Skip to content

Formative Assessment

Pattern · All patterns
Evidence · 42 claims (21 for, 20 mixed, 1 against) · 48 studies (15 causal, 12 quant-synthesis, 8 review, 5 qualitative, 4 associational, 3 theoretical, 1 design), q1–q4 · 17 of 48 report an effect size · 31 claims rest on one study

Description

A reusable elicit–interpret–act–reobserve policy, qualified by what a response can establish and by whether a relevant next action is available.

Principle

This reusable policy instantiates the formative assessment principle.

Learner state, objective and valued goal

This is a reusable design model. A study result supplies a bounded observation; it does not reveal the state or future outcome of a new learner.

Before selecting a configuration, record the target task and representation, task-specific prior experience, access and language constraints, time available, observed responses and the exact help provided. Preserve an unknown rather than replacing it with a generic label such as novice. Ask the learner what they want to accomplish and why it matters. Record that valued goal separately from the designer's learning objective; agreement is an observation to elicit, not an assumption. If they diverge, negotiate the task or purpose and retain the unresolved difference.

Define the intended capability, a representative assessment task and scoring instrument, the criterion chosen for this design, and the assessment horizon. Agree what response would warrant changing support and how to check later independent performance. Criteria are local decisions, not universal mastery thresholds. If the brief only asks for better learning, elicit these fields before selecting an exact dose or forecasting a gain.

Keep the record in this form: response under stated conditions → uncertain state hypotheses → discriminating observation → next activity → response at the stated horizon. Difficulty, accuracy, speed, confidence and engagement are different observations. None alone proves learning. Confidence in particular is not a reading of the state: across 22 meta-analyses, self-evaluations agreed with measured ability only moderately (mean r = .29), more closely on familiar, domain-specific tasks, and in one undergraduate course the lowest performers were grossly overconfident — Prior knowledge is needed for accurate self-assessment [~M]. Performance during learning also misleads about the horizon: with prose passages, restudying beat recall on a test five minutes later and lost to it at two days and one week — Retrieval practice improves long-term retention [+S]. So the response that settles a branch is the one taken at the stated horizon, not the one taken during the session. Record a learner's changed purpose or constraints as well as a changed performance.

Observation and adaptation

This diagnostic policy is a design proposal; synthesis-level effects do not validate its specific branches. Elicit an unassisted explanation or application aligned with the intended capability, plus the learner's reasoning and confidence. Record the prompt, response mode, language demand, stakes and assistance. Interpret the response using explicit rival hypotheses before selecting feedback; a checkpoint without a usable next action is not this pattern.

Four findings bear on the steps, though none of them tests these branches. Elicit with an unassisted recall or production attempt where the target allows it: across 159 effect sizes from 61 studies, a practice test beat restudying, more so at delays of a day or more and when feedback followed, and with no feedback and initial performance at or below 50% the benefit was about zero — Retrieval practice improves long-term retention [+S]. Act with information about the task and how to do it, not a verdict or praise: across 435 educational studies high-information feedback (d = 0.99) far outran reinforcement or punishment (d = 0.24), and in a cross-setting meta-analysis more than a third of feedback interventions lowered performance, effectiveness declining as feedback drew attention toward the self — Feedback improves learning by a medium amount on average, but its effect varies widely with the information it carries, and a substantial share of feedback effects are negative [+S]; Feedback works best at task and process levels [+S]. Direct the correction at the errors: in a word-pair experiment, giving the correct answer after an incorrect retrieval greatly raised retention a week later, while feedback after correct responses made little difference even when learners were unsure; in classroom quizzing studies, quizzes followed by corrective feedback showed a larger testing effect (g = 0.537 against 0.374, a between-study comparison that an earlier meta-analysis did not replicate) — Feedback enhances retrieval practice [+S]. Make the next action contingent on the response: in one-to-one long-division tutoring, support raised after a failure and lowered after a success beat fixed high, moderate or no support at immediate and one-month tests, with eight learners per condition — Contingent scaffolding improves learning more than fixed or absent support [+S]. Before the feedback, asking the learner to judge their response against the criterion is the self-monitoring step that two syntheses hold makes feedback usable; it is a mechanism argument, with no effect size recorded — Self-monitoring improves self-regulation and supports better learning decisions [+M].

