Comparison of Fed-GLMM accuracy relative to pooled analysis between different data splitting strategies.

Using randomly extracted small sub-datasets (n = 100,000) from the EHR of a single facility, we compare the accuracy of Fed-GLMM with different data splitting strategies. Two splitting strategies were attempted and compared: random splitting and our proposed clustering-based splitting introduced in S1 Table. A sub-dataset was split into 5 subsets by both strategies. The absolute relative bias was calculated as the difference between the corresponding Fed-GLMM estimates and those given by the pooled analysis in absolute percentage. A total of 50 randomly extracted sub-datasets were used in the evaluation. For all coefficients of interests, clustering-based splitting resulted in negligible bias compared with the random splitting strategy. Abbreviations: NH—Non-Hispanic; LEP—Limited English Proficiency. (JPG)