Omitted Variable Bias¶

Correct for the distortion in a regression coefficient when a left-out variable both causes the outcome and correlates with an included regressor, so the estimate absorbs the omitted effect as the signable product of two relationships.

Core Idea¶

Omitted variable bias (OVB) is the systematic distortion in ordinary least-squares regression coefficients that arises when the model excludes a variable that both (a) is a true determinant of the outcome and (b) is correlated with one or more included regressors. When these two conditions hold, the included-variable coefficient absorbs a portion of the omitted variable's effect: the OLS estimate of the coefficient on included regressor X₁ equals its true direct coefficient β₁ plus the product of the omitted variable's true coefficient β₂ and the regression slope of the omitted variable X₂ on X₁ (denoted δ), so the bias is β₂ · δ. The direction of the bias — whether the estimate is inflated or attenuated — can be read qualitatively from the signs of β₂ and δ without knowing their magnitudes, giving the analyst a sign-of-bias diagnostic from substantive domain knowledge alone. The canonical illustration is the ability-bias problem in Mincerian wage regressions: unobserved ability is positively correlated with both schooling (δ > 0) and wages (β₂ > 0), so OLS estimates of the return to schooling are upwardly biased; Card, Krueger, Angrist, and co-authors used twin-difference and quarter-of-birth instrumental-variable designs to estimate bias-corrected returns, typically finding OLS inflated by 10–30%. Uncorrelated omitted variables — variables that determine the outcome but are orthogonal to all included regressors — do not produce OVB; they increase residual variance and reduce precision but leave included coefficients unbiased. The correction strategies are substrate-specific: instrumental variables exploiting variation in X₁ uncorrelated with the omitted variable, fixed-effects models that difference out time-invariant unobserved heterogeneity, regression discontinuity designs, randomisation, and sensitivity analyses that bound the magnitude of omitted-variable effect needed to overturn a conclusion (the E-value of VanderWeele and Ding 2017).

Structural Signature¶

Sig role-phrases:

the fitted OLS regression — the model whose coefficient on an included regressor X₁ is the object whose trustworthiness is in question
the omitted determinant — a left-out variable X₂ that is a true cause of the outcome (β₂ ≠ 0)
the correlation channel — the association δ between the omitted variable and the included regressor; the conduit through which contamination flows
the two-condition gate — bias occurs only if the omission both determines the outcome and correlates with a regressor; orthogonal determinants only cost precision, leaving coefficients unbiased
the β₂·δ bias term — the coefficient mixture: the estimate equals the true direct effect plus the omitted effect flowing through the correlation
the readable sign (engineered guarantee) — bias direction (inflate vs. attenuate) follows from the signs of β₂ and δ from domain knowledge alone, no data on the missing variable needed
the channel-breaking remedies — instruments (variation in X₁ uncorrelated with X₂, killing δ), fixed effects (differencing out a time-invariant omission), discontinuity/randomisation that sever the channel by construction
the sensitivity bound (what it deliberately leaves implicit) — when no correction is available, the reverse question "how large must β₂·δ be to overturn the conclusion?" (E-value, Rosenbaum bounds) converts an unmeasured threat into a robustness verdict

What It Is Not¶

Not caused by every left-out variable. Omission biases a coefficient only when both conditions hold: the variable determines the outcome and correlates with an included regressor. A determinant that is orthogonal to the regressors produces no OVB at all; the vast cloud of unmeasured variables failing either condition is harmless to the point estimate.
Not a loss of precision. OVB contaminates the point estimate — it shifts the coefficient away from its true value. An orthogonal omitted determinant does something different: it inflates residual variance and widens the standard error while leaving the coefficient unbiased. The two failures must be kept apart; only the correlated omission moves the estimate itself.
Not a bias of unknown direction. Because the contamination equals β₂·δ — the omitted variable's effect on the outcome times its correlation with the regressor — its sign is readable from the signs of those two relationships alone, using domain knowledge and no data on the missing variable. The analyst can state whether the estimate is inflated or attenuated, not merely that it is "wrong."
Not confounding in general, nor Simpson's paradox. OVB is the specific regression-coefficient rendering of confounding, expressed through the OLS β₂·δ formula. Confounding is the substrate-independent causal structure; Simpson's paradox is its dramatic discrete-contingency version. Treating OVB as identical to either loses the formula — the readable sign and the bound — that is its distinctive content.
Not measurement error or simultaneity. These are sibling sources of endogeneity under the same OLS-assumption-violation umbrella, but distinct mechanisms: measurement error attenuates through a noisy regressor, simultaneity through reverse causation. OVB specifically flows through a correlated, outcome-relevant omission, and the instrument that fixes one need not fix the others.

