NCJ Number
79610
Journal
Evaluation Review Volume: 5 Issue: 5 Dated: (October 1981) Pages: 639-670
Date Published
1981
Length
32 pages
Annotation
Using data from a pretrial release study, data analysis techniques developed to transform the criterion variable to a scale that permits the development of a good-fitting model are illustrated.
Abstract
Data analysts are faced with the problem of developing models that describe how a set of independent or carrier variables affects dependent or criterion variables. A good model should be capable of breaking the criterion variable into a fit and a residual part, such that (1) the fit contains as few parameters as possible; (2) the estimation of the parameters is not highly dependent on a small set of observations; (3) the parameters and variables comprising the fit have a meaningful interpretation; and (4) when using least-squares procedures, the residual can be described as independently and identically distributed random variables from a Gaussian or normal distribution. In considering the data analysis techniques, several exploratory techniques that use relationships between spread and location parameters to suggest power transformations of the criterion variable are reviewed. Data from a study of pretrial release illustrate the techniques. An examination of residuals is shown to complement the search for a good transformation. The effect that independent or carrier variables have on untransformed criterion variables is examined for power, logit, and arcsin transformations of the criterion variable. The discussion shows how various transformations imply different relationships between criterion and carrier variables. Inverse transformations are briefly discussed. Graphic and tabular data, 11 notes and 14 references are provided. (Author abstract modified)