(157) Standardizing to Specific Target Populations in Distributed Networks: A Proof of Concept

Publication Number: 1202

Background: In distributed networks, a coordinating centre works with multiple data partners (i.e. “nodes”). The centre distributes an analytic protocol and/or code, the nodes implement the protocol, and results are returned to the coordinating center for compilation and dissemination. These results can be difficult to interpret when treatment effect estimates differ across nodes of the distributed network, especially when nodes represent distinct populations.

Objectives: To assess whether specifying target populations improves interpretability of findings from distributed networks in a simulated example.

Methods: We simulated distributed networks of four nodes consisting of 10,000, 20,000, 40,000, and 80,000 individuals across 1,000 replicates. Within nodes, we simulated a binary outcome Y and binary treatment X. Four confounders (C1-C4) differed in distribution across the nodes and were associated with the probability of Y and logit-linearly with the probability of X. Depending on scenario, the model for the probability of Y could be linear or log-linear and the effect of X on Y could be uniform or C2 and C3 might interact with X. We estimated the within-node effect of X on Y using inverse probability of treatment weights (IPTW) and used inverse variance pooling (IVP) on these within-node estimates in a “naïve” fashion. We also specified target populations (the full network, the smaller three nodes, or the smallest node) and used inverse odds of sampling weights (IOSW) to standardize the node-specific estimates to these new targets (and then apply IVP). These results were contrasted with gold standard estimates obtained from directly analyzing the target population with IPTW.

Results: When no heterogeneity existed, IVP standardized estimates still accurately estimated effects in the target populations. When heterogeneity existed, specifying a target population and reweighting each node's estimate to resemble a common target reduced differences in estimates between nodes (e.g. within-node mean risk differences of -3.50%, -1.25%, 1.00%, and 3.25% across the nodes became mean risk differences of -0.21%, -0.30%, -0.27%, and -0.25% when targeting the three smallest nodes ). Naïve IVP resulted in estimates that were close to the full population target, though they still differed slightly (mean naïve RD of 1.48% vs gold standard RD of 1.60% vs IOSW RD of 1.62%; mean naïve RR of 1.146 vs gold standard RR of 1.152 vs IOSW RR of 1.155).

Conclusions: In simulations, specifying target populations eased comparison of the estimates produced by nodes of a distributed network and allowed examination of specific network components. Future work exploring this strategy in real-world data is necessary to understand its feasibility and impact.