Scope of Application¶

Omitted variable bias lives wherever regression-based inference is drawn from observational data; its reach is bounded by that formalism — every habitat runs the same OLS machinery, so the β₂·δ decomposition, the sign-of-bias diagnostic, and the remedy menu carry untranslated. (The discrete-contingency Simpson's version, the latent-state and spurious-correlation versions, and the broader pattern belong to the parent prime, confounding, not here.)

Econometrics — the canonical home: the ability-bias problem in returns-to-schooling, treated identically in Greene, Wooldridge, and Mundlak, with instruments and twin designs supplying the bias-corrected estimate.
Epidemiology — the same object as confounding adjustment, carrying unmeasured-confounding sensitivity tools (the E-value, Rosenbaum bounds) to ask how strong an omission must be to overturn an association.
Education research and policy evaluation — selection on unobservables in non-randomised programmes, where an omitted determinant of both participation and outcome biases the estimated effect.
Empirical political science and sociology — the endogeneity worry in cross-country and cross-unit regressions, an omitted institutional or cultural variable contaminating the included coefficient.
Health-services and labour research — wage-gap and observational treatment-effect estimates, where a correlated unmeasured determinant inflates or attenuates the reported effect.
Causal inference for A/B testing and machine learning — covariate imbalance when randomisation fails, the regression rendering of confounding reappearing in experiment analysis and post-stratification.

Clarity¶

Naming omitted variable bias turns a vague unease about regression — "can I trust this coefficient?" — into a precise, two-pronged test. The concept makes clear that a left-out variable is harmless unless it satisfies both conditions at once: it must influence the outcome and be correlated with an included regressor. That sharpens a distinction practitioners routinely blur — between omitted variables that bias coefficients and omitted variables that merely cost precision. A determinant of the outcome that is orthogonal to the regressors inflates residual variance but leaves the estimates unbiased; only the correlated determinant contaminates them. Knowing which kind one faces tells the analyst whether the worry is about the point estimate or merely about the standard error.

Its most useful clarity is that the bias has a legible sign. Because the contamination is the product of the omitted variable's effect on the outcome and its association with the included regressor, an analyst who knows only the signs of those two relationships — from substantive domain knowledge, with no data on the missing variable at all — can say whether the reported coefficient is inflated or attenuated, and therefore whether the true effect is even larger or smaller than estimated. This converts an unmeasured threat into a directional argument a referee can interrogate, and it reframes the remedy: an instrument is recognizable as exactly a source of variation in the regressor that is uncorrelated with the omitted variable, and a sensitivity analysis becomes the sharp question "how strong would the missing confounder have to be to overturn this conclusion?" The concept thus separates three things a casual reading of a regression conflates — a coefficient that is biased, one that is merely imprecise, and one that is robust to plausible omissions — and gives each its own diagnostic.

Manages Complexity¶

The threat that an observational regression coefficient is untrustworthy is, in the raw, unbounded: any of indefinitely many unmeasured variables might be contaminating the estimate, and without a structure for thinking about them the analyst is left with a diffuse, ungovernable anxiety and no way to say which of the infinite possible omissions matter. Omitted variable bias collapses that open-ended worry to a two-condition gate. A left-out variable can distort a coefficient only if it both determines the outcome and correlates with an included regressor; every variable failing either condition can be set aside as harmless to the point estimate. This immediately partitions the unmanaged space of "things I didn't measure" into three tracked categories with distinct consequences — variables that bias the estimate (both conditions hold), variables that merely cost precision (determine the outcome but are orthogonal to the regressors, inflating residual variance while leaving coefficients unbiased), and variables irrelevant to both — so the analyst reads off whether a given omission threatens the coefficient or merely the standard error, rather than treating all unmeasured variables as an undifferentiated cloud of doubt.

