Bestimmung der Reliabilität psychometrischer Testverfahren in der Psychiatrie mittels der kanonischen Korrelation

Test results (raw scores) are composed of an unknown true score and an error term. The error term can be estimated by means of test reliability which is defined by the ratio of true variance and obtained variance. Different estimates of reliability either based on single measurements (e. g. Cronbach's coefficient, split half reliability, Kuder Richardson method) or two measurements (test/retest, inter- or intrarater reliability) are available. Parallel test reliability depends on the correlation of two different tests obtained in one session. Canonical correlation methods allow an extension of the parallel test situation and split half technique. Two or more tests are performed in a sample of subjects. Randomized subsets are correlated using canonical correlation technique. The objective of this study is to estimate the homogeneity of test batteries. 94 patients (64 f, 30 m; age: 54 - 89 ys.) supposed to have dementia were tested using the clocktest (CT, scores: 1 - 5), MMSE (mini mental state examination) and SKT (Syndrom Kurztest). Four (i, j: 1 - 4) subsets of 20 patients each were determined by random and the following characteristics were calculated: Empiric correlation coefficient for n = 94 (R), canonical correlation coefficient (Rcan), eigenvalues (EV) and redundancy (Rnd) of corresponding variable sets. The results of canonical analysis showed canonical correlation coefficients in order of 0.8 to 0.9 (p-values < 0,001). This high internal consistency can be interpreted as a measure of reliability of the test batteries. In conclusion, canonical correlation based on parallel tests splitted in subsets gives information on consistency, i. e. reliability, of test batteries in addition to conventional correlation methods.