Interpret from the response, not from impressions of the learner. In one within-subjects experiment, 33 Dutch secondary teachers judged students' text comprehension more accurately when they had only performance cues (scores on diagnostic tasks) than when they had only student cues such as interest in the topic, and more accurately with performance cues alone than with both mixed — Teachers judge more accurately with diagnostic cues [+M]; Non-diagnostic cues hinder teacher judgment accuracy [+M]. The same teachers ranked which cues predict comprehension only moderately well (mean rank correlation .55), often counting a student's interest in the topic as diagnostic, and a brief explanation of which cues are diagnostic did not measurably raise accuracy, which was already fairly high — Teachers are partly aware which cues are diagnostic [~W]; A brief cue-diagnosticity intervention did not improve accuracy [~W]. In an earlier think-aloud study, teachers judging anonymous students' diagrams mostly used the diagrams (78% of the cues they named) but still guessed at characteristics of the unseen student for half of them — Teachers fabricate student cues when judging from diagrams [~W]. So the record of a reading should name the performance evidence it rests on; a reading that names none is an impression to check, not a diagnosis. The learner's own reading of which answers were wrong is no substitute for scoring: in a word-pair experiment, learners judged 22% of their wrong answers correct and, when allowed to skip feedback, skipped it after 15% of their errors while still viewing it after about a third of their correct answers — Learners' correctness self-assessments are imperfect [~M]; Learners' feedback-skipping decisions diverge from the optimal rule [~M]. Decide which responses get corrective feedback from the scored response, not from the learner's sense of having got it right.

Observation Rival interpretations and discriminating observation Proposed next action
An explanation omits an important relation A conceptual misconception and an expression/language barrier can produce this. Compare two cases varying one relevant quantity, ask for a prediction and justification, and allow a diagram or accessible oral response. If the relation still fails across response modes, give a targeted contrast and invite revision. If reasoning is coherent in another mode, address expression/access rather than reteach the concept wholesale.
A revised answer becomes correct Answer copying and revised understanding both fit. Ask for a new unassisted case and justification, then repeat at the intended delay. If only the feedback example succeeds, retain targeted support and another independent attempt. If the new case succeeds, reduce that support provisionally and check later.
A learner stays silent or guesses Uncertainty about the concept and evaluative threat/disengagement are different possibilities. Ask privately about purpose and constraints and compare a low-stakes accessible attempt with the original response. Adjust the conditions when they obstruct expression; negotiate purpose when the task lacks learner significance. Do not interpret participation alone as competence.

For repeated force-only explanations, ask how acceleration changes when mass varies at the same force, and ask why. Connect the task to a learner-valued aim such as explaining bicycle braking while separating the teacher's conceptual target. Agree a reasoning rubric for the relation, an unassisted new braking example after feedback today and another next week. Those are local design choices; the source does not establish their optimality. If the target, purpose or starting explanation is missing, elicit it before declaring a misconception or choosing feedback.

For the third row, one design-based study in two community dance centres, with no comparison condition, observed youth sharing more elaborate, exploratory representations after a facilitator reframed the check as a conversation rather than a show, and drawing on movement, music and vernacular language to connect physics ideas to their own experience — Multimodal, culturally grounded formative assessment supports engagement and agency [~W]. That makes the reframing and the wider response modes a feasible move against evaluative threat, not a tested effect, and the researchers' existing relationships with the sites may be part of why it worked.