The sharper compression is that the bias, once a variable clears the gate, is not an unknown of unknown direction but a quantity with a legible sign read from domain knowledge alone. Because the contamination equals the product β₂ · δ — the omitted variable's effect on the outcome times its association with the included regressor — the analyst who knows only the signs of those two relationships, with no data on the missing variable whatsoever, can state whether the reported coefficient is inflated or attenuated, and therefore whether the true effect lies above or below the estimate. A whole regression's worth of worry about one suspected confounder reduces to tracking two sign bits and reading their product. This turns an unmeasured threat into a directional argument a referee can interrogate, and it organises the entire remedy space around the same small parameter set: an instrument is recognisable as precisely a source of variation in the regressor that is uncorrelated with the omitted variable (breaking δ), a fixed-effects design as one that differences out a time-invariant omitted determinant, and a sensitivity analysis as the bounded question "how large would β₂ · δ have to be to overturn the conclusion?" Instead of re-deriving, for each observational study, an open-ended account of everything that might be wrong, the analyst runs the two-condition gate, signs the product, and reads off both the direction of the bias and the shape of the fix — a potentially infinite-dimensional threat reduced to a two-term formula with a decidable branch structure.

Abstract Reasoning¶

The signature move is signing the bias from domain knowledge with no data on the missing variable. Because the contamination equals β₂ · δ — the omitted variable's effect on the outcome times its association with the included regressor — the analyst who knows only the two signs can state whether the reported coefficient is inflated or attenuated, and therefore whether the true effect lies above or below the estimate. The characteristic inference runs from substantive belief about the world to the direction of the error: unobserved ability raises both schooling (δ > 0) and wages (β₂ > 0), so the product is positive and the OLS return to schooling is upwardly biased — the true return is smaller than reported. The move converts an unmeasured threat into a directional argument a referee can interrogate, requiring the analyst to track just two sign bits rather than the magnitude of anything unmeasured.

The gating move runs first and is a two-condition boundary test that partitions everything left out of the model. A left-out variable distorts a coefficient only if it both determines the outcome and correlates with an included regressor; the analyst reasons from those two conditions to sort every omission into three tracked categories with distinct consequences — variables that bias the estimate (both conditions hold), variables that merely cost precision (a true determinant orthogonal to the regressors, inflating residual variance while leaving coefficients unbiased), and variables irrelevant to both. The inference is from "does this omission clear both conditions?" to "does it threaten the point estimate or merely the standard error?" — and the move's discipline is that it lets the analyst set aside the infinite cloud of unmeasured variables and worry only about the ones that pass the gate.

The interventionist move reads the remedy off the same β₂ · δ structure rather than reaching for a generic fix. Recognizing that the bias flows through the correlation channel δ, the analyst infers what each correction must do to it: an instrument is recognizable as precisely a source of variation in the regressor that is uncorrelated with the omitted variable (breaking δ); a fixed-effects design is one that differences out a time-invariant omitted determinant; a discontinuity or randomization severs the channel by construction. The inference runs from which term of the bias formula a method neutralizes to whether it actually removes this particular bias — so the choice of correction is derived from the structure of the contamination, not applied by rote.

Finally there is a sensitivity / robustness move that bounds the threat when no correction is available. Reasoning from the bias formula in reverse, the analyst asks the sharp counterfactual question: how large would β₂ · δ have to be — how strong an outcome effect and how strong a correlation — to overturn this conclusion? The inference runs from the observed estimate and its margin to a threshold on the missing confounder's strength (the E-value, Rosenbaum-style bounds), converting "maybe something is omitted" into "any omitted variable weak enough to be plausible cannot flip the sign." This separates three states a casual reading conflates — a coefficient that is biased, one that is merely imprecise, and one that is robust to all plausible omissions — and assigns the third its own defensible verdict.

