Fifty-seven patients with dementia (mean age 74 years) were treated for 16 weeks with 2 different doses of a test drug or placebo. On day 0 no significant differences (Kruskal-Wallis H test) were found between the 3 treatment groups on psychometric test scores or the total scores on 3 rating scales (Modified Crichton Geriatric Behaviour Rating Scale, Sandoz Clinical Assessment - Geriatric (SCAG) scale and Brief Cognitive Rating Scale (BCRS). The score differences (the total scores on day 112 minus the total scores on day 0) on SCAG, BCRS and psychometric tests were subjected to a series of one-way analyses of variance. As no significant difference was found between the groups, they were combined into a single group. In this group the relationship between total scores on SCAG and BCRS and psychometric test scores was analysed. Setwise multiple regression analyses (with the psychometric test scores as the independent variables) showed that at most 34% and 45% (respectively) of the variance in the SCAG ratings was accounted for by the results from a small subgroup of tests (dealing with visuo-motor speed and verbal memory) on day 0 and day 112. The results of a somewhat bigger subgroup of tests (dealing with visuo-motor speed, verbal memory and general intelligence) explained 58% of the variance in the BCRS ratings both on day 0 and after 16 weeks. Canonical correlation analysis using scores on the 2 rating scales and the psychometric test results was applied complementarily to make the inferences as valid as possible.(ABSTRACT TRUNCATED AT 250 WORDS)