Should history be built on content knowledge or on generic historical-thinking skills?

Verdict

Content-rich history teaching works, on history. That is the finding, it is replicated across eight randomised trials, and it is smaller than what its advocates claim and larger than what its critics concede.

The reason this topic matters beyond history is that history is where the archive's most attractive hypothesis — knowledge is the durable lever for reading comprehension — should cash out, because history is the largest block of teachable general knowledge in a school day. The trials say it cashes out only over years, and not at all within a semester.

Meanwhile the rival organising principle for the subject, generic historical thinking taught as a transferable skill, has no adequately controlled positive result at all. Its best studies use volunteer teachers or no comparison group; the one that randomly selected which teachers got the training found nothing on its own aligned writing measure.

What the evidence shows

Cabell 2025 is the measure-alignment experiment this archive has been waiting for, and it is brutal. Same children, same semester, same construct, two instruments. Curriculum-aligned social studies knowledge moved 0.93 SD. The Woodcock-Johnson Social Studies subtest moved 0.00 SD, p = .999. The one apparently significant standardized effect — WJ Picture Vocabulary at 0.09 — collapsed to 0.01 (p = .702) once the authors deleted three items whose words (canoe, moccasins, loom) the curriculum explicitly taught. METHODOLOGY's nominal correction for researcher-designed measures is 2×. On a one-semester dose here it is not a discount; it is the entire effect.

Eight trials, one significant standardized reading result, and it did not replicate. Vaughn 2013's Gates-MacGinitie 0.21 is the single positive. The same team's own direct replication two years later returned nulls on both comprehension measures, the grade-11 version returned nulls, and the 48-school effectiveness trial returned 0.14 ns. That is a failed replication inside one research programme, which under this archive's replication rule caps how far the transfer claim can be pushed.

The only doses that have ever moved standardized reading are measured in years. Kim's three-year spiral gives 0.11-0.12; Grissmer's six-year Core Knowledge lottery gives ITT 0.24. This is consistent with the background-knowledge verdict that knowledge is a slow lever — but note the awkward detail: Kim's trial moved mathematics by 0.12-0.16, the same amount as reading. No knowledge-transfer mechanism predicts that. Something more general — engagement, teacher effort, time on academic text — is a live alternative explanation.

The skills side has no controlled positive result. Reisman's is the famous one and it is a non-randomised, developer-led, five-school study with unmodelled clustering whose author attributes the reading effect to time-on-task rather than to disciplinary transfer. De La Paz 2014 reports pre-post growth with volunteer teachers and no credible counterfactual. Steiss 2026 is the one design that randomly selected which teachers received the training, and it found nothing — with the positive d = 0.29 appearing only among the self-selected teachers who came back for a second year. And Nokes' 2×2 factorial separates the two ingredients directly: giving students multiple documents raised content knowledge; explicitly teaching them the sourcing heuristics did not.

Hereditarian-lens assessment

Risk: low, and this is one of the better-identified topics in the archive. Two grade-A school-randomised trials and a kindergarten admissions lottery carry the verdict; within each, treated and untreated children have the same expected distribution of ability-relevant alleles.

Where the lens still does work:

• The gap-closing claim fails under the lens, and it fails in the trials too. Content-rich instruction compounds advantage rather than closing gaps. Cabell's effect on aligned knowledge ran 0.53 at −1 SD entering vocabulary, 0.92 at the mean and 1.31 at +1 SD; English learners 0.58 against native speakers' 0.99. Wanzek 2014's low-pretest students got no benefit at all. Reisman's struggling readers gained facts but not comprehension. This is exactly the pattern METHODOLOGY predicts when instruction removes an environmental bottleneck: the children with more prior knowledge and vocabulary extract more from more knowledge.
• Tyner & Kabourek is the most-cited source in this area and it is observational. Time allocation is a school and teacher decision correlated with everything. Its own internal check is the strongest argument against it: science is content-rich too, and 30 extra minutes of daily science predicts −0.01 SD. A background-knowledge account predicts a positive science coefficient. The authors' reverse-causation robustness check is good and rules out one confound; it does not create identification.
• Recht & Leslie is an existence proof, not a lever. High-knowledge poor readers out-recall low-knowledge good readers on a topic-matched passage. That is near transfer inside a known domain, from n = 64, unreplicated in 35 years. It shows knowledge matters; it does not show that teaching knowledge raises general reading, and the trials above are the direct test.