Evidence-bounded expectation

Kingston and Nash's corrected synthesis reports a science mean of 0.19 with 95% CI 0.06–0.31 across 17 effects; the erratum supersedes the original 0.09 science estimate: assessment-for-learning observation and comparison frame [+M]. It also removes an inappropriate comparison between two formative approaches. The synthesis supports a bounded program-level expectation and a need for better research; it cannot forecast this learner's response, distinguish these diagnostic branches or establish next-week transfer.

Before comparing another synthesis or applying an effect-size label, record assignment and baseline comparability, learner/context characteristics, exact treatment and control, outcome alignment and independence, score-range/standardization constraints, horizon and uncertainty. Preserve unknown fields. A correction removes one known error; it does not establish that all included estimates are comparable. Report raw magnitude alongside a justified context-sensitive interpretation; do not rank interventions or infer individual probabilities from these pooled values. The contingent step has its own limit: the evidence that support adjusted to each response beats fixed support comes mainly from one-to-one tutoring, and the review behind it reports that genuine contingency is scarce and hard across a whole class — Contingent scaffolding improves learning more than fixed or absent support [~S]; so in a class the act step is more often a decision about groups than about each learner. Contingent teaching can be learned: in an intervention study with 30 prevocational social-studies teachers working with small groups, those in a scaffolding programme showed more complete cycles of contingent teaching afterwards than the other teachers — Programme teachers showed more complete contingent-teaching cycles [+W]; that is a change in teacher behaviour, and the claim records no effect on students' learning. The 0.4–0.7 standard deviation gain often quoted for regular classroom formative assessment is a second-hand figure traced to Black and Wiliam's 1998 review — Regular classroom formative assessment can raise achievement by 0.4–0.7 SD [~W]; the later meta-analyses of controlled studies recorded here found smaller averages (d = .29 and .20, below), so plan from those. The formative loop's causal pathway remains a hypothesis to test against subsequent responses.

Fitting the design to a situation

Ask for five facts when the brief omits them: the length of the cycle (minutes, days, a unit), the item format of the check, whether feedback can follow every check, how many learners each teacher must read, and who or what reads the responses (the teacher, the learner, peers or software).

