Correspondence Regarding Changes in the Incidence of Retinal Vascular Occlusions After COVID-19 Diagnosis

Abstract

topic 2 supporting these meta-analytic results (we observed lower diabetic retinopathy [DR] prevalence in pregnant women with T2D than those with type 1 diabetes [T1D]).We agree that our included studies had small sample sizes and this condition has been reported to cause bias. 3 Especially for studies with 0 events, we understand that continuity corrections in meta-analyses can be problematic, as mentioned by Zhou and Shen and the literature. 3,4 However, we avoided a need for continuity correction to stabilize variances in our meta-analyses by choosing to use the Freeman-Tukey double-arcsine transformation (FTT) method for our random-effects model. 4 We ran an additional generalized linear mixed model (GLMM), specifically, a random intercept mixed-effects logistic regression model to confirm and compare with our pooled estimates from the FTT method. In all but 1 outcome (progression from nonproliferative diabetic retinopathy to proliferative diabetic retinopathy), the 2 methods were comparable (eTable 8 in the Supplement). 1 Although we note that a GLMM method would remove much of the concern around sampling error and bias, one downside to the GLMM is a lack of transparency regarding the weights given to each study. Thus, we decided to present the FTT results and forest plots so that our readers can focus on the clinical message and not be distracted by methodological arguments. In that vein, while prediction intervals can provide a useful concept for interpreting published pooled estimates of effect size of one treatment compared with another in the context of clinical decision-making for an individual patient, 5 we omitted them because if a clinician wanted to predict their presentday patients' risk of developing DR or predict its progression in pregnancy, they may be better served to consider only studies with excellent quality, validated DR grading schemes, and consistent methodology in a population with the same type of diabetes and access to the same level of (present day) diabetes care. Using such criteria, this would limit the analysis to only 4 T1D studies (grand total of 82 participants). We plan to produce a table of meta-analysis results from FTT and GLMM methods that include prediction intervals and look forward to the opportunity to present these results.