Letters to the Editor
Reply: Dichotomization and Manipulation of Numbers
Dear Editor:
Dr Hutchinson is quite correct that we need not worry about the assumptions of parametric tests (for example, t-tests, analyses of variance, and related techniques) if we use randomization tests. However, I believe that his recommendations to recode data when there are floor or ceiling effects, or pathology at both extremes of a variable, misses the mark. First, recoding does not eliminate the problem; it only masks it. If a scale has a floor (or ceiling) effect, it means that it is not accurately tapping the attribute in question. People may have less (or more) of whatever is being measured, but the scale is unable to differentiate among them, owing to insufficient items at the extremes. It is a tenet of psychometric theory that reliability is directly related to a scale’s ability to discriminate among people (1). Recoding values does not solve the problem of unreliability at the extremes; it merely disguises it by assuming that all people below the floor (or above the ceiling) actually have the same score.
The second problem is that recoding may distort any relation between or among variables. Recoding and randomization tests may be able to answer the question of statistical significance, but they do not help us understand these relations and may even change them. If we retain the original values, then we are able to use other techniques, such as nonlinear regression, to better model what is going on.
Hutchinson is absolutely correct, though, to caution researchers to be concerned about whether their dependent variable truly represents what is of interest. Scores on a test are simply scores on a test; they are not reality.
References
1. Streiner DL, Norman GR. Health measurement scales: a practical guide to their development and use. 3rd ed. Oxford: Oxford University Press; 2003.
David L Streiner, PhD, CPsych
Toronto, Ontario
|
