What Reading Alone Misses: Evidence on Peer Interaction, Collaboration Scripts, Collaborative Inhibition, Peer Instruction, and Vicarious Learning

🌏 中文版

This is Part 20 of the "Writing Articles That Teach" series and the first part of its fifth section, "who you learn with, and who can't read this." The first four sections dealt with one reader and one article: what goes on the page, what the reader is asked to do, what the reader brings, and what happens after some time. This section brings other people back in: classmates, discussion partners, people who leave comments, and people who for various reasons can't read the article at all.

The question comes from a gap that kept showing up in earlier parts. In the classroom meta-analyses covered in Part 8, active-learning classes mostly included small-group discussion and responses from the instructor. In the problem-first studies covered in Part 11, versions with small groups had larger effects (except in interventions lasting only a few hours, where working individually went with a larger increase). Both parts noted that these ingredients don't carry over to someone reading an article alone. This site's course guides and this series are written to be read alone: the article pages have no comment section, and readers have nowhere to talk to each other. So this part goes back to two questions. What does the evidence for learning together is better actually compare, and where does the benefit come from? And what does a person reading alone miss, and how much of it can an article give back?

Whether to find someone to read with is the learner's decision; the Learning How to Learn series on this site covers methods learners can use themselves. This part takes the writer's side. Comment sections, forums, and study groups themselves are covered in a later part; how much teaching others helps the person doing the teaching is covered in Part 22.

How the sources were chosen, and their biases. The source list came from scanning the literature first and then choosing what to read closely, with at least one study per direction whose result was opposite or null. There are 14 references, each read in full in the version I had, with an extraction record kept for each. The biases are clear. All are English-language journals. The participants are children, adolescents, and university students in school. The interaction was arranged by researchers or teachers. The two meta-analyses of collaboration scripts come from the same research group and mostly include the same papers; the two papers on vicarious learning come from the same first author's team. For argumentative dialogue and Knowledge Building, the only full texts available were one review that computes no effect sizes and two small studies. None of the studies looked at adults who read long articles online on their own.

How far these studies carry over

Before carrying the studies below over to one person reading a web article, keep five differences in mind.

1. "Together" is compared against many different things. Some studies compare with working alone, some with a control group that did nothing and was only tested twice, some with one-on-one learning with an adult, some with individual learning on a computer, and some with people recalling separately and pooling the results. For each number, check which comparison sits behind it.
2. The participants are students. The largest meta-analysis of peer interaction includes only ages 4 to 18; the two collaborative-learning meta-analyses draw mostly on school and university courses; Peer Instruction and vicarious learning studies used university students.
3. The time is short. In the peer-interaction meta-analysis, most studies had a single interaction.
4. The interaction is arranged. There is a task, a time limit, and a teacher or researcher deciding who works with whom. A self-learner finding someone to talk to is a different situation.
5. The outcomes are tests or amounts recalled. They do not measure whether someone understood a long article.

What can be borrowed is which ingredients go along with the benefit and when the benefit disappears. Applying them to a person reading an article alone is my inference throughout, and the text marks it.

The conclusions first

How much better is learning together on average?

The peer-interaction meta-analysis: what it compares

The 2020 meta-analysis by Harriet Tenenbaum and colleagues looks at one kind of interaction only: spontaneous two-way conversation between untrained peers. Structured or trained programs (cooperative-learning programs, peer tutoring) are excluded. It covers 62 papers and 71 samples, with participants aged 4 to 18.

Compared with other learning conditions, the peer-interaction groups scored higher by an average g = 0.40 (g is the difference between two group means divided by the standard deviation, so 0.40 is about four tenths of a standard deviation). The authors call it small but meaningful in the results and small to medium in the discussion.

Read this average together with three things:

• The comparison groups are mixed together: children doing the same task alone, control groups that did no task and were tested twice, and children learning one-on-one with an adult. Looking only at studies where the comparison was other children, g = .45.
• Against one-on-one with an adult, no difference was detected: g = −.30, an interval from −.70 to .10, and only 10 comparisons. The two versions of the paper word this differently. The author manuscript's abstract says peer interaction was as effective as one-on-one with an adult, while its discussion says one-on-one with an adult was more effective; the published version's abstract (readable on ERIC) changes this to "peer interaction was not more effective than child–adult dyadic interaction." Because the interval crosses 0, the most one can say is that peers were not better; one cannot say they were worse, or that they were equal.
• The studies vary a lot: results differ widely from study to study. No difference was detected between subgroups by age, by pairs vs. larger groups, by gender composition, or by testing during vs. after the interaction.

By domain, effects were significant for conservation concepts, scientific reasoning, and general reasoning; memory, spatial reasoning, and creativity were not significant. The original studies on memory and creativity mostly used a group whose members worked separately and then pooled their results as the comparison, which is exactly the situation in the collaborative inhibition section below.

Computer-supported collaboration: smaller numbers in randomized experiments

The 2018 meta-analysis by Juanjuan Chen and colleagues covers experiments and quasi-experiments on computer-supported collaborative learning (CSCL) from 2000 to 2016. The comparison most relevant here is the first category: both sides use computers, one side collaborates and the other learns individually, 84 studies.

The collaborating groups scored higher on knowledge tests by an average g = 0.42, on skills by 0.64, and on affect (questionnaires on course evaluation, self-efficacy, and so on) by 0.38.

But 0.42 pools all designs. Taken separately, randomized experiments show only 0.27, and quasi-experiments 0.54. The abstract does not mention this gap.

One more item in the same analysis bears on online readers: when one group discussed in a forum and the other collaborated face to face, no difference in knowledge tests was detected, with only 8 comparisons and a wide interval.

