Two Different Reasons Treatment Support Can Be Missing
A marginal structural analysis needs treatment options to be available for the histories relevant to its target intervention. Yet a treatment group can be absent from a sample for two quite different reasons: the treatment may truly be unavailable for a particular history, or it may be available but so rare that the sample happened to contain no such treatment. These cases can look identical in the observed data, but their implications differ. A true absence is a failure of support; a rare but possible treatment creates an estimation problem that may be severe even when the formal support condition holds.
For a longitudinal setting, write \(H_t\) for the observed covariate and treatment history immediately before treatment \(A_t\), and let \(g_t(a\mid h)=P(A_t=a\mid H_t=h)\). For a target regime that assigns treatment \(a_t\) at history \(h_t\), positivity requires that the assigned treatment have positive probability at the histories the regime can encounter in the target population. The condition is an almost-everywhere population statement, not a claim that every treatment-history combination must appear in every finite sample. As in other identification conditions for marginal structural models, positivity is considered alongside consistency and the relevant exchangeability assumptions; it does not replace them.
“Structural” does not mean merely that a fitted model estimated a probability of zero. It refers to the underlying treatment mechanism or a genuine restriction of the treatment options. For example, a treatment may be prohibited by protocol for a subgroup, or a clinical rule may deterministically rule it out. Conversely, a finite dataset may omit treated individuals from a subgroup even though the true probability of treatment there is positive.
Why Finite Samples Cannot Prove a Zero Probability
Consider \(n\) independent individuals in a fixed, prespecified history stratum, each with the same treatment probability \(p\). Let \(X\) be the number who receive the treatment. Then \(X\sim\operatorname{Binomial}(n,p)\), so
For every finite \(n\), this probability is positive whenever \(p<1\), including when \(p\) is positive but small. For example, if \(p=0.01\) and \(n=20\), then the probability of observing no treated individuals is \(0.99^{20}\), about \(0.818\). Thus an empty treatment cell is quite compatible with a positive treatment probability.
The same argument explains a deeper limitation: with any finite sample size, probabilities arbitrarily close to zero make an observed zero count arbitrarily likely. Observed data can reveal that the sample contains no examples of a treatment in a subgroup. They cannot, without additional assumptions or external design knowledge, prove that the underlying probability is exactly zero. This is an inferential limitation, not a defect that a more elaborate diagnostic can remove.
Worked Example: A Zero Count With a Positive Treatment Probability
Suppose treatment is given independently to \(n=20\) people in a prespecified stratum, with a common probability \(p=0.10\). Calculate the probability of seeing no treated people.
Here \(X\sim\operatorname{Binomial}(20,0.10)\). Therefore,
A sample with no treated people occurs about 12.2% of the time under this positive treatment probability. Consequently, the observed zero is concerning for estimation but is not proof of structural nonavailability. The calculation relies on independent observations with the same probability in the specified stratum; it would not directly describe a heterogeneous or clustered sample.
What a Zero Count Can Say: An Exact Bound
Under the same independent, common-probability binomial model, a zero count can be paired with an exact one-sided confidence bound. This quantifies how large \(p\) could plausibly be given the count, at a chosen confidence level. It does not decide whether \(p\) is exactly zero.
The definition gives an endpoint for every possible count, including \(x=n\). In the special case \(x=0\), the beta distribution has parameters \((1,n)\), whose cumulative distribution function is \(1-(1-u)^n\). Setting this equal to \(1-\alpha\) gives
The endpoint has at least \(1-\alpha\) repeated-sampling coverage. To see why, fix a candidate probability \(p_0\) and test \(H_0:p=p_0\) against smaller probabilities by rejecting when \(P_{p_0}(X\le x)\le\alpha\). Since \(P_p(X\le x)\) decreases as \(p\) increases, for \(x<n\) the equation \(P_{U(x)}(X\le x)=\alpha\) defines the boundary: values \(p_0>U(x)\) are rejected, while \(p_0\le U(x)\) are not. For \(x=n\), no value is rejected by this lower-tail test, matching \(U(n)=1\). Inverting these tests gives the stated upper endpoint. Under the true \(p\), the probability of rejection is at most \(\alpha\): the rejection outcomes form a lower-tail set of binomial counts whose probability under \(p\) is no greater than \(\alpha\). Hence \(P_p\{p\le U(X)\}\ge 1-\alpha\). This argument establishes coverage for the random endpoint defined at every possible outcome, not only for a zero count. \(\square\)
Worked Example: Interpreting the 95% Bound After Zero Treated People
Suppose the stratum contains \(n=20\) independent observations and none received treatment. Use a one-sided 95% Clopper–Pearson upper endpoint, so \(\alpha=0.05\).
Substitute \(n=20\) and \(\alpha=0.05\):
Under the binomial assumptions, the upper endpoint is about 0.139. The sample is compatible, at this confidence level, with a treatment probability as high as roughly 13.9%; the zero count does not establish that the probability is zero. This calculation also shows why “we observed none” is not equivalent to “the probability is zero.” It is not a universal bound for a continuous history or a heterogeneous stratum: its interpretation depends on the stated sampling model.
