This article argues that, across different psychological contexts, the methods of data collection, treatment, and analysis in word association tests have hitherto been inconsistent. We demonstrate that this inconsistency has resulted from inadequate control, in previous studies, of certain important variables including the basis of norm comparisons, and we present a principled method for collecting, scoring, and analysing association responses, to address these issues. The method is evaluated using test and retest data sets from 16-year-old and over-65-year-old twins (n = 636), which enable us to (a) compare samples matched for key environmental variables, (b) assess the transferability of norming information between age cohorts, and (c) evaluate the reliability of the scoring protocols. We find systematic differences in the association behaviour of the two age cohorts, indicating the importance of evaluating data only against norms lists that are matched to the target population. Individual association behaviour is found to be consistent across test times, both in terms of response stereotypy and response type.