta-analyses, we directly established that equal instructional time and content exposure were provided to treatment and control students. Quasi-experiments were included only if the authors reported a pretest learning measure and grades were not used as a measure of learning. We also operationalized and coded studies for reliability. In addition, we applied sophisticated data analysis procedures, including winzoring outliers, adjusting quasi-experiments for clustering effects, correcting for small sample size bias, and examining possible publication bias.

This is not to say, however, that writing about science, social studies, or mathematics always enhanced learning in the studies reviewed here. A negative effect was obtained in 18% of our studies. The variables coded in this review offered limited insight though into why these studies were not successful. A post hoc comparison of studies with negative and positive effects revealed that these investigations were similar on most of the coded variables. The only variables that should have theoretically enhanced writing-to-learn effects and were 15% or more less common in studies with negative outcomes were professional development (21% less common), number of days writing activities were applied (21% less common), metacognitive prompting (17% less common), and writing activities that promoted higher levels of learning (activities that promoted evaluation were 25% less common). These differences were not pervasive enough to make any of these variables a singular negative catalyst across studies. It is possible that one or more of these variables did exert a strong negative influence in a specific investigation, or they may have operated in tandem to produce negative effects across multiple studies.

In order to better understand why writing-to-learn studies do or do not produce positive learning effects, future research needs to (1) better describe studies, (2) expand the variables examined, and (3) systematically examine the relationship between writing-to-learn outcomes and variables that potentially moderate these effects. We provide specific examples below.

Two possible reasons for why some studies may produce negative results include that the target writing activities were poorly constructed or applied. In the current review, it was not possible for us to evaluate either of these variables, as researchers did not consistently provide enough information to determine all of the mental operations writing activities were designed to provoke (e.g., rehearsal, organization, elaboration, transformation, goal setting) or how students actually used them while learning. Such rich descriptions of these and other variables (e.g., instructors, participants, context, content to be learned, and assessment procedures) are needed to better understand why a writing-to-learn study is or is not successful.

We were also not able to examine all of the variables we considered important to the success of writing as a learning activity, as they simply were not studied by researchers. For instance, it is reasonable to expect that students’ capability as writers might influence the success of a specific writing activity, particularly when the writing task is complex, even when learners are capable of applying a writing activity, though, they may fail to do so if they do not value writing or learning (Holliday et al., 1994). Measures of such attributes were virtually nonexistent in the 56 studies reviewed here. We encourage researchers to expand the variables in their studies to include additional ones that are theoretically important to the success of writing-to-learn.

Researchers should also take a more active approach to examining if promising variables predict the success of writing-to-learn activities. For example, girls are generally better writers than boys, and students from more affluent families tend to write better than ones from poorer families (Graham, 2018). Researchers can determine if these variables are related to learning outcomes by making gender and socioeconomic status independent variables in studies. Other potentially promising variables such as teachers’ motivations or their views about the acceptability of writing activities can be examined by correlating teachers’ scores on such measures with how much students’ learn. Such analyses were rarely conducted in the studies reviewed here.

The Factors That Create Variability in Writing-to-Learn Effects Are Uncertain

While writing generally enhanced learning in this meta-analysis, there was considerable variability in effects (ESs ranged from 1.67 to −0.74). Over 70% of the variability in effects across studies was due to factors other than chance. Because we anticipated that writing-to-learn effects would be variable, we constructed a number of theoretically derived moderator analyses. Despite our prediction that the moderators tested would account for excess variability in observed outcomes, this was not the case. There were no statistically significant difference by content area (science, social studies, and mathematics), grade (elementary, middle, and high school school), features of writing activities (i.e., types of writing, promotion of analysis and interpretation, inclusion of metacognitive prompting, and highest level of learning promoted), features of instruction (professional development, teaching writing-to-learn activities, instructor, number of days of instruction, and number of days writing activities were applied), features of assessments (highest level of knowledge assessed, type of test, who designed the assessment, alignment of assessment with content, and the match between the level of knowledge assessed and level of knowledge the writing activity was designed to promote), or study quality.

