Handwriting vs Typing: The Famous Study Was Corrected, the Replications Failed, and the Meta-Analysis Still Finds an Effect

Three things are true about this question at once, and almost every article you will find asserts one of them and hides the other two.

1. The famous study everyone cites had its statistics corrected four years after publication, and the authors themselves say the informative effect sizes are much smaller than they printed.
2. Two well-powered direct replications failed to find the learning advantage.
3. A 2024 meta-analysis of 24 studies and 3,005 people still finds a real advantage for handwriting, g = 0.248.

These do not contradict each other. Working out why is the most useful thing this article can do, because the reason is a general lesson about how to read replication failures.

The famous study

Mueller and Oppenheimer (2014), “The Pen Is Mightier Than the Keyboard,” in Psychological Science, ran three studies. Students watched lectures and took notes by hand or on a laptop, then answered factual and conceptual questions.

The reported pattern: no difference on factual questions, a handwriting advantage on conceptual questions, and a mechanism - laptop users transcribed more verbatim and wrote more words, which the authors argued reflected shallower processing.

The mechanism findings were large and consistent. Verbatim overlap, laptop against longhand: 14.6% vs 8.8% in Study 1, t(63) = -3.77, p < .001, d = 0.94; 12.11% vs 6.9% in Study 2, d = 1.12; 11.6% vs 4.2% in Study 3, d = 1.68.

The corrigendum

Here is what almost nobody citing this study mentions.

Psychological Science published a four-page corrigendum in 2018 - volume 29, issue 9, pages 1565-1568.

It corrected several categories of error. The degrees of freedom printed with the headline results were the interaction error degrees of freedom rather than the condition error degrees of freedom; the corrected values are around 4, not 55 and 89, “due to the use of mixed and random-effects analyses.” Effect sizes “were reported as η_p² values when they were actually η² values.”

And the sentence that matters most, from the authors: “more informative estimates of effect sizes would be much smaller, ranging from .03 to .11.”

After correction, the flagship Study 1 conceptual result reads F(1, 4.09) = 8.05, p = .046. Study 2 reads F(1, 6.60) = 19.65, p = .003. Study 3’s key contrast becomes t(105) = 2.86, p = .005, d = 0.56.

The authors state the results “remain qualitatively the same.” That is a fair characterisation of direction. It is not a fair characterisation of strength: the headline finding moved from an apparently robust result to p = .046 on roughly four error degrees of freedom, with the authors themselves recommending much smaller effect-size estimates.

If you quote the 2014 numbers - and you will find them quoted everywhere - you must quote the corrigendum with them.

The replications

Morehead, Dunlosky and Rawson (2019), in Educational Psychology Review, ran two experiments with 193 and 222 participants. The handwriting advantage did not replicate. The sharper fact from that paper: in Experiment 2, performance did not reliably differ even from a group that took no notes at all.

Urry and colleagues (2021), in Psychological Science, ran a preregistered direct replication. This study is widely misdescribed, so here is what it actually was: 88 authors, a single site - Tufts University’s Medford campus - 145 recruited and 142 analysed. It was a class-linked project in which the undergraduate authors collected the data. It was not a many-labs or multi-site replication.

Its results: factual g = -0.08, 95% CI [-0.41, 0.25]; conceptual g = -0.13, 95% CI [-0.45, 0.20]. Null in both directions.

Urry and colleagues also ran a mini meta-analysis of eight studies of the same paradigm: total g = 0.04, 95% CI [-0.13, 0.20], z = 0.46, p = .645.

But they reproduced the mechanism decisively: word count g = -0.91, verbatim overlap g = -0.78. Typists really do write more and transcribe more.

The reversal

Flanigan, Wheeler, Colliot, Lu and Kiewra (2024), in Educational Psychology Review, meta-analysed 24 studies across 21 articles, 49 effect sizes, 3,005 participants.

Pooled effect favouring handwriting: g = 0.248, p < .001, 95% CI [0.181, 0.315].

Heterogeneity: Q(48) = 128.95, p < .001, I² = 64.39%. Publication bias: Egger’s test t(47) = 0.550, p = 0.585; published versus unpublished Q(1) = 1.831, p = 0.176 - no evidence of bias.

And the moderator that explains most of the story. Whether students reviewed their notes: Q(2) = 7.94, p = 0.019.

Note review	Hedges’ g	95% CI
Yes	0.421	[0.287, 0.554]
No	0.208	[0.124, 0.292]
Optional	0.195	[0.067, 0.323]

Note that the no-review subgroup is still significant. Review amplifies the effect; it does not create it.

Typing wins decisively on one thing: note volume, g = 0.919, p < 0.001, 95% CI [0.679, 1.160]. Typists capture far more.

