ABSTRACT
In this chapter, the author argues that he can make causal analyses with ordinal data. He describes two procedures for doing so and applies them to various models and measurement condition. The author examines the advantages and disadvantages of employing ordinal data in correlation and regression formulas in the context of causal analysis. Roderick Bell's solution to the problem was to substitute category scores into correlation and partial correlation formulas, which, in effect ignored the level of measurement. Leo A. Goodman and William H. Kruskal suggest computing a weighted average of contingent measures of association, the average being interpreted as a partial correlation coefficient. In many cases the magnitude of gamma is so large that the partialling techniques do not reduce the total correlation close enough to zero to produce an unambiguous test for spuriousness. Combined with the investigator's substantive knowledge of a problem, ordinal partials based on these measures should provide useful tools for testing causal models.
