Many of the most important causal questions in medicine involve loops. Body weight raises diabetes risk, and diabetes then lowers body weight. Physical inactivity raises depression risk, and depression then deepens inactivity. Bidirectional Mendelian randomization (MR) is the main tool for studying such loops in human populations. It treats inherited genetic variants as natural experiments, called instruments, and runs the analysis once in each direction. The method outputs two numbers, one for each direction. It does not say what those numbers mean. The real system is a dynamical process that settles over a lifetime, while the data are cross-sectional snapshots taken once that process has settled, a state called equilibrium. We propose to treat the pair of MR estimates as a causal abstraction: a formal, lossy summary of a fine-grained dynamical model. We will prove exactly which interventions the summary answers correctly, and build diagnostics for when it does not.
The central claim is that a genetic instrument acts like a small, sustained "shift" in one equation of a feedback system. Under three conditions (a unique stable equilibrium, local linearity over the range of genetic variation, and instruments that target a single equation), the MR ratio equals the effect of clamping the exposure at equilibrium. In that regime the two directional estimates together determine a loop gain, the extra amplification or damping the loop adds to any intervention. The loop gain is what is needed to predict how much a drug of known strength will actually move its target once the feedback responds. Outside that regime the abstraction breaks in specific, detectable ways. The project delivers the correspondence theorem, a summary-statistic estimator of loop-adjusted intervention effects, a set of failure diagnostics, and a benchmark of feedback systems with known ground truth.
Why this sits between the two directions
The work on causal framework fragmentation provides the vocabulary but stops short of loops, and the relevant papers play different roles here.
Comparing Causal Frameworks builds a framework-neutral counterfactual language. It proves that potential-outcome models are abstractions of structural models, but only for acyclic systems. We extend the spirit of that result to cycles for one restricted class of interventions.
Abstraction between Structural Causal Models reviews how a fine model maps onto a coarse one. It shows that structure-preserving abstractions cannot collapse a feedback loop into a simpler directed summary. This result constrains us. Our abstraction will be deliberately non-structure-preserving, and it will be exact only on sustained shift interventions and clamps, not on arbitrary ones. We read that paper's negative result as saying such a restriction is necessary, not as ruling the project out.
A Unifying Framework for Causal Modeling With Infinitely Many Variables describes differential-equation systems purely by how interventions map to outcomes. That supplies our fine-grained layer. The paper never connects this description to instruments or summary statistics, and that connection is the gap we fill.
On the bidirectional side, three lines of work provide the estimators and the tension we resolve.
GSMR estimates pairwise effects in both directions from genome-wide association study (GWAS) summary statistics, with a filter (HEIDI-outlier) that removes pleiotropic variants, meaning variants that affect the outcome through some other route. It reports, for example, that one standard deviation of BMI gives a 3.29-fold increase in type-2 diabetes risk, and that diabetes lowers BMI.
MR.RGM fits Bayesian linear networks that allow cycles among many traits. It assumes a static equilibrium and notes that its identifiability conditions are fragile.
Cyclic Causal Discovery from Continuous Equilibrium Data learns cyclic signaling graphs from equilibrium proteomics. It proves that the graph structure is recoverable but the quantitative parameters are not.
Set against the MR literature's claims of point identification, that non-identifiability result is the tension our theorem is meant to explain. An instrument with a known single target is exactly the kind of mechanism perturbation that restores parameter identifiability. Identifiable Exchangeable Mechanisms makes a related point in a different setting: variation in mechanisms across environments identifies causal direction. We read a genotyped population as a mixture of such environments.
Core Claim
Take a fine-grained system $\dot{x} = f(x,y) + \gamma_x G_x$, $\dot{y} = h(x,y) + \gamma_y G_y$. Here $x$ and $y$ are the two traits, $f$ and $h$ are their unknown dynamics, $\dot{x}$ and $\dot{y}$ are their rates of change over time, $G_x$ and $G_y$ are genotype scores that each enter only one equation, and $\gamma_x$ and $\gamma_y$ are the sizes of those genetic effects.
