Why High-Dimensional Overlap Is Difficult to See
Overlap is a property of the joint covariate distribution, not simply a collection of separate comparisons. With a binary treatment \(A\) and a covariate vector \(X\), conditional positivity requires that both treatment groups be possible at covariate values in the target population. In finite data, however, the joint covariate space can be thinly populated even when every covariate looks well represented on its own.
This becomes especially important as the number of measured covariates grows. A graph of one covariate at a time may show broad overlap, while particular combinations of covariates occur in only one treatment group or in very few people overall. The practical task is therefore to look for evidence of sparse or separated regions using several complementary views, and to distinguish what those views can establish from what they cannot.
The distinction between population support and observed support matters. An absent treatment group in a small sample may reflect sampling variability, not a structural impossibility. Conversely, observing a few people from each group does not guarantee that the data contain enough information for stable comparisons in that part of the covariate space.
Why Support Becomes Sparse as Dimensions Accumulate
A simple count illustrates the problem for categorical covariates. If there are \(p\) binary covariates, they define \(2^p\) possible joint cells. If \(n_1\) treated observations were spread evenly across those cells, the average treated count per cell would be \(n_1/2^p\). In real data, the distribution is rarely even, so some cells can be much sparser than this average.
For continuous covariates, the same issue appears when we define a local neighborhood around a point \(x\). Suppose the control-group covariates have a smooth density \(f_0\) near \(x\), and consider a ball of radius \(r\) in \(p\) dimensions. The expected number of controls in the ball is \(n_0\) times the probability that one control falls there. For a small ball, that probability is approximately its volume times the local density:
Here \(v_p\) is the volume of the unit ball in \(p\) dimensions. For a fixed small radius, the factor \(r^p\) shrinks rapidly as \(p\) increases. Thus a local region that contains many opposite-treatment observations in a low-dimensional analysis may contain few or none when more covariates are considered jointly. This is one reason that nearest-neighbor comparisons and finely divided subgroups can become unstable in high dimensions.
Worked Example: Expected Sparsity Across Binary Covariates
Suppose a hypothetical sample contains 120 treated and 120 control individuals, and six binary pretreatment covariates define the joint cells. Assume, just for this calculation, that each treatment group is distributed equally over all possible cells.
There are \(2^6=64\) cells. The average treated count per cell is
The same average applies to controls. An average below two per group means that even under the favorable assumption of equal distribution, most cells cannot be expected to contain many observations from each group.
For any specified cell, the probability that it receives no treated individuals is \((63/64)^{120}\), because each treated individual has probability \(63/64\) of falling outside that cell. Numerically,
Under independent sampling of the two groups, the probability that this cell is occupied by at least one treated and at least one control individual is approximately
So even under equal cell probabilities, a specified cell has about a 27.9% chance of being empty in at least one group. This is a sampling illustration, not a claim that the population lacks positivity. Unequal cell probabilities would generally make the sparsity problem more severe in less common cells.
Use Projections as Screens, Not Proof
A low-dimensional projection examines only part of \(X\), such as a pair of continuous covariates or a covariate against an estimated propensity score. Such plots are useful because they can reveal broad separation, unusual clusters, and regions deserving closer inspection. They are not a guarantee of joint overlap: omitting dimensions can make groups look similar after their separation in the omitted dimensions has been hidden.
Worked Example: Matching Margins but Separated Joint Locations
Consider a hypothetical sample with two covariates already expressed in fixed scale units. The treated observations are at \((0,0)\) and \((2,2)\); the controls are at \((0,2)\) and \((2,0)\). This small example is designed to show how separate projections can be misleading.
For the first covariate, both groups have values 0 and 2. The same is true for the second covariate. Therefore, each one-dimensional projection has identical observed values in the two groups. But in the joint plane, every treated observation is at least two units from its nearest control:
The treated point \((2,2)\) also has nearest-control distance 2. The median nearest opposite-treatment distance for these four observations is therefore 2 units. The projections alone do not reveal this joint separation. In a larger data set, a plot of each covariate pair can help expose such patterns, but inspecting every pair still does not fully describe support in a space with many dimensions.
The example’s four points do not establish a population support gap. They show why marginal overlap is not sufficient evidence of joint overlap, and why projections should be combined with other checks.
Three Complementary Checks
No single summary captures all aspects of high-dimensional support. A useful assessment combines projections, counts in meaningful subgroups, and distances to the opposite treatment group. The covariates and summaries should be chosen using the design and subject-matter context, not by searching for a result that looks reassuring.
Plot or summarize important covariates and selected pairs by treatment group. Look for broad gaps, clustering, and tails represented mostly by one group. Treat a reassuring projection as a screen, not proof about the full joint space.
For categorical covariates, tabulate treatment-specific counts for prespecified combinations. For continuous covariates, use clinically or substantively meaningful categories cautiously, or define neighborhoods whose scale and purpose are stated explicitly.
For each observation, calculate its distance to the nearest observation in the opposite group using a stated scaling or distance measure. Report the distribution, not just its average, and inspect observations with unusually large distances.
For example, standardized Euclidean distance between observations \(i\) and \(j\) can be defined using prespecified covariate scales \(s_k\):
The nearest opposite-treatment distance for observation \(i\) is the smallest \(d(i,j)\) among observations \(j\) in the other treatment group. The resulting distances can be summarized separately for treated and control observations, for example by the median, upper quantiles, and the largest values. Units depend on the chosen scaling; with standardized covariates, the distance is a dimensionless summary, not a probability or a direct measure of causal bias.