The authors include one sentence in their practical recommendations:

Last, note that CSCL is not a panacea for all kinds of learning. It might be only suitable for some learning content and activities.

That sentence rests on the observation that some studies lasted only a few weeks, not on an analysis. The authors also say task complexity could not be determined from the original studies, so the idea that tasks must be complex enough is only someone else's claim in this paper. The source of that claim is the theory paper in the next section.

Where the benefit comes from: the structure of the interaction

Being told to reach consensus

In Tenenbaum and colleagues' moderator analysis, studies whose task instructions asked peers to reach consensus had g = .61; those that didn't had g = .17, with a lower bound of only .02.

This is a comparison between studies, not random assignment of consensus instructions within one study. What can be said is that in the studies with larger benefits, peers were asked to talk their way to a shared answer.

Collaboration scripts

A collaboration script is a set of prompts that arranges how people learning together interact: assigning roles, or asking them to explain, ask questions, or make arguments to their partner.

Freydis Vogel and colleagues' 2017 meta-analysis and Anika Radkowitsch and colleagues' 2020 update both compare collaboration with a script against collaboration without one. Both exclude comparisons of collaboration with individual learning, so they do not answer whether learning together beats learning alone; they answer whether collaboration should be given structure.

• Domain knowledge: d = 0.20 in Vogel and g = 0.24 in Radkowitsch. Both are small effects.
• Collaboration skills: larger effects, but inflated by two outlier studies; without them, Vogel gets 0.65 and Radkowitsch 0.59.
• Motivation: measured only by Radkowitsch, 9 independent effects, g = 0.13, not significant. The authors' abstract calls this the first counterevidence against the criticism that scripts hurt motivation, while the conclusion says "the existing empirical evidence might still not be enough to refute this criticism completely."

The two are not independent pieces of evidence. When I compared their lists of included studies, 20 of the 21 papers with data rows in Vogel's appendix table also appear in Radkowitsch (Vogel's text says 22 papers were included, one of which has no data row in the appendix table). Radkowitsch's first search round carried over Vogel's criteria and exclusion decisions directly, and both papers list Vogel and Fischer as authors.

Vogel's abstract says scripts are especially effective when they prompt learners to build on their partner's contributions or come with content-specific scaffolding. In the body, neither test is significant, with power of only about one in ten, and the authors write in the discussion that "they do not constitute strong evidence."

As for why structure is needed, Paul Kirschner and colleagues, in the theory paper in the next section, cite a 2007 study by Weinberger and colleagues: "learners typically do not engage in effective collaboration processes without guidance." I did not read the original study.

Transaction costs: theory says the task must be complex enough

Paul Kirschner, John Sweller, and colleagues' 2018 paper extends cognitive load theory to collaboration and calls the result collaborative cognitive load theory. It is a theory paper with no data of its own.

Its argument is that when people collaborate, material one person can't hold can be spread over several people's working memories, but communicating and coordinating carry costs of their own (the authors borrow an economics term, transaction costs). Collaboration only pays when the benefit exceeds the cost. The first of its nine principles:

Effective collaboration occurs when a task is complex enough to justify the extra time and effort involved in the necessary transactional activities. If a task is not complex enough, unnecessary transactional activities will cause extraneous cognitive load and will, thus be detrimental to learning.

"Complex enough" is relative to an individual's working memory, so it is tied to prior knowledge. The authors say that when most group members already know the material, collaboration is unnecessary, and a high-level group may even show expertise reversal (support that helps novices can hurt experts; covered in Part 12). The conclusion takes it to the final step: "This decision could very well be to not use collaborative learning."

These are claims. The experiment supporting the claim that groups are more efficient when the task is complex enough is cited by the authors from a 2011 study by F. Kirschner and colleagues, which I did not read.

Where doing it together is worse: collaborative inhibition

The 2024 meta-analysis by Xiaochun Luo and colleagues covers collaborative inhibition: when several people recall together, they remember fewer items than the same number of people recalling separately with the non-duplicate items pooled. It covers 77 papers and 140 studies.

Recalling together loses to recalling separately and pooling, d = 0.54. After trim-and-fill added studies that may not have been published, it fell to 0.40; the authors read this as little sign of publication bias.

The inhibition doesn't show up everywhere:

• It was strongest among strangers recalling together (d = 0.75); no inhibition was detected between couples or friends.
• No inhibition was detected in mixed-gender pairs.
• With memory monitoring added (for example, having recallers judge how well they remember), no inhibition was detected, but all 12 of those effect sizes came from unpublished theses.

This paper measures the group's output at the moment of recalling together, and only for item memory such as word lists. It does not measure how much each person learned afterwards. The authors mention in their future directions that collaboration also has benefits (fewer errors when recalling together, better individual recall after collaborating) but do not analyze them. So it cannot be used to say that reading alone is better for memory. What can be said is that together doesn't help every kind of task, and for one common task, doing it together produces less.

Taking it apart: Peer Instruction

The meta-analyses above say the benefit follows structure. Peer Instruction is a method that writes the structure down as fixed steps and has been studied a lot, so we can look step by step at where the benefit lands.

The procedure and the evidence for each step

The 2015 literature review by Trisha Vickrey and colleagues covers studies in university STEM courses that followed Mazur's steps. It is a narrative review with no pooled effect size; the studies below are all reported through it, and I did not read the original studies.

Mazur's seven steps are: pose a conceptual question → give time to think → each student answers alone → students try to convince their neighbors → students answer again → the instructor looks at the answer distribution → the instructor explains the correct answer.