Knowledge Transfer¶

Within the substrate of regression-based inference from observational data, omitted variable bias transfers as mechanism, and across many fields, because every one of them runs the same OLS machinery. It is the canonical home concern in econometrics (the ability-bias example in returns-to-schooling, treated identically in Greene, Wooldridge, and Mundlak); in epidemiology it is the same object as confounding adjustment, with unmeasured-confounding sensitivity analyses (the E-value, Rosenbaum bounds) attached; in education research and policy evaluation it is selection on unobservables in non-randomised programmes; in empirical political science and sociology it is the endogeneity worry in cross-country and cross-unit regressions; in health-services and labour research it underlies wage-gap and observational treatment-effect estimates; and in causal inference for A/B testing and machine learning it is covariate imbalance when randomisation fails. Across all of these the transfer is literal because the formalism is shared: the β₂·δ decomposition (the included coefficient becomes its direct effect plus the omitted variable's effect flowing through the included-omitted correlation) holds unchanged, so the sign-of-bias diagnostic from domain knowledge alone carries, the two-condition gate (a determinant of the outcome and correlated with a regressor) carries, and the remedy menu carries — instruments that break δ, fixed effects that difference out a time-invariant omitted determinant, regression discontinuity and randomisation that sever the channel by construction, and sensitivity bounds that ask how strong an omission must be to overturn the conclusion. The vocabulary travels untranslated because it is regression vocabulary used by all these fields at once; what moves is not an analogy to regression but regression, applied to different substantive questions.

Beyond the regression formalism the structure recurs, but as a shared abstract mechanism expressed in other machinery, and the cross-domain weight belongs to the parent, not to OVB's named formula. The general pattern — a common cause that influences both the putative cause and the outcome contaminates the estimated effect — is the prime confounding, and OVB is precisely its regression-coefficient rendering. The same underlying structure surfaces, in different formalisms, as Simpson's paradox (its dramatic discrete-contingency version), as hidden-variable / latent-state bias (its factor-model and time-series version), and as spurious correlation (its unconditional version). These are not separate phenomena OVB transfers to — they are co-renderings of one causal structure, and when a cross-domain lesson is needed it is confounding (with hidden_variable and selection_bias nearby) that carries it, because that is the substrate-independent statement. The home-bound cargo is everything that makes OVB specifically the regression rendering: the OLS coefficient-mixture form, the β₂·δ product and its two readable signs, the residual-variance-versus-bias distinction, and the instrument-as-uncorrelated-variation device — none of which has a referent outside a fitted regression. So importing "omitted variable bias" into a setting with no regression coefficient — calling any unmodelled influence an "omitted variable" — is analogy at best, borrowing the contamination shape while leaving behind the formula that makes the bias signable and boundable. The disciplined position is that the formalism-bound concept transfers wherever regression is run, while the deeper cross-domain recurrence is the work of the confounding prime it instantiates (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

Take the textbook returns-to-schooling regression. The true wage model is (log) wage = β₁·schooling + β₂·ability + error, but ability is unobserved, so the analyst regresses wage on schooling alone. The fitted coefficient does not recover β₁; it equals β₁ + β₂·δ, where δ is the slope from regressing the omitted ability on schooling. Both signs are known from substantive knowledge: ability raises wages (β₂ > 0) and more able people acquire more schooling (δ > 0), so the bias β₂·δ is positive — the OLS return to schooling is inflated. Put numbers to it for illustration: if the true return β₁ is 0.07 (7% per year) and β₂·δ works out to 0.02, OLS reports 0.09, overstating the causal return by roughly 29%. Angrist and Krueger's quarter-of-birth instrument and twin-difference studies, exploiting schooling variation unrelated to ability, recover lower, bias-corrected returns — consistent with OLS being inflated on the order of 10–30%.

