This paper explains why a technical tool used to correct for hidden confounding—called a treatment bridge function—can be read as a kind of balancing weight. Proximal causal inference is a framework that tries to recover causal effects when some confounders are unobserved by using two sets of observed proxy variables: treatment confounding proxies (variables related to treatment and the hidden confounders, but not the outcome once you condition on treatment and confounders) and outcome confounding proxies (variables related to the outcome and the hidden confounders, but not to treatment once you condition). Bridge functions are functions that solve certain integral equations and are central to identification in this framework, but until now their role has felt abstract and hard to inspect.

The authors show by algebraic rearrangement that the equation defining a treatment bridge function is equivalent to a balancing condition. In this balance view the weights are functions of the treatment confounding proxies and ordinary covariates, while the quantities being balanced are the outcome confounding proxies and covariates. Because the weights are built from one set of proxies but force balance on the other, the authors call this a cross-proxy balancing structure. They also show how this perspective explains several existing identification arguments: some proofs amount to showing that balance on the outcome proxies implies balance on the unobserved confounders, a property the paper terms balance propagation.

Why this matters: recasting bridge functions as balancing weights links proximal causal inference to a large literature on covariate balance and weighting. That connection makes the mechanism for bias correction more interpretable. It also leads to practical implications: estimating a treatment bridge function can be seen as a form of balancing-weight estimation, and many proximal estimators (including those that use outcome bridge functions) can be written in an outcome-weighted form. Using these representations the authors give conditions under which common proximal estimators are numerically equivalent, and when augmented estimators reduce to simpler inverse-probability–type forms or even coincide exactly under linear assumptions.