Worked Example: Subgroup Counts Reveal a Thin Cell
In a hypothetical sample, three binary covariates define eight joint subgroups. The following table gives treated and control counts. Each row is a distinct combination, with 1 indicating presence of the corresponding characteristic.
| Covariate combination | Treated count | Control count |
|---|---|---|
| 000 | 18 | 17 |
| 001 | 12 | 11 |
| 010 | 10 | 9 |
| 011 | 8 | 7 |
| 100 | 14 | 13 |
| 101 | 10 | 9 |
| 110 | 6 | 8 |
| 111 | 2 | 6 |
The treated counts sum to 80 and the control counts sum to 80. Every observed subgroup contains both groups, but the 111 subgroup has only eight people. Its observed treated fraction is
This subgroup is much thinner than the others and warrants attention: estimates that rely on comparisons within it would be based on just two treated observations. The counts do not establish the true treatment probability in that subgroup, and the observed fraction \(0.250\) is not proof of a population positivity violation. They do show that the sample offers limited information there.
Interpreting Distance Summaries and Diagnosing Problems
Distance summaries are useful only when their construction is transparent. A distance can change substantially when a variable is rescaled, when correlated variables are counted repeatedly, or when a different metric is used. Standardizing each variable addresses differences in measurement units, but it does not remove dependence among covariates. A distance measure should therefore be described along with the variables included, their scales, and the treatment group used as the comparison set.
Look beyond the mean distance. A small mean can conceal a few observations with very large nearest opposite-treatment distances, while a large mean may be driven by a small number of unusual cases. Report group-specific distributions and examine the observations in the upper tail in context. Also check whether the conclusion is sensitive to reasonable choices of scales or distance definitions. Such sensitivity is a warning that a single numerical summary is not a robust description of support.
For subgroup counts, define categories before examining outcomes and avoid creating so many combinations that nearly every cell is empty. Broad groups can conceal finer gaps; excessively fine groups can manufacture apparent sparsity from limited data. Counts are most informative when the combinations correspond to important covariate patterns and the resulting group sizes are shown directly.
How the Checks Fit Together
The diagnostics answer different questions. Projections make broad patterns visible but discard dimensions. Subgroup counts show whether particular categorical combinations have representation from both groups, but depend on how groups are defined. Distance summaries consider multiple covariates at once, but depend on the metric and scale and can be hard to interpret in very high dimensions. Agreement among these checks strengthens the case that the observed data provide useful support; disagreement is a reason to investigate, not to select whichever diagnostic is most favorable.
These checks complement propensity-score comparisons rather than replace them. A score is a one-dimensional summary and can be convenient for displaying broad treatment-group separation, but similar scores do not guarantee that every covariate pattern is well represented. Direct examination of covariates can expose patterns hidden by a compressed summary. Conversely, a large collection of direct comparisons can be difficult to summarize, which is why a small set of well-chosen projections, counts, and distances is more useful than an unstructured search through every possible view.
Evidence of sparse support has practical consequences: estimates in thin regions may rely heavily on extrapolation or a small number of observations. It does not, by itself, dictate a particular restriction or define a new target population. Any decision to restrict the analysis requires an explicit rule and an account of the population that remains. The present diagnostics are intended to characterize the design and locate uncertainty, not to choose a restriction based on observed outcomes.
Common Mistakes
- Equating marginal overlap with joint overlap. The wrong reasoning is that each covariate overlaps separately, so their joint combinations must overlap. The correction is to inspect combinations or multivariate distances; separate projections can hide dependence and joint separation.
- Treating an empty sample cell as proof of zero population probability. A cell may be empty because it is rare and the sample is limited. Report the observed count as a finite-sample diagnostic, not a population certainty.
- Reporting one distance average as a complete diagnosis. An average can conceal a problematic tail and is affected by scales and metric choice. Show the distribution, explain the metric, and examine large distances.
- Assuming that every nonempty cell is adequate. A cell with one observation in one group is technically represented in the sample but may provide very little information for a comparison.
- Searching many projections until one appears reassuring. Select views based on important covariates and report the scope of the inspection. A few favorable projections cannot establish support in every dimension.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- With five binary covariates and 96 treated observations, what is the average treated count per joint cell under an equal-distribution calculation? What does that average fail to establish?
- Why can two treatment groups have identical one-dimensional covariate distributions but poor joint overlap?
- A joint subgroup contains 3 treated and 9 control observations. Calculate its observed treated fraction and state one conclusion the count supports and one it does not support.
- Why should a distance diagnostic report more than the mean nearest opposite-treatment distance?
- State one strength and one limitation of low-dimensional projections for assessing high-dimensional support.
Solutions. (1) Five binary covariates define \(2^5=32\) cells, so the average is \(96/32=3\) treated observations per cell. This does not establish that every cell has three observations, nor that the population treatment probability is positive in every cell. (2) The groups can have matching distributions on each coordinate while occupying different combinations of coordinates; the four-point example above demonstrates this. (3) The observed treated fraction is \(3/(3+9)=0.250\). The small total count signals limited sample information, but it does not prove that the population probability of treatment is \(0.250\) or that positivity fails. (4) The mean can hide a tail of observations far from the other group; the distribution and large distances should also be examined. (5) Projections can reveal broad gaps and clusters in an interpretable view, but they omit dimensions and therefore cannot establish joint overlap on their own.