This study aims to investigate to what extent grammatical variables are related to each other in order to predict the discourse-based classification of English relative clause extraposition (RCE) construction with the help of a total 597 corpus data. In doing so, an extraposed relative pronoun is assumed to function as a cohesive device, thereby classifying RCE construction into four types according to discourse manner: Discourse-Old-Old-RCE, Discourse-Old-New-RCE, Discourse-New-Old-RCE, and Discourse-New-New-RCE. This discourse-based classification tells us which information is prevalent within an extraposed relative clause as well as a subject NP. Multinomial logistic regression is used for determining which one among the indefiniteness/definiteness of a subject NP, the predicate type (i.e., passive, presentative intransitive, and predicative complement of copula be), and/or grammatical weight (i.e., subject NP-to-extraposed RC length ratio) primarily has the inclination for each type of RCE construction. The definiteness of a subject NP, passive voice, and positive value of the length ratio are critical predictors to bring about Discourse-Old-Old-RCE and Discourse-Old-New-RCE, whereas the indefiniteness of a subject NP is a main predictor of Discourse-New-Old-RCE and Discourse-New-New- RCE. Interestingly, only Discourse-New-Old-RCE taking up the largest portion in this corpus study has a negative correlation with the length ratio. In other words, both a small number of constituents within a subject NP and heavy information within an extraposed relative clause are strongly associated with the occurrence of Discourse-New-Old-RCE. Therefore, the conclusive statement is drawn that the grammatical heaviness within a restrictive relative clause becomes a powerful trigger for the extraposition. (Chodang University)