Linearize at the population equilibrium. Write $b$ for the local sensitivity of $y$'s equilibrium to $x$, and $a$ for the local sensitivity of $x$'s equilibrium to $y$. Solving the equilibrium equations then gives the following shifts when a sustained intervention adds $\delta$ to $x$'s mechanism:
$$\Delta x = \frac{\delta}{1-ab}, \qquad \Delta y = \frac{b\,\delta}{1-ab}.$$
This holds when the equilibrium is stable, which in this linearization means $ab<1$.
The Wald ratio from $G_x$ (the instrument's effect on $y$ divided by its effect on $x$) recovers $b$. The reverse ratio recovers $a$. Because the equilibrium is unique, $b$ also equals the effect of clamping $x$. This is the correspondence we will prove in general for multi-trait networks with a unique stable equilibrium: the pairwise instrument ratio equals the clamp effect.
What the pair of estimates adds beyond each single one is the factor $1/(1-\hat a\hat b)$. This factor converts a pharmacological dose $\delta$ into the change it actually produces once the loop has settled.
For weight and diabetes, $a<0$ and $b>0$, so the factor is below one. A weight-lowering mechanism therefore achieves less than its nominal strength, because falling diabetes risk removes a force that was pushing weight down. That is a specific, testable prediction that no single-direction analysis makes.
The abstraction fails when the conditions fail, and each failure has a signature. If the loop has not settled at the age of measurement, the estimated ratio drifts with age. If the dynamics are nonlinear and people sit at different equilibria, the estimate is an average of local slopes and does not transfer to large interventions. If an instrument enters both equations, the pleiotropy problem returns, and the loop factor inherits bias from both directions.
Research Program
The theoretical component has three results. The first is a correspondence theorem stating when the equilibrium cyclic structural model is an exact abstraction of the dynamical model on the set of sustained shift and clamp interventions. The second is a counterexample family showing that the abstraction fails for transient interventions, which makes precise where structure-preserving abstraction breaks. The third is an extension to multi-trait networks, where loop-adjusted effects come from $(I-B)^{-1}$. Here $B$ is the matrix of direct effects, with entry $B_{ij}$ giving the direct effect of trait $j$ on trait $i$, and $I$ is the identity matrix.
The methodological component is an estimator of loop-adjusted dose effects from summary statistics. It propagates uncertainty from both directions using the delta method or posterior samples from MR.RGM. It also reports the stability condition $\hat a\hat b<1$ as a quantity with its own credible interval. Three diagnostics accompany the estimator. The equilibrium diagnostic compares ratios estimated in age-stratified summary statistics, which can be computed from UK Biobank individual data. The local-linearity diagnostic tests whether ratios for individual variants depend on the size of each variant's effect. The target-specificity diagnostic reuses outlier filtering in both directions at once.
The empirical component is a benchmark of simulated feedback systems with known ground truth, plus two real-data tests.
Baselines
The naive reference point ignores the loop. It fits an acyclic graph with PC or GES from causal-learn and applies the standard acyclic rules for computing intervention effects, or equivalently it runs one-direction inverse-variance-weighted MR and reads the result as a dose effect. This is what practitioners do by default. Beating it only requires showing that loops matter at all, so it is the floor, not the bar.
The strongest pairwise competitor is GSMR, run in both directions on identical summary statistics with the same variant selection and outlier filtering. GSMR is the best-established pleiotropy-robust bidirectional estimator. Its reported BMI→T2D estimate of 3.29-fold per standard deviation is the published reference against which our reanalysis of that pair must be checked for consistency.
The strongest multi-trait competitor is MR.RGM, given the same data and the same MCMC budget. Its direct-effect matrix is composed through $(I-B)^{-1}$. For its accuracy on simulated networks, no reported result was found on Lacuna; the file states only that it achieved higher true-positive and lower false-positive rates than OneSampleMR. We therefore rerun it on our simulations. MR.RGM plus matrix inversion is in effect the linear version of our estimator. Beating it means one of two things: giving better-calibrated dose predictions in regimes where it is misspecified, or correctly flagging when its output should not be trusted.
On the proteomics data, the competitor is the cyclic equilibrium model of Cyclic Causal Discovery, which reports a negative log-evidence of 72.9 ×10³ against 79.2 ×10³ for the acyclic reconstruction by Sachs et al. We will reimplement it on the same eight conditions.
