Latvia’s language question, asked of the data rather than of the debate

Russian-medium schooling is not where Latvia’s language gap lives. The gap is between students whose home language matches their school and students whose home language does not, and the reform now under way moves children from one side of that line to the other.

Dr Milos Kankaras | ORCID 0000-0002-3190-7751 | Center for Psychology, Podgorica | PISA 2022 country notes, Baltic states | August 2026

Download the full note (PDF)

Summary

Latvia is completing a transition to Latvian as the sole language of instruction in state-funded schools. PISA 2022 was administered while Russian-medium schooling still existed at scale: 28.6% of the assessed cohort sat the test in Russian. That makes this the last cycle in which the question can be asked of real data rather than argued from first principles.

The question everyone assumes PISA answers is whether Russian-medium schools underperform. It does not answer that, because the answer turns out to be no.

What the data shows.

• Students assessed in Russian scored 481.0 points, against 484.3 for students assessed in Latvian. The difference is -3.3 points with a standard error of 4.8, which is indistinguishable from zero.
• Adjusting for socio-economic status changes nothing: -2.6 points, standard error 4.0. Russian-medium and Latvian-medium schooling produced statistically identical mathematics results.
• The real language gap is elsewhere. Students who speak a language at home other than the one they were tested in scored 28.8 points below those whose home and test language match (4.3), and 22.3 points below once socio-economic status is held constant.
• The gap that dominates everything is neither: students in village and rural schools scored 39.8 points below city students, 24.9 net of socio-economic status.

The policy implication is uncomfortable and precise. The mismatch penalty is real and the medium-of-instruction penalty is not, so a reform that moves Russian-speaking children out of Russian-medium schools does not remove a disadvantage they were suffering; it moves them from the matched group into the mismatched group. Whether that is the right decision is a question about national cohesion, not about mathematics scores, and this note takes no view on it. What the note can say is that the mathematics-attainment argument for the reform is not supported by Latvia’s own PISA data, and neither is the mirror-image argument that Russian-medium schooling was holding children back.

Why this is the question for Latvia

Almost every country’s PISA commentary reaches for the same three variables: money, gender and immigration. Latvia’s structural question is none of them. It is a language-of-instruction reform affecting roughly a quarter of the school-age population, legislated and now being implemented, and PISA 2022 is the only recent source of comparable evidence on the students it affects.

It also happens to be a question the standard international indicators get wrong. The OECD publishes an indicator for students who speak a language at home other than the language of the assessment. In most European systems that indicator identifies a minority studying in the majority language, and it behaves like a disadvantage marker. In a country that administers PISA in the minority language as well, a Russian-speaking child in a Russian-medium school reports speaking the test language at home, and the structural divide the debate is about disappears from the indicator entirely.

This note therefore uses two different language variables and keeps them apart: the language a student was assessed in, which is the medium of instruction, and whether that language matches the one spoken at home. In Latvia they point to different children and to opposite conclusions.

The data, and how it was handled

5,373 students in 225 schools, representing a weighted population of 16,833 15-year-olds, from the OECD’s PISA 2022 student and school Public Use Files. National mean in mathematics: 483.2 points (SE 2.0).

Achievement is imputed rather than measured, so estimates are pooled by Rubin’s rules over all ten plausible values. The sample is clustered in schools, so standard errors come from the 80 Fay-adjusted replicate weights the OECD ships for the purpose. Omitting either understates the uncertainty, usually by a factor of two or more.

Source: OECD, PISA 2022 Results (Volume I and II) Country Note: Latvia, published 5 December 2023, https://www.oecd.org/en/publications/pisa-2022-results-volume-i-and-ii-country-notes_ed6fbcc5-en/latvia_f58bca29-en.html (read 6 August 2026)

Medium of instruction: no gap

3,674 students were assessed in Latvian and 1,583 in Russian. The two groups are close in socio-economic terms, at -0.00 and -0.03 on the PISA index, and closer still in performance.

The confidence interval on the difference comfortably includes zero, and adjustment for socio-economic status leaves it there. This is a null result, and null results are worth publishing when a policy argument depends on them.

Two cautions keep this honest. Sampling, not the reform, decides who is in which group here: this is a comparison of two school systems as they existed in 2022, not an estimate of what happens to a child moved between them. And a null result is not proof of equivalence; it means that if a difference exists, the data cannot distinguish it from zero at this sample size, and the interval rules out anything larger than roughly six points in either direction.