Findings from the moderator analyses reported above were not completely consistent with Bangert-Drowns et al. (2004). They found that grade level was related to variability of effects, with middle school studies producing smaller effects than elementary, high school, and college students. Differences in outcomes in the two meta-analyses are likely due to the addition of new middle school studies in this review that produced more positive effects. They further found that writing activities that prompted metacognition were more effective than ones that did not. It is not readily clear why similar effects were not obtained in the current review. This was probably not due to the inclusion of college students in Bangert-Drowns et al., as the average effects for college (0.47; k = 5) and school-aged students (0.40; k = 12) were similar.

We also thought that it was possible that specific design features of the experiments in the 56 studies reviewed would account for excess variability in effects. This was also not the case, as there was no statistically significant differences between measures of study quality (peer review, type of experimental design, reliability of measures, ceiling/floor effects for measures, pretest equivalence, teacher effects controlled, level of attrition, and more than one instructional group in each condition) or the type of treatment/control comparison (writing vs. no writing; more writing vs. less writing).

While our findings from the moderator analyses are consistent with the proposition that the effects of writing-to-learn are generally constant across different disciplines, grade level of students, writing activities, instructional procedures, types of outcomes, and design features of studies, the effects of writing on learning are more situational than these findings imply for three reasons. One, as noted earlier, not all studies produced a positive effect and the reasons for why this was the case were not clear. Two, our moderator analyses involved 56 studies. It is possible that we did not have enough statistical power to detect differences that did exist. For instance, the variable that produced the largest effect, argumentative writing (0.42), and the one that resulted in the smallest effect, attrition greater than 10% (0.05), were based on just seven and nine studies, respectively. The impact of argumentative writing was more than double at least one of the other types of writing, graphical representation, included in its moderator analysis, whereas attrition greater than 10% was 7 times smaller than its contrast, attrition less than 10%. As additional studies accumulate over time, it is possible that future meta-analyses with more statistical power will be more successful in identifying factors that moderate writing-to-learn effects.

As more studies become available, it should also be possible to examine the interactions between potential moderators. For example, argumentative writing may be differentially effective depending on students’ grade. There were not enough studies in this review to examine such interactions. Instead we had to concentrate on main effects, analyzing grade and type of writing separately. Consequently, it is possible that the moderator variables examined in this review may account for variability in writing-to-learn effects but only in terms of their relationship with other variables. This proposition should be tested in future meta-analyses.

Three, and most important, the substantial variability in the effects obtained in this meta-analysis provided the strongest evidence that writing-to-learn effects are not constant across all situations. If we are to identify the sources of this variability, we need to isolate and test additional theoretically derived moderators. As indicated in the previous section, this requires that not only must researchers provide richer descriptions of their studies so that additional moderators can be coded in future meta-analyses, but they must also purposefully expand the moderators included in their investigations. We illustrate a broad array of features that may moderate writing-to-learn effects by drawing on a model proposed by Graham (2018). This is not the only model of writing that could be used for this purpose, but it is the only model available that conjointly describes the social and cognitive aspects of writing.

In the current meta-analysis, it was especially difficult to identify contextual variables that might moderate writing-to-learn effects. As Klein (1999) noted, context is poorly described in such studies. Graham’s (2018) model provided a useful organizing structure for more fully describing context and identifying new moderates, as it specified seven contextual features that shape and constrain how writing, including writing-to-learn, operate in a writing community: (1) purposes for writing (goals, values, norms, audiences, and motivations for writing as well as the stances and identities writing fosters), (2) community members (teachers, mentors, writers, collaborators, and readers as well as their roles, responsibilities, participation, familiarity with community practices, and the value and power within the community), (3) writing tools (tools for writing including paper and pencil and digital tools), (4) writing actions (sanctioned and typical practices used to accomplish writing purposes), (5) written products (text completed and in production, plans, source and content material, and modeled text), (6) physical and social environment (physical and digital locales, relations among community members, and social practices for writing), and (7) the collective history of the writing community (past and current events that shape how writing is enacted in the community, which are shaped by broader social-cultural, political, institutional, and historical forces).