Two ablations isolate our two ingredients. The first applies the loop correction without the diagnostics. The second applies the diagnostics to unadjusted single-direction estimates.
Why not just…?
Why not just use two-stage least squares on simultaneous equations? In the linear, stable, single-target regime this is a real threat. The loop factor is algebra that simultaneous-equation estimators already imply, and we will not claim otherwise; MR.RGM is our linear estimator. What those estimators do not supply is a statement of which interventions their coefficients answer when the true system is dynamical and nonlinear, or any warning when that answer is wrong. The case for the proposal therefore rests on the diagnostics. If, in simulation, the diagnostics do not flag equilibrium, linearity, or target violations better than chance, the project reduces to "run MR.RGM and invert the matrix," and we will report that.
Why not just unroll the loop over time into an acyclic graph? GWAS data contain no time series. The proteomics data of Cyclic Causal Discovery were collected only at equilibrium. The dynamical treatment of time in A Unifying Framework for Causal Modeling runs into an infinite-regress problem when defining minimal causes. In our simulations we do have trajectories, so we include a time-unrolled estimator as an oracle. This shows how much information equilibrium data lose relative to full trajectories.
Why not just run off-the-shelf causal discovery? The main algorithms in causal-learn assume acyclicity. In the multi-environment setting of Identifiable Exchangeable Mechanisms, FCI, GES, NOTEARS, and DirectLiNGAM were reported to perform near chance. We include PC and GES as the naive baseline, so this failure is measured directly rather than assumed.
Proposed Experiments
Simulation benchmark. Each simulated system has two to six traits. The dynamics come from four families: linear, saturating, threshold (thermostat-like), and systems with slow relaxation, meaning they have not settled at the age of measurement. Loop gains span damping and amplification. Genotypes are simulated with realistic effect-size distributions and a controllable fraction of variants acting on two equations. From each system we generate GWAS-style summary statistics at several simulated ages. Ground truth is the equilibrium response to sustained dose-specified interventions, computed by integrating the dynamics. The primary metric is the error in predicted equilibrium response, with interval coverage as a secondary metric. For the diagnostics, the metric is detection AUROC against known violations.
The decisive check has two parts. In the linear regime, loop-adjusted predictions should beat one-direction MR by a margin that grows with $|ab|$, and match MR.RGM-plus-inversion. In the violated regimes, the diagnostics should fire on exactly the systems where every estimator's dose prediction is wrong.
Proteomics test. We treat the known-target perturbations in the Sachs et al. (2005) data as instruments. The model is fit on seven of the eight conditions and predicts the equilibrium mean shifts in the held-out condition. We compare against the acyclic reconstruction and the reimplemented cyclic model. The two-cycles reported in Cyclic Causal Discovery (Mek–PKC, Akt–Erk, Mek–PKA) are where loop correction should matter most.
Real GWAS pairs. For BMI and T2D, the prediction is a loop factor below one. For inactivity and MDD, the reported reverse odds ratios are small, 1.01 and 1.03 in the inactivity–depression study, so the prediction is a factor near one. Neuroticism and postpartum depression serve as a negative control: the reverse effect was null in the postpartum depression study, so the factor must be indistinguishable from one. A clearly non-unit factor there would indicate that the pipeline manufactures loops.
For the binary diseases in these pairs, we will state explicitly that the linearization is taken on the liability scale, the latent continuous risk underlying a yes/no diagnosis. We will treat conclusions about those diseases as less secure than the simulation results.
References
- Comparing Causal Frameworks: Potential Outcomes, Structural Models, Graphs, and Abstractions
- Abstraction between Structural Causal Models: A Review of Definitions and Properties
- A Unifying Framework for Causal Modeling With Infinitely Many Variables
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- MR.RGM: an R package for fitting Bayesian multivariate bidirectional Mendelian randomization networks
- Cyclic Causal Discovery from Continuous Equilibrium Data
- Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning
- Causal-learn: Causal Discovery in Python
- Dissecting the causal association between social or physical inactivity and depression: a bidirectional two-sample Mendelian Randomization study
- Causal effects of neuroticism on postpartum depression: a bidirectional mendelian randomization study