Boundaries & what critics say

• Almost every positive number here is on a researcher-designed, treatment-aligned measure, produced by the people who built the curriculum. The PACT trials are developer-led; so is Reisman; so is De La Paz. The two independent standardized results in the whole cluster are Grissmer's state tests and Cabell's Woodcock-Johnson, and they point in opposite directions at different doses.
• PACT's effectiveness trial produced a larger knowledge effect (0.46) than its developer-supported efficacy trials (0.31-0.36) — a clean counterexample to this archive's efficacy-to-effectiveness decay through-line. Two caveats: the outcome is the developers' own aligned measure throughout, and PACT classes were 0.13 SD ahead at baseline.
• Grissmer 2023 remains an unpublished working paper three years on, with no critique and no replication, and it introduces a novel bias correction of the authors' own devising for "two previously undiscovered sources of bias inherent in kindergarten lotteries." The archive's highest-value curriculum result therefore depends on a non-standard correction nobody has independently vetted.
• Chronology versus thematic sequencing is unevidenced. There is no causal literature on how to order a history curriculum. Strong opinions exist; trials do not.
• No causal evaluation of England's 2014 knowledge-led national curriculum exists. A whole national curriculum was rewritten around this question and left no identification strategy behind.
• A defender of historical thinking would say the trials test the wrong outcome — that disciplinary reasoning is worth teaching for its own sake, not as a reading intervention. That is a fair position, and it is the position historical source work evaluates on its own terms.

Practical guidance

• Build history as a coherent, cumulative, text-heavy content curriculum, and run it for years. Every measurable transfer effect in this topic required three years or more.
• Justify it on history knowledge. That is the outcome it reliably delivers, at roughly two days of teacher PD and a set of units, and it is a worthy outcome in its own right.
• Do not promise a reading-comprehension gain inside a year. Eight trials; one significant result; it did not replicate. If a programme is sold to you on a one-year comprehension effect, ask which test.
• Use multiple documents rather than a textbook — it raised content knowledge in the one factorial that isolated it — but do not expect the explicit heuristic instruction bolted on top to add content learning.
• Do not buy generic historical-thinking PD as the organising principle. The best-identified study found nothing on its own aligned measure, and the apparent effect lived entirely in teacher self-selection.
• Plan for the strong to gain more. Knowledge-rich instruction compounds prior knowledge. If closing gaps is the goal, this is not the lever, and pretending otherwise sets the programme up to be judged on a criterion it will fail.
• Do not cut science to buy social-studies minutes on the strength of Tyner & Kabourek. The same paper finds science time predicting nothing, which should make anyone cautious about reading its social-studies coefficient causally.

Open questions

• Is Grissmer replicable? It is the single most valuable result in the archive's curriculum area, it is unpublished, uncritiqued, unreplicated, and rests on a bespoke bias correction. Replicating it is the highest-value study in this topic.
• Why did Kim's content curriculum move mathematics as much as reading? Either the mechanism is not knowledge transfer, or knowledge transfer is not specific — and both readings change the recommendation.
• Does the knowledge effect ever show on a standardized measure at a dose under three years? Nothing has, and nobody has mapped the dose–response curve between one semester (0.00) and six years (0.24).
• Chronological or thematic? Unevidenced by any causal design.
• What happens to low-prior-knowledge students? Three studies find they gain least or not at all, and nobody has tested a version designed for them.