While it is highly unlikely that researchers will describe all of the contextual features described above, this model provides a blueprint for better contextualizing writing-to-learn studies. It also draws attention to multiple features of context that may moderate writing-to-learn effects. We encourage researchers to not only describe such features in future studies (especially purposes for writing, content to be learned, and common practices used to promote learning and writing) but also increasingly test the effects of these moderators in their investigations.

As noted in the previous section, studies in the current meta-analysis provided little to no information about what students actually did when applying writing-to-learn activities, nor did they devote much attention to identifying cognitive factors that affected these mental processes. Graham’s (2018) model described the cognitive activities students use as they write, including writing-to-learn (Graham, 2019). According to this model, students engage in a variety of activities while writing-to-learn, including making decisions about what to do and learn as well as how much effort to commit. These decisions are driven by students’ beliefs (e.g., value and utility of the writing activity and content to be learned, perceptions of competence as a writer and learner, and identity as a writer and learner), which in turn fuels effort. This provides the impetus for drawing on resources, including content knowledge, from long-term memory as well as content presented in class, as writers use the control processes of attention, working memory, and executive control to initiate, direct, and sustain writing processes (conceptualization, ideation, translation, transcription, and reconceptualization) that engage students in thinking that facilitates learning (e.g., rehearing, elaborating, evaluating, culling, connecting, and organizing ideas). These actions are further affected by psychological, physical, and biological factors including emotions, personality traits, and physical states.

As with context, Graham’s (2018) model provided a useful description of cognitive factors that are likely to moderate writing-to-learn effects, including students’ goals, beliefs about writing and learning, current knowledge of content, functioning of control mechanisms (including students’ success in using the thinking processes evoked by writing activities), writing capabilities, emotional reactions to writing and learning, personality traits, and physical state (e.g., stress). Again, we do not expect that researchers will address the moderating effects of all of these factors in every future study, but some of them should become routine fixtures in these investigations (e.g., prior knowledge, writing ability, and description of the cognitive activities writing was designed to evoke), and all of them should increasingly become more common in future research. These same variables (goals, knowledge and beliefs, cognitive processing, writing capabilities, emotions, personality traits, and physical states) also influence teachers’ actions (Graham & Harris, 2018) and provide additional candidates for moderating variables.

Implications for Theory

Cognitive theories of writing-to-learn contend that writing elicits both implicit and explicit use of cognitive operations and structures that facilitate learning (e.g., Galbraith & Baaijen, 2018; Klein, 1999; Klein & Boscolo, 2016). Evidence from the current meta-analysis provides support for the proposition that explicitly controlled cognitive processes promote learning. All studies reviewed used writing activities that prompted students to use cognitive operations to facilitate learning. To illustrate, 77% of studies encouraged students to apply writing activities involving analysis and interpretation, metacognitive prompting, or both. In the remaining studies, students used writing to identify important information, to organize the information, or both. We cannot draw any conclusions about the specific cognitive mechanisms that facilitated learning or the impact of genre on writing-to-learn, though, as none of our moderator analyses were statistically significant and writing activities and their use was not described consistently in sufficient detail to permit a finer grained analysis of specific mental operations (e.g., rehearsal, elaboration, comparisons, organization). Nor can we draw specific conclusions about implicit processing of information, as the studies reviewed did not purposefully assess such learning.

The current review also provides support for a social-cultural view of writing-to-learn, demonstrating that writing is a tool students can use in a goal-directed fashion to successfully construct meaning in content classes. Our findings also support the social-cultural tenet that writing effects are not automatic (Smagorinsky, 1995) but vary. Positive effects were not obtained in 18% of studies reviewed here, and there was considerable variability in effects across studies. Evidence that writing-to-learn effects are more likely in classrooms where writing is valued and sanctioned was not substantiated, as we did not find that larger effects were obtained in studies where writing was more commonly used to promote learning.

