1G-06 - Cross-Site Imputation for Recovering Variables without Individual Pooled Data

Background: In multi-site studies, it is common for some sites not to have recorded key variables – such as important confounders. Although it is theoretically possible to use data from sites with recorded observations to impute the missing values, this process becomes challenging when data pooling is not feasible due to logistic or legal constraints.

Objectives: We propose a new multiple imputation framework — cross-site imputation — to recover any variables across sites without the need to pool individual-level data. The approach is illustrated with a practical example of recovering confounders across Swedish hospitals.

Methods: Cross-site imputation involves four main steps: 1) key variables are identified that are not recorded at one or more sites, 2) fitting a prediction model for the target variable using sites with observed data, 3) transferring estimated model parameters to sites with missing data, and 4) imputing the target variable using these predictions. We applied this approach to a multi-site study using Swedish registry data on children born between 1997 and 2014 across five major hospitals. The hypothetical study aimed to estimate the causal effect of maternal antidepressant use during pregnancy on offspring neurodevelopmental conditions. For illustrative purposes we pretend that parental psychiatric history—a key confounder—was never recorded at two hospitals. We compared three approaches: (i) an unadjusted pooled analysis, (ii) a federated analysis adjusting for the confounder at three hospitals and not at the other two, and meta-analyzing nonetheless, and (iii) cross-site imputation to recover the missing confounder at all hospitals.

Results: Without adjusting for parental psychiatric history (i), the odds of neurodevelopmental conditions diagnosis were 1.76 (95% CI: 1.67, 1.86) times higher in offspring exposed to maternal antidepressant use during pregnancy compared to unexposed offspring. When only adjusting for available data estimate at each hospital (ii), the association attenuated to an odds ratio of 1.29 (95% CI: 1.22, 1.36). Using cross-site imputation (iii), we incorporated all hospitals with full adjustment, further attenuating the odds ratio to 1.17 (95% CI: 1.11, 1.24). The latter estimate is in line with existing estimates in the literature.

Conclusions: As multi-site studies become increasingly common in pharmacoepidemiology, and federated analysis becomes a norm, cross-site imputation offers a practical solution for recovering unrecorded variables in settings where data sharing is restricted, improving causal inference.