Why an Empty Treatment Cell Needs Interpretation
Suppose a clinical dataset contains no patients in a particular history group who received a treatment of interest. That observation matters for positivity, but it does not by itself explain the absence. The treatment may be prohibited for that group by an eligibility rule, or it may be possible but uncommon and simply not observed in the sample. These are different problems: collecting more data can reveal a rare treatment, but cannot create observations of an action that the assignment rule makes impossible.
The distinction is especially important in longitudinal studies. At decision time \(t\), let \(H_t\) denote the measured history available before treatment, and let \(A_t\) denote the treatment assigned at that time. For a discrete treatment option \(a\), write \(g_t(a\mid h)=P(A_t=a\mid H_t=h)\). Positivity for a target action requires this conditional probability to be positive in the histories relevant to that action. The issue here is how to recognize when it is exactly zero in the population assignment mechanism, rather than merely estimated as zero from a finite sample.
For a binary treatment, requiring support for both options at a history means \(0<P(A_t=1\mid H_t=h)<1\). If a target concerns only one option, support for that option is the relevant requirement. In either case, positivity is a property of the assignment distribution in the population, while an empty cell is a feature of the realized sample.
Eligibility Rules and Deterministic Assignment
Let \(E_t\) be an eligibility indicator determined before treatment, with \(E_t=1\) for eligible patients and \(E_t=0\) for ineligible patients. Suppose a protocol deterministically assigns \(A_t=0\) whenever \(E_t=0\). If ineligible patients occur with positive probability, then the treatment option \(A_t=1\) has zero probability in that group. This is structural, not a small-sample approximation.
Proof. Since \(E_t\) is part of the pre-treatment history, the stratum \(E_t=0\) is a set of histories. Its probability is positive by assumption. The deterministic rule gives \(P(A_t=1\mid E_t=0)=0\), so the required positive probability for action 1 fails on a positive-probability set of histories. Therefore structural positivity fails for that action. \(\square\)
A clinical contraindication can produce this pattern, but the word “structural” describes the assignment mechanism, not necessarily a biological impossibility. A hospital policy may make an action unavailable in its observed setting even if another setting could assign it. Conversely, a clinical rule written in a protocol is not enough to establish a zero if clinicians sometimes depart from it. Check the actual mechanism and the data-generating setting being studied.
Worked Example: A Treatment Reserved for Eligible Patients
A study follows 100 patients at a treatment decision. Eligibility is recorded before assignment. There are 35 ineligible patients, all assigned standard care (\(A=0\)). Among 65 eligible patients, 26 receive the treatment of interest (\(A=1\)) and 39 receive standard care.
For ineligible patients, the empirical conditional proportion treated is \(0/35=0\). Suppose the protocol confirms that treatment is not permitted when \(E=0\). Then \(P(A=1\mid E=0)=0\) under the protocol, and \(P(E=0)=35/100=0.35>0\). The result above shows a structural positivity failure for treatment \(A=1\) on the ineligible stratum.
Among eligible patients, the observed treatment proportion is \(26/65=0.40\). This number is an estimate of the eligible-stratum assignment probability, not proof of its exact population value. The key difference is that the protocol rules out treatment for ineligible patients, while the nonzero count among eligible patients confirms that treatment occurred there. If the target were to assign treatment to ineligible patients, the observed assignment mechanism supplies no such cases.
Deterministic rules can also arise from treatment algorithms, resource allocation, or protocol thresholds. For example, if a high dose is assigned only when a pre-treatment laboratory value exceeds a fixed threshold, histories below the threshold may have zero probability of receiving that dose. Whether that is structural for the analysis depends on whether the threshold is truly enforced in the population and whether the history variable captures the rule.
Worked Example: A Protocol Threshold for High Dose
A protocol assigns a 20-mg dose only when a pre-treatment kidney-function measure \(K\) is at least 60. Patients with \(K<60\) are assigned 10 mg. Suppose the study population includes 120 patients with \(K<60\), and none receives 20 mg.
Define \(A=1\) for 20 mg and \(A=0\) for 10 mg. The protocol states \(A=0\) whenever \(K<60\), so \(P(A=1\mid K<60)=0\). Because \(P(K<60)=120/N>0\), positivity for the 20-mg option fails over a positive-probability history set.
The zero is not repaired by fitting a more flexible model to the same assignment mechanism: the model cannot learn the outcome under 20 mg for these histories from treatment observations that cannot occur. Also, if clinicians sometimes override the protocol, the deterministic conclusion must be revised; then the relevant question is whether the true probability is positive, perhaps very small.
Why Sparse Nonzero Probabilities Can Look Like Zeros
In a finite sample, even a positive treatment probability can produce no treated patients in a stratum. Consider \(n\) independent patients from the same history stratum, each with treatment probability \(p>0\). The number treated has a binomial distribution with parameters \(n\) and \(p\). Thus the probability of seeing no treated patients is \((1-p)^n\), which is positive whenever \(p<1\), even though treatment is possible.
