Abstract
Purpose :
Optical biometry parameters may not be measurable before cataract surgery due to a variety of factors such as cataract density and patient compliance. Thus, in statistical analysis involving these parameters, missing data may limit the use of methods that require complete cases. This study assesses the ability of multiple imputation by chained equations (MICE) to yield valid inference for missing anterior chamber depth (ACD), lens thickness (LT), and white to white (WTW) optical biometry values from a cataract outcomes database under various induced missingness conditions.
Methods :
Eyes with complete axial length (AL), keratometry (K), ACD, LT, and WTW measurements were identified and eyes with prior ocular or refractive surgery were excluded. ACD, LT, and WTW values were set to missing in a simulation study. Missing at random (MAR) and missing not at random (MNAR) mechanisms were employed using missing rates that varied from 5% to 90%. 10,000 imputations were performed for each missingness condition using the 2l.lmer method from the R package “mice”. The imputation model included all non-missing optical biometry values, as well as history of dry eye, mature cataract, and history of corneal disease. Intercept-only linear mixed effects models with random intercepts for individual were used to calculate pooled means and standard errors according to Rubin’s rules. Raw bias (RB) and percent bias (PB) of means, and differences in 95% confidence interval width (CIW) were used to assess imputation performance.
Results :
1,635 eyes from 1,065 individuals were identified. RB was generally minimal for all variables under all conditions, with the most extreme RB of -0.096 for LT under the 75% MNAR condition, with a corresponding PB of 2.1%. There were minimal differences in CIW between complete and imputed data with the largest magnitude of difference observed for LT under the 5% MNAR condition (-0.011).
Conclusions :
In this study, RB was minimal under most conditions, and PB for all variables under all conditions was less than 5%, an upper limit for acceptable performance. The MICE algorithm is able to yield valid inference of missing ACD, LT, and WTW values in the presence of a hierarchical data structure under MAR and MNAR with at least 90% of cases containing missingness. Further study is needed to assess imputation performance on missing AL and K values, and the effect of imputed values on lens power calculation accuracy.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.