Tutorials › Causal Inference › Positivity for Transport and Sampling Overlap

Transportability and Generalizability · Tutorial 935 of 972

Positivity for Transport and Sampling Overlap

Learn how to state, check, and interpret positivity conditions when transporting causal effects from a source study to a target population.

Advanced 10 min read

What You'll Learn

  • Define transport positivity using source and target covariate support
  • Distinguish transport positivity from treatment positivity within the source
  • Connect support overlap to identification of a target-population effect
  • Derive the covariate density-ratio identity used in transport weighting
  • Recognize how absent or sparse overlap affects conclusions and estimates
  • Diagnose common mistakes when checking positivity

Why Positivity Matters for Transport

A causal effect estimated in a study population does not automatically describe a different target population. Transport formulas use covariates to connect the two populations: the study must contain information about the covariate profiles that occur in the target. If some target profiles never appear in the source, the study cannot directly tell us what the causal effect is for those profiles. This requirement is called positivity for transport, or sampling overlap.

The issue is not merely whether the source and target populations have similar average covariates. They may have different distributions and still have adequate overlap. The key question is whether every covariate profile represented in the target has a corresponding profile represented in the source, at least in the population support relevant to the estimand.

Definition: Let \(S=1\) denote membership in the source study and \(S=0\) denote membership in the target population, and let \(X\) be the measured covariates used for transport. Transport positivity holds when every covariate value with positive probability in the target also has positive probability in the source: $$ P(X=x\mid S=0)>0 \ \Longrightarrow\ P(X=x\mid S=1)>0. $$ For continuous covariates, the corresponding condition is that the target covariate distribution is supported within the source covariate support: \(f_T(x)>0\) implies \(f_S(x)>0\), apart from sets of probability zero under the target distribution.

Here \(f_T\) and \(f_S\) denote the covariate densities in the target and source populations. In a combined population where \(S\) marks source membership, the same condition can be written \(P(S=1\mid X=x)>0\) for every \(x\) occurring in the target. The formulation is about support, not equality of distributions: the target may place more weight on some profiles than the source does, provided those profiles are represented in the source.

Transport Positivity and Treatment Positivity

Two distinct positivity requirements can arise in one transport analysis. Transport positivity concerns whether the source represents the target's covariate profiles. Treatment positivity concerns whether the source contains observations under each treatment being compared within the relevant covariate profiles. One condition does not guarantee the other.

For example, a source study could include participants across the full range of target ages, satisfying transport positivity, but assign every older participant to the same treatment. It would then lack treatment positivity for comparing treatments among older participants. Conversely, a randomized trial could include both treatment arms at every age represented in the trial while including no participants above age 70; if the target includes people older than 70, treatment positivity within the trial does not solve the transport support problem.

To state the role of transport positivity precisely, let \(Y(a)\) be the potential outcome under treatment \(a\), and define the source conditional causal effect as \(\tau_S(x)=E[Y(1)-Y(0)\mid X=x,S=1]\). A common target estimand is the target average treatment effect, \(\psi_T=E[Y(1)-Y(0)\mid S=0]\). Under consistency, valid identification of the treatment effect within the source at each \(x\), and conditional exchangeability across study participation, \((Y(1),Y(0))\perp S\mid X\), the target effect can be written as:

$$ \psi_T = E[\tau_S(X)\mid S=0] = \int \tau_S(x) f_T(x)\,dx. $$

This expression averages source-identified conditional effects over the target covariate distribution. The exchangeability and treatment-identification conditions are substantive requirements in their own right; positivity is a separate support condition that makes the needed source conditional effects available wherever the target average needs them. Without positivity, the formula may call for \(\tau_S(x)\) at a target value of \(x\) for which the source provides no observations.

