Why Logistic Fitting Has a Special Connection to Overlap Weights
Overlap weights give more emphasis to units whose covariates make either treatment assignment plausible, and less emphasis to units whose assignment is nearly deterministic. Their practical appeal is not only that the weights are bounded: when the propensity score is estimated by an ordinary logistic regression, the fitted model’s likelihood equations imply an exact balance result in the observed sample. This is a finite-sample algebraic property, not a general guarantee that every covariate is balanced or that a causal effect is identified.
Let \(Z_i\) be the binary treatment indicator for unit \(i\), with \(Z_i=1\) for treated and \(Z_i=0\) for control. Let \(X_i\) denote that unit’s pretreatment covariates, and let \(e(X_i)=P(Z_i=1\mid X_i)\) be the propensity score. The earlier material on the overlap population estimand establishes the target population; here the focus is how logistic estimation yields weights that balance specified model terms in the sample.
The Logistic Model and Its Score Equations
Write \(f(X_i)\) for a vector of specified model terms, including a constant term for the intercept. These terms may include raw covariates, transformations, or interactions. Under the logistic propensity model,
The fitted score is \(\hat e_i=e_{\hat\beta}(X_i)\). For independent binary treatment indicators with this conditional logistic likelihood, the log-likelihood is
If the model is fit by unpenalized maximum likelihood, a finite interior maximum exists, and the fitting algorithm converges to it, its score with respect to \(\beta\) is zero. Differentiating gives
This equation is the key to the balance result. It holds for the included model terms because they appear in the likelihood score. Correct specification of the treatment-assignment model is not required for this particular sample algebra: the fitted maximum-likelihood equations still set those score components to zero. Correct specification and causal assumptions matter for interpreting a causal estimand, but they are distinct from this balance identity.
Deriving Exact Balance
For any component \(f_j(X_i)\) of the model-term vector, subtract the control-group overlap-weighted sum from the treated-group overlap-weighted sum:
The right-hand side is zero by the corresponding logistic score equation. The intercept is one of the components, so the same calculation with \(f_j(X_i)=1\) shows that the total overlap weight is equal in the treated and control groups. Dividing the equal weighted sums for \(f_j\) by these equal, positive total weights yields equal weighted means.
The result is sometimes called exact balance because the equality follows from the fitted score equations, rather than from a large-sample approximation. In numerical computation, small differences can remain because of rounding or convergence tolerances. The result concerns weighted sums and means in the estimation sample; it is not a claim that the groups have identical distributions or that every possible feature of their covariates is balanced.
Worked Example: Calculating an Individual’s Overlap Weight
Suppose a hypothetical logistic model uses one covariate \(x\) and has fitted linear predictor \(\hat\eta(x)=-1.2+0.8x\). Calculate the propensity score and overlap weight for a unit with \(x=1\), considering both possible treatment groups.
The fitted linear predictor is \(\hat\eta(1)=-1.2+0.8(1)=-0.4\). Therefore,
If the unit is treated, its overlap weight is \(1-\hat e(1)=1-0.4013=0.5987\). If the unit is a control, its overlap weight is \(\hat e(1)=0.4013\). These are alternative weights for the two possible treatment statuses, not two weights assigned to the same observed unit. The weight is largest near a fitted propensity of 0.5 and declines toward zero as the fitted propensity approaches either extreme.
A Finite-Sample Balance Calculation
The identity can be seen directly in a small hypothetical sample. Suppose the propensity model includes an intercept and a binary covariate \(X\), with a separate fitted log-odds for each value of \(X\). The following treatment counts produce finite maximum-likelihood estimates in both covariate groups:
| Covariate value | Treated count | Control count | Fitted propensity |
|---|---|---|---|
| \(X=0\) | 1 | 3 | \(1/4=0.25\) |
| \(X=1\) | 3 | 1 | \(3/4=0.75\) |
Worked Example: Verifying Exact Balance for a Binary Covariate
In the hypothetical sample above, the fitted probability in each group is its observed treated fraction. Calculate the overlap-weighted totals for treated and control units separately at each covariate value.
For \(X=0\), the treated unit receives weight \(1-0.25=0.75\), so the treated weighted total is \(1(0.75)=0.75\). Each of the three controls receives weight \(0.25\), giving a control weighted total of \(3(0.25)=0.75\).
For \(X=1\), each of the three treated units receives weight \(1-0.75=0.25\), giving \(3(0.25)=0.75\). The one control receives weight \(0.75\), also giving a total of \(0.75\). Thus the weighted sum of \(X\) is \(0.75\) in each group: only the \(X=1\) units contribute to that sum. The total weight in either group is \(0.75+0.75=1.50\), so the weighted mean of \(X\) is \(0.75/1.50=0.50\) for both groups.
This balance follows within each \(X\) category here because the model is saturated for the binary covariate. In a model that includes only selected terms, the general guarantee is equality for those terms, not necessarily equality within every category or on every transformation.
Which Features Balance, and Which Do Not?
