Keep the Children That Don't Move: Shift-Aware Markov Blanket Selection

Methods that pick "causal" features to resist spurious correlations disagree on a basic question. Should a model use only the direct causes of the label, or also its direct effects? Effects such as a symptom of a disease, or a visual trace left by an object's class, are often the most predictive inputs available. They are also the inputs most likely to break when the world changes. One group of methods keeps only causes. Another keeps causes, effects, and the other causes of those effects (together called the Markov blanket, the smallest feature set that makes the label independent of everything else). Each group has evidence that it wins. We propose that both are right in different regimes, and that a model can tell which regime it is in from multi-environment data. An effect of the label should be kept exactly when the mechanism that generates it from the label stays fixed across environments. It should be dropped when that mechanism is the thing that shifts.

The resulting selector is a local, per-feature invariance test on each candidate effect, phrased in a causal graph. It returns a "stable blanket": the causes, the effects whose generating mechanisms are invariant, and the co-causes needed to condition on those effects safely. When no effect mechanism shifts, it should match full-blanket accuracy. When effect mechanisms shift, it should match causes-only robustness. In the mixed regimes that real data usually presents, it should beat both. We test this on tabular causal benchmarks with known graphs and on image benchmarks where the spurious feature (color in Colored MNIST) is itself best read as an effect of the label whose mechanism changes across environments.

The disagreement this resolves

The causes-only position has three sources. DRCFS tests each feature for a direct effect on the target using double machine learning. Causal filter selection in microarray data uses three-way information terms to discard downstream descendants. MIMB proves that intersecting blankets found in several interventional datasets leaves exactly the target's parents. Related evidence comes from Causal Foundation Models Perform Better without Post-treatment Variables, which shows that feeding downstream variables to pretrained causal models raises treatment-effect error by up to 72%.

The full-blanket position comes from CMBRL and Causal Markov Blanket Representation Learning for Out-of-distribution Generalization. Both learn latent variables with an identifiable variational autoencoder (a VAE whose latents are recoverable given environment labels). They then select parent, child and spouse latents, and their ablations show that dropping children loses accuracy.

Neither side states when its choice is correct. MIMB's intersection strips every child, including stable and highly informative ones. CMBRL's d-separation proof assumes the environment touches only spurious latents, so it silently keeps a child whose mechanism the environment changes. Our claim is that this single structural fact decides which side wins, and that it can be measured.

Core Claim

Let $E$ be an environment index, $\mathrm{PA}(Y)$ the label's parents, and $\mathrm{CH}(Y)$ its children. Call a child $C$ stable if $C \perp E \mid \mathrm{PA}(C)$, where $\mathrm{PA}(C)$ includes $Y$ and $C$'s other parents (its spouses with respect to $Y$). Let $C_s$ be the largest set of stable children that is closed in the following sense: no member has an unstable child of $Y$ among its own parents. Closure is needed because conditioning on such a parent would open the collider path $Y \to C' \leftarrow E$. The stable blanket is

$$S^\star = \mathrm{PA}(Y) \,\cup\, C_s \,\cup\, \mathrm{SP}(C_s),$$

where $\mathrm{SP}(C_s)$ are the non-$Y$ parents of the stable children.

We will prove that, if $E$ does not act directly on $Y$ and the graph has no hidden confounders and is faithful (every independence in the data reflects a missing path in the graph), then $Y \perp E \mid S^\star$. We will also prove that $S^\star$ is the most informative subset of the blanket with this property. The information it adds over parents alone, $I(Y; C_s, \mathrm{SP}(C_s) \mid \mathrm{PA}(Y))$, is the predicted accuracy gain over causes-only selection. The information it gives up relative to the full blanket is the information carried by unstable children. That second quantity is exactly what turns into out-of-distribution error under CMBRL-style selection.

The decision requires only low-order tests: each child is conditioned on $Y$ and its few spouses, never on the whole blanket. This matters because high-order conditional independence tests are the known sample bottleneck (Multi-Source Causal Feature Selection).

Research Program

The work has three parts.

Theory. We first establish the result above. We also characterize which children are lost when environments are few, using the environment-diversity lower bounds of Is Spurious Correlation Removal Always Learnable? as the constraint that tells us when per-child tests cannot be powered.

Tabular selector. We find a pooled blanket and split it into parents and children. The splitting uses MIMB's cross-environment intersection to name parents and CLIMB-style description-length scoring to orient the rest. For each child we then test invariance of its residual given $Y$ and its spouses across environments. In nonlinear settings the residuals are cross-fitted, reusing the orthogonal-score machinery of DRCFS so that the test stays valid when the regressions are flexible. We apply closure iteratively. A final test of $Y \perp E \mid S^\star$ serves as a self-check and a fallback trigger: if it fails, the selector reverts to parents only.

Latent selector for images. On top of CMBRL's learned latents, we replace the "keep all children" step with the same per-child invariance test on latent children.

We treat hidden confounding as an explicit scope limit. It is the open problem that every bridging paper in this area shares (Gap 1 in the literature). Optimal adjustment sets in graphs with hidden variables (O-sets) assume a known graph, so we do not attempt that extension here.