The evidence for each step:

• Answering alone first: in one survey study, 82% of students preferred to answer on their own before discussing. But the authors state that none of the studies in that section report learning gains, and the importance of an individual vote still needs research.
• What share of correct first answers calls for discussion: the authors report that Crouch and Mazur 2001 found the largest improvement on the revote when about half the class answered correctly at first, with substantial improvement still between 35% and 70%. On this basis the authors recommend peer discussion in that range, a hint or explanation first below 35%, and going straight to the explanation above 70%. The same paragraph also says students may still benefit from discussion when few answered correctly, and ends with "Regardless of the proportion of correct answers on the initial vote, students seem to benefit from peer discussion."
• What the discussions contain: they report that in 361 discussions recorded by James and Willoughby 2011, 62% included incorrect or unanticipated ideas. The authors conclude that "the improvements observed between the first and second vote may overestimate student understanding."
• Discussing reasons or discussing answers: they report that in Knight and colleagues 2013, groups asked to discuss their reasoning showed higher reasoning quality. This measured reasoning quality and was not linked to learning outcomes.
• The instructor's explanation: they report that in Smith and colleagues 2011, peer discussion plus the instructor's explanation produced larger gains than either alone.

The improvement here is mostly the change from the first vote to the second on the same question, not a later test.

Before and after discussion: where the answers go

Jonathan Tullis and Robert Goldstone's 2020 study analyzes data from six of the authors' own psychology and educational psychology courses: 208 students and 86 questions. For each question, students answered alone and rated their confidence, discussed with a neighbor, and answered again. It is a before-and-after comparison on the same question, with no group that skipped discussion; the authors also say they do not intend to evaluate long-term effects.

After discussion, the odds of a correct answer were 1.57 times the odds before, and accuracy went from 65% to 72%. Two of the six classes did not reach significance.

Where the answers moved is more useful than the average:

Students who answered wrong and picked the same answer as their neighbor stayed wrong 91% of the time, and their confidence still went up. Most of the students who switched from wrong to right came from the cell where they disagreed with their neighbor.

Do students just follow whoever is more confident? When peers disagreed, final answers went with the more confident peer only 58% of the time. With both peers' confidence controlled, students whose own first answer was right had 4.48 times the odds of keeping it compared with students whose first answer was wrong. Confidence itself also mattered, but whether the answer was right predicted the outcome beyond confidence. The authors interpret this as new understanding arising in the discussion; the data do not measure that directly, because the discussions were not analyzed.

The authors also note that 78% of students who answered correctly at first had a neighbor with the same answer, against only 43% of those who answered wrong. They infer that students sit next to people at a similar level and suggest that "Peer instruction could potentially benefit from randomly pairing students together." Random pairing was not tested.

For a person reading an article alone, my inference from this table is that the benefit of Peer Instruction concentrates on the moments when the person next to you picked a different answer. Someone reading alone has no such person. And when students met someone with the same wrong answer, only 9% switched to the right one after discussion, and their confidence went up.

Watching someone else being taught

Another line of research asks whether watching someone else being taught gives you the benefit of being taught.

Michelene Chi and colleagues' 2008 experiment recruited 70 university students who had not taken university physics to learn problems applying Newton's laws. There were five conditions: solving problems one-on-one with a tutor; pairs watching a video of someone else being tutored while solving the same problems; pairs solving the problems together without the video; one person watching the video and solving the problems; and one person solving the problems from the textbook. About a week later everyone took a test individually. The tutoring group was not randomly assigned (the first 10 people recruited all went into it).

Pre-to-post gains on deep problem-solving steps:

• Tutored one-on-one 21.3%, pairs watching together 17.03%; no difference was detected between the two. This was not an equivalence test, the groups had only 10 and 20 people, and the tutoring group was still numerically higher.
• Pairs solving together without the video, 7.19%.
• One person watching the video 8.91% and one person reading the textbook 3.39%; neither group's pre-to-post gain was significant.
• Pairs watching together vs. one person watching, p = .055, a marginal difference.

Two post hoc splits are worth noting. People who watched videos of good tutees gained 21.9%; those who watched poor tutees gained only 7.8%. Of the 10 people who watched alone, the 4 who more often wrote problem-solving steps on the whiteboard, paused and rewound the video, and asked themselves questions gained 18.4%, which the authors compare with 21.3% for the tutees and 17.0% for the pairs watching together; the other 6, more passive, gained 2.6%. No statistical test was done on this split. The authors' correlational analysis also found that what observers learned was more often related to the tutees' responses and not related to the tutor's own explanations.

So what can be said is: watching passively alone produced almost no learning; watching in pairs while solving the same problems produced gains close to those of the tutees themselves. One cannot say that watching alone doesn't work.

The same team's 2017 paper reanalyzes data from Muldner and colleagues 2014: pairs watching a video of a tutoring dialogue learned more on 11 transfer questions than pairs watching the same tutor explain to the camera. In both viewing conditions the pairs were required to discuss and fill in one shared workbook, and there was no condition of watching alone; the effect size printed in the paper doesn't match the test statistic, so it is not cited here. Dialogue videos and written dialogue are covered more fully in Part 22.

Argumentative dialogue and Knowledge Building: what is available

When planning this part, I meant to include two lines often cited in support of learning together: having learners argue with each other, and Knowledge Building. After reading the full texts I could get, neither line can support a conclusion that it works or that it doesn't.

Argumentative dialogue

Chrysi Rapanta and Mark Felton's 2021 review covers 143 studies from primary school to university and looks at how "learning to argue" is taught in classrooms. The authors state that they do not measure effectiveness: "our goal was not to measure the effectiveness of dialogue on students' learning (such a goal would have been impossible given the variety of methodological approaches)." Of the 143 studies, 57% are descriptive. From associations among study characteristics the authors derive four patterns (whole-class low-structure, two kinds of small-group high-structure, and one-to-one dialectic), and they say these associations cannot support causal inferences.

