How do I include non-English studies in a literature review?

English-only searching produces a systematic bias, and the direction of that bias depends on the field. In areas where a phenomenon is studied mainly in one region, the international journals carry the exportable subset while the substantive body of work sits in national publications. A review that misses it can be confidently wrong.

The scale of the omission is documented. Across 300 randomly sampled Cochrane reviews, 27 percent explicitly excluded non-English studies, 35 percent stated no language criteria at all, 39 percent searched for non-English studies, and only 15 percent included any, and in a sample of 123 Campbell reviews, 108 excluded non-English studies in advance with few giving a justification (Neimann Rasmussen and Montgomery, 2018, p. 2). It also matters to the result. In a review of 50 meta-analyses, adding non-English studies changed the effect estimate in more than half, making it more positive in five cases and less positive in 16, usually with lower precision (Neimann Rasmussen and Montgomery, 2018, p. 2). The same source is careful about the reason: comparisons of English and non-English publications are inconsistent, some finding better design or reporting in English, others finding no significant difference, so include on eligibility and appraisal, not on an assumption that the non-English literature is systematically better or worse.

Searching in another language is more tractable than it looks. The terms you need appear in the abstracts and keyword lists of the few non-English papers you already have, or in the translated titles that many national databases provide. Two or three terms per concept is usually enough to run a search you can then screen with translation.

The databases matter as much as the terms. International indexes cover national journals unevenly, and regional platforms often hold literatures that are simply absent elsewhere. A subject librarian will know which ones apply to your field, and asking is faster than discovering them one at a time.

Removing the language filter from an English database is not the same as searching in the language. Non-English work is indexed less rigorously in the international databases, so the recommendation is to run non-English search terms, build a list of the major non-English journals in your field, and use networks of investigators to find work that is not indexed anywhere (Amano et al., 2016, pp. 5-6). Develop each strategy the way you would an English one, checking that the terms are complete, balancing sensitivity against specificity, and translating the optimized strategy across databases rather than retyping it (Bramer et al., 2018, p. 531). Record the translated terms, the databases and interfaces, the dates, and every adaptation you made.

Where you genuinely cannot cover a language, the answer is not to pretend the literature does not exist. Say what you excluded and why, and where possible say what the exclusion might have cost. That sentence protects you and tells the next researcher where to look.

What the exclusion costs can be specific. Population estimates for Taiwan's fairy pitta existed only in traditional Chinese and were left out of a global assessment, Chinese-language reports of avian influenza in pigs went unnoticed by international organizations at first, and Wetlands International Argentina produced more than 20 technical conservation publications over 20 years of which only two were available in English (Amano et al., 2016, pp. 3-5). If your field has an equivalent, name it in the limitation.

Is machine translation good enough for screening?

For screening, yes. Deciding whether a paper is about your topic and roughly what it found is well within what current translation handles. For extracting a finding you will cite, it is not, and that step needs someone who reads the language or a checked professional translation.

The failure mode at the extraction stage is quiet. Technical terms with a fixed disciplinary meaning get rendered into ordinary words, hedged claims lose their hedging, and a conditional finding becomes a general one. None of these look like errors on the page.

A practical compromise for a small number of important papers is to have the specific passages checked rather than the whole article. Verifying four paragraphs is a favor a colleague can do in twenty minutes.

Arrange that help before screening starts rather than when you hit the first paper. Review authors report cost, time, funding, and access to translators, translation software, and English abstracts as the recurring barriers, and training, practical guidance, and access to non-English databases as what actually helps (Neimann Rasmussen and Montgomery, 2018, pp. 7-8). Recruiting speakers of Spanish, Portuguese, Chinese, and French covers a large share of the non-English scientific record, with quality checks by native speakers where you can get them (Amano et al., 2016, pp. 5-6). Write into your protocol who translates titles and abstracts, who verifies full-text readings, how disagreements are settled, and where translation costs are logged.

How do I report a language restriction in my methods?

State which languages you searched and screened in, and say plainly that other-language work may be under-represented. If your topic has a substantial literature in a language you excluded, name it. A declared limitation is acceptable, an undeclared one looks like a claim of completeness you cannot support.

Where translation was used, say so and say at which stage. "Abstracts were screened using machine translation, and included full texts were read in the original with assistance from a native speaker" is a methods sentence a reviewer can evaluate.

Record the counts too. Knowing how many records were in other languages, and how many survived screening, turns a vague caveat into a quantified one, and it costs nothing if you logged it as you went.