Implications for Practice

This meta-analysis provides current evidence that science, social studies, and mathematics teachers can reasonably expect that asking their students to write about content material will enhance learning. Writing-to-learn improved students’ comprehension and application of content knowledge, including the learning of material that was not the immediate focus of instruction (as assessed by more distal measures of learning).

Based on the findings from this review and how writing-to-learn treatments were instantiated, we provide the following observations for practice. One, writing is a useful learning tool across grades, and should be applied accordingly. Two, students can apply writing in multiple ways to improve content learning. This includes using writing to promote knowledge/comprehension, analysis/synthesis, and evaluation of content through informational, argumentative, and journal writing that involves metacognitive reflection but may or may not involve analysis and interpretation. Three, writing-to-learn activities can be profitably integrated into science, social studies, and mathematics classes as a common and frequent element of instruction. In the studies reviewed here, students used writing to promote learning an average of 30 days per study. Four, the incorporation of writing into content learning is not always effective, and some features of writing activities may require special scrutiny. While types of writing (informational, argumentation, journal, and graphical representation) and metacognitive prompting (present/not present) did not moderate writing-to-learn effects, graphical representation and no metacognitive prompting failed to produce statistically significant effects.

As with any research-supported practice, caution must be exercised when teachers implement writing-to-learn practices in their classes. First, we recommend that teachers carefully match writing activities with their goals for promoting content learning. It is important that students have the prerequisite writing skills needed to use such writing activities to support learning, which may require that teachers provide instruction on how to use the writing activity (Cross, 2008). Second, teachers should not automatically assume that writing activities that were effective in the research studies reviewed here will automatically be effective in their classrooms. The conditions that exist in a research study and a teacher’s class will not be identical. When using writing-to-learn activities, we encourage teachers to monitor if writing activities achieve the desired goals, and to make necessary adjustments if this is not the case. Moreover, the effective use of writing-to-learn assumes that teachers are adequately prepared to use these procedures. Many teachers indicate they are not prepared to do so (Ray et al., 2016).

Limitations and Additional Research Recommendations

Our meta-analysis located only 56 studies that met our inclusion/exclusion criteria. It is notable that no studies examined writing-to-learn in Grade 12. The number of studies that used narrative or argumentative writing tasks for learning was limited. Only three studies involved digital writing tools, and it was rare for researchers to examine if the same writing activity was effective across multiple studies. Additional high-quality intervention research is needed to draw a fuller and more nuanced picture of the effects of writing on learning.

The findings from our review must further be tempered by the fact that we were unable to draw any conclusions about the effects of writing on students’ learning over time. Maintenance data on learning effects were virtually nonexistent, and this must be rectified in future studies.

With the exception of grade level, we were not able to draw any conclusion about the effects of writing-to-learn on different types of students. This was because researchers inconsistently described participant characteristics, and they rarely provided outcome data for different groups. This must be rectified in future studies. At a minimum, findings should be reported by grade, gender, race, socioeconomic status, prior knowledge, and writing skills.

While we conducted a comprehensive search for studies, it is impossible to locate every pertinent investigation. We do not believe that missing studies were a significant problem though, as out post hoc analytic strategies (i.e., trim and fill and the Begg and Mazumdar rank correlation test) did not suggest that bias in the studies reviewed was evident.

Finally, while formal scientific review (peer review vs. no peer review) and study quality (e.g., study design, attrition) did not moderate obtained effects in this review, researchers need to ensure that future studies are even more sound than the studies reviewed here, and that all procedures are adequately described. This includes better addressing teacher effects and reliability of measures as well as decreasing/eliminating N-of-1 investigations and administering treatment fidelity measures in every investigation.