Hongli Gao and colleagues' 2022 experiment had 32 pairs of university students discuss how artificial intelligence will affect humans, online, with a computer agent for 25 minutes. For one group the agent only offered different viewpoints; for the other it followed each viewpoint with two questions, "do you agree with me?" and "could you give the reasons to support your viewpoint?", and also gave prompts encouraging participation. The group with the argumentative prompts stated fewer opinions and had fewer question-and-answer exchanges, and its discussion stayed closer to the agent's keywords. Expressions of disagreement averaged 0.13 in both groups, and no difference was detected in discussion depth or in change of understanding from before to after. In the authors' words, "the content-independent scaffolding in the experimental condition did not provoke more debates." The experiment had no learning test, and for the one significant self-rating item, the paper doesn't make the direction of the scale clear.

The scan turned up three meta-analyses of argumentation-based instruction (Koçoğlu and Kanadlı 2024, Fu and Zhan 2025, Ramallosa and colleagues 2022). These studies exist, but I have not read a legally available full text of any of them, so their numbers are not cited.

Knowledge Building

Knowledge Building is an approach proposed by Scardamalia and Bereiter in which a group of learners post ideas in a shared online note space, Knowledge Forum, build on each other's notes, and then synthesize several notes into one.

Gaoxia Zhu and colleagues' 2021 study revised the activities three times based on student feedback in a 23-student master's course in education, and looked at changes in the network of who read whose notes and in note content. The density of the reading network rose from one focal week to the next, and the share of elaborated theorizing and referencing notes increased. It is design-based research: one class, no control group, no achievement test, no statistical tests. The authors write: "We would expect some degree of increasing reading network density as a simple result of the course progression (i.e., even with no changes to the design)."

For this line I found no meta-analysis and no independent evaluation from outside the group; I have not read the full text of the two handbook chapters or of Chen and Hong's 2016 review of thirty years of research. So this part has no basis at all for saying whether Knowledge Building improves learning.

What students themselves think

Yunfeng Wei and colleagues' 2024 survey asked 202 students in an introductory psychology course in spring 2021, during the pandemic and in hybrid teaching.

• Is studying more effective in a group or alone? 102 vs. 100; no difference was detected.
• Is recalling more effective in a group or alone? More chose the group (about 120 vs. 82).
• For exam preparation, about 112 preferred studying alone and about 60 preferred a group.
• 151 students, about three quarters, rarely or never attend study groups (I summed this from the table).

The authors' summary is that "students believe recalling in study groups is beneficial, but they prefer to study individually and rarely use study groups, indicating their actual and preferred study behaviors are not consistent with their beliefs." The abstract says students generally see collaboration as beneficial, but the question on studying was actually split down the middle. Broken down by group, students who often attend study groups mostly thought group study was more effective and also preferred it; those who rarely attend leaned the other way. The gap between belief and preference mainly appears once everyone is pooled.

The main reason for choosing to study alone was that groups are distracting. Among the strategies listed for group study, teaching each other, quizzing each other, asking others, and comparing notes were never listed by anyone for studying alone.

The survey measures beliefs and habits, not learning, and it can't show whether collaboration actually helps.

What reading alone misses

Putting the results above together, what follows is my inference, with the basis given after each point.

What is missing

1. An answer different from your own. The gains in Peer Instruction concentrate on cases where neighbors disagreed (Tullis and Goldstone); among the group-study strategies students listed, quizzing each other, comparing notes, and asking others don't happen when you study alone (Wei and colleagues).
2. Having to explain your reasons to someone, and having to reach a conclusion together. Peer interaction had larger effects when consensus was required (Tenenbaum and colleagues, a comparison across studies); discussing reasoning produced higher reasoning quality than discussing answers (reported by Vickrey).
3. Someone arranging the structure of the interaction. With a script, collaboration gained a little in domain knowledge and showed a larger difference in collaboration skills (Vogel, Radkowitsch); the theory paper cites a study saying students typically don't collaborate effectively without guidance.
4. Someone to work with, or at least something that gets you doing things. When watching someone else being taught, people who watched in pairs and solved problems as they went learned more, and people who watched passively alone learned almost nothing; but the 4 people who watched alone and more often wrote steps and questioned themselves also gained close to the tutees (Chi and colleagues 2008; small sample, post hoc split, no test).

What isn't necessarily missing

• The cost of coordinating. Reading alone doesn't pay it. When the task isn't complex enough, or the reader already knows the material, collaboration may not pay off (Kirschner and colleagues, theory).
• Distraction while studying. This is the main reason students choose to study alone (Wei and colleagues).
• Recall output. Groups recalling together actually produced less (Luo and colleagues), but that isn't a measure of learning.

What an article can give back is only part of the first point: the author can write, in the answer, another answer a reader might pick, explain why it looks reasonable, and say where it goes wrong. There is a risk. In Tullis and Goldstone's data, meeting the same wrong answer as your own raised confidence. So the other answer, once written down, has to be refuted right away, not merely set alongside. Writing out the wrong idea, saying plainly that it is wrong, and then giving the correct explanation is itself supported by meta-analyses of refutation text; see Part 1 and Part 13. What no study has tested is using text in place of a real person who disagrees. The second and third points need another person, and an article can't provide one. For the fourth, an article can ask readers to write steps as they read and to solve before reading on, but no study has compared whether that is enough.

This points the same way as the ICAP framework covered in Part 8: by definition, an article's static questions reach at most "constructive," and reaching "interactive" requires a dialogue partner (the original allows a computer agent; whether the AI chat sidebar on this site's article pages counts, Part 8 leaves open, and this part doesn't count it either). What Chi and colleagues 2008 tested was precisely the hypothesis of active, constructive, and interactive observing.

