Non-linear calibration curves allow researchers to assess the relationship between the predicted and observed probability of an outcome. This can be achieved via the use of a restricted cubic spline in a logistic regression model relating the predicted probabilities to the observed binary outcome. The present simulation study shows that using the predicted probabilities directly (the default in R functions available) leads to incorrect calibration curves that suggest miscalibration of correctly specified models. Further, the degree of the suggested miscalibration depends on the degree of non-linearity or interaction present. Better performance is achieved by first logit transforming predicted probabilities.