If the brief says… Then change… Basis
Unit-based school course (pre-k-12) Place a written check with feedback inside or between units (a medium cycle) rather than only at the end; expect a small overall gain Assessment for learning improves achievement [+M]
Teachers are new to formative assessment Add professional development on interpreting responses, or a computer-based formative system, to the rollout; both did better than other approaches in one synthesis. Where teachers support small groups, include practice in the diagnose–adjust–check cycle itself. Ask teachers to record the lesson changes the responses led to Assessment for learning improves achievement [+M]; Teachers changed lessons in response to formative information [+W] (teacher reports from one programme); Programme teachers showed more complete contingent-teaching cycles [+W] (teacher behaviour, one study; no student outcome recorded in the claim)
The teacher judges each learner's understanding from their work (marking, reading checks, conferences) Give the teacher the learner's performance on the task and base each judgment on it; leave out or set aside impressions such as interest, effort or behaviour, and note which piece of work each judgment rests on. Do not count on a short briefing about which cues matter to raise accuracy Teachers judge more accurately with diagnostic cues [+M]; Non-diagnostic cues hinder teacher judgment accuracy [+M]; Teachers fabricate student cues when judging from diagrams [~W]; A brief cue-diagnosticity intervention did not improve accuracy [~W] (Dutch secondary teachers judging text comprehension, in experimental tasks with anonymous students)
No feedback can follow the check (it is scored later, or not returned) Prefer multiple-choice items, which learners answer correctly far more often; switch to short-answer items once the correct answer will follow, where they came out slightly ahead Short-answer checks outperform multiple-choice mainly when feedback follows [~M] (laboratory studies reported in one review chapter)
The goal includes durable recall (facts, vocabulary, the steps of a procedure) Make the check a recall attempt, return the correct answer after each error rather than after every item, decide what counts as an error by scoring rather than by letting learners skip feedback when they feel sure, and judge the branch by a check a day or more later Retrieval practice improves long-term retention [+S]; Feedback enhances retrieval practice [+S]; Learners' correctness self-assessments are imperfect [~M]; Learners' feedback-skipping decisions diverge from the optimal rule [~M] (one word-pair experiment)
The check will use multiple-choice items Always return the correct answer; immediate or end-of-set feedback both work, and without it chosen lures are reproduced later Feedback after multiple-choice tests halves lure intrusions [~M]
Feedback after each check will be brief Make it say what to change in the task or the method, not only whether the answer was right Feedback improves learning, varying with its information [+M]; Feedback works best at task and process levels [~M]
There is time before a new topic Open with a short, low-stakes pretest on the coming material, followed by teaching that answers it; a pretest the teaching does not answer is wasted time Pretesting enhances learning [+M] (laboratory studies with undergraduates, retention up to a week; not a diagnostic pretest used to plan teaching)
The class is large and the work is extended (essays, projects) Add a peer-review round between draft and final version; freeform review is as effective as scripted Peer assessment improves performance [+M]; Structured criteria are not shown to improve peer assessment [~M]
Learners will self-assess before revising Give them a rubric and a modelled example of self-assessment first; do not accept untrained self-ratings as accurate, least of all from the learners who are furthest behind; expect a small gain in strategy use rather than in scores. Telling learners how far their self-assessment missed their actual score does not by itself improve their choice of the next task Self-assessment accuracy is low without training [~M]; Rubric-referenced self-assessment improved work [~M]; Prior knowledge is needed for accurate self-assessment [~M]; Self-assessment improves self-regulated learning [+M]; Self-assessment feedback reduced task-selection accuracy during learning [-M] (one experiment: choices during learning were less accurate with such feedback, d = 0.46, most of it negative)
Writing course with AI scoring available Run several draft–diagnose–revise cycles, with the teacher reviewing the diagnosis with learners; plan for more teacher review time, not less, and judge the arrangement after several cycles rather than the first, since in the one study the gap on diagnostic accuracy, feedback comprehension, revision quality and self-assessment accuracy widened across three cycles Diagnostic review gave larger writing gains [+W]; Diagnostic review class scored higher on process indicators [+W]; Group-by-cycle interactions for formative outcomes [+W] (two intact classes of 48, one quasi-experiment)
An LLM or other AI system will diagnose learners' work Use it where the work states the reasoning explicitly: ask learners to show their steps, and route vague or partial responses to a teacher. Keep feedback on reasoning and synthesis with a person LLM diagnosis drops sharply on vague evidence [~M] (a benchmark of 18 models on handwritten mathematics work, not a classroom test); High-level feedback should remain a human responsibility [~W] (a conceptual argument)
One-to-one tutoring Make each next hint depend on the last response, and let the tutor see the learner's work as it happens Contingent scaffolding improves learning [+M]; Teachers changed lessons in response to formative information [~W]
Teachers lack time to give feedback (many students, three-dimensional tasks) Choose fewer, aligned checks that each feed a planned decision, and share the reading with peers or software OpenSciEd formative practice limited by grading and time [~W]; Faculty time limits targeted formative feedback [~W]
A large lecture with a polling system, checking several times within one session Interrupt the lecture with multiple-choice questions, show the class's response distribution, and follow it at once with a related question or a discussion; decide the number of checks locally and record what changed after each Wireless keypads convert a passive audience to active learners [+W] (the authors' account of their own lectures, no comparison); the frequency is an untested proposal

The evidence by question

Further evidence, not yet read against this model

The seventeen claims restored here on 2026-10-01 were read against the model on 2026-10-15 and sit in the model above. Of the seventeen claims the orphan-claim linker added under ## Claims (read 2026-10-19), thirteen now sit in the interpretation step, the evidence-bounded expectation, the situation rows or the evidence-by-question list; these four were read and do not change the model:

Design Decisions

How large a gain should a designer plan for, and over what cycle?