Where the sources disagree

Peer interaction vs. one-on-one with an adult. Tenenbaum and colleagues' author manuscript says in its abstract that the two are equally effective; the results paragraph of the manuscript says both things in the same paragraph (one-on-one with an adult beat peer groups, and then that the two did not differ), and the discussion says one-on-one with an adult was more effective; the published version's abstract changes this to say peer interaction was not more effective. All three rest on the same numbers: g = −.30, with an interval that crosses 0.

Do observers learn as well as tutees? Chi and colleagues' 2008 abstract says observers who watched together learned as effectively as the tutees, and the 2017 abstract says "they can learn as well as the tutees." Both rest on the absence of a significant difference, with no equivalence test and samples of 10 vs. 20; in 2008 the tutoring group was still numerically higher.

Is watching together better than watching alone? Chi and colleagues' 2008 conclusion says four results all support the hypothesis of active and interactive observing. Of the four, watching together vs. watching alone was p = .055; the active solo watchers were only 4 people with no test; and the high- vs. low-interaction pairs were a post hoc split.

Scripts and motivation. Radkowitsch and colleagues' abstract calls the analysis the first counterevidence against the claim that scripts hurt motivation; the conclusion says the evidence is not yet enough to refute that claim completely. The estimate of 0.13 is not significant; after adding studies that may be missing, it is a significant 0.16.

The threshold for discussion. Vickrey and colleagues recommend peer discussion when 35% to 70% of first answers are correct, and in the same paragraph say students may still benefit when few answer correctly and that the threshold needs more research. In Tullis and Goldstone's data, improvement after discussion appears at every level of difficulty, but a few questions saw accuracy drop after discussion.

Cultural differences in collaborative inhibition. Across all of Luo and colleagues' data, inhibition was smaller in Eastern samples than in Western ones, and the difference was significant; among published studies only, the difference was not significant. The abstract reports only the former.

How large the effect of computer-supported collaboration is. Chen and colleagues' abstract reports the positive pooled effects for each category; split by study design, randomized experiments are much smaller, and in the category of collaborating with vs. without computers, randomized experiments show only 0.09, not significant.

Comparing with this site's practice

This site's writing spec for course guides (course-guide-contract.md) requires a self-check at the end of each main section with an answer or criteria, and requires concept guides to point out common misconceptions. Set against the steps of Peer Instruction:

What you can do

1. Don't write "research shows learning together works better" in an article. If you cite it, state the comparison and the participants, for example "peer interaction among 4- to 18-year-olds, compared with working alone or with control groups tested twice." Basis: Tenenbaum and colleagues' comparison groups are mixed, and Chen and colleagues' randomized experiments show only 0.27.
2. In a self-check answer, write out another answer, then refute it. Pick the wrong answer readers are most likely to choose, and write why it looks reasonable and where it goes wrong. Basis: Tullis and Goldstone's answer flows and the finding that confidence rose when students shared a wrong answer; writing out the wrong idea and saying plainly that it is wrong is supported by meta-analyses of refutation text (see Part 13). What no study has tested is text standing in for a neighbor who really disagrees; that part is this site's judgment.
3. Put questions where readers' answers are likely to split. Basis: Crouch and Mazur as reported by Vickrey, where discussion produced the largest improvement when about half the class answered correctly at first. The author can't see readers' accuracy, so the choice is a judgment call; this is my inference.
4. Give people reading together a question they must reach a shared answer to. If the article is material for a study group, add after the question: "Each of you write an answer first, then talk until you agree." Basis: Tenenbaum and colleagues' consensus moderator (children, a comparison across studies) and the first four steps of Peer Instruction. No study has tested this with adult study groups.
5. Don't make "find someone to read with" a recommendation for every article. Basis: collaboration has coordination costs and may not pay off when readers already know the material (theory); recalling together produces less (Luo and colleagues); most students prefer to study alone (Wei and colleagues).

Three things to do tonight

1. Pick one self-check in your own article and add two sentences at the start of the answer: "A common other answer is X, because…; it goes wrong at…". An example from this article's own content (written for this article): the question "Peer interaction has g = 0.40. Can we say reading an article together means learning four tenths of a standard deviation more?" The answer first says "Many people would say yes, because it's the average from a meta-analysis," then "No: the participants were 4 to 18, most studies had a single interaction, the comparison groups mix working alone, being tested twice, and one-on-one with an adult, and none of the studies involved adults reading an article."
2. Find a question that only asks for a conclusion ("Is A better than B?"), turn it into one that asks readers to write down their reasons, and give criteria in the answer: a good reason names the comparison and the conditions. The basis is the finding reported above that discussing reasoning produced higher reasoning quality than discussing answers; that was classroom discussion, not one person writing.
3. If someone is reading a series with you, pick one question in one part. Each of you writes an answer first without showing the other; if the answers differ, talk until you both agree, then open the correct answer. That is the first four steps of Peer Instruction plus the final explanation.

