(SLA-012) Handling Missing Laboratory Values across Various Missing Data Settings in Electronic Health Records

Background: Missingness among laboratory data may introduce selection bias in electronic health records (EHR). Understanding different missing settings and choosing imputation methods are important for causal inference.

Objectives: To evaluate the impact of handling missing laboratory values under various missing data settings in EHR.

Methods: We used a case example: a retrospective cohort study under target trial emulation adapted from DAPA and EMPA-CKD trial using data from the Chang Gung Research Database, a multi-institutional EHR in Taiwan. We compared the risk of hyperkalemia between SGLT2 inhibitors (SGLT2is) and GLP-1 RAs in patients with chronic kidney disease (CKD) level at stage III-stage V, using GLP-1 RAs as the reference group. We compared three settings of missing laboratory values: HbA1c only, UACR only and both missing. We selected HbA1c because it is routinely monitored in patients initiating SGLT2is or GLP-1 RAs, whereas UACR was chosen as it is tested only when specific symptoms arise. We applied three imputation approaches: complete case analysis (CCA), imputation of mean value and multiple imputation (MI) using Multivariate Imputation by Chained Equations (MICE).

Results: We included a total of 9,002 patients with CKD level at stage III-stage V. The missing rate was 3%, 37%, and 41% for Hba1c only, UACR only and both missing, respectively. In Hba1c only missing setting, the result of CCA suggested no difference in the risk of hyperkalemia (HR=1.01; 95% CI, 0.85-1.20), similar to the result of mean value imputation (HR=1.04; 95% CI, 0.87-1.23) and MI using MICE (HR=1.03; 95% CI, 0.87-1.23). In UACR only missing setting, the result of CCA differed from the result of mean value imputation (HR=0.87; 95% CI, 0.76-0.98) and MI using MICE (HR=0.90; 95% CI, 0.79-1.03). In both missing settings, the result of CCA also differed from the result of mean value imputation (HR=0.89; 95% CI, 0.78-1.00) and MI using MICE (HR=0.92; 95% CI, 0.80-1.04).

Conclusions: When the missing rate is high and missingness is related to the occurrence of specific symptoms, imputing these missing laboratory values is important to causal inference. Further evaluation, including subgroup analyses before and after imputation using different methods, is warranted.