EFFECT OF DICHOTOMIZING CONTINUOUS VARIABLES IN REGRESSION MODELS by

Jose Francisco Cumsille. Effect of Dichotomizing Continuous Variables in Regression Models (Under the Direction of Dr. Shrikant I. Bangdiwala). The dichotomization of continuous variables is a very common practice when people analyze data. Some of the consequences of such dichotomization are well known: loss of information, grouping people in the same group when they are different, loss of power of the statistical methods, underestimation of the correlation coefficient, among others. The dichotomization is not only a methodology issue; this practice can have a strong impact in the interpretation of the empirical results. One of the most popular tools to analyze data is the regression model. It is very common to see the results from a particular linear regression model, where one or more continuous variables have been dichotomized. The objectives of this dissertation are to study the effect on the model structure of categorization of continuous variables in multiple linear regression models, and also to study the effect on the exposure (main effect) of categorizing a continuous confounding variable in multiple linear regression models and also in multiple logistic regression models. These considerations imply that the outcome can be continuous or binary and the confounding variable must be continuous. About the main effect, we study the situation in which it is also a continuous variable. In order to evaluate the consequence in terms of the model structure, an analytic approach was used, whereas to evaluate the impact on the measure of association between the outcome and the exposure, a simulation approach was used for both the linear and logistic regression situations. The effect on the measure of association was evaluated by using a variety ofmethodologies.-