Mapped back: The wage-on-schooling model is the fitted OLS regression; ability is the omitted determinant (β₂ ≠ 0); its correlation with schooling is the correlation channel δ. Ability clears the two-condition gate (affects wages and correlates with schooling), so the estimate carries the β₂·δ bias term. Signing that product from domain knowledge alone is the readable sign, and the quarter-of-birth instrument is a channel-breaking remedy — schooling variation uncorrelated with ability, killing δ.

Applied / In Practice¶

Hormone-replacement therapy (HRT) is the classic epidemiological cautionary tale. Large observational cohorts, notably the Nurses' Health Study, found that postmenopausal women taking HRT had substantially lower rates of coronary heart disease, and for years HRT was prescribed partly for cardioprotection. But the estimate was contaminated: women who took HRT were, on average, wealthier, leaner, more health-conscious, and better connected to medical care — traits that independently lower heart-disease risk and correlate with taking HRT ("healthy-user bias"). When the Women's Health Initiative randomized women to HRT or placebo (results 2002), it found HRT did not protect the heart and even raised risk for some outcomes. Randomization severed the link between treatment and the omitted health-status variables, overturning the observational conclusion.

Mapped back: The observational HRT–CHD regression is the fitted OLS regression; underlying health-consciousness/socioeconomic status is the omitted determinant, which both lowers heart disease and correlates with taking HRT — clearing the two-condition gate through the correlation channel. Because both relationships bias the estimate toward apparent protection, the readable sign was negative (spuriously beneficial). The WHI randomization is the channel-breaking remedy that severs the treatment–confounder correlation by construction, exposing the true near-null-to-harmful effect.

Structural Tensions¶

T1: Readable sign versus unknowable magnitude (half an answer, cheaply won). OVB's celebrated economy is that the direction of the bias reads off two domain-knowledge sign bits — the signs of β₂ and δ — with no data on the missing variable at all. This converts an unmeasured threat into a directional argument a referee can interrogate. But sign without magnitude is only half a verdict: knowing the return to schooling is "inflated" does not say whether the true effect is 90% of the estimate or near zero, nor whether a large bias could flip the conclusion entirely. The very feature that makes OVB tractable — that the sign needs no magnitudes — is what leaves it unable to say how much, so the cheap directional diagnostic must be supplemented by sensitivity bounds (E-values) to become an actual bound on the effect. The elegance and the limitation are the same fact. Diagnostic: Does the conclusion depend only on the direction of the bias (where the sign diagnostic suffices) or on its size (where two sign bits are silent and a magnitude bound is required)?

T2: A decidable gate versus speculative inputs (reasoning about the variable you never measured). The two-condition gate — a variable biases the coefficient only if it both determines the outcome and correlates with an included regressor — is what lets the analyst set aside the infinite cloud of omissions and worry only about the ones that pass. But the gate's inputs are, by construction, facts about an unmeasured variable: whether it affects the outcome and whether it correlates with a regressor are exactly the relationships there is no data on. So the filter that looks crisply decidable is applied using substantive belief and assumption, not measurement, and a wrong prior about whether a variable clears the gate wrongly dismisses a real confounder or wastes effort on a harmless one. The tension is that the gate's discipline depends on judgments about relationships the analyst has, and can have, no direct evidence for. Diagnostic: Is there a defensible substantive basis for believing this omitted variable does (or does not) both affect the outcome and correlate with the regressor — or is the gate being decided on a guess about an unmeasured relationship?

T3: Single-confounder clarity versus multiple-omission tangling (signs that do not compose). The β₂·δ diagnostic is beautifully clean for one suspected omitted variable: sign the two relationships, read the product. But real models omit many variables at once, and their biases sum — with individual terms of opposite sign — so the net contamination need not share the direction of any single confounder, and the readable-sign logic does not compose. Two omissions each "inflating" and one "attenuating" can net out to anything, and a confident sign argument built on the one confounder the analyst happened to think of can be reversed by a second they did not. The tension is that the diagnostic's clarity lives in the single-variable case while observational reality is multi-confounder, and aggregating sign bits across omissions is exactly what the clean formula does not license. Diagnostic: Is there plausibly a single dominant omitted confounder (sign logic applies) or several with conflicting sign contributions (where the net bias direction is not readable from any one of them)?

