Why Positivity Matters
An observational comparison is informative about a causal effect only when the data contain relevant information about both exposure options for the kinds of people in the target population. If a certain covariate profile always receives the exposure, the data cannot show what would have happened to people with that profile without it. If almost nobody with that profile receives one option, the comparison may be possible in principle but poorly supported in practice.
This requirement is called positivity. It is one of the assumptions needed to identify causal effects from observational data. Positivity is about the exposure assignment process conditional on covariates; it is not a claim that the exposed and unexposed groups have identical covariate distributions.
The conditional probability \(P(A=1\mid L=l)\) is sometimes called the exposure probability at profile \(l\). Positivity does not require this probability to be one-half, nor does it require the two exposure groups to be equally large. It requires only that both options have a nonzero probability for each relevant profile. In practice, probabilities very close to zero or one can still cause serious problems, even though they meet this mathematical condition.
The condition depends on the full covariate profile, not just on each covariate considered separately. For example, the exposed and unexposed groups might each contain people from both age categories and both risk categories, while one particular age-by-risk combination appears in only one group. That joint profile lacks positivity.
Positivity Among the Identification Assumptions
Write \(Y(1)\) and \(Y(0)\) for the potential outcomes under exposure and no exposure. For a causal comparison to be identified from observational data using covariates \(L\), three assumptions are commonly used. They are stated here to show the distinct role of positivity.
- Consistency: for an individual whose observed exposure is \(A=a\), the observed outcome equals that individual's potential outcome under that exposure: \(Y=Y(a)\).
- Conditional exchangeability: \(Y(a)\) is independent of \(A\) given \(L\), for \(a=0,1\). In words, within covariate profiles, exposure assignment carries no remaining information about the potential outcome.
- Positivity: each exposure option has positive probability at every covariate profile in the target population.
Exchangeability concerns whether the measured covariates adequately account for confounding. Positivity concerns whether both exposure groups exist, in the population, at each relevant profile. One assumption does not imply the other: perfect control of confounding cannot supply observations for an exposure option that is never assigned at a profile.
Why a Missing Exposure Option Blocks Identification
For discrete covariates \(L\), the mean potential outcome under exposure \(a\) in the target population can be written as a weighted average across covariate profiles. Under consistency and conditional exchangeability, each profile-specific mean can be replaced by an observed outcome mean among people with that exposure and profile. The resulting identification formula is:
Here the profile weights \(P(L=l)\) refer to the target population. Positivity is needed because the formula requires information about \(E[Y\mid A=a,L=l]\) for each exposure option at each target-relevant profile.
The key step can be seen profile by profile. Exchangeability gives \(E[Y(a)\mid L=l]=E[Y(a)\mid A=a,L=l]\). Among people with \(A=a\), consistency gives \(Y=Y(a)\), so the conditional mean equals \(E[Y\mid A=a,L=l]\). This substitution is supported by observed data only when \(P(A=a\mid L=l)>0\). If that probability is zero, there are no people with that exposure at the profile, and the observed data cannot determine the missing conditional mean.
Thus, if positivity fails, the causal effect for the full target population is not identified by this adjustment argument, even if the other assumptions hold. A regression model may still produce a number by extrapolating into the unsupported profile, but that number depends on the model's extrapolation rather than on observed comparisons there.
Worked Example: A Treatment Rule Creates a Structural Violation
Suppose, hypothetically, a clinic assigns an intervention according to baseline severity: all patients classified as severe receive it, and all patients classified as mild do not. The target population includes both severity groups. The following counts summarize 100 patients under this rule.
| Baseline severity | Exposed, \(A=1\) | Unexposed, \(A=0\) | \(P(A=1\mid L)\) under the rule |
|---|---|---|---|
| Mild | 0 | 60 | 0 |
| Severe | 40 | 0 | 1 |
For mild patients, the exposure probability is \(0/60=0\), so the probability of no exposure is \(60/60=1\). For severe patients, the exposure probability is \(40/40=1\), so the probability of no exposure is \(0/40=0\). Both profiles violate positivity: neither has both options available.
The problem is not solved by having 100 patients or by measuring severity accurately. There is no observed outcome for an exposed mild patient or an unexposed severe patient. A comparison of intervention and no intervention across the full target population would require unsupported counterfactual information at both profiles. Under this assignment rule, the violation is structural: the rule itself makes one option impossible in each group.
Check the Joint Profile, Not Just Margins
A marginal comparison ignores some covariates. Such a comparison can look supported even when exposure assignment is deterministic within combinations of covariates. Positivity must be assessed conditional on the covariates used for the causal identification argument, considered jointly.
