This paper investigates the effects of data transformation on nonlinearity by means of a simulation analysis based on empirical threshold models for the unemployment rate. Unemployment rate series are particularly suitable because they exhibit a number of interesting features: business cycle asymmetries, persistence, long memory and seasonality. The main finding is that evidence of nonlinearity is not independent of the form in which data are analysed and that most data transformations result in a loss of nonlinearity. This is particularly the case for seasonal adjustment transformations, which remove not only seasonality but also nonlinear features, as shown for the commonly applied Census X12 method.