Baselines

Naive reference: ERM. Empirical risk minimization on all features, trained on pooled environments with the same model class. On Colored MNIST, CMBRL reports 10.3% out-of-distribution accuracy for ERM (CMBRL). On CelebA worst-group accuracy, Regulating Internal Alignment Flows reports 52.52% for ERM.

Causes-only: MIMB. MIMB's parent set is the strongest exponent of the causes-only position and the most direct competitor on graphs with interventions. It reports F1 often at or above 0.95 on ALARM and recovers all four plausible parents on the Educational Attainment data. We run it on identical environment splits with the same test family and the same significance level. We beat it if the stable blanket reaches equal worst-environment accuracy and higher average accuracy wherever stable children exist. On the Educational Attainment data, we expect no gain unless education has stable children, so this dataset doubles as a negative check.

Full blanket: CMBRL. CMBRL is the strongest full-blanket method on images. The two CMBRL papers report 74.1% and 61.8% Colored MNIST out-of-distribution accuracy respectively, and 87.5% on PACS is reported in the second. Because the two Colored MNIST figures disagree, we rerun CMBRL ourselves with identical encoders, latent dimension, environment tags and tuning budget. Our selector differs only in the child-selection step, which makes this comparison the key ablation. We beat it with a significant gain on Colored MNIST and no loss on PACS.

Strongest graph-free image baseline: Alignment-Gated Suppression (AGS). AGS reports 93.95% worst-group accuracy on CelebA and 76.09% on BAR, without group labels. On CelebA, we count a result as competitive if it is within noise of 93.95% while also returning an interpretable selected set. We do not claim to beat AGS there.

For Waterbirds, no reported absolute worst-group number for these baselines was found on Lacuna; we will run them ourselves.

Ablations. Three ablations isolate the key ingredient. The first keeps all children (CMBRL-equivalent). The second keeps stable children but drops closure. The third keeps stable children but omits their spouses, which should reintroduce collider bias.

Why not just…?

Why not just keep the full blanket and let pooling across environments sort it out? If an environment acts on a child's mechanism, the full blanket is not invariant, and no amount of pooled training fixes that. The d-separation guarantee in CMBRL simply assumes this case away. Colored MNIST is a natural instance: color is generated from the label with a flip rate that changes per environment. The decisive check is the per-child ablation on semi-synthetic ALARM and ANDES environments in which we intervene on one child at a time. Full-blanket worst-environment risk should rise with the strength of that intervention, while ours should stay flat.

Why not just use parents only, as MIMB's intersection does? This approach is safe but leaves accuracy on the table whenever children are stable. CMBRL's ablations already show that children carry discriminative signal. In regimes where all child mechanisms shift, we expect causes-only to tie with us. We will report that tie openly rather than hide it.

Why not just pick the feature subset with the best leave-one-environment-out validation? This is the most serious threat, because it needs no graph at all. Its weaknesses are cost and noise. The search over subsets is combinatorial; greedy versions are cheaper but can be fooled by spouse–child interactions, where a child helps only together with its spouse. With few environments, held-out accuracy is also noisy. We run greedy forward selection and backward elimination scored by worst held-out environment, using the same environments and model class, and with a model-fit budget matched to ours. If it matches our worst-environment accuracy at 3–5 environments, the graph structure is not earning its place, and we will scale the claim back to interpretability and test efficiency.

Why not just scale the pretrained backbone or use IRM? The Effect of Model Size on Worst-Group Generalization finds that larger models narrow but do not close the worst-group gap. Separately, Is Spurious Correlation Removal Always Learnable? shows that invariant risk minimization (IRM) fails below a sample threshold set by environment diversity. We include IRM and GroupDRO (group distributionally robust optimization) as reference points, not as main competitors.

Proposed Experiments

Tabular. On ALARM (37 nodes) and ANDES (223 nodes), we generate environments with hard and soft interventions. Each draw targets a controlled mix of parents-of-parents, children of $Y$, and spouses, never $Y$ itself. Sample sizes per environment run from 100 to 5,000, and the number of environments from 2 to 8. On the Sachs network and on synthetic nonlinear additive-noise graphs, we vary the share of shifting children from 0 to 1. Metrics are F1 of the selected set against the true stable blanket, the number of tests performed, and accuracy and worst-environment log-loss on held-out environments that apply fresh interventions.

The decisive check is a crossover. As the share of shifting children rises, full-blanket error should rise, causes-only error should stay flat but high, and stable-blanket error should track the lower envelope of the two. A further sample-size study should show where per-child tests lose power, and whether that point agrees with the environment-diversity scaling from Zhou et al.

Images. We run Colored MNIST, PACS, Waterbirds and CelebA with CMBRL's latent pipeline. Colored MNIST is the sharpest test: if the color latent is labeled a child, our test should reject it where CMBRL keeps it, and out-of-distribution accuracy should rise above our CMBRL rerun. PACS is the guard against over-pruning, since style shifts there may leave label-caused shape latents stable. A loss on PACS would show the test discards useful children.

Real data. The Educational Attainment data, split by distance to the nearest four-year college, checks that the selector does not invent children where MIMB's four parents are already the plausible answer.