Limits of this evidence

• No study looked at adult self-learners reading long articles. None of these sources compared adults reading a web article alone with reading it together.
• The participants are students: the peer-interaction meta-analysis includes only ages 4 to 18; collaborative learning draws mostly on school and university courses; the Peer Instruction, vicarious learning, and student survey studies all used university students.
• The interactions are short: in the peer-interaction meta-analysis, most studies had a single interaction; Chi and colleagues' 2008 intervention was limited to two hours.
• The moderator analyses compare across studies: consensus, the various script moderators, and Chen and colleagues' split by study design are not random assignment within a study and cannot be read causally. Radkowitsch and colleagues state this; what Vogel and colleagues say is that the power of their moderator analyses was only one or two in ten.
• The same research groups: the two script meta-analyses share most of their papers and authors; the two vicarious learning papers come from the same first author's team.
• The two Peer Instruction papers each have gaps: Vickrey and colleagues is a narrative review whose numbers are all reported secondhand; Tullis and Goldstone has no group without discussion, the authors were the course instructors, and the measure is answering the same question again.
• Collaborative inhibition measures group output, not learning, and only item memory.
• Versions: for Tenenbaum and colleagues I read the author manuscript; of the published version I read only the public abstract, and all numbers in the text and appendix come from the manuscript. For Chi and colleagues 2017 I read the published version on the authors' lab website.
• My inferences appear in six places: carrying the research over to one person reading an article; what the concentration of Peer Instruction's benefit on disagreement means for readers; the whole section "What reading alone misses"; that writing out another answer can give back part of it; the verdicts in the comparison with this site's practice; and "What you can do" and "Three things to do tonight." Each is marked in the text.
• Known to exist, but without a legally available full text I have read: Jeong and colleagues' 2019 meta-analysis of CSCL in STEM; Nokes-Malach and colleagues' 2015 review on when it is better to learn together; Marion and Thorley's 2016 meta-analysis of collaborative inhibition (the moderator of group size is only in this one; for familiarity, Luo and colleagues' subgroups are available); Balta and colleagues' 2017 meta-analysis of Peer Instruction; the original studies by Smith and colleagues 2009 and Crouch and Mazur 2001 (read only as reported by Vickrey and colleagues); the three meta-analyses of argumentation-based instruction; Chen and Hong's 2016 review of Knowledge Building; and the handbook chapters by Scardamalia and Bereiter and by Andriessen and Baker.
• Secondhand citations: Crouch and Mazur 2001, Smith and colleagues 2009 and 2011, James and Willoughby 2011, Knight and colleagues 2013, and Nicol and Boyle 2003 as reported by Vickrey and colleagues; Weinberger and colleagues 2007 and F. Kirschner and colleagues 2011 as reported by Kirschner and colleagues; Muldner and colleagues 2014 as reported by Chi and colleagues 2017.
• Gaps in reading: I did not read the supplementary materials of Chen and colleagues or Luo and colleagues (study lists, funnel plots); for most papers I read only the figure captions, not the figures.

Appendix: design and numbers for each study

The main text keeps one or two numbers per item; the rest are here.

Tenenbaum, Winstone, Leman, Avery (2020), Journal of Educational Psychology 112(7). Random-effects model, Hedges' g. 62 papers, 71 samples, ages 4 to 18; studies through 2017, English only; trained peer programs and studies without a comparison group excluded. Overall g = 0.40 (95% CI 0.27–0.54), Qw(70) = 344.23. Fail-safe N = 2,669; trim-and-fill unchanged when trimming from the left, 0.42 from the right. Consensus: required .61 (.42–.81, k = 39), not required .17 (.02–.33, k = 33), Q = 11.74 (k totals 72 because one paper was split into two entries). Comparison: the author manuscript uses two categories, other children .45 (.32–.58, k = 68) and adults −.30 (−.70 to .10, k = 10). Pre-post vs. posttest-only: no difference detected. Timing of the posttest: during the interaction .68, delayed .41, immediately .10 (k = 15), Q(2) = 4.85, p = .18, no significant difference between groups; by my calculation from the degrees of freedom, that Q corresponds to p ≈ .09, so the two numbers don't match and can only be copied as printed. Ages 11–18 .62, ages 4–10 .33 (p = .10); groups larger than two .61, pairs .34 (p = .17); the k values for gender composition are the reverse of what the methods section says. Domain Q(8) = 11.05, p = .20: conservation .98, scientific reasoning .45, general reasoning .39, spatial conservation .37, mathematical reasoning .34, moral .31 (not significant), memory .25 (not significant), spatial reasoning −.03 (not significant), creativity −.61 (not significant, k = 2). In Table 2 of the author manuscript, Diehl and Stroebe 1987 on brainstorming has g = −1.86.

Chen, Wang, Kirschner, Tsai (2018), Review of Educational Research 88(6). Web of Science plus Google Scholar, English-language journals 2000–2016, requiring an independent control group and confirmed pretest equivalence. 356 papers, 425 studies, in four categories. Category 1 (collaboration vs. individual learning, both with computers), 84 studies, 11,684 people: knowledge 0.42 (0.32–0.53, k = 73, QT significant), skills 0.64 (k = 17), affect 0.38 (k = 26). Category 1 knowledge by design: randomized 0.27 (0.10–0.44, N = 25), quasi-experimental 0.54 (N = 36), unspecified 0.39; by sample size, under 100 people 0.45, 250 or more 0.25; by duration, under 1 month 0.36, 1 month to 1 semester 0.52, 1 semester or more 0.40; researcher-made tests 0.45, standardized tests 0.29 (k = 6). Category 2 (collaborating with vs. without computers), knowledge 0.45, of which randomized experiments 0.09 (−0.17 to 0.35, N = 16, not significant). Category 3, adding tools: basic online discussion vs. face to face 0.52 (−0.66 to 1.70, k = 8, not significant); adding strategies: instruction and guidance (mainly collaboration scripts) 0.41 (k = 25), peer feedback 0.35, role assignment 0.36, teacher guidance 0.34 (not significant). Publication bias was checked only with funnel plots. Several statements in the text don't match the tables (for example, basic online discussion is described as a medium effect).

