Chapter 1 the study of collocations

 
3.6.2  Analyses for the Hypotheses 
 
 
The analyses performed for testing the two hypotheses are described 
below. 
 
i)  Analysis for Hypothesis 1 
 
 
229

---

 
To address Hypothesis 1, that there are stable patterns of development in 
the acquisition of collocations across proficiency levels, non-parametric 
Kruskal-Wallis tests were performed on the data.  Kruskal-Wallis is the non-
parametric equivalent of ANOVA.  Due to the wide range of types of 
collocations that were used in the analysis of the essay data, and the unequal 
number of tokens for each of the types of collocations tested in the translation 
and the blank-filling tests, the data were not expected to be normally 
distributed and thus non-parametric statistics were considered suitable to 
address the first hypothesis.  Previous studies on collocations used non-
parametric statistics too (see Zhang 1993). 
 
For the free production essay data that resulted from the analysis of the 
students' essays, Kruskal-Wallis tests followed by Dunn's multiple comparison 
procedures were performed on the mean tokens of each of the 37 collocation 
types used by subjects in each group. 
 
For the cued production translation data non-parametric Kruskal-Wallis 
tests followed by Dunn's procedure were performed on the mean accuracy of 
responses to each of the 6 types of collocations repeated across the three 
groups. 
 
For the cued production blank-filling data non-parametric Kruskal-
Wallis tests followed by Dunn's procedure were performed on the mean 
accuracy of responses to each of the 11 types of collocations repeated across the 
three groups. 
 
230

---

 
Implicational scaling analysis for each of the three sets of data was also 
performed in order to reveal any acquisition orders.  The frequency data were 
used for this analysis.  For each implicational scale the Guttman's coefficients of 
reproducibility and scalability were then calculated (Hatch & Lazaraton 1991:204) 
in order to test the validity of the scales and the scalability of the items on the 
scales.