T4: Controlling for the omission versus over-controlling (the fix that makes new bias). The natural remedy OVB suggests is to include the missing determinant, and more generally to add controls until the worrying correlation is absorbed. But controls are not monotonically good: conditioning on a collider (a common effect of the regressor and the outcome) opens a spurious path, and conditioning on a mediator (a variable on the causal pathway from regressor to outcome) removes part of the very effect being estimated. So the reflex the OVB framing encourages — throw in more covariates to kill δ — can manufacture collider bias or mediator bias worse than the omission it cures. The tension is that OVB pushes toward inclusion while good causal-graph reasoning insists some variables must be left out, and the formula that diagnoses omission does not by itself tell you which additions are safe. Diagnostic: Is the candidate control a genuine confounder (a common cause, safe to include) or a collider/mediator (whose inclusion introduces new bias) — and does the OVB framing pressure toward adding it regardless?

T5: Channel-breaking remedies versus their own untestable premises (trading a known bias for an unverifiable assumption). The remedy menu — instruments, fixed effects, discontinuity, randomization — is derived elegantly from which term of β₂·δ each neutralizes. But every correction short of randomization buys the removal of OVB with a fresh, often untestable assumption: an instrument requires the exclusion restriction (it affects the outcome only through the regressor), which cannot be verified from data; fixed effects remove only time-invariant omitted determinants and leave time-varying ones; regression discontinuity identifies only a local effect at the cutoff. The tension is that the cure for an unmeasurable bias imports its own unverifiable premises, so "corrected" estimates rest on assumptions as strong as the confounding they displace — the analyst trades a bias they can at least sign for an identifying assumption they must simply assert. Diagnostic: Does the chosen remedy's identifying assumption (exclusion restriction, time-invariance, local validity) actually hold here, or has one untestable premise merely been swapped for another?

T6: Autonomy versus reduction (a regression rendering or the confounding prime). Within regression-based inference OVB transfers as full mechanism — the β₂·δ decomposition, the sign diagnostic, the two-condition gate, and the remedy menu carry untranslated across econometrics, epidemiology, policy evaluation, and A/B testing, because all run the same OLS machinery. But its cross-domain content reduces: the general structure — a common cause influencing both the putative cause and the outcome contaminates the estimated effect — is the prime confounding, and OVB is precisely its regression-coefficient rendering, with Simpson's paradox (the discrete-contingency version), hidden-variable bias (the latent-state version), and spurious correlation (the unconditional version) as co-renderings, and selection_bias nearby. The home-bound cargo is everything formula-specific: the OLS coefficient mixture, the signable β₂·δ product, the residual-variance-versus-bias distinction, the instrument device — none of which has a referent outside a fitted regression. The tension is between a formalism-bound concept that owns regression practice and the recognition that its portable causal content is confounding. Diagnostic: Resolve toward confounding (with Simpson's, hidden_variable, selection_bias as neighbors) when the setting has no regression coefficient; toward named omitted variable bias when a fitted OLS coefficient absorbs a correlated omission through the β₂·δ channel.

Structural–Framed Character¶

Omitted variable bias sits in the middle of the structural–framed spectrum — best read as mixed: a formal statistical property whose core is a real substrate-neutral causal structure, but which is bound to the regression-estimation practice and named as a defect. On evaluative_weight it is mildly loaded in the estimate-quality sense — "bias" names a distortion, a coefficient shifted away from its true value — but this is a technical property of an estimator, not a moral verdict on any person, so it sits just above a purely neutral mechanism. On human_practice_bound the signal is genuinely mixed, which fixes the placement: the underlying confounding — a common cause influencing both the putative cause and the outcome — is a causal fact of the world that holds observer-free, but "omitted variable bias" specifically is a property of a fitted regression, an artifact of the human analytical practice of OLS estimation, and dissolves where no coefficient is being estimated. On institutional_origin it is neutral-to-structural: the β₂·δ formula and the OLS machinery are mathematical facts of a method, not artifacts of a survey or agency. On vocab_travels it scores low: the OLS coefficient mixture, the signable β₂·δ product, the residual-variance-versus-bias distinction, and the instrument device have no referent outside a fitted regression. And on import_vs_recognize the transfer is layered — within regression-based inference it ports literally as mechanism across econometrics, epidemiology, policy evaluation, and A/B testing (all running the same OLS machinery), but importing "omitted variable bias" into a setting with no regression coefficient is analogy, borrowing the contamination shape while leaving behind the formula.

