When Auxiliary Variables Add Identification

Why an Auxiliary Variable May Matter

A fully observed auxiliary variable can contribute information about outcomes that are missing, but its presence alone does not reveal those outcomes. The crucial questions are how the auxiliary variable relates to response and to the outcome, and what assumptions connect those relationships. Without a link between observed and unobserved outcomes, two different complete-data distributions can agree on every recorded value.

Let \(Y\) be an outcome that may be missing, and let \(R=1\) if \(Y\) is observed and \(R=0\) otherwise. Let \(Z\) be a fully observed auxiliary variable and \(X\) denote other fully observed covariates. The observed data include \((X,Z,R)\), and include \(Y\) only when \(R=1\). The target might be the marginal distribution of \(Y\), its mean, or a conditional quantity such as \(E(Y\mid X=x)\).

Two distinct roles are useful. First, \(Z\) may help make a conditional missing-at-random assumption plausible: after conditioning on \(X\) and \(Z\), response is independent of the outcome. Second, a shadow variable can support identification under a missing-not-at-random mechanism by being related to the outcome while having no direct association with response once the outcome and \(X\) are held fixed. These roles use different assumptions; an auxiliary variable is not automatically both.

Conditional MAR: Identification by Stratification

Suppose the assumption is \(Y\perp R\mid X,Z\): within each value of the observed covariates and auxiliary variable, response does not depend on \(Y\). Also require positivity, \(P(R=1\mid X=x,Z=z)>0\) wherever the population has support. Then the outcome distribution among respondents in a stratum equals the outcome distribution in that stratum in the full population:

Because \(X\) and \(Z\) are observed for everyone, their population distribution is also observed. The law of total expectation therefore gives

For discrete variables, the integral is a sum over strata. This calculation requires the stated conditional independence and positivity; observed data alone cannot establish that respondents and nonrespondents have the same outcome distribution within a stratum. Conditioning on a genuinely outcome-predictive \(Z\) may make the assumption more defensible than conditioning on \(X\) alone, but it does not prove it.

Response Association Alone Does Not Identify Outcomes

A variable can be associated with response and still leave the missing outcomes undetermined. Consider hypothetical data where \(P(Z=0)=P(Z=1)=0.5\), response rates are 0.8 for \(Z=0\) and 0.4 for \(Z=1\), and 25% of respondents have \(Y=1\) in each stratum. The observed positive-outcome mass is

The missing mass is 0.10 in the first stratum and 0.30 in the second. One completion of the data can assign \(Y=0\) to every nonrespondent, giving \(P(Y=1)=0.15\). Another can assign \(Y=1\) to every nonrespondent, giving \(P(Y=1)=0.15+0.10+0.30=0.55\). Both completions have the same observed distribution of \(Z\), response, and respondent outcomes. In both, \(Z\) is associated with response. Yet the full outcome prevalence differs substantially. Without a further assumption, the observed association does not select between the completions.

Shadow Variables and an MNAR Identification Equation

A shadow-variable strategy targets settings where response may still depend on the missing outcome after conditioning on \(X\). The assumptions are that \(Z\) is related to \(Y\) given \(X\), and that response is independent of \(Z\) given the outcome and covariates:

The second condition is an exclusion restriction: \(Z\) may be associated with response because it is associated with \(Y\), but it has no additional relationship with response at fixed \(Y\) and \(X\). These conditions are not implied by observing \(Z\). Moreover, identification needs a uniqueness condition, not just the exclusion restriction.

To see the identification argument, fix \(X=x\), and define \(g_1(y,z\mid x)=f(y,z,R=1\mid X=x)\), the observed respondent subdensity. Write \(h(z\mid x)=f(z\mid X=x)\), which is observed because \(Z\) is complete. Define the response odds in the nonrespondent direction as

Under the exclusion restriction, the response probability at fixed \(Y,X\) does not vary with \(Z\). Therefore the missing-outcome subdensity is the respondent subdensity multiplied by \(w(y,x)\). Integrating over \(y\) yields the observed marginal distribution of \(Z\):

The left side is the observed density of \(Z\) among nonrespondents. The equation says that it must be reproduced by weighting respondent outcome-and-\(Z\) records by their outcome-specific nonresponse odds. If this equation has a unique admissible solution \(w\), the missing-outcome distribution follows from

The term \(1\) restores the respondent mass, and \(w\) adds the nonrespondent mass. In general, uniqueness is an assumption about the information in the \(Z\)-to-\(Y\) relationship. In a finite problem, a corresponding linear system must have enough independent equations to determine its unknown weights. If the system has multiple admissible solutions, the observed data do not uniquely identify the target under these assumptions.

What to Check, and What the Result Does Not Establish

An identification argument should make the auxiliary variable’s role explicit. For conditional MAR, ask whether the covariates and \(Z\) plausibly capture the outcome-related reasons for response, and whether every target-relevant stratum has respondents. For a shadow-variable argument, ask whether \(Z\) carries information about \(Y\) given \(X\), whether the exclusion restriction is scientifically credible, and whether the identification equation has a unique admissible solution.

Several failure modes matter in practice:

• No positivity: if a stratum has no respondents, its outcome distribution cannot be learned by conditional MAR adjustment from that stratum.
• Weak information: a shadow-variable system may be formally unique but nearly singular. Small changes in observed quantities can then produce large changes in estimated weights and outcomes.
• Insufficient variation: if \(Z\) does not distinguish outcome-related patterns, the integral equation may have multiple solutions. Association in the marginal data is not a guarantee of the needed uniqueness.
• Unjustified assumptions: the conditional MAR assumption and the shadow-variable exclusion restriction concern unobserved relationships. Good fit to the observed data cannot, by itself, validate them.

The main alternatives answer different questions. Conditional MAR identifies through outcome comparisons within observed covariate strata. A shadow-variable approach can identify under an MNAR mechanism, but relies on its exclusion and uniqueness assumptions. With neither kind of identifying structure, the auxiliary variable may still support descriptions or bounds, but it does not generally determine a single missing-outcome distribution. In some settings it can improve precision without changing what is identified.

An auxiliary variable adds identification when its measured relationships and the assumed structure together rule out competing complete-data distributions. It does not do so merely by being complete, predictive, or statistically associated with response. The conclusion is always conditional on the assumptions used to make the link.