The dementia claim we could market, and will not

Here is a sentence I could legally publish: a 2020 meta-analysis in Neuropsychology Review found that bilinguals reported the first symptoms of Alzheimer's disease 4.7 years later than monolinguals, with a 95% confidence interval of 3.3 to 6.1 years.

That sentence is true. Brini and colleagues reviewed 16 cross-sectional studies covering 4,671 people, and the 4.7 is what came out of the 7 of them that reported age at symptom onset.

We sell a language app. Putting that line above a download button would cost me nothing and probably convert. I'm not going to, and this post is the reasoning, with all the numbers, so you can check whether I'm being honest or performing it.

The ad I could run

The worst version first, because vague self-criticism is cheap and a specific draft is not.

Every clause survives a fact-check. The number is real, it's from a peer-reviewed synthesis, and I haven't said the app causes anything. The verb "arrived" is passive on purpose, which is how the genre works.

The deception is entirely in what sits between the two sentences. Sentence one is about people who grew up speaking two languages. Sentence two is about you, at 38, opening an app on the train. Nothing in the literature connects them, and the ad works precisely because a reader assumes something does.

The question worth answering is whether the number is about anyone who would ever download this.

What the meta-analysis actually reports

Brini's team reports five results. The marketing version quotes one.

Brini et al. 2020 · five results, one quoted

Outcome	Result	Reads as
Age at AD symptom onset	4.7 years later95% CI 3.3 to 6.1, k = 7	Significant, and the one everyone prints
Age at dementia diagnosis	3.3 years later95% CI 1.7 to 4.9, k = 13	Significant, and smaller
Age at MCI diagnosis	3.2 years later95% CI −3.4 to 9.7, k = 4	Interval crosses zero. Nothing shown
Disease severity at diagnosis	Hedges' g = 0.0595% CI −0.13 to 0.24, k = 12	No difference
Risk of developing dementia	OR 0.8995% CI 0.68 to 1.16, k = 5, p = 0.289	No reduction in risk

Now the number that decided this for me. Alongside the confidence interval for age at diagnosis, the authors report a 95% prediction interval of −1.9 to 8.5 years.

A confidence interval describes how precisely you have pinned down an average. A prediction interval describes the range a new study would be expected to land in.

Brini's team spells out what theirs means: "we could expect that in some 95% of all populations comparable to those in this meta-analysis..., the association between bilingualism and dementia may be strong, while in others, this association may be absent or may even tend to be in the opposite direction." Roughly 75% of the observed variance across studies reflected real differences between them rather than sampling noise.

Negative 1.9 is inside the range. A next study finding bilinguals diagnosed almost two years earlier would sit comfortably within what this evidence base predicts. That is the shape of the 3.3.

Delayed onset, unchanged risk

The title says it, and the people quoting it rarely repeat the second half: "Bilingualism Is Associated with a Delayed Onset of Dementia but Not with a Lower Risk of Developing it."

Across six longitudinal prospective studies with 4,227 participants, five of which produced a usable risk estimate, bilingualism was not associated with a reduced risk of dementia. The odds ratio was 0.89, interval 0.68 to 1.16, two-tailed p of 0.289.

Split the two designs and the pattern gets sharper. Van den Noort and colleagues, reviewing 34 studies in Behavioral Sciences in 2019, put it directly: "In most prospective studies, no association between bilingualism and the delay of the onset of dementia was found while in the majority of retrospective studies an association between bilingualism and the delay of the onset of dementia seemed to exist."

It lists the failures individually. Crane 2010 found neither spoken nor written Japanese in midlife affected late-life cognitive decline. Sanders 2012 found no association between non-native English speakers and dementia or AD. Zahodne 2014 found bilinguals had better baseline memory and executive function, and no protective effect. Lawton 2015 found no difference in mean age of dementia diagnosis at all. Kowoll 2015 is recorded there as showing no cognitive-reserve effect, though that is the reviewers' verdict rather than the study's own headline. The review's conclusion, having found evidence on both sides: "to date, no firm conclusions can be drawn."

Retrospective studies read the charts of people who already walked into a memory clinic. Prospective studies enrol healthy people and wait. When those two designs disagree this cleanly, the usual suspect is who ends up in the clinic and when, which is a fact about referral.

The other side deserves its strongest version. Anderson, Hawrylewicz and Grundy's 2020 meta-analysis in Psychonomic Bulletin & Review argues the discrepancy comes from "conflating incidence rates of dementia and the age at which the symptoms first appear", finds a moderate effect on age of onset (Cohen's d = 0.32) against weaker evidence on incidence (d = 0.10), and states their results "cannot be explained by SES, education, or publication bias." Two teams, two readings, same split. Their disagreement is about which side of it matters, and both agree the incidence effect is the weak one.

The word every study uses, and we would drop

Now the sentence that ends the argument, and it is in the paper the whole field is built on.