A Density-Ratio Justification

The connection between support and weighting follows directly from a change of density. If \(f_T(x)>0\) only where \(f_S(x)>0\), then for any integrable function \(g\) whose target expectation exists:

$$ E_T[g(X)] = \int g(x)f_T(x)\,dx = \int g(x)\frac{f_T(x)}{f_S(x)}f_S(x)\,dx = E_S\left[\frac{f_T(X)}{f_S(X)}g(X)\right]. $$

The ratio \(f_T(x)/f_S(x)\) reweights source covariate profiles to represent their relative frequency in the target. If \(f_S(x)=0\) while \(f_T(x)>0\), this ratio is not defined there, and the source expectation cannot recover the target contribution from that region. This is why positivity is required for the usual transport weighting identity, not simply a technical preference.

When source and target membership are represented together and \(q=P(S=1)\), Bayes' rule gives the relationship:

$$ \frac{f_T(x)}{f_S(x)} = \frac{q}{1-q} \frac{P(S=0\mid X=x)}{P(S=1\mid X=x)}. $$

Thus, positive probability of source participation at target profiles is also what keeps the inverse odds of sampling finite in the population. The constant \(q/(1-q)\) does not affect a normalized weighted mean, but a very small source participation probability can still produce very large weights.

Worked Example: Checking Support in Two Covariate Strata

Suppose a hypothetical trial records whether participants are in a lower- or higher-risk covariate stratum. There are 80 lower-risk and 20 higher-risk participants among the 100 source participants. In the target population, 50% are lower-risk and 50% are higher-risk. The trial estimates conditional treatment effects of 4 points in the lower-risk stratum and 10 points in the higher-risk stratum. Assume the treatment effect within each source stratum is identified, and the conditional exchangeability condition needed for transport is credible.

Check transport positivity. Both strata have positive source representation: \(80/100=0.80\) and \(20/100=0.20\). Both also occur in the target, each with probability \(0.50\). Therefore, the target strata are represented in the source and transport positivity holds for these two covariates.

Calculate the target effect. Standardize the conditional effects to the target distribution:

$$ \psi_T = (0.50)(4\text{ points})+(0.50)(10\text{ points}) = 2+5 = 7\text{ points}. $$

The transported effect is 7 points. By contrast, averaging with the source distribution would give \((0.80)(4)+(0.20)(10)=3.2+2=5.2\) points. That is a different population average. The difference illustrates why overlap permits transport but does not make the source and target covariate distributions interchangeable.

Overlap Can Hold While Weights Become Unstable

Positivity is a population condition, but its practical assessment uses finite samples. A source profile may have positive population probability and still be absent from a modest study by chance. Even when every target profile is technically represented, profiles that are rare in the source and common in the target can receive large weights. Such weights make estimates sensitive to a small number of observations and can increase uncertainty.

Worked Example: Reweighting With Unequal Sampling Odds

Consider another hypothetical source and target with two covariate strata. The source proportions are 0.80 lower-risk and 0.20 higher-risk; the target proportions are 0.50 and 0.50. Suppose the estimated conditional effects are 4 and 10 points, respectively. Both strata occur in both populations, so transport positivity holds.

The density ratio for the lower-risk stratum is \(0.50/0.80=0.625\), and for the higher-risk stratum it is \(0.50/0.20=2.5\). Weighting the source stratum proportions by these ratios gives contributions \(0.80(0.625)=0.50\) and \(0.20(2.5)=0.50\). They sum to 1, as they should for these normalized target weights.

The weighted effect is therefore:

$$ \frac{(0.80)(0.625)(4)+(0.20)(2.5)(10)} {(0.80)(0.625)+(0.20)(2.5)} = \frac{2+5}{0.50+0.50} = 7\text{ points}. $$

The higher-risk stratum receives a larger weight because it is more common in the target than in the source. The estimate is calculable because that stratum is present in the source. If it were only extremely rare rather than absent, the weight could be very large and the estimate less stable; positivity alone does not guarantee good precision.

What a Support Violation Means for Identification

If a target covariate profile has no source representation, the data do not identify the source conditional effect at that profile. Conditional exchangeability across participation does not fill in the missing information: it relates source and target potential outcomes at the same covariate value, but it does not determine the potential outcomes at a value where no source observations exist. Any answer for that region would require additional information or assumptions beyond the observed source outcomes and target covariates.

