A locality gradient in the that / zero complementizer alternation: structural position modulates predictor importance

The that / zero complementizer alternation in English complement clauses has played an important role in testing theories of sentence production. This study revisits this alternation through a fine-grained analysis of the high-frequency complement-clause-taking verb think , which strongly favors complementizer omission. Drawing on the information locality framework and the discourse-level account of cue effectiveness decay, we ask whether an analogous locality gradient can be observed at the clause level: predictors located closer to the complement-clause onset are expected to exert stronger effects on that / zero choice. We analyzed 3,274 instances from the written subsection of COCA, using imbalance-aware modeling and likelihood-based measures of predictor contribution, with alternative metrics serving as robustness checks. In line with previous multi-verb work, the results indicate that once lexical variation is controlled, complexity-related predictors, particularly complement-clause length and intervening material between the matrix verb and the complement clause, are among the strongest predictors. More importantly, the results provide evidence for a locality gradient: predictors closest to the verb–complement boundary make the largest contributions to model fit, with their predictive power systematically decreasing as their structural distance from this juncture increases.