Bialystok, Craik and Freedman, Neuropsychologia 2007, 45(2), 459-464. Their abstract opens: "This study examined the effect of lifelong bilingualism on maintaining cognitive functioning and delaying the onset of symptoms of dementia in old age."

Lifelong. The sample was 184 memory-clinic patients, 91 monolingual and 93 bilingual, drawn from 228 referrals. The bilingual group spoke 25 languages between them, most commonly Polish, Yiddish, German, Romanian and Hungarian, which tells you what kind of life produced them.

One correction while we are here, since this is the number that travels. The abstract says "4 years". York University's own release on the study gives the group means: 71.4 years for the monolingual group, 75.5 for the bilingual. That difference is 4.1 years. Figures between 4 and 5.5 circulate for this literature, some of them from later cohorts rather than from this paper, so if you are citing Bialystok 2007 specifically, 4.1 is the arithmetic its own institution published.

Read that group against a Verbamor customer. Ours is an adult, monolingual for three or four decades, studying Spanish a few hours a week, on a device, in a country where they'll keep speaking English at work tomorrow. There's no study in this literature measuring that person's dementia outcomes, because a study like that would need to start now and report in 2065.

So the inference an ad needs runs: lifelong childhood or immigration bilingualism is associated with later symptom onset, therefore adult app study delays your symptoms. That step is doing all the persuasive work and none of the evidence supports it. It is the same error as the age 17.4 finding, where a real statistical result gets stretched over a population it never measured.

What happens to the studies that find nothing

One more reason to hold this evidence at arm's length, about the field rather than any single result.

De Bruin, Treccani and Della Sala went looking in 2015 for what happens to bilingualism studies that do not find an advantage. They took conference abstracts from 1999 to 2012 on bilingualism and executive control and checked which ones later reached publication. Their finding, verbatim: "Studies with results fully supporting the bilingual-advantage theory were most likely to be published, followed by studies with mixed results. Studies challenging the bilingual advantage were published the least." They add that this "was not due to differences in sample size, tests used, or statistical power", and that a funnel-plot test showed asymmetry too.

Then there is what happens with enough people to settle it. Nichols and colleagues at Western University tested 11,041 people on 12 executive tasks, in Psychological Science in 2020. Bilinguals beat monolinguals on one test. Monolinguals beat bilinguals on four. Every difference vanished once the groups were matched on age, education, socioeconomic status and language. The one surviving unmatched advantage, on Digit Span, was a coefficient of 0.05, "one twentieth of 1 standard deviation", with a Bayes factor supporting the null.

The detail I have not seen quoted anywhere sits deeper in that paper, and it is the most useful thing in it for reading claims like mine. Nichols ran 64 model variants per task. On Double Trouble, a Stroop variant and "one of the tasks most frequently used in bilingual-advantage research", the unmatched sample produced a significant result "50% of the time, depending on the set of regressors included."

A coin flip, decided by which covariates an author picked. Nichols also tested whether bilingualism buffers age-related decline, and the group-by-age interaction showed "no relationship" in either sample.

Executive function isn't dementia, and I won't pretend one refutes the other. What it does establish is that the proposed mechanism, the lifelong executive workout that cognitive reserve is supposed to be built from, doesn't show up at population scale. A claim needs both its outcome and its mechanism to hold. This one has a contested outcome resting on a mechanism that 11,041 people couldn't detect.

What we will say instead

So the rule here, written down so someone can hold me to it: Verbamor will not claim, imply, or arrange words to suggest that using this app affects anyone's risk of dementia, Alzheimer's, cognitive decline, or brain aging. Not in an ad, not in an App Store description, not in a blog post that quotes a number and lets the reader finish the sentence.

Three things follow. We don't quote the 4.7 or the 4.1 in any promotional context. We don't use "cognitive reserve" as a product benefit. And when we write about this research, as in the brain-aging study of 86,149 people, the limits go in the same paragraph as the finding rather than in a footnote at the bottom.

Honesty about the cost: I think this claim would work. Health framing converts, the citation is real, and a competitor running it would face no regulator that I know of. Publishing the reasoning is partly how I make the refusal expensive to walk back later, which is a less noble motive than it sounds and I'd rather say it than not.

What Verbamor does is narrow enough to describe without any of this. You record a lesson or import audio, it pulls out the words and phrases you met, and schedules them so they are still there in six months instead of six days. That is an ordinary memory problem, the one the forgetting curve describes. The honest benefit is that you remember what your tutor taught you, which is a smaller promise than a healthier brain and one I can show you evidence for.

If a reason to keep going is what you actually came for, starting at 40 has better-evidenced arguments behind it than this one, and none of them require a claim about your neurology in 2065.

If you have seen a language app make this claim, the test is one word. Check whether the study behind it says "lifelong". They almost all do.

---

Sources

Every load-bearing claim Verbamor makes is traced to its paper on the research page.