Multiple Regression in L2 Research: A Methodological Synthesis and Guide to Interpreting R² Values

Abstract

Multiple regression is a family of statistics used to investigate the relationship between a set of predictors and a criterion (dependent) variable. This procedure is applicable in a variety of research contexts and data structures. Consequently, and similar to quantitative traditions in sister‐disciplines such as education and psychology (see Skidmore & Thompson, 2010), second language researchers have turned increasingly to multiple regression. The present study employs research synthetic techniques to describe and evaluate the use of this procedure in the field. Five hundred and forty‐one regression analyses (K = 171) were coded for different models, variables, procedures, reporting practices, and overall variance explained (R2). Summary results reveal a number of inconsistencies (e.g., model types) as well as a lack of transparency (e.g., missing/unreported reliability estimates; see Larson–Hall & Plonsky, 2015). The distribution of R2 values (median = .32) is described to facilitate utilization and interpretation of regressions models. We also provide specific, empirically grounded recommendations for future research.

On Bilingual Aptitude for Learning New Languages: The Roles of Linguistic and Nonlinguistic Individual Differences

“…Due to the ceiling effect, the LLAMA–E model reported in Table should be interpreted with some caution, although when we also conducted the same analysis as a Tobit regression model (see Appendix S1 in the Supporting Information online), which accounts for the ceiling effect, we obtained the same pattern of results as seen in Table . The standard regression model explained 44% of the variance in LLAMA–E scores, which approached the standard considered to be a large amount of explained variance ( R 2 = .50), according to Plonsky and Ghanbar's () benchmarks. It also differed significantly from a model with no predictors.…”

“…The regression model for LLAMA–F, aptitude for grammatical inferencing, met all assumptions for linear regressions and is summarized in Table . This model also accounted for a moderate amount of variance in LLAMA–F scores (33%), falling in between the designations of small ( R 2 = .20) and large ( R 2 = .50) amounts of variance (Plonsky & Ghanbar, ). The model also significantly differed from a model with no predictors.…”

“…We follow Plonsky and Ghanbar's () benchmarks for L2 acquisition research that R 2 = .20 represents a small amount of variance explained and R 2 = .50 and above corresponds to a large amount of variance.…”

Exploring the Dynamic Nature of Second Language Listeners’ Perceived Fluency: A Mixed‐Methods Approach

“…The model shows that L2 listeners’ fluency perception was predicted primarily by speed of delivery and secondarily by clause‐internal pausing. According to Plonsky and Ghanbar's () field‐specific benchmark, the explained variance could be considered relatively large ( R 2 > .50), suggesting that our L2 listeners greatly relied on temporal information (speed, breakdown) in their fluency judgements.…”

“…Our quantitative findings indicate that L2 speakers’ perceived fluency is mainly associated with the speed of delivery and pausing behaviors within clauses. The magnitude of the link between acoustic information and L2 fluency judgements is relatively large ( R 2 > .50) (Plonsky & Ghanbar, ). The results here serve as additional supportive evidence for the significant role of temporal features in L2 fluency judgements.…”