Worked Example: An Unsupported Target Stratum

In a hypothetical study, \(X\) is binary. Every source participant has \(X=0\), while every target member has \(X=1\). The source trial has both treatment arms at \(X=0\), and its observed data are compatible with mean potential outcomes \(E[Y(0)\mid X=0,S=1]=10\) and \(E[Y(1)\mid X=0,S=1]=14\). The target covariates are known, but target outcomes are not observed.

Transport positivity fails because \(P(X=1\mid S=0)=1\), while \(P(X=1\mid S=1)=0\). Consider two possible data-generating accounts for the target stratum. In the first, the target mean potential outcomes are 10 under control and 12 under treatment, so its effect is \(12-10=2\) points. In the second, they are 10 under control and 18 under treatment, so its effect is \(18-10=8\) points.

Both accounts agree with the observed source outcomes at \(X=0\), and both agree with the observed target covariates. Since no source participant has \(X=1\) and no target outcomes are observed, the available data cannot distinguish the two target effects. The target average effect is therefore not identified from these data and the stated information alone. The arithmetic cannot be repaired by assigning a finite weight to a stratum with zero source probability.

How to Diagnose Positivity in Practice

A positivity assessment should be tied to the covariates that define the transport formula and to the target population for which the effect is sought. It is not enough to compare overall sample sizes or a few marginal averages. The source may cover the target's age range but fail to represent particular age-and-severity combinations. Conversely, different marginal distributions do not by themselves establish a violation if all target combinations are represented.

1
Specify the target covariate profiles.
Identify the covariates used to condition the transport assumptions and determine which values or combinations occur in the target.
2
Compare source and target support.
Check whether each target profile is represented in the source. For continuous covariates, consider joint coverage, not only separate ranges for each covariate.
3
Assess the degree of overlap.
Among profiles represented in both populations, examine whether source representation is sparse enough to generate unstable estimates or highly variable weights.
4
Keep the two positivity checks distinct.
Separately assess whether the source contains the treatment groups needed to identify conditional effects. Adequate sampling overlap does not imply adequate treatment assignment overlap.

Common Mistakes

  • Equating different distributions with no overlap. A source and target can have different covariate proportions while sharing support. Positivity asks whether target profiles are represented, not whether the distributions match.
  • Checking only marginal ranges. Separate coverage of age and disease severity does not ensure coverage of their joint combinations. A target combination may be absent from the source even when both marginal ranges overlap.
  • Confusing transport positivity with treatment positivity. Source participants may cover the target covariates but lack one treatment arm at some profiles. These are separate conditions and must be checked separately.
  • Treating a large estimated weight as proof of a population-level violation. A large weight signals sparse source representation and potential instability; a structural positivity violation means the source probability is zero. Finite data may make the distinction hard to determine.
  • Claiming the target effect is identified because target covariates are available. Target covariates show which profiles need to be represented. They do not reveal the causal effects for profiles absent from the source.
Key takeaway: Transport positivity requires source representation for every covariate profile that contributes to the target estimand. It enables the source conditional effects to be averaged or reweighted over the target distribution. A support gap obstructs identification at that profile; sparse overlap can make estimation unstable even when strict positivity holds.

Check Your Understanding

Answer each question by distinguishing support, identification, and estimation where relevant.

  1. A target population includes covariate strata with probabilities 0.70 and 0.30. The source includes both strata, with probabilities 0.95 and 0.05. Does transport positivity hold? What practical concern may remain?
  2. Explain why overlap in the marginal distributions of age and disease severity may fail to establish overlap in their joint distribution.
  3. A source study represents every target covariate profile, but has no treated participants in one profile. Which positivity condition fails?
  4. In the density-ratio identity, what happens at a target profile where the source density is zero and the target density is positive?
  5. Why does knowing the target covariate distribution not, by itself, identify the target causal effect when a target profile is absent from the source?
Need help applying this method to your own data?

Our team supports biostatistics, statistical programming and pharmacometric analyses. Send us a request and we will reply by email.

Request an analysis
Clinical Biostats