Positivity Within Covariate Strata

Why Treatment Must Be Possible Within Covariate Strata

Suppose an observational study compares outcomes for people who received a treatment with outcomes for those who did not. If treatment choice depends on covariates such as baseline severity, a comparison within relevant covariate strata can help separate treatment differences from differences in who received treatment. But this comparison needs data from both treatment groups in the strata represented in the target population. If treatment is never given to people with a particular covariate profile, their outcomes under treatment cannot be learned directly from treated people with that profile.

This requirement is called conditional positivity. It is a condition on treatment probabilities given covariates, not a requirement that every specific covariate value occur with positive probability. That distinction is especially important for continuous covariates: an exact value can have probability zero while the conditional probability of receiving either treatment at that value is well-defined and positive.

The phrase almost every allows exceptional covariate values that have probability zero under the target distribution. For discrete covariates, it means that every stratum with positive target-population probability must have positive probability of each treatment level. For continuous covariates, conditional probabilities are understood through a regular conditional distribution; the condition is on \(P(A=a\mid X=x)\), not on the probability of the joint event \(A=a, X=x\).

Positivity is target-specific. If the target population includes covariate profiles that do not occur in the study population, the study cannot establish positivity for those profiles merely by having both treatments somewhere in its sample. The relevant question is whether each treatment level is possible across the covariate values represented in the target population.

Checking Positivity in Discrete Strata

When covariates define a small number of strata, positivity can be stated and inspected stratum by stratum. The following hypothetical counts illustrate the check. They are sample observations, so their treatment fractions are empirical estimates of the underlying conditional probabilities, not proof of those probabilities.

A marginal treatment proportion can conceal a failure of conditional overlap. For example, an overall sample may contain many treated and untreated people even though all high-risk individuals are treated and all low-risk individuals are untreated. The total counts look balanced, but there is no within-stratum comparison. Positivity must therefore be checked conditional on the covariates used to define the target comparison, not merely by reporting the overall treatment share.

How Positivity Supports Identification

Let \(Y(1)\) and \(Y(0)\) be the potential outcomes under treatment and no treatment, and let \(Y\) be the observed outcome. For a binary treatment, the average treatment effect in the target population is \(E[Y(1)-Y(0)]\). The connection between positivity and identification is clearest when it is considered together with two other assumptions: consistency, which links the observed outcome to the potential outcome under the treatment actually received, and conditional exchangeability, \(Y(a)\perp A\mid X\), for each treatment level \(a\).

Conditional exchangeability says that, at a given covariate value, treatment assignment is independent of the potential outcome under each treatment level. Consistency says that if a person receives level \(a\), their observed outcome equals \(Y(a)\). Under these assumptions, and where \(P(A=a\mid X=x)>0\), the conditional mean of \(Y(a)\) can be recovered from observed outcomes among people with \(A=a\) at that covariate value:

The first equality follows from conditional exchangeability. The second follows from consistency among those receiving treatment level \(a\). Positivity matters because it makes the conditional outcome distribution for level \(a\) available at the target covariate values in the regular-conditional-distribution sense. If that level has zero conditional probability at a target value, the observed data do not provide outcomes from that treatment group there.

For discrete covariate strata \(x\), averaging the identified conditional means over the target distribution gives the standardization formula. This is the same logic whether the target is an average potential outcome or a contrast of two such averages.

Continuous Covariates and Conditional Probabilities

With a continuous covariate, it is generally incorrect to demand a positive probability for the exact joint event \(A=a,X=x\). For a continuously distributed \(X\), \(P(X=x)=0\) at an individual point. Positivity instead asks whether the conditional treatment probability \(P(A=a\mid X=x)\) is positive for almost every \(x\) in the target distribution. This is a conditional probability statement, not a claim that a positive number of sampled people have exactly the value \(x\).

Diagnostics, Comparisons, and Limitations

In practice, positivity is assessed by examining treatment representation across covariate profiles that matter for the target population. For discrete covariates, cross-tabulations can reveal strata with no treated or no untreated observations. With continuous covariates, plots or summaries of treatment probabilities across the covariate distribution can reveal regions where one treatment group is sparse. These are diagnostics, not automatic tests of the assumption: finite samples can have no observations in a group even when its population probability is positive, and a model-based estimate may look smooth while relying on strong modeling choices.

The main design alternative is random assignment with a positive assignment probability for each treatment level. It can provide treatment representation without relying on observational treatment choices, although realized finite samples can still be imbalanced by chance. In observational studies, conditional positivity concerns treatment availability within covariate profiles; a large overall sample or a favorable marginal treatment ratio is not a substitute. Regression adjustment or extrapolation can produce predictions in sparsely represented regions, but those predictions do not create direct information about outcomes under an unobserved treatment level.

A frequent mistake is to reason, “There are treated and untreated individuals in the full sample, so positivity holds.” The correction is to inspect the conditional treatment probabilities across the target covariate distribution. Another mistake is to say, “No exact continuous covariate value has positive probability, so positivity cannot hold.” The correction is that the condition concerns \(P(A=a\mid X=x)\) almost everywhere, not \(P(A=a,X=x)\). Finally, positivity is an assumption about the treatment assignment distribution in the target population; observed counts can inform its plausibility but cannot prove it.

Conditional positivity is necessary for identifying treatment-specific means by the conditional comparisons described here, but it is not sufficient for causal identification. Consistency and conditional exchangeability are distinct requirements. Positivity can also be target-dependent: restricting the target to covariate profiles with treatment representation may change the population whose effect is being described. Such a change should be stated as a change in the estimand, not presented as though it resolved the original question without consequence.