Radkowitsch, Vogel, Fischer (2020), ijCSCL 15. Computer-supported collaboration with vs. without a script; comparisons with individual learning excluded. 49 papers, 53 studies, 5,616 people, ages 6–64, 36 studies in higher education. Motivation g = 0.13 (−0.01 to 0.27, 9 independent effects), 3PSM 0.14, trim-and-fill 0.16 (0.01–0.31). Domain learning 0.24 (0.16–0.33, k = 54), I² = 46%; recall 0.27, application 0.21. Collaboration skills 0.72 (0.44–1.01, k = 23), I² = 82%; 0.59 without two outlier studies. No study measured coordination skills. Neither the number of types of activities prompted nor their combination was a significant moderator. Limitations the authors list: script designs vary enormously, most studies don't report what learners actually did, all moderator comparisons are between studies, and most interventions were short.

Vogel, Wecker, Kollar, Fischer (2017), Educational Psychology Review 29. Searched in 2013; 22 papers, 24 studies, 2,825 people, average age 20; average learning time with the support about 4 hours. Intervals in the tables are 90% CIs and p values are one-tailed. Domain knowledge d = 0.20 (0.04–0.36, k = 33), unchanged by trim-and-fill. Collaboration skills 0.95 (0.63–1.27, k = 12), 0.65 without outliers; the funnel plot is skewed, 0.85 after trim-and-fill. The moderators of transactivity, script level, and content support were all non-significant for domain knowledge, with power of 8–16%. The meta-regression on learning time was not significant. Subject area was significant only for collaboration skills.

Kirschner, Sweller, Kirschner, Zambrano (2018), ijCSCL 13(2). Theory paper with no data. Nine principles of collaborative cognitive load: task complexity, task guidance and support, domain expertise, collaboration skills, team size, team roles, team composition, prior task experience, and prior team experience. The authors call these hypotheses still to be tested.

Luo, Zhao, Tang, Xiao, Liu (2024), Behavioral Sciences 14(9). Chinese and English databases through November 2023; 77 papers (15 theses), 140 studies, 185 effect sizes, 19,492 people. d = 0.54 (0.49–0.58), I² = 73%; 0.40 (0.33–0.51) after trim-and-fill added 36 studies on the left. Gender pairing: same gender 0.64, mixed −0.11 (not significant). Relationship: couples −0.19 and friends 0.18 (both not significant), acquaintances 0.45, strangers 0.75. Culture: Eastern 0.47, Western 0.76; published studies only, Q = 2.31, p = 0.13. Memory monitoring: without 0.66, with 0.14 (not significant; all 12 effect sizes from theses). Age appears only in the discussion: older adults −0.09 (not significant). Several numbers differ between the text and the tables.

Vickrey, Rosploch, Rahmanian, Pilarz, Stains (2015), CBE—Life Sciences Education 14(1). Narrative review of university STEM studies that followed Mazur's steps; the supplementary table lists fifty-some studies (my count from the table). By the authors' count, 84% of the cited studies were in lower-level courses and 75% in classes of 50 or more. Other studies reported: in Crouch and Mazur 2001, normalized FCI gains were about twice those of traditional lecturing (a longitudinal comparison); in Zingaro 2014, final exam scores were 4.4% higher, no difference detected (p = 0.10); Lasry 2008 detected no difference between clickers and flashcards; in Perez 2010, students who saw the vote distribution were 30% more likely to switch to the most common answer; in Smith 2009's genetics course, 77% of students who answered wrong at first and right after discussion went on to answer an isomorphic question correctly.

Tullis and Goldstone (2020), Cognitive Research: Principles and Implications 5. Six courses, 208 people, 1,657 responses, 86 questions, mixed-effects models. Odds ratio 1.57 (95% CI 1.31–1.87); 88% of students improved or stayed the same in accuracy. Per-class paired t tests before vs. after had d from 0.31 to 1.74, with two classes at p = 0.06 and 0.07. Answer flows: 603 wrong first answers (343 disagreeing, 260 agreeing), 1,054 correct first answers (232 disagreeing, 822 agreeing); 5% switched from right to wrong, 28% from wrong to right. Confidence when agreeing on a wrong answer went from 5.45 to 6.60. Each additional point of one's own confidence multiplied the odds of keeping one's answer by 1.25; each point less of the partner's confidence by 1.19. Always choosing the more confident peer's answer would give 69% accuracy, against an actual 72%; choosing the right answer whenever either partner had it would give 80%. Metacognitive resolution rose overall (d = 0.24), but only two of six classes were significant and two went the other way.

Chi, Roy, Hausmann (2008), Cognitive Science 32(2). 70 university students: 10 in tutoring, and 10 pairs or 10 individuals in each of the other four groups; one tutor with a PhD in physics; participants first read textbook chapters 1–4 until they answered 80% correctly on each; the intervention was limited to 2 hours; the posttest came about 7.4 days later. No difference among the five groups on shallow steps. ANCOVA effect sizes on deep steps against reading alone: tutoring .815, watching together .613, collaborating .326. Effect sizes of pre-post gains: tutoring 1.063, watching together 0.883, collaborating 0.405, watching alone 0.291, reading alone 0.296. High-interaction watching pairs (difference in the two partners' contributions under 15%), 6 pairs, gained 26%; low-interaction pairs, 4 pairs, 4%. Several F values seem not to match their p values or d values (for example F = 6.448 with p = .044), and the 10 watching pairs were analyzed as 20 individuals.