Derivation. For each patient \(i\), let \(I_i=1\) if treated and \(I_i=0\) otherwise. Under the stated assumptions, \(P(I_i=0)=1-p\), and the indicators are independent. The event of zero treated patients is the event \(I_1=0,\ldots,I_n=0\). Independence therefore gives
Since \(p>0\), treatment has population support in the stratum, even though the displayed event may occur in the sample. \(\square\)
Worked Example: A Rare Treatment With No Observed Cases
Suppose a treatment has probability \(p=0.02\) in a particular history stratum, and 40 patients from that stratum are observed. Under independent sampling with a common probability, the probability that none is treated is
So even though the treatment probability is nonzero, there is about a 44.6% chance of an empty treated cell in a sample of this size. The probability of observing at least one treated patient is \(1-0.4457=0.5543\), about 55.4%. Seeing no treated patients here is therefore quite compatible with a sparse but nonzero assignment probability. It is not proof of a structural failure.
The calculation assumes a common \(p\) and independent observations within the stratum. If treatment probabilities vary across patients even after defining the stratum, or if sampling is dependent, this particular binomial calculation need not apply. The general lesson remains: the observed count is evidence about support, not a direct definition of it.
Diagnostics: What the Data Can and Cannot Establish
An empty cell should prompt an audit rather than an immediate conclusion. First verify that the history is measured before treatment and that the stratum is not defined by a post-treatment variable. Then check the denominator: an empty treatment cell among many eligible patients differs from a stratum containing no eligible patients at all. Review clinical eligibility criteria, treatment protocols, formularies, and actual practice. Tabulate treatment counts by clinically meaningful pre-treatment strata, including the number eligible and the number assigned each option.
If a fitted treatment model produces estimated probabilities of exactly zero or one, check whether that reflects a genuine deterministic rule, sparse data, or a modeling artifact. A model can output extreme fitted probabilities because the observed data separate treatment groups; the fitted zero is not automatically a population fact. Likewise, a nonzero fitted probability does not establish that treatment support exists in a region where the protocol forbids the action.
For inverse-probability weighting, the distinction has a direct computational consequence. A weight for action \(a\) uses a factor \(1/g_t(a\mid H_t)\). If the true probability is zero for a history, this factor is undefined for that action and history. If the probability is small but positive, the factor is finite but can be large; the resulting weights may be unstable. Weight truncation can limit the influence of large estimated weights, but it does not create observations or turn a structural zero into positive support.
| Feature | Structural zero | Sparse but nonzero probability |
|---|---|---|
| Population assignment probability | Exactly zero for an action on a positive-probability history set | Positive, possibly very close to zero |
| Possible explanation | Eligibility restriction or deterministic assignment rule | Uncommon treatment choice or limited sample size |
| Meaning of an empty observed cell | Consistent with the rule, but assess the actual mechanism | Can occur by chance; its probability depends on sample size and treatment probability |
| What more data can do | Cannot produce the prohibited action under the same rule | May yield treated observations, though not guaranteed in a finite sample |
Common Mistakes and Limits
- Calling every empty cell a structural zero. A zero observed count is a sample fact. To claim a structural failure, establish that the population probability is exactly zero on relevant histories, for example from a verified deterministic rule.
- Calling a very small probability zero. A probability such as \(0.02\) is positive. It can generate severe finite-sample instability, but it is not an exact support failure.
- Assuming protocol language proves actual deterministic assignment. Check adherence, exceptions, and whether the eligibility information was available before the decision. Deviations may make the probability positive.
- Treating a model’s fitted zero as definitive. Sparse data or model separation can cause boundary estimates. Examine counts and assignment procedures instead of interpreting a fitted value in isolation.
- Assuming a larger sample fixes every problem. More observations can help distinguish a rare action from a sample zero, but if the action is ruled out by the assignment mechanism, the same design still contains no support for it.
The appropriate response depends on the source of the gap. For a structural zero, one may need to narrow the target population to histories where the action is available or define a different intervention that respects the clinical rule. For a sparse but positive probability, additional data or a carefully chosen analysis may improve precision, although large weights and limited effective information can remain concerns. Neither choice should be made by relabeling an empty cell: the clinical question, assignment process, and target action must be considered together.
Check Your Understanding
Answer each question, then compare with the solutions.
- A protocol assigns treatment only to eligible patients. Ineligibility is measured before treatment and occurs in 20% of the population. What does this imply for positivity of treatment among ineligible patients?
- In a stratum, the true treatment probability is \(0.05\), and 10 independent patients are observed. Calculate the probability of seeing no treated patients.
- A study has no treated patients in one stratum, but the protocol allows treatment there. Does the observed zero establish structural failure? Explain.
- A fitted propensity model gives probability zero for a treatment in a group. Name two checks needed before concluding that this is structural.
Solutions. (1) Treatment has conditional probability zero among ineligible patients under the stated deterministic rule. Since that group has probability \(0.20>0\), positivity for treatment fails on a positive-probability history set. (2) The probability is \((1-0.05)^{10}=0.95^{10}\approx0.5987\), so an empty treated cell is not surprising even though the true probability is positive. (3) No. The empty cell may be a sampling zero; check the assignment mechanism and sample size, because permission to treat does not establish either a positive probability or an exact zero. (4) Check the treatment counts and denominators in that group, and review eligibility rules and actual clinical practice. Also assess whether model separation or specification caused the fitted boundary value.