Why Small Positive Probabilities Still Matter
A treatment probability can satisfy formal positivity and still cause unstable estimation. The simplest calculation makes the distinction clear. In a stratum where treatment \(A\) is Bernoulli with probability \(p>0\), the inverse-probability contribution for the treated treatment level is \(W=I(A=1)/p\). Its expectation is one, but its second moment and variance are
These identities follow directly from \(I(A=1)^2=I(A=1)\). As \(p\) approaches zero, the variance increases without bound, although it is finite for every fixed \(p>0\). At \(p=0\), \(W\) is not defined for the target treatment level at all: no observed individual can represent that level in the stratum. This is the practical distinction between unstable information under near-positivity and absent support under structural nonpositivity.
Worked Example: Rare Treatment and Large Weight Variability
Suppose the probability of treatment in a relevant stratum is \(p=0.02\). The treated-level inverse-probability contribution is \(1/p=50\) for a treated person and zero for an untreated person.
Its expectation is \(0.02(50)+0.98(0)=1\). Its second moment is \(0.02(50^2)=50\), so its variance is \(50-1=49\). The probability that a sample of 100 independent people contains no treated individual is
Thus even though positivity holds in this stratum, there is about a 13.3% chance of observing no treated person in a sample of 100, and a treated observation carries a contribution of 50 before other weight factors are considered. This example illustrates both finite-sample sparsity and the large variability associated with a small positive probability. It does not imply that every weighted estimator has variance 49; the calculation concerns this single contribution.
What to Diagnose, and What Requires Other Knowledge
Observed-data diagnostics can identify empirical warning signs: empty treatment cells in recorded histories, very small estimated treatment probabilities, and extreme weights. These findings show that the available sample may not support the intended weighted comparison well. They do not, by themselves, distinguish a true probability of zero from a small positive probability, nor do they verify that a fitted treatment model is correct. A fitted probability is a model-based estimate, not direct proof of the population mechanism.
Structural nonavailability may instead be established through design information, protocol restrictions, or well-supported substantive knowledge about the treatment mechanism. Such evidence must match the target population and histories under study. A rule that applies only at one time or in one population cannot automatically establish a structural zero elsewhere.
| Finding | What it supports | What it does not establish |
|---|---|---|
| No treated observations in a recorded subgroup | The sample has an empty treatment cell; estimation there is unsupported empirically. | That the population probability is exactly zero. |
| Very small fitted treatment probabilities or large weights | A warning that estimation may be unstable and sensitive to a few observations. | That the fitted model is correct, or that the issue is structural. |
| A binding treatment rule documented for the relevant subgroup | Potential evidence of a structural treatment restriction. | That positivity holds for other histories or target populations. |
Several responses are possible, but they answer different questions. Restricting the target population or changing the target regime can avoid treatment levels that are unavailable in some histories, at the cost of changing the scientific question. A model may produce estimates where data are sparse by relying on assumptions about how treatment probabilities or outcomes behave across histories; those estimates then depend more heavily on extrapolation. Truncating or otherwise modifying extreme weights can reduce the influence of large contributions, but it does not create observations where none exist or make a structural zero positive. Each choice should be reported as a change in assumptions, target, or estimator—not as proof that the original support problem has disappeared.
Common Mistake and Practical Limits
A common mistake is to see zero treated people in a subgroup and conclude, “positivity is structurally violated.” The reasoning confuses a sample count with a population probability. Under the binomial model, \(P(X=0)=(1-p)^n\) is positive for every \(p<1\); therefore, an empty cell can occur even when \(p>0\). The correction is to call the finding an observed empty cell, investigate design and substantive restrictions, and quantify uncertainty under an appropriate sampling model. If that model is not credible—for example, because observations are dependent or have different treatment probabilities—the binomial bound above should not be applied without modification.
There is no data-only test that reliably proves an exact zero against every arbitrarily small positive probability from a finite sample. Conversely, failure to find an empty cell does not establish that probabilities are safely large: a finite sample can contain a few treated observations even when the corresponding probabilities are extremely small. The practical question is therefore not only whether support is formally positive, but also whether the available data provide enough information for the intended weighted comparison.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a stratum of 15 people, no one receives treatment. Does that observation prove structural nonpositivity? Explain.
- For \(n=10\) and \(p=0.05\), calculate the probability of observing no treated people.
- If \(p=0.10\), calculate the expectation and variance of \(I(A=1)/p\).
- With zero treated people among \(n=20\), what does the one-sided 95% Clopper–Pearson upper endpoint say, and what does it not say?
Solutions. (1) No. It establishes an empty sample cell, but a positive probability can also generate zero observations; design evidence or additional assumptions are needed to assert a structural zero. (2) \(P(X=0)=(1-0.05)^{10}=0.95^{10}\approx0.5987\), so a zero count is likely despite positive treatment probability. (3) The contribution is 10 with probability 0.10 and zero otherwise. Its expectation is \(0.10(10)=1\); its second moment is \(0.10(100)=10\); hence its variance is \(10-1=9\). (4) \(U(0)=1-0.05^{1/20}\approx0.1391\). Under the independent, common-probability binomial model, this is a one-sided 95% upper confidence endpoint; it does not prove that the true probability is positive or zero.