Building an Identification Argument From Assumptions to Contrast

From Assumptions to an Identifying Contrast

An observed difference in average outcomes between treated and untreated people is not automatically a causal effect. The groups may differ in pretreatment characteristics that also predict the outcome. The identification argument here makes the target explicit, states what links potential outcomes to observed data, and derives the observed-data quantity that equals the target under those assumptions.

Consider a binary treatment \(A\), where \(A=1\) denotes treatment and \(A=0\) denotes no treatment, an outcome \(Y\), and a vector \(X\) of pretreatment covariates. Write \(Y(1)\) and \(Y(0)\) for the outcomes the same unit would have under treatment and no treatment, respectively. Only \(Y=Y(A)\) is observed for each unit.

The aim is to express each mean \(E[Y(a)]\) using observed outcomes and the target population distribution of \(X\). The result is a covariate-standardized contrast: compare the observed conditional means for treatment and no treatment at each covariate value, then average those differences over the target population.

Assumptions for the Argument

The following assumptions are stated for the target population. The outcome must have finite means under both treatment levels. The other three assumptions are the substantive links between the causal target and the observed data.

A commonly stated, stronger version of conditional exchangeability is joint independence, \((Y(0),Y(1))\perp A\mid X\). Joint independence implies each of the componentwise conditions above. The converse does not generally hold: separate independence of each potential outcome does not guarantee independence of the pair. The derivation here uses the componentwise conditions, not an assertion that the two formulations are equivalent.

The Identification Derivation

Define the observed conditional mean \(m_a(x)=E[Y\mid A=a,X=x]\). For a discrete \(X\), averaging over its distribution means summing across its values; for a continuous \(X\), it means integrating with respect to that distribution. The derivation applies to each treatment level \(a\).

The first equality is the law of iterated expectations. Conditional exchangeability justifies the second: within \(X\), the conditional mean of \(Y(a)\) is the same among those assigned \(A=a\) as in the full covariate stratum. Positivity ensures that this treatment group exists with positive probability at the relevant covariate values, so the conditional mean given \(A=a,X\) is available there. Consistency justifies the third equality, because in the group with \(A=a\), the observed outcome is \(Y(a)\). These steps also require the finite-mean condition so that the expectations are well-defined.

Apply the result once for \(a=1\) and once for \(a=0\), and subtract. This yields the identifying expression for the target contrast:

This is an identification result, not by itself an estimation recipe or a guarantee of precise estimates. In a sample, analysts estimate the conditional means and average their difference over a sample representing the target covariate distribution. Sampling uncertainty and choices about how to estimate those means are separate questions.

Worked Applications

What to Check, and What the Formula Does Not Guarantee

The derivation is short, but each assumption does a distinct job. Consistency links an observed outcome to a potential outcome for the treatment received. Exchangeability links the potential outcome mean among those receiving a treatment to the corresponding mean in the full covariate stratum. Positivity makes both treatment-specific conditional means available throughout the target support. None of these roles can be replaced merely by calculating a difference in observed averages.

• Check covariate support. Look for target-population covariate values with little or no representation in one treatment group. Empty strata violate positivity; very sparse strata can make estimates unstable even when strict positivity holds.
• Keep the target distribution explicit. The averaging distribution in \(E_X\{m_1(X)-m_0(X)\}\) is the target population’s distribution of \(X\). Changing that distribution can change the contrast.
• Distinguish assumptions from diagnostics. Observed data can reveal some lack of overlap, but they cannot establish that all relevant confounders were measured or that conditional exchangeability is true. The assumptions require substantive judgment as well as data checks.
• Do not mistake an estimate for identification. A fitted model may produce a number when some treatment-covariate combinations are absent. Producing that number does not restore positivity or make its extrapolation an implication of the assumptions.

The main alternative to standardization is the unadjusted observed mean difference, \(E[Y\mid A=1]-E[Y\mid A=0]\). That contrast can equal the target effect under additional conditions, such as unconditional exchangeability, but conditional exchangeability given \(X\) alone does not generally imply that equality. Standardization instead uses the conditional means and averages them using the target distribution of \(X\). More elaborate adjustment methods may estimate the same identifying expression in different ways; they do not remove the need for the identification assumptions.

A common mistake is to say, “The treated mean is higher, so treatment caused the difference,” or to call a crude difference an adjusted effect simply because covariates were recorded. The correction is to name the target contrast, establish the relevant assumptions, and show how the observed-data expression follows from them. Also avoid saying that componentwise independence is “equivalent” to joint exchangeability: joint independence implies componentwise independence, but the reverse implication is not generally valid. The componentwise conditions are sufficient for this derivation.