The one portable structural skeleton is confounding — a common cause that influences both the putative cause and the outcome contaminates the estimated effect — with hidden_variable and selection_bias nearby and Simpson's paradox, latent-state bias, and spurious correlation as co-renderings in other formalisms. That skeleton is genuinely substrate-independent and is the real-world causal structure OVB makes precise. But it does not pull OVB toward the structural pole, because confounding is exactly what OVB instantiates from its umbrella as the regression-coefficient rendering, not what makes "omitted variable bias" itself travel: the cross-domain reach belongs to confounding, while the β₂·δ formula, the two readable signs, and the instrument-as-uncorrelated-variation device stay bound to the regression formalism. Its character: an evaluatively near-neutral formal statistical property with a genuinely structural confounding core, held in the mixed band because it is the regression-estimation rendering of that core — practice-bound and formula-bound where the confounding prime it instantiates is neither.

Structural Core vs. Domain Accent¶

This section decides why omitted variable bias is a domain-specific abstraction and not a prime — why its portable causal content is a parent prime while its formula stays bound to regression.

What is skeletal (could lift toward a cross-domain prime). Strip the regression and a thin relational structure survives: a common cause that influences both the putative cause and the outcome contaminates the estimated effect, so an association absorbs a portion of the common cause's influence to the extent the two are linked. The portable pieces are abstract: a target relationship whose estimate is in question, a hidden common cause, and a contamination that flows through the link between them. This skeleton is genuinely substrate-portable, which is why the catalog carries it as the parent confounding that OVB instantiates — the substrate-independent causal structure — with hidden_variable and selection_bias nearby, and Simpson's paradox (the discrete-contingency version), latent-state bias (the factor-model version), and spurious correlation (the unconditional version) as co-renderings in other formalisms. But it is the core OVB shares with those co-renderings, not what makes OVB the distinctive thing it is.

What is domain-bound. What is distinctive is the regression formalism, and none of it survives extraction. The fitted OLS regression whose coefficient is in question; the β₂·δ bias term — the included coefficient equals its true direct effect plus the omitted variable's effect flowing through the included-omitted correlation; the two readable signs that make the bias direction legible from domain knowledge alone; the two-condition gate (a determinant of the outcome and correlated with a regressor) that partitions omissions into biasing, precision-costing, and irrelevant; the residual-variance-versus-bias distinction; and the instrument-as-uncorrelated-variation device (plus fixed effects, regression discontinuity, E-value/Rosenbaum sensitivity bounds) — none of these has a referent outside a fitted regression. The decisive test: importing "omitted variable bias" into a setting with no regression coefficient — calling any unmodelled influence an "omitted variable" — borrows the contamination shape while leaving behind the formula that makes the bias signable and boundable. Remove the OLS machinery and there is no omitted variable bias in particular, only the bare confounding parent.