Chi, Kang, Yaghmourian (2017), Journal of the Learning Sciences 26(1). A reanalysis of data from Muldner and colleagues 2014: 10 tutees, 40 observers (same-gender pairs), on the concept of diffusion. In both viewing conditions the pairs had to fill in one shared workbook, with no time limit. On 11 transfer questions, tutees and dialogue observers each improved from pre to post (d = 1.07 and 1.08, within-group gains); monologue observers p = .26. Dialogue vs. monologue t(38) = 2.05; the effect size printed in the paper doesn't match the t value, so it is not cited. Dialogue observers referred to the tutee more often than to the tutor (8.70 vs. 4.00). The abstract says the behaviors "mediated" learning, while the text says the sample was too small for a mediation analysis.

Rapanta and Felton (2021), Educational Psychology Review 34. Integrative review of 143 studies (1997 to November 2020, K-16, conference papers excluded); 82 descriptive, 37 quasi-experimental, 24 experimental. The four patterns cover 108 studies. The abstract says three patterns; the text has four.

Gao, Xu, Yang, Hu (2022), Frontiers in Psychology 13. 64 university students (9 men, 55 women), 16 pairs per condition. Opinion statements 3.13 vs. 6.88 (d = −1.29); stage-one questions and answers 3.31 vs. 8.00 (d = −0.84); share of valid viewpoints 0.70 vs. 0.57; no difference detected in number of viewpoints, sentences, categories, or depth. Self-rated comprehensive understanding of the topic 4.13 vs. 3.84, but the questionnaire numbers option 1 as strongly agree, and the paper doesn't say whether it was reverse-scored. A dozen or so t tests with no correction for multiple comparisons.

Zhu, Raman, Xing, Slotta (2021), IJETHE 18. 23 master's students in education, a 12-week course, with weeks 3, 6, and 9 as focal weeks; 259 notes in Knowledge Forum. Reading network density 0.57, 0.74, 1.09 (multiple edges allowed). The emotion questionnaire analyzed only the 8 people who completed it both times; facial expression analysis covered 3 people; 9 were interviewed. Percentages of note types were read off the figures; the text gives no numbers.

Wei, Soderstrom, Meade, Scott (2024), Behavioral Sciences 14(11). 202 people, average age 20.1. Studying effective 102 vs. 100 (χ² p = 0.89); recalling effective χ²(1) = 6.42, p = 0.01; exam preference χ²(2) = 51.13; study-group frequency: often about 11, sometimes about 40, rarely about 95, never about 56. Reasons for studying alone: distraction, about 92 people. The most common solo strategy was rereading (69.8%), with spaced repetition at only 4.0%; the most common group strategy was teaching each other or discussing (44.1%). Frequent vs. rare study-group users: study beliefs 38/13 vs. 64/87 (p < 0.001); recall beliefs not significant; exam preference 32/15 vs. 28/97.

References

Each was read in full in the version I had, with an extraction record kept; versions and parts I could not read are noted in parentheses.

• Tenenbaum, Winstone, Leman & Avery (2020), How Effective Is Peer Interaction in Facilitating Learning? A Meta-Analysis (read as the author manuscript in the King's College London repository, peer-reviewed but not the final version; of the published version, only the abstract on ERIC).
• Chen, Wang, Kirschner & Tsai (2018), The Role of Collaboration, Computer Use, Learning Environments, and Supporting Strategies in CSCL: A Meta-Analysis (read as the published version in the Open University of the Netherlands repository; supplementary materials not read).
• Radkowitsch, Vogel & Fischer (2020), Good for learning, bad for motivation? A meta-analysis on the effects of computer-supported collaboration scripts (published version, CC BY).
• Vogel, Wecker, Kollar & Fischer (2017), Socio-Cognitive Scaffolding with Computer-Supported Collaboration Scripts: a Meta-Analysis (published version, CC BY).
• Kirschner, Sweller, Kirschner & Zambrano (2018), From Cognitive Load Theory to Collaborative Cognitive Load Theory (published version on PMC, CC BY).
• Luo, Zhao, Tang, Xiao & Liu (2024), Meta-Analysis of Collaborative Inhibition Moderation by Gender, Membership, Culture, and Memory Monitoring (published version on PMC, CC BY; supplementary materials not read).
• Vickrey, Rosploch, Rahmanian, Pilarz & Stains (2015), Research-Based Implementation of Peer Instruction: A Literature Review (published version on PMC, including the supplementary table).
• Tullis & Goldstone (2020), Why does peer instruction benefit student learning? (published version on PMC, CC BY).
• Wei, Soderstrom, Meade & Scott (2024), Metacognition About Collaborative Learning: Students' Beliefs Are Inconsistent with Their Learning Preferences (published version on PMC, CC BY).
• Rapanta & Felton (2021), Learning to Argue Through Dialogue: a Review of Instructional Approaches (published version, CC BY; references only skimmed).
• Gao, Xu, Yang & Hu (2022), The double-edged sword effect of argumentative scaffolding on group discussion in an adaptive discussion system (published version, CC BY).
• Zhu, Raman, Xing & Slotta (2021), Curriculum design for social, cognitive and emotional engagement in Knowledge Building (published version, CC BY).
• Chi, Roy & Hausmann (2008), Observing Tutorial Dialogues Collaboratively: Insights About Human Tutoring Effectiveness From Vicarious Learning (published version, free to read on the publisher Wiley's page; appendices A and B not read).
• Chi, Kang & Yaghmourian (2017), Why Students Learn More From Dialogue- Than Monologue-Videos: Analyses of Peer Interactions (read as the published version on the first author's lab website).
• On this site: What the Evidence for Learning by Doing Measures, How Much Guidance Should a Project in a Teaching Article Give?, Writing a Teaching Article for Readers at Different Levels, Writing the "Common Misconceptions" Part of a Teaching Article, Writing Questions and Wording in a Teaching Article, Learning How to Learn, Harvard CS50 AI Week 0 guide.