Diagnosing Bias From Omitted Feedback Loops

Why an Omitted Return Path Can Change an Intervention Prediction

A model can include the direct pathway from a drug target to a biological response and still mispredict the intervention effect if a downstream response feeds back to the target or an upstream process. The essential question is not simply whether a feedback mechanism exists. It is whether the intervention changes a state that travels around the omitted return path, and whether that return path changes the output being predicted.

An open-loop representation allows the intervention to produce a response but does not send that response back through the pathway that generated it. A closed-loop representation includes the return path. Comparing these two predictions helps diagnose the direction and possible scale of distortion attributable to the omitted loop. The calculations below use local, scaled changes from a baseline, rather than absolute biological concentrations.

A Local Model and Its Assumptions

For a fixed intervention, suppose \(d\) is its direct effect on state \(x\), before the return path is applied. The mediator \(z\) responds locally in proportion to \(x\), and then affects \(x\) in return:

Here \(b\) and \(r\) are signed local gains. Their units need not be identical: the product \(rb\) must be dimensionless because it maps a change in \(x\) back to a change in \(x\). Substituting the mediator equation gives \(x=d+rbx=d+Lx\). If \(L\ne 1\), the closed-loop response is therefore:

Under the open-loop boundary assumption, the return contribution is removed, so \(x_{\mathrm{open}}=d\). The signed difference between predictions is:

These expressions are local approximations. They assume that the intervention and response are in the range where the gains are approximately constant; that the same baseline, intervention, and other model inputs are used in both predictions; and that the represented return path is the only feedback path being compared. The algebraic calculation also assumes a quasi-steady response: it does not describe the transient trajectory or establish dynamic stability. If the loop is interpreted as repeated contributions \(d+Ld+L^2d+\cdots\), that geometric series converges when \(|L|<1\). The closed-form algebraic solution itself requires only \(L\ne1\), but values near 1 can yield a very large response and should prompt scrutiny of the local approximation and supporting evidence.

Reading the Sign of the Loop

When \(L>0\), the return contribution has the same sign as the change it responds to: the loop is reinforcing locally. When \(L<0\), the return contribution opposes that change: the loop is balancing locally. These labels describe the sign of the complete path, not the sign of any single link. For example, two negative links can produce a positive loop gain.

A positive loop gain below 1 in magnitude amplifies the magnitude of a direct response in this simple model, because \(1-L\) is positive and less than 1. A negative loop gain attenuates the response relative to the open-loop prediction, because \(1-L>1\). The endpoint need not follow the same pattern as \(x\): an endpoint may depend on both \(x\) and the mediator, and those contributions can reinforce or cancel one another.

Worked Comparisons Under Different Loop Signs

When the Endpoint Changes the Diagnosis

A loop can substantially change an internal state yet have little effect on a particular output, or it can shift an endpoint through a mediator even when the state response is attenuated. For an endpoint represented locally as \(y=cx+qz\), the two predictions are \(y_{\mathrm{closed}}=cx_{\mathrm{closed}}+qz_{\mathrm{closed}}\) and \(y_{\mathrm{open}}=cx_{\mathrm{open}}+qz_{\mathrm{open}}\). Their difference, not the change in \(x\) alone, determines whether omitting the loop distorts that endpoint.

In the first worked example, \(x\) became more negative with feedback, while the endpoint became more positive because the mediator contributed positively to \(y\). The endpoint comparison therefore has to retain the mediator's direct contribution. If the decision depends on a biomarker, clinical outcome, or safety measure, trace the loop all the way to that specified output rather than using an internal state as a substitute.

Diagnosis, Alternatives, and Limitations

A practical diagnosis begins with a causal trace: identify the direct intervention path, follow the response to the proposed mediator, and ask whether a plausible return path reaches an upstream state that influences the intervention response or endpoint. Then assign the signs of the links in the return path and determine the sign of their product. Finally, compare open- and closed-loop predictions using the same intervention and endpoint definition. A loop is decision-relevant when its inclusion changes the predicted output enough to matter for the intended use; a loop's mere biological existence does not establish that it must be modeled in detail.

A reduced-form alternative is to adjust an empirical dose-response relationship so it matches observed intervention data without representing the return mechanism explicitly. This may be adequate for interpolation in a well-characterized setting, but it does not reveal whether the correction will hold when the mechanism or context changes. A mechanistic closed-loop model makes the return path explicit, at the cost of needing evidence for its links and gains. An open-loop model is simpler and may be adequate when the loop is weak, the endpoint is insensitive to it, or the return path is outside the question's relevant timescale.

The linear calculation is a diagnostic, not a substitute for biological validation. Gains can vary with state, drug exposure, time, or disease context; saturation and delays can change the response; and multiple loops can interact. A large predicted amplification near \(L=1\) may signal genuine sensitivity, but it may also indicate that a local linear approximation is being extrapolated beyond its useful range. The algebra cannot establish that a proposed path exists, quantify uncertainty in its gains, or determine transient dynamics. Those judgments require evidence and a model suited to the question.