lang-data-analysisUse when planning or auditing the analysis of a Language (LSA) manuscript so the evidence credibly supports the theoretical claim. Covers quantitative modeling (mixed-effects in R), phonetic measurement, corpus statistics, and the analytic trail from glossed data or judgments to the generalization. Improves the analysis chain; it does not fabricate results.
Independent GitHub snapshot · Automated classification · Original documentation governs
At Language the analysis exists to make the theoretical claim credible — not to display technique or notation. A cross-subfield, double-anonymous reviewer will ask whether the evidence actually warrants the generalization and whether uncertainty is handled honestly. Where the work is quantitative, Language now expects properly specified models (typically mixed-effects models in R) rather than by-subject t-tests or raw counts; where it is analytic, it expects the pattern to be demonstrable from the glossed data. This skill stress-tests the analysis chain in the idiom of your work.
Language rewards a generalization shown through more than one window — e.g., an experimental effect corroborated by a corpus trend, or judgments backed by text frequencies. When windows disagree, say so and explain the discrepancy rather than hiding the inconvenient one.
| Referee writes… | The Language-specific fix |
|---|---|
| "No random effects / pseudoreplication." | fit the justified mixed model; cluster by subject and item |
| "Significance without effect size." | report estimates + intervals in interpretable units |
| "The stat model doesn't match the design." | align random-effects structure with the sampling |
| "Analysis doesn't rule out the alternative." | show it derives attested and blocks unattested cases |
Orienting heuristics; confirm against the current author pages. Language increasingly expects that a quantitative claim rests on a model appropriate to the clustered, repeated-measures nature of linguistic data — the modal avoidable failure is pseudoreplication (ignoring by-speaker or by-item structure). Illustrative: a paper claims a durational contrast "is significant (p < .01)" from 1,200 tokens produced by 8 speakers, analyzed as if independent. A referee writes "pseudoreplication." The fix refits a mixed-effects model with by-speaker and by-word random intercepts and slopes, reports the estimate (an illustrative 12 ms, 95% CI ~4–20), and notes two speakers who show no effect — turning a fragile claim into a credible, bounded one.
Language asks for a properly specified model, which is a claim about a fitted object,
not about a paragraph. Fit it and report from it. Full map:
execution-with-mcp.
mixed (continuous responses) and melogit / meglm (binary and
categorical ones) carry the crossed by-subject and by-item structure the design justifies;
icc states how much clustering there actually is, which is the number a reviewer needs
when the sample's non-independence is the objection.bootstrap for intervals where the asymptotics are thin — small
fieldwork samples and unbalanced cells, both routine here.holm or benjamini_hochberg across a family of contrasts.
A typological or corpus paper testing many predictors at once needs this stated, not
assumed.etable / margins so the effect sizes and intervals in the prose are
the fitted ones.Where a server is not connected, adapt the ../../resources/code/ skeleton and say so —
never report a number you did not compute. This bridge touches the quantitative strand
only; analytic and historical arguments are made from the glossed examples themselves.
Treat this skill as an executable review pass, not a prose hint. First lock the empirical generalization, evidence base, warrant, and theoretical payoff; then judge whether the manuscript answers the venue's real reader: linguists across subfields who value grounded analysis, transparent and checkable evidence, and careful, appropriately scoped generalizations.
claim / evidence / risk / manuscript location rows so the next agent can
edit rather than rediscover the issue.resources/official-source-map.md has
been checked and the manuscript has one concrete fix for the largest venue-specific risk.【Claim under test】from theory-building
【Primary evidence】the analysis that carries the claim
【Model】mixed-effects structure matches the design? [Y/N/NA]
【Uncertainty】effect sizes + intervals reported? [Y/N]
【Convergence】corroborated across windows? [Y/N/NA]
【Confirmatory vs. exploratory】labeled where relevant? [Y/N]
【Next】lang-data-and-transparency
../../resources/external_tools.md — R / lme4 / brms, Praat, corpus tooling../../resources/official-source-map.md — Language quantitative and evidence expectations