Does dual-language immersion work, and is the gain in the second language or in English?

Verdict

Immersion is the rare education topic where a lottery exists, and the lottery result is good — but it answers a different question from the one parents are asking.

Parents choose immersion so their children will speak another language. The causal evidence measures English reading, English maths and English science. It finds that immersion students do slightly better in English than lottery losers and no worse in the content subjects — which demolishes the standard objection to immersion (that instructional time spent in another language is stolen from the curriculum) without confirming the standard promise (that the children come out bilingual).

So: moderate-support, at medium confidence. The English result is well identified and unreplicated. The second-language result does not exist.

What the evidence shows

The lottery is the entire epistemic basis of this topic and it is worth being precise about why. Immersion enrolment is the most self-selected decision in education: the families who apply are more educated, more motivated, more likely to be bilingual already, and more likely to move house for a school. Every non-lottery immersion estimate is contaminated by that, and by a second filter nobody talks about — weaker students differentially leave immersion, so the surviving immersion cohort has been selected twice. Steele's design removes both by comparing applicants who won a lottery to applicants who lost one. Note what happens to the effect size when it does: the observational syntheses report +0.33 to +0.45 on English reading; the lottery reports +0.13 to +0.22. That is the standard deflation, right on schedule.

The null on maths and science is the most useful finding in the topic. The objection to immersion has always been arithmetic: an hour taught in Mandarin is an hour not taught in English, and content has to suffer. Under randomisation, it does not. Nothing is lost. This is a genuinely surprising result and it is the one a founder should lean on.

The reclassification result is a trap for anyone evaluating on a short horizon. Umansky & Reardon found that two-language students are reclassified as English-proficient more slowly through elementary school and then finish ahead by the end of high school on reclassification, English proficiency and academic threshold passage. An evaluation stopping at grade 3 concludes immersion damages English acquisition. The same students, followed to grade 12, say the opposite. Sign flips within the same children are much harder for selection to explain than level differences, which is why this is the most informative observational study in the cluster.

Nobody has measured the second language under randomisation. This is not a small gap; it is the entire purpose of the programme. The best available answer comes from Genesee's review of a self-selected literature: immersion students reach high functional proficiency, clearly above conventional foreign-language classes, with receptive skills approaching native levels and productive skills that stay measurably non-native. Usable fluency, not a native speaker. That is a reasonable expectation to set — but it is set on grade-D evidence.

Hereditarian-lens assessment

Risk: low for the verdict, and this is one of the few topics in the archive that earns it honestly. The verdict rests on an oversubscription lottery, where winners and losers are drawn from the same applicant pool. Whatever heritable ability and family motivation the applicants carry, both arms carry it equally. Valentino & Reardon's preference-matching design attacks the same confound from a different angle and reaches the same answer, which is the closest thing this cluster has to replication.

Where the lens still bites:

• Everything except the lottery is presumptively confounded, including the entire Canadian French immersion literature that gave immersion its reputation. Families choose it; weaker students leave it. Both filters push the estimate up.
• Rolstad 2005 declines on principle to weight by study quality, which is a defensible meta-analytic position and a fatal one here. Its pooled numbers inherit every confound in the underlying corpus and are graded D on that basis, not on the topic.
• Bruton's CLIL critique is the same argument in European clothes. CLIL students are higher-attaining and better-supported before they start and often get more hours; the literature then credits the method. There is no CLIL lottery anywhere.
• The lottery's external validity is bounded by who applies. A lottery makes the comparison fair; it does not make the sample general. These are the children of families who queued for an oversubscribed immersion programme.

Boundaries & what critics say

• One lottery, one district, one decade. Portland is the whole grade-B evidence base. No second immersion lottery of comparable scale has been published, which caps confidence at medium under the single-source rule.
• "Immersion raises English achievement" is a headline that hides the mechanism. The effect is 0.13-0.22 SD across five to nine years of schooling — real, cheap in the sense that the programme was going to be staffed anyway, and small. It is not a transformation, and the study cannot say whether it comes from metalinguistic awareness, from better teachers self-selecting into immersion strands, or from peer composition among lottery winners.
• The bilingual-education literature is about English learners, not about elective immersion. Slavin, Cheung, Rolstad and Valentino are answering "what is the best way to educate a child who does not speak English?", which is a different question from "should an English-speaking child learn Mandarin?". Applying one to the other is common and wrong; they are included here because they are the only large bodies of evidence that bear on whether partner-language instruction costs English, and on that specific question they converge with the lottery.
• Cheung & Slavin's own update is the best argument against over-reading their earlier number. Their 2012 synthesis reports that the largest and longest evaluations — including the only multiyear randomised evaluation of transitional bilingual education — found no difference by the end of elementary school, and concludes that quality of instruction matters more than language of instruction. That is the deflation curve happening inside a single research programme.
• Immersion is expensive and hard to staff. Nothing in the evidence base addresses whether a school can find enough teachers fluent enough to teach fractions in the partner language, which is the binding constraint in practice.

Practical guidance

• Adopt immersion as a multi-year, whole-strand structure or not at all. Every effect in this topic is measured after five to nine years. There is no evidence for a light or partial version.
• Use the null to answer the objection. Under randomisation, maths and science did not suffer. That is the single most useful sentence to have available when a board asks what the English-language curriculum is giving up.
• Expect a small English-reading gain, not a large one. Budget 0.13-0.22 SD by grade 8, not the +0.45 the observational syntheses advertise.
• Set the L2 expectation honestly: high functional fluency, non-native production. And say out loud that this expectation rests on grade-D evidence, because no lottery has ever tested it.
• If you serve English learners, do not judge the programme on elementary reclassification rates. Two-language students reclassify slower early and better late. Committing to immersion means committing to a metric that looks worse for several years before it looks better.
• Discount European CLIL evidence heavily when deciding. It is a selection literature with no randomised or lottery-based evaluation behind it.

Open questions

• How much of the partner language do immersion students actually learn? No lottery has measured it. This is the largest and most embarrassing gap in the topic, because it is the reason the programme exists.
• Does the Portland result replicate? A second immersion lottery — Utah, Texas, or any oversubscribed district — would move this topic from moderate-support/medium to a much firmer place, and none has been published.
• What produces the English gain? Metalinguistic transfer, teacher selection into immersion strands, and peer composition are all consistent with the data and have different implications for whether the effect survives scale-up.
• Does 90:10 differ from 50:50? The dose–response question is untouched by any credible design.
• Does the effect hold for children who are not from immersion-applying families? A lottery generalises to its applicant pool and no further.