The exactness applies to the columns actually used in the logistic regression. If the model includes \(1\) and \(X\), it guarantees equal overlap-weighted means of the intercept and \(X\). If it also includes \(X^2\), it guarantees equality of the weighted second moment. If it includes an interaction \(X_1X_2\), it guarantees equality of that product’s weighted mean. Any linear combination of included terms is balanced as well, since a weighted sum of quantities with equal means also has equal means.
However, balancing a mean does not imply that the full distributions match. For example, equal means of \(X\) do not generally imply equal means of \(X^2\), equal variances, or equal proportions above a threshold. Those additional features are guaranteed only if represented by terms in the fitted model (or happen to balance for other reasons). Balance of included terms is therefore a targeted diagnostic, not a universal certificate of covariate similarity.
Worked Example: Adding a Nonlinear Covariate Term
Suppose a hypothetical fitted propensity model uses \(f(X)=(1,X,X^2)\) and has linear predictor \(\hat\eta(x)=-0.6+0.4x+0.2x^2\). First calculate the fitted score and the two possible overlap weights at \(x=0,1,2\).
At \(x=0\), \(\hat\eta=-0.6\), so \(\hat e=1/(1+\exp(0.6))\approx0.3543\); the treated and control weights are approximately \(0.6457\) and \(0.3543\). At \(x=1\), \(\hat\eta=-0.6+0.4+0.2=0\), so \(\hat e=0.5000\) and both possible weights are \(0.5000\). At \(x=2\), \(\hat\eta=-0.6+0.8+0.8=1.0\), so \(\hat e=\exp(1)/(1+\exp(1))\approx0.7311\); the treated and control weights are approximately \(0.2689\) and \(0.7311\).
If this model is fit by converged, unpenalized maximum likelihood with an intercept, its score equations guarantee equal weighted means of \(X\) and \(X^2\) in the two treatment groups, as well as equal total weights. The calculations at the three covariate values alone do not establish those sample-level equalities; they describe individual fitted scores. If \(X^2\) were omitted, the fitted model would guarantee balance of \(X\), but not of \(X^2\).
Assumptions, Diagnostics, and Failure Modes
Three kinds of conditions should be kept distinct. First, the exact balance theorem is conditional on the fitting procedure: the treatment model is logistic, it has an intercept, it is fit by ordinary unpenalized maximum likelihood, and the fit has a finite converged solution. Complete or quasi-complete separation can prevent a finite maximum-likelihood estimate; nonconvergence means the score equations may not be sufficiently close to zero. Penalization, constraints, or other fitting procedures generally change the estimating equations, so the exact result need not hold.
Second, the algebraic result does not by itself identify a causal effect. A causal interpretation still requires appropriate causal conditions, including consistency, conditional exchangeability given the measured pretreatment covariates, and positivity for the target population. The propensity model must also be suitable for the intended causal analysis. Exact balance of selected terms cannot repair unmeasured confounding, incorrect treatment or outcome measurement, or a lack of relevant covariate overlap.
Third, inspect the data and the fitted analysis. Confirm the model converged and check whether the maximum-likelihood fit is finite. Compute treated and control overlap-weighted means for every included term; differences should be zero up to numerical tolerance. Also inspect scientifically important covariates and transformations not included in the model, since they have no automatic guarantee. Examine the propensity-score distribution and the effective information remaining after weighting. Bounded overlap weights avoid arbitrarily large individual weights, but a small or highly concentrated weighted sample can still yield imprecise estimates.
Overlap weights differ from inverse probability weights for the average treatment effect, which target a different population and can become large near propensity-score extremes. They also differ from matching, which constructs comparisons by pairing or selecting units rather than by this likelihood-score identity. These approaches answer different design questions; none is universally preferable. Overlap weights emphasize the population with covariate profiles compatible with either treatment, and their exact balance applies to included logistic terms. They do not promise balance of omitted features, and they do not turn observational treatment assignment into random assignment.
Common Mistake
Choose pretreatment covariates and any transformations or interactions needed to represent the balance features of interest.
Use the fitted probabilities only after checking convergence and whether a finite unpenalized maximum-likelihood estimate exists.
Use \(1-\hat e_i\) for treated units and \(\hat e_i\) for controls; verify included-term balance and inspect other relevant features separately.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- What condition on the logistic model’s terms is needed for the total treated and control overlap weights to be equal?
- A fitted model includes an intercept and \(X\), but not \(X^2\). Which of these terms is guaranteed to have equal weighted means?
- Why does an exact finite-sample balance result not establish that the weighted comparison is causal?
- For a treated unit with fitted propensity score \(0.20\), what is its overlap weight? What would the weight be for a control unit with the same score?
Solutions. (1) The model must include an intercept, and the unpenalized logistic maximum-likelihood fit must have a finite converged solution; the intercept score equation gives equal total weights. (2) The intercept and \(X\) are balanced; \(X^2\) is not guaranteed because it is omitted. (3) Balance is an algebraic property of included fitted-model terms. Causal interpretation additionally requires appropriate conditions such as consistency, conditional exchangeability, and positivity. (4) The treated weight is \(1-0.20=0.80\); the control weight is \(0.20\).