Citation

Bruehning E, Ernst G. Proc. IRCOBI 1988; 16: 35-46.

Copyright

(Copyright © 1988, International Research Council on Biomechanics of Injury)

DOI

unavailable

PMID

unavailable

Abstract

Data used in empirical research and, in particular, in traffic and accident research, very frequently do not have the (quantitative) interval scale format which is required by most statistical methods. Instead, they are merely ordinal or nominal scaled: measurements and observations are qualitative (e.g. Age in age groups; sex; type of vehicle; accident severity -killed, seriously injured, slightly injured). The categories of such a nominal scaled variable only stand in the relation "not equal to". This means that we cannot make any statements about rank orderings or distances and that the range of mathematical operations allowed with such data is different from that available for quantitative data. Thus there is an obvious need for analysis methods which are suited to such data. A further problem constantly confronted in accident research is "complexity": i.e. Problems are not simple enough to permit an explanation of some variable y in terms of one other variable x. Analysis procedures for the relations between nominal scaled variables usually rely on contingency tables. But conclusions which were valid for cross tabulations of two variables often have to be considerably modified as soon as a third (control) variable is introduced. We then have to face the question: if we want control by a third variable, why not introduce a fourth for even more control, etc? Why not make a multivariate analysis using all the variables that are considered to be theoretically relevant. The subject of this contribution is statistical methods on the basis of loglinear and logit models which offer the possibility of multivariate analyses while depending only on realistic assumptions about the scale types of the variables.