Why this does not clear the prime bar. A prime's vocabulary travels and its transfer is recognition of the same mechanism, not analogy. OVB's transfer is bimodal. Within regression-based inference it moves as full mechanism, literally — the β₂·δ decomposition, the sign-of-bias diagnostic, the two-condition gate, and the remedy menu carry untranslated across econometrics, epidemiology, education and policy evaluation, political science, health-services research, and causal inference for A/B testing and ML, because every one of those runs the same OLS machinery (what moves is not an analogy to regression but regression, applied to different questions). Beyond the regression formalism the structure recurs only as co-renderings of one causal structure in other machinery — Simpson's paradox, hidden-variable bias, spurious correlation — not as OVB transferring to them; these are alternative renderings of confounding, and when a cross-domain lesson is needed it is confounding (with hidden_variable, selection_bias) that carries it, because that is the substrate-independent statement. The genuinely portable structure is not OVB but confounding, of which OVB is the regression-coefficient rendering. So the cross-domain reach belongs to the parent; the disciplined move is to carry confounding when the setting has no regression coefficient, and reserve "omitted variable bias" for where a fitted OLS coefficient absorbs a correlated omission through the β₂·δ channel. It clears the domain-specific bar comfortably for regression practice, but its only substrate-spanning content is already carried, in more general form, by the prime it instantiates.

Relationships to Other Abstractions¶

Current abstraction Omitted Variable Bias Domain-specific

Parents (2) — more general patterns this builds on

Omitted Variable Bias presupposes Endogeneity Domain-specific

Omitted-variable bias presupposes endogeneity because its two-condition gate makes an included regressor correlated with the regression error.
Omitted Variable Bias is a decomposition of, conditional Confounding Prime

Omitted-variable bias exposes confounding when the omitted determinant is a pre-treatment common cause of the included regressor and outcome; correlation arising otherwise is excluded.

Hierarchy paths (17) — routes to 10 parentless roots

Omitted Variable Bias → Endogeneity → Regression → Signal Extraction

Show alternative paths (16)

Not to Be Confused With¶

Confounding (the parent prime). The substrate-independent causal structure — a common cause influencing both the putative cause and the outcome contaminates the estimated effect. Omitted variable bias is the regression-coefficient rendering of confounding, expressed through the β₂·δ formula. Tell: is there a fitted OLS coefficient absorbing a correlated omission (OVB), or the bare causal structure in any formalism (confounding)? (Treated more fully as the umbrella it instantiates in Structural Core vs. Domain Accent.)
Simpson's paradox. The discrete-contingency co-rendering of the same confounding structure — an association that reverses sign when a subgroup variable is conditioned on. It is a co-rendering of confounding in contingency tables, not a use of the OVB regression formula. Tell: is the phenomenon a shift in a fitted continuous coefficient (OVB) or a reversal of a rate/association across strata of a table (Simpson's)?
Measurement error & simultaneity. Sibling sources of endogeneity under the same OLS-assumption-violation umbrella, but distinct mechanisms: measurement error attenuates a coefficient through a noisy regressor, simultaneity biases through reverse causation. OVB flows specifically through a correlated, outcome-relevant omission, and an instrument that fixes one need not fix the others. Tell: does the bias arise from a mis-measured regressor (measurement error), a two-way causal loop (simultaneity), or a left-out correlated determinant (OVB)?
Selection bias. Distortion arising from who is in the sample — non-random selection into the data — rather than from a variable left out of an otherwise-representative model. OVB is a nearby but distinct member of the bias family: the contamination is a left-out determinant, not a filtered sample. Tell: is the estimate corrupted by which units were observed (selection bias) or by which variables were modeled (OVB)?
Multicollinearity. High correlation among included regressors, which inflates standard errors and destabilizes coefficients but leaves them unbiased. OVB requires the correlated variable to be omitted and outcome-relevant; multicollinearity is a precision problem among variables that are in the model. Tell: are the correlated variables both in the regression (multicollinearity, a variance problem) or is the culprit correlated-and-left-out (OVB, a bias problem)?
Collider bias / over-controlling. Bias created by including a variable — conditioning on a common effect of regressor and outcome (a collider) opens a spurious path, and conditioning on a mediator removes part of the real effect. This is the opposite hazard to OVB: OVB comes from wrongly leaving out a common cause, collider bias from wrongly putting in a common effect. Tell: was a needed common cause omitted (OVB) or a collider/mediator wrongly included (collider/over-control bias)?

Neighborhood in Abstraction Space¶

Omitted Variable Bias sits in a sparse region of the domain-specific corpus (97^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (309 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12