(145) Imputation of Missing Diagnosis Date and Implications for Analyses in Registries: A Simulation Study

Publication Number: 1190

Background: Patient registries provide real world evidence on treatment patterns and patient characteristics, but data is often incomplete and may not be missing at random. In the Alexion-sponsored generalized Myasthenia Gravis (gMG) Registry, which began enrolling patients in 2019, partially missing dates of diagnosis are imputed to July 1 when day and month are missing to limit information loss. However, the magnitude of bias that may be introduced by imputing or ignoring missing data has not been explored in this setting.

Objectives: To evaluate the impact of incomplete date of diagnosis (month and day missing) on bias in determining diagnosed disease duration prior to treatment and to compare simulated results to results from the gMG registry.

Methods: Data were simulated to create a complete population of 100 patients similar to the gMG registry. Missingness of diagnosis date was simulated assuming that older dates were more likely to be missing day and month: 30-50% missing below the 25th percentile, 25-30% missing in the 25th to 75th percentile and 5-15% missing above the 75th percentile. Each simulation was performed 100 times and summary statistics were calculated for those with complete data, and again with imputation of missing month and day to July 1 of the year of diagnosis. The % difference in averages of descriptive statistics were compared to summary statistics for the original simulated population with complete data. Results were also compared to findings from the gMG registry.

Results: In the simulated complete population, mean ± standard deviation (SD) disease duration from diagnosis to treatment was 4.16 ± 5.96 years. Under the conditions of nonrandom missingness using complete case analysis, estimates of mean bias for time from diagnosis to treatment initiation ranged from -8.6% to -18.2% across the scenarios evaluated. A greater percentage of missingness in older dates was associated with a larger magnitude of bias. Imputation for month and day of year reduced the mean bias to < 1%. The patterns observed in simulated data were similar to those observed from the gMG registry, where the mean ± SD time from diagnosis to treatment using a complete case approach was 2.7 years ± 4.64; n=43 but with imputation increased to 5.5 years ± 8.95; n=63.

Conclusions: Missing data may impact bias and precision of estimates. However, these results provide empirical evidence to support the use of imputation when day and month are missing but year of diagnosis is available within the gMG registry. Findings also suggest that when older dates are missing, bias may be greater. Our findings highlight importance of assessing patterns of missingness which can influence interpretation of results in observational data