Source-linked AI summary
Linear models and linear mixed effects models in R with linguistic applications
Bodo Winter
TL;DR
Linguistic data contain repeated subject and item observations, challenging independence assumptions and allowing effects to vary across subjects and items. The paper introduces linear mixed-effects models in R using voice pitch and politeness, showing that politeness lowered pitch by about 19.7 Hz.
Problem
Repeated observations violate independence, while politeness effects may differ across subjects and items.
Method
The paper uses mixed models with subject and item random effects, random slopes, and likelihood ratio tests to model variation and assess effects.
Results
19.7 Hz ± 5.6: politeness affected pitch, lowering it significantly (χ2(1)=11.62, p=0.00065).
Takeaways & Limitations
Mixed models account for subject and item variation simultaneously while retaining the full data rather than averaging over either source.
Takeaways & Limitations
Mixed models can still violate independence when important fixed effects are missing.
Abstract
from arXiv · showhide
This text is a conceptual introduction to mixed effects modeling with linguistic applications, using the R programming environment. The reader is introduced to linear modeling and assumptions, as well as to mixed effects/multilevel modeling, including a discussion of random intercepts, random slopes and likelihood ratio tests. The example used throughout the text focuses on the phonetic analysis of voice pitch data.
Part 2: A very basic tutorial for performing
The tutorial presents mixed models as a way to model subject- and item-specific dependence by combining fixed effects with structured random effects. Using voice-pitch data, it demonstrates politeness and gender effects, likelihood-ratio testing, random slopes, and reproducible reporting.
- Why mixed models: Repeated responses from the same subject or item are non-independent, so mixed models account for both sources of variation in one analysis.Individual voice-pitch differences affect all responses from a subject, while items can also systematically differ.
- Why mixed models: Mixed models combine fixed effects with random effects that structure the error term and represent idiosyncratic variation due to subjects or items.This mixture distinguishes mixed models from fixed-effects-only models with a general, unstructured error term.
- Voice-pitch example: In the voice-pitch example, polite speech lowers pitch by about 19.7 Hz, with politeness affecting pitch (χ2(1)=11.62, p=0.00065).The model describes frequency as pitch measured in Hertz, with higher values indicating higher pitch.
- Voice-pitch example: Adding gender shows that males and females differ by about 109 Hz, while the intercept becomes 256.846 Hz for females in the informal condition.Including gender also reduces subject-associated random variation because gender had previously been confounded with subject variation.
- Model comparison: Likelihood-ratio tests compare a null model without the factor of interest against a model containing it, but joint comparisons cannot identify which added factor explains significance.Comparing an intercept-only model with a full model containing attitude and gender tests their combined difference rather than either effect separately.
- Random effects: Random-intercept models allow subject- and item-specific baselines while holding politeness effects constant, whereas random-slope models allow those effects to vary.The example finds politeness coefficients negative across subjects and items, indicating a consistent downward pitch tendency despite individual variation.