  • Default: plan for a small gain, and run the cycle within or between units rather than only at the end. A meta-analysis of 33 controlled US K-12 studies found d = .29 overall, with larger effects for a medium cycle within or between units (d = .52) and for formal evidence such as written feedback on quizzes (d = .40); a second K-12 meta-analysis of 13 usable studies found a weighted mean of .20, against the .40–.70 often cited — Assessment for learning improves achievement [+S]
  • Tested with: US and other K-12 classrooms in mathematics, literacy, arts and science (effects .32 in English language arts, .17 in mathematics, .09 in science in Kingston & Nash 2011).
  • Not settled: the cycle-length and evidence-type figures are meta-regression results from one synthesis, not experiments that varied them. How often to check within a lesson has no wiki evidence.

What should the feedback after a check say?

What should the check ask, and does it need feedback?

Should evidence be elicited before teaching as well as after?

  • Default: a short pretest on the coming material, followed by the instruction, helps. Learners pretested on to-be-read material, answering essentially at floor, retained it better than those given equivalent extra study time, and failed retrieval attempts followed by feedback improved later learning — Pretesting enhances learning [+S]
  • Tested with: two laboratory studies, with educational passages, trivia facts and word pairs.
  • Not settled: whether a diagnostic pretest used to plan teaching gives the same benefit; the claim page's boundary conditions (alignment with instruction, low stakes) are stated without cited studies.

When should the feedback come?

Should learners assess their own work, or each other's?

Who interprets the evidence and adjusts, and what do teachers need?

Elements

Source verification and open tests

Kingston & Nash (2011), 2015 publisher erratum. Corrected content-area table checked; full primary-study assignment and measurement details remain unverified in this run.

The examples are local design instances, not additional observations from those studies. Test the proposed interpretation and activity branches with new learner responses; compare competing designs under matched conditions before claiming a causal or predictive advantage. Reassess the model when the observations disagree with it.

