Adjusting Intervals for Random-Text Samples

Corpus linguistic research routinely violates a fundamental assumption of Binomial statistical models: that instances are independently sampled from an infinite population. Instead, corpus data are typically drawn from contiguous text samples, where instances within the same text are more likely to share characteristics than instances from different texts. This chapter addresses whether standard Binomial methods remain valid under such cluster sampling conditions and proposes principled adjustments to maintain statistical rigour. The chapter presents a variance recalibration method that measures case interdependence by comparing predicted subsample variance (assuming random sampling) with observed Binomial subsample variance across texts. When observed variance exceeds predicted variance, the method adjusts effective sample size downward, widening confidence intervals to reflect reduced statistical certainty. Conversely, when corpus constraints limit variability, the method can legitimately shrink intervals. Three detailed examples demonstrate the approach: interrogative clause proportions, clause-per-word rates (the inverse of clause length), and shall/will modal auxiliary alternation in DCPSE. The analysis reveals that ‘balanced’ corpora with diverse genres cause rates per text to not approximate closely to the Binomial distribution, recommending researchers focus on genre-specific variation patterns when possible. For small samples with many singletons, the chapter introduces a partitioning and pooling strategy. The proposed methods integrate seamlessly with existing statistical techniques while preserving the power and flexibility of the Binomial model.