How all of this can be true

Four things reconcile them, and each is stated from the papers themselves.

They estimate different things. The Urry mini meta-analysis pools eight near-direct replications of one paradigm - video lecture, immediate test, no review period. Flanigan pools 24 studies across classroom and laboratory, immediate and delayed testing, review and no review. A null on one paradigm is not a null on the construct.

The intervals nearly touch. This is imprecision, not contradiction. Urry’s pooled total is g = 0.04 with an upper bound of 0.20. Flanigan’s no-review subgroup is g = 0.208. The replication set’s ceiling sits essentially on the meta-analytic estimate. Urry had roughly 68 and 74 participants per condition; detecting an effect of g ≈ 0.25 with 80% power needs something closer to 250 per arm. The direct replications are not evidence of zero. They are underpowered to see 0.2.

The replications used the weakest cell of the design. Formal review produces g = 0.421; no review produces g = 0.208. The replications tested the no-review, immediate-test condition.

The mechanism replicates; the specific claim does not. Urry reproduced the transcription difference decisively. But Flanigan found that achievement type - factual versus conceptual - was not a moderator, Q(2) = 0.63, p = .730. So the pooled literature supports a small general handwriting advantage while failing to support Mueller and Oppenheimer’s specific claim that the benefit is concentrated in conceptual understanding. Typists transcribe more; that simply does not cash out as the predicted selective conceptual deficit.

The 2024 brain study that measured no memory at all

In 2024 a study circulated very widely as evidence that handwriting improves memory. It did not measure memory.

Van der Weel and Van der Meer (2024), in Frontiers in Psychology, recorded EEG from 36 analysed participants while they handwrote or typed. The outcomes were EEG connectivity, coherence, event-related synchronisation and band power. There was no memory test, no recall measure and no learning outcome of any kind. Participants wrote words they already knew, from Pictionary cards, with no requirement to encode anything.

The authors nonetheless wrote: “We urge that children, from an early age, must be exposed to handwriting activities in school.”

Pinet and Longcamp (2025) published a General Commentary in the same journal. Their objection, verbatim: “the protocol did not include any learning. Participants repeatedly wrote well-known words without any requirement for memory encoding, preventing any conclusions in terms of learning.” They also note that the analysis was performed “exclusively on the difference between handwriting and typing,” which “precludes any conclusion from being drawn on either condition on its own,” that typing was constrained to a single right index finger - artificially suppressing typing-related connectivity - and that the typing condition lacked screen feedback while handwriting had normal visual feedback.

We found no reply from the original authors.

Any sentence of the form “a 2024 brain study found handwriting improves memory” is false. We are naming it because it is the most widely circulated claim on this subject.

What this leaves you with

Handwriting confers a modest advantage: about g = 0.25. That is small. It is real, it survives publication-bias testing, and it is not what the popular version claims.

The advantage roughly doubles if you actually review your notes. That is the largest lever in the data, and it is about what you do afterwards rather than what you write with.

Typing captures far more material - g = 0.919 on note volume. If your goal is a complete record, type.

The mechanism story is only half right. Typists do transcribe more verbatim. That difference does not produce the specific conceptual-over-factual deficit the original study proposed.

Do not cite the 2014 statistics without the 2018 corrigendum.

This article is about the note-taking method. The separate question of what laptops do in a lecture hall - including the effect on other students - is covered in laptops in the classroom, which states that the handwriting advantage did not survive direct replication. That remains accurate: Morehead and Urry are direct replications and they did fail. The meta-analysis discussed here is not a direct replication, and both findings stand.

For what reading on a screen does to comprehension, see screen reading vs paper. For a disfluency intervention that did not survive testing, see Sans Forgetica.

Limitations of this article

We read the Morehead sample sizes in the Flanigan meta-analysis data table, not in the Morehead paper’s own method section. Note that Flanigan lists Urry as N = 145 - the recruited figure - where the paper analysed 142.

Flanigan’s meta-analysis does not engage Morehead or Urry as replication failures; it characterises the field as inconsistent. It also includes no moderator for laboratory versus classroom and none for study quality or randomisation, and its heterogeneity is substantial at I² = 64.39%.

One small discrepancy inside the corrigendum: its explanatory text gives the corrected Study 1 error degrees of freedom as 4.01 while the corrected result line reads F(1, 4.09). We print the result line as published.

A third study comparing longhand, keyboard and stylus with 97 participants found no significant effect, F(2, 268) = 1.294, p = .276. We could not resolve its journal, volume and DOI to our citation standard and have therefore not listed it as a source or leaned on it.

No preregistered direct replication has been published since the 2024 meta-analysis.