of doing regression, either linear regression or logistic regression. A lot of times, regression is used to test the association between the covariate of interest and outcome. In this case, one is more interested in the regression coefficent than the prediction equation. One should use a multivariate regression model to get an unbiased regression coeffient by controlling for other confounders. I don't know about the design mentioned here. But univariate analysis is usually not good enough.