Claims

Key Sources

  • Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education: Principles, Policy & Practice, 5(1), 7–74. doi:10.1080/0969595980050102
  • Kingston, N., & Nash, B. (2011). Formative assessment: A meta-analysis and a call for research. Educational Measurement: Issues and Practice, 30(4), 28–37. doi:10.1111/j.1745-3992.2011.00220.x
  • Lee, H., Chung, H. Q., Zhang, Y., Abedi, J., & Warschauer, M. (2020). The effectiveness and features of formative assessment in US K-12 education: A systematic review. Applied Measurement in Education, 33(2), 124–140. doi:10.1080/08957347.2020.1732383
  • Wisniewski, B., Zierer, K., & Hattie, J. (2020). The power of feedback revisited: A meta-analysis of educational feedback research. Frontiers in Psychology, 10, 3087. doi:10.3389/fpsyg.2019.03087
  • Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. doi:10.1037/0033-2909.119.2.254
  • Panadero, E., Jonsson, A., & Botella, J. (2017). Effects of self-assessment on self-regulated learning and self-efficacy: Four meta-analyses. Educational Research Review, 22, 74–98. doi:10.1016/j.edurev.2017.08.004
  • Kostons, D., van Gog, T., & Paas, F. (2012). Training self-assessment and task-selection skills: A cognitive approach to improving self-regulated learning. Learning and Instruction, 22(2), 121–132. doi:10.1016/j.learninstruc.2011.08.004
  • Zell, E., & Krizan, Z. (2014). Do people have insight into their abilities? A metasynthesis. Perspectives on Psychological Science, 9(2), 111–125. doi:10.1177/1745691613518075
  • Double, K. S., McGrane, J. A., & Hopfenbeck, T. N. (2020). The impact of peer assessment on academic performance: A meta-analysis of control group studies. Educational Psychology Review, 32(2), 481–509. doi:10.1007/s10648-019-09510-3
  • Pratt, M. W., & Savoy-Levine, K. M. (1998). Contingent tutoring of long-division skills in fourth and fifth graders: Experimental tests of some hypotheses about scaffolding. Journal of Applied Developmental Psychology, 19(2), 287–304. doi:10.1016/s0193-3973(99)80041-0
  • van de Pol, J., Volman, M., & Beishuizen, J. (2010). Scaffolding in teacher–student interaction: A decade of research. Educational Psychology Review, 22(3), 271–296. doi:10.1007/s10648-010-9127-6
  • Richland, L. E., Kornell, N., & Kao, L. S. (2009). The pretesting effect: Do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied, 15(3), 243–257. doi:10.1037/a0016496
  • Kornell, N., Hays, M. J., & Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989–998. doi:10.1037/a0015729
  • Karpicke, J. D. (2017). Retrieval-Based Learning: A Decade of Progress. Learning and Memory: A Comprehensive Reference, 2nd edition, Volume 2. https://doi.org/10.1016/B978-0-12-809324-5.21055-9
  • Yao X-C and Fan L (2026) From automated scoring to learning diagnosis: a mechanism study of AI-supported formative assessment in English writing. Front. Psychol. 17:1905455. https://doi.org/10.3389/fpsyg.2026.1905455
  • van de Pol, J., Rowan, E., Janssen, E., & van Gog, T. (2024). Effects of availability of diagnostic and non-diagnostic cues on the accuracy of teachers' judgments of students' text comprehension. Metacognition and Learning. https://doi.org/10.1007/s11409-024-09383-0
  • van de Pol, J., van den Boom-Muilenburg, S. N., & van Gog, T. (2021). Exploring the relations between teachers' cue-utilization, monitoring and regulation of students' text learning. Metacognition and Learning. https://doi.org/10.1007/s11409-021-09268-6
  • van de Pol, J., Volman, M., Oort, F., & Beishuizen, J. (2014). Teacher scaffolding in small-group work: An intervention study. Journal of the Learning Sciences. https://doi.org/10.1080/10508406.2013.805300
  • Hays, M. J., Kornell, N., & Bjork, R. A. (2010). The costs and benefits of providing feedback during learning. Psychonomic Bulletin & Review. https://doi.org/10.3758/PBR.17.6.797
  • Raaijmakers, S. F., Baars, M., Paas, F., van Merriënboer, J. J. G., & van Gog, T. (2019). Effects of self-assessment feedback on self-assessment and task-selection accuracy. Metacognition and Learning. https://doi.org/10.1007/s11409-019-09189-5
  • Kim, Y., Jin, H., Doh, H., Kim, E., Jung, D., Kim, S., Choi, K., Son, J., & Kim, J. (2026). Benchmarking large language models for diagnosing students' cognitive skills from handwritten math work. arXiv. https://doi.org/10.48550/arxiv.2504.00843
  • Venetsanos, D. T. (2026). A tripartite feedback framework for AI-assisted assessment of complex reports in higher education. AI Educ., 2, 26. https://doi.org/10.3390/aieduc2030026
  • Shute, V. J., & Kim, Y. J. (2013). Formative and stealth assessment. In J. M. Spector et al. (Eds.), Handbook of Research on Educational Communications and Technology. Springer. https://doi.org/10.1007/978-1-4614-3185-5_25
Page metadata
Field Value
type pattern
id formative-assessment
aliases keypad-quizzing-interrupts-lecture-for-formative-assessment
title Formative Assessment
description A reusable elicit–interpret–act–reobserve policy, qualified by what a response can establish and by whether a relevant next action is available.
status review
grain_size lesson
author formative assessment tradition
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 3 — black-1998, clark-2012, burnstein-2007

Frontmatter is the page's machine-readable record — see the Schema & Guide.