We read with interest the editorial by Hoffer and associates and appreciate the authors’ efforts to standardize methods for studies assessing the accuracy of intraocular lens (IOL) power calculations. However, we consider that their recommendations may not always be appropriate, depending on the context of a specific study, and have concerns with their criticisms of our paper. Specifically, the authors cite the paper as an example of studies that are compromised by not optimizing IOL constants such that the mean error (ME) is zero, utilize the desired target refraction to assess predictability, do not exclude bilateral eyes, and include multiple IOLs.
In our original paper, as well as in our reply to Savini and associates, we addressed the reason for not optimizing the IOL constants in our study for an ME of zero. Adjusting the ME is considered questionable when analyzing a subset of axial lengths, as in this series of eyes with long axial lengths in patients with high and extreme myopia. Formulas that maintain accuracy for short and long axial lengths without requiring different constants for each subgroup are desirable for clinical practice. In addition, a previous study that attempted to optimize lens constants for long eye datasets found this approach to be less predictable than axial length modification.
We agree that the target refraction desired by the surgeon is of no consequence when evaluating the prediction errors of IOL power calculation. Therefore, in our study we did not compare target refraction results, as mistakenly stated by Hoffer and associates.
Although it is preferable to avoid using both eyes from a single patient whenever the dataset is large enough, in the context of limited data (postrefractive, long and short eyes), use of the fellow eye is considered acceptable when appropriate statistical analysis is applied. Indeed, as we have already explained in our response to the letter from Savini and associates, we addressed this issue in our study by using the Wilcoxon test to compare medians from paired data and by using intraclass correlation coefficients to rule out the existence of high correlation between fellow eyes.
Similarly, including more than one model IOL is acceptable when limited data are available as long as this limitation is clearly acknowledged. When we optimized the data of a single model (MA60MA; Alcon) for the Holladay 1 (SF = 26), SRK/T (Aconst = 139), HofferQ (pACD = 26), and Haigis (a0 = 10, a1 = 0.4, a2 = 0.1) formulas, we found the median absolute prediction errors to be similar to the optical constants and inferior to the ULIB constants, the axial length adjustment method, and the Barrett Universal formula.
We acknowledge the importance of Hoffer and associates’ article and appreciate this opportunity to explain the rationale for our study design in response to what we feel are misleading criticisms of our paper.