Worked Example: Marginal Overlap Hides a Joint Violation
Consider a hypothetical service that assigns exposure deterministically using both location and baseline risk. There are 80 people, split evenly across four location-by-risk profiles.
| Location | Risk | Exposed | Unexposed | \(P(A=1\mid \text{location, risk})\) |
|---|---|---|---|---|
| Urban | Low | 20 | 0 | 20/20 = 1 |
| Urban | High | 0 | 20 | 0/20 = 0 |
| Rural | Low | 0 | 20 | 0/20 = 0 |
| Rural | High | 20 | 0 | 20/20 = 1 |
At each location considered alone, there are 20 exposed and 20 unexposed people: both exposure options appear in urban areas and in rural areas. At each risk level considered alone, there are also 20 exposed and 20 unexposed people. Those marginal counts suggest overlap.
But the joint profiles tell a different story. For urban, low-risk people, \(P(A=1\mid L)=20/20=1\); for urban, high-risk people, it is \(0/20=0\). The rural profiles have probabilities \(0/20=0\) and \(20/20=1\), respectively. Each joint profile violates positivity. Therefore, marginal overlap in each covariate does not support a conditional comparison across the combined profiles.
The causal implication depends on the target. A target population containing all four profiles requires comparisons the data do not support. Restricting the target to a subset of profiles is a different causal question, and it must be specified and justified as such; it does not restore support for the excluded profiles.
Structural and Practical Positivity Problems
A structural violation occurs when the assignment process makes an option impossible for a profile, so its population probability is exactly zero. A policy may prohibit a treatment for a group, or a service may be available only to people meeting a rule. The opposite exposure option may likewise be impossible, giving a probability of one for exposure.
A practical positivity problem occurs when both options are possible, but one has a very small probability in some profiles. This is sometimes called near-violation of positivity. It matters because a finite sample may contain very few people, or none, receiving the uncommon option. Estimates based on those few people can be highly sensitive to individual outcomes and may have substantial uncertainty.
Population positivity and observed counts should not be confused. If a sample contains no exposed people in a profile, that fact alone does not prove the population probability is zero; the option may simply be rare. Conversely, a single exposed person does not establish that the comparison is well supported. The assignment mechanism, scientific context, target population, and available data all matter when judging support.
Worked Example: A Positive but Very Small Exposure Probability
Suppose, hypothetically, that in a remote-area profile each eligible person has probability \(0.02\) of receiving an exposure and probability \(0.98\) of not receiving it. Positivity holds in the population because \(0<0.02<1\). Now suppose a sample contains \(n=50\) people from this profile, with exposure assignments behaving independently at that probability.
The probability that none of the 50 is exposed is:
So there is about a 36.4% chance that this sample will contain no exposed person from the profile, despite population positivity. The expected number exposed is \(50(0.02)=1\), which is also very sparse support. A causal comparison at this profile would be fragile: in many samples, the exposed outcome mean cannot be calculated at all, and in samples with only one or two exposed people it would be highly uncertain.
This calculation illustrates a practical problem, not a structural one. The exposure option remains possible, but data may be too scarce to support a stable comparison. The probability is not evidence that the exposure effect is zero or that the causal assumptions are satisfied.
What Positivity Does and Does Not Say
Positivity is a support condition, not a balance condition. It does not guarantee that exposed and unexposed people have similar covariate distributions, and it does not remove confounding. Conversely, adjustment for confounding cannot repair a structural absence of one exposure option within a profile. These are separate concerns.
Positivity is also tied to the target population. If the target includes profiles where only one exposure option occurs, the full target causal comparison is unsupported by the observed assignment process. A scientifically meaningful target restricted to profiles with both options may avoid that particular issue, but it changes the estimand and the population to which the conclusion applies. That restriction should not be made silently just to obtain a comparison.
In applied work, positivity is therefore a question to ask before interpreting adjusted causal comparisons: for every relevant covariate profile, could people like this have received either exposure option? If the answer is no, the data do not support the full comparison. If one option is merely rare, the comparison may be possible in principle but weakly supported.
Common Mistake
- Wrong reasoning: “Both exposure groups include urban and rural people, so positivity holds.” This checks location only and overlooks the joint location-by-risk profiles.
- Correction: Check the probability of each exposure option conditional on the full covariate profile used for the causal analysis. Positivity requires both options within every profile represented in the target population.
- Wrong reasoning: “There are zero exposed people in this sample profile, so the population probability must be zero.” A finite sample can miss a rare exposure option.
- Correction: Distinguish a true structural zero from sparse sample data. Both can limit analysis, but they are not the same claim about the population assignment process.
Check Your Understanding
Answer each question by considering the target population and the full covariate profile.
- For a binary exposure, state the positivity condition in terms of \(P(A=1\mid L=l)\).
- A treatment is never given to patients with a particular baseline condition. If those patients are part of the target population, what kind of positivity problem occurs, and why?
- Why can overlap in each of two covariates considered separately fail to establish positivity conditional on both covariates jointly?
- A sample has no exposed people in one profile, but the exposure probability in the population is believed to be positive. Is this necessarily a structural violation? Explain.
- How does restricting the target population to profiles with both exposure options change the interpretation of a causal comparison?