Home language: a real and substantial gap

Now ask the different question. Rather than which language the school teaches in, ask whether that language is the one the student speaks at home.

This is the gap the reform interacts with. Adjusting for socio-economic status leaves 22.3 points, and holding the language of assessment constant as well leaves it essentially unchanged, which tells us the mismatch penalty is not simply Russian-medium schooling under another name. It is what happens to a child whose school language is not their home language, in either medium.

The size is worth stating plainly: it is roughly a quarter of a standard deviation on the PISA scale, and it applies to 10.0% of the cohort as measured in 2022. A reform that moves Russian-speaking children into Latvian-medium schools increases that share.

Can this comparison be trusted? A measurement audit

Test scores are placed on a common scale by design. Questionnaire indices are not. Comparing an index across groups assumes the items mean the same thing in every group being compared, and that assumption is testable. It is usually not tested.

For Latvia the audit is not an afterthought. If the two language groups do not answer the questionnaire items in the same way, then any statement comparing their school experience, including the ones a ministry would most want to make about how the reform is being received, rests on nothing. The construct tested here is sense of belonging at school, which is the index most likely to be quoted during a language-of-instruction transition.

Measurement invariance of the sense of belonging at school block across language of assessment. N = 5,196. Estimator MLR. Decision rule after Chen (2007).

Read the belonging figures against the achievement figures and the picture sharpens. On attainment the two media of instruction are indistinguishable; on belonging they are also close, and both sit below the OECD average. Whatever the reform changes, it is not starting from a system in which Russian-medium students were alienated and Latvian-medium students were thriving.

What follows

Three statements are supported by the analysis above and are narrow enough to defend under questioning.

• The attainment case for the language reform is not in Latvia’s PISA data. Russian-medium and Latvian-medium schooling produced statistically identical mathematics results in 2022, before and after adjustment for socio-economic status.
• The home-school language mismatch is a real disadvantage of about 22.3 points net of socio-economic status, and the reform mechanically increases the number of children in that category. Transitional language support is therefore not a courtesy, it is the mitigation for a measured effect.
• The largest gap in Latvian schooling is geographic, not linguistic. Village and rural schools sit 24.9 points below city schools net of socio-economic status, which is larger than any language contrast in this note.

The question this note cannot answer, and which is the obvious next piece of work, is what happens to mismatched students over time. A mismatch penalty measured at 15 says nothing about whether it narrows with years of exposure. That needs either a longitudinal design or the 2025 and 2029 cycles read as a sequence, with the link errors carried properly.

What this note does not claim

• No trend statement. Comparing 2022 with an earlier cycle requires the published link error for that cycle pair to be carried in the variance. No such comparison is made here.
• No causal claim. Every difference is an association measured at one point in time. “Net of ESCS” means one measured index is held constant, not that other things are equal.
• The invariance cascade is fitted without the replicate-weight design, which is standard practice and is stated rather than left implicit.
• The language of assessment is a proxy for the medium of instruction, not a direct measure of it. It is a good proxy in Latvia, where the two aligned by school in 2022, but it would not be in a system where students can choose an assessment language independently of their school.
• PISA 2022 predates the completion of the transition. This note describes the system that existed, which is the point, but it cannot observe the reform’s effects.

Method and reproducibility

Data: OECD PISA 2022 Public Use Files, downloaded from webfs.oecd.org on 6 August 2026. Point estimates pooled over ten plausible values by Rubin’s rules; sampling variance from 80 Fay-adjusted replicate weights with a Fay factor of 0.5, computed per plausible value and averaged; imputation variance inflated by (1 + 1/M). Invariance tested as a staged configural, metric, scalar and strict cascade in lavaan with full-information estimation for the rotated questionnaire design, plus a categorical sensitivity re-run on pairwise-present data.

Subgroups are derived from variables the OECD codes identically in every participating system, so this note was produced without country-specific recoding. Where a country’s own sampling strata carry a more policy-relevant structure, that structure is used instead and the note says so.

Enquiries about reproducing the analysis are welcome at milos@centerforpsychology.me. Commissioning the equivalent for another country is consulting work, handled by AdriaMont Consulting DOO at milos@adriamont.me.

All country notes in this series