Feedback and formative assessment

Verdict

Feedback is the most double-edged tool in teaching. The famous numbers — Hattie's d=0.79, Black & Wiliam's 0.4–0.7 for formative assessment — are both dead: the first was cut to 0.48 by Hattie's own coauthored reanalysis and falls to g≈0.07–0.17 in the only risk-of-bias-graded school-age synthesis; the second was never computed from anything and deflates to +0.09 when the developer's own program was tested at scale on national exams. Meanwhile more than a third of measured feedback effects are negative — the only major method with a stable backfire rate.

What survives is a precise and useful moderator structure, stable across 30 years: feedback that carries task information helps; feedback that directs attention to the self hurts. That structure — not any headline effect size — is the actionable finding. The verdict is mixed because the method genuinely runs both directions; confidence is high because multiple grade-A/B at-scale RCTs and the rigor-graded synthesis all converge on the same picture.

What the evidence shows

Feedback per se:

The rigor gradient is the story: serious-bias studies g=0.62 → moderate 0.20 → lowest-bias 0.07. Feedback in schools is real but small, positive mainly in early primary and math, and ~zero for secondary students. The backfire conditions are consistent: praise and person-comments, discouragement, ego-involving normative grades, and feedback on very complex tasks.

Formative assessment as a program:

The trajectory tracks rigor perfectly: 0.4–0.7 (asserted) → 0.20 (weak meta) → 0.09 (at-scale RCT) → ~0 (US commercial variants). The one mildly working variant (EEF's) cost £1.20/pupil/yr — so even 0.09 would be excellent value if real — and its distinguishing ingredient was monthly peer accountability, not the assessment techniques themselves.

Hereditarian-lens assessment

Risk: low — the verdict rests on randomized designs with national-exam outcomes. Two lens-relevant notes: (1) Butler's exception — high achievers anticipating more grades kept performing — plus the ego-threat moderator suggest grading regimes interact with stable (partly heritable) traits, plausibly harming fragile/low-ability students most; (2) the praise literature's perverse targeting loop (adults give the most inflated praise to exactly the low-self-esteem children it harms) is direction-plausible but replication-fragile (Mueller & Dweck vs Li & Bates) — treat person-praise avoidance as a free "do not," not a quantified lever.

What critics say / limits

• Added 2026-07-30, with a verification failure recorded honestly. A citation-graph audit flagged Mandouit & Hattie 2023, "Revisiting 'The Power of Feedback' from the perspective of the learner" — Hattie re-examining his own 2007 feed-up/feed-back/feed-forward model from the receiving end. We could not read it. It is paywalled in Learning and Instruction with no open-access or repository copy and no abstract deposited with Crossref or Semantic Scholar; the reported response frequencies ("how to improve" 51%, error flagging 50%, "done well" 41%, feed-forward 35%) come from secondary reporting and are not verified against the paper, and we could not establish its sample size or design. It is recorded in the archive as disposition: unresolved — a known unread lead, deliberately not cited in the evidence table above, because a paper we could not read cannot be evidence. It is kept rather than dropped so the gap stays visible. It changes nothing here, and would not even if verified: it measures what students say they want, which is a preference, not an effect. Its only interest is directional corroboration from a different evidence type — students ask for exactly the task-and-improvement information that the Kluger & DeNisi moderator structure says works, and not for the person-evaluation that backfires. Note the irony worth keeping in view: this is the same author whose d = 0.79 headline was cut to 0.48 by his own coauthored reanalysis, so a Hattie paper agreeing with us is not independent support.
• The at-scale nulls (CASL, MAP) were powered for ~0.25, so effects of ~0.1 can't be excluded — "null at scale" and "+0.09 at scale" are statistically compatible.
• McMillan et al. argue Kingston & Nash's 0.20 is not credible as an underestimate — but name no alternative number; uncertainty runs both ways within ~0.1–0.3.
• Durability is essentially unmeasured: virtually no feedback study has a ≥1-year follow-up.

Practical guidance

• Give task information: what was wrong, why, and what to do next. The three reliable ingredients: correct-solution information, progress-relative-to-prior-work, and high information content.
• Never person-evaluate: no intelligence praise, no ego-involving comparisons, no discouragement. Praise the work and the process sparingly if at all.
• Comments without grades on formative work. Adding the grade neutralizes the comment. Save grades for genuinely summative moments.
• Slash marking volume. Written marking consumes hours weekly with no causal evidence; mark less, mark better, and give class time to respond to it.
• Buy the cheap version of formative assessment (exit tickets, whiteboards, quizzing — which is really retrieval practice) and expect ~0.1, not 0.5. Skip commercial interim-testing data platforms; they change nothing.

Open questions

• Whether the EFA +0.09 is a real effect (p=.09, unreplicated) or noise.
• Whether comment-only marking's benefit survives at scale (the biggest marking RCT lost its outcome arm to COVID).