Modeling percentage change: a potential linear mirage

The percentage change over time in a variable such as serum creatinine or blood pressure is a potentially appealing outcome measure for longitudinal studies of prognosis or therapy. We demonstrate that when the baseline value of the variable used in the calculation of the percentage change outcome is included in a linear regression model as an independent predictor, a quadratic relationship between the baseline level and the predicted outcome level of this same variable is inevitably masked by the linear regression model. Misinterpretation of the percentage change model has resulted in conflicting results regarding the value of serum creatinine as a prognostic marker in lupus nephritis. We illustrate the problem and resolve the conflict using data on 87 patients with lupus nephritis. A straightforward alternative approach to the use of percentage change as an outcome is proposed: it is not subject to the problems of interpretation noted in the percentage change model.