The Bland-Altman limits of agreement (LoA) method is almost universally used to compare two measurement methods when the outcome is continuous, despite warnings regarding the often-violated strong underlying statistical assumptions. In settings where only a single measurement per individual has been performed and one of the two measurement methods is exempt (or almost) from any measurement error, the LoA method provides biased results, whereas this is not the case for linear regression.
Thus, our goal is to explain why this happens and illustrate the advantage of linear regression in this particular setting. For our illustration, we used two data sets: a sample of simulated data, where the truth is known, and data from a validation study on the accuracy of a smartphone image-based dietary intake assessment tool.
Our results show that when one of the two measurement methods is exempt (or almost) from any measurement errors, the LoA method should not be used as it provides biased results. In contrast, linear regression of the differences on the precise method was unbiased.
The LoA method should be abandoned in favor of linear regression when one of the two measurement methods is exempt (or almost) from measurement errors.