Bounding Individual Influence Contributions

Why Bound Individual Contributions?

A large inverse-probability weight can make one observation dominate a weighted estimating equation. Clipping the weight at an upper threshold $c$ puts a direct ceiling on the weight multiplier. The resulting contribution bound is useful, but it answers a narrower question than “is the estimator stable?” Its strength also depends on the score being multiplied, the sensitivity of the estimating equation, and whether the contributions being bounded include all sources of estimation uncertainty.

The discussion below uses independent, identically distributed observations and treats the weights as known functions of each observation. The same algebra applies conditionally when weights are regarded as fixed, but uncertainty from estimating weights requires additional terms in the estimator’s first-order representation.

From a Weight Bound to a Contribution Bound

Definition: Let $O_i$ be observation $i$, let $W_i\geq 0$ be its weight, and let $\psi(O_i;\beta)\in\mathbb{R}^p$ be its estimating function. For an upper threshold $c>0$, define the clipped weight and clipped contribution by $$ W_i^{(c)}=\min(W_i,c), \qquad Z_i^{(c)}(\beta)=W_i^{(c)}\psi(O_i;\beta). $$ The clipped weighted estimating equation is $n^{-1}\sum_{i=1}^n Z_i^{(c)}(\beta)=0$. Its population target $\beta_c$ satisfies $E[Z_i^{(c)}(\beta_c)]=0$, when such a target exists.

Throughout the vector bounds, $\|\cdot\|_2$ denotes the Euclidean norm, and for a matrix $B$, $\|B\|_{\mathrm{op}}=\sup_{\|x\|_2=1}\|Bx\|_2$ is the induced operator norm. Suppose that $\|\psi(O_i;\beta_c)\|_2\leq K$ almost surely. Since $0\leq W_i^{(c)}\leq c$, homogeneity of the Euclidean norm gives the observation-level bound

\|Z_i^{(c)}(\beta_c)\|_2 = W_i^{(c)}\|\psi(O_i;\beta_c)\|_2 \leq cK.

This is a bound on the vector entering the estimating equation, not yet a bound on how much the fitted parameter changes when an observation changes. That second step involves the local sensitivity of the equation to $\beta$.

Worked Example: Clipping a Scalar Contribution

Suppose an observation has raw weight $W_i=7$, threshold $c=3$, and scalar estimating function $\psi(O_i;\beta_c)=-2$. Before clipping, its contribution is $7(-2)=-14$. After clipping,

W_i^{(c)}=\min(7,3)=3, \qquad Z_i^{(c)}(\beta_c)=3(-2)=-6.

If every observation’s estimating function at $\beta_c$ has absolute value at most $K=2$, then every clipped contribution has absolute value at most $cK=3(2)=6$. This observation attains that bound. The calculation bounds the magnitude of one summand; it does not determine the variance of the sum or the uncertainty of the fitted parameter.

How the Bound Propagates to the Estimator

To relate a score contribution to the estimator, assume the estimating equation is differentiable near $\beta_c$, that differentiation and expectation can be interchanged under suitable domination, and that its sensitivity matrix

A_c = E\!\left[ \frac{\partial Z_i^{(c)}(\beta_c)}{\partial\beta^T} \right]

is nonsingular. Also assume a consistent, regular solution $\hat\beta_c$ admits the usual first-order linearization; in particular, the independent contributions satisfy a central limit theorem and the empirical derivative converges to $A_c$. These conditions yield

\sqrt{n}(\hat\beta_c-\beta_c) = -A_c^{-1}\frac{1}{\sqrt{n}}\sum_{i=1}^n Z_i^{(c)}(\beta_c) +o_p(1).

The minus sign comes from solving the first-order expansion of the estimating equation. Because $E[Z_i^{(c)}(\beta_c)]=0$, the first-order influence contribution for one observation is $-A_c^{-1}Z_i^{(c)}(\beta_c)$. Its norm is at most $\|A_c^{-1}\|_{\mathrm{op}}cK$. If contributions are instead written as centered terms $Z_i^{(c)}-E[Z_i^{(c)}]$, their norm is at most $2cK$, by the triangle inequality. At the target defined above the mean is zero, so the centered and uncentered score coincide.

Result: Suppose $Z_i^{(c)}(\beta_c)$ are independent and identically distributed, $E[Z_i^{(c)}(\beta_c)]=0$, $A_c$ is nonsingular, and $\|\psi(O_i;\beta_c)\|_2\leq K$ almost surely. Then the trace of the first-order covariance matrix obeys $$ \operatorname{tr}\!\left\{ A_c^{-1}\operatorname{Var}\!\left(Z_i^{(c)}(\beta_c)\right)A_c^{-T} \right\} \leq \|A_c^{-1}\|_{\mathrm{op}}^2c^2K^2. $$ Here $A_c^{-T}=(A_c^{-1})^T$. Under the regularity conditions for the linearization, this is the trace of the asymptotic covariance of $\sqrt{n}(\hat\beta_c-\beta_c)$.

Proof. For any mean-zero random vector $Z$ with finite second moment and fixed matrix $B$, the trace of $B\operatorname{Var}(Z)B^T$ equals $E[\|BZ\|_2^2]$. Apply this with $Z=Z_i^{(c)}(\beta_c)$ and $B=A_c^{-1}$. By the induced operator norm,

\begin{aligned} \operatorname{tr}\!\left\{ A_c^{-1}\operatorname{Var}(Z_i^{(c)})A_c^{-T} \right\} &=E\!\left[\|A_c^{-1}Z_i^{(c)}\|_2^2\right]\\ &\leq \|A_c^{-1}\|_{\mathrm{op}}^2 E\!\left[\|Z_i^{(c)}\|_2^2\right]\\ &\leq \|A_c^{-1}\|_{\mathrm{op}}^2c^2K^2. \end{aligned}

The last inequality uses the almost-sure contribution bound. This proof uses the Euclidean norm both for the vector contribution and for the operator norm; changing norms requires corresponding norm-conversion constants. $\square$

Worked Example: A Vector Covariance Bound

Consider a two-dimensional estimating equation at its target with clipped contributions $Z_i^{(c)}=(2,4)$ or $(-2,-4)$, each with probability $1/2$. Their mean is zero and their covariance matrix is

\operatorname{Var}(Z_i^{(c)}) = \begin{pmatrix} 4&8\\ 8&16 \end{pmatrix}.

Suppose the sensitivity matrix is $A_c=\operatorname{diag}(2,1)$, so $A_c^{-1}=\operatorname{diag}(1/2,1)$ and $\|A_c^{-1}\|_{\mathrm{op}}=1$. The transformed influence contributions, ignoring the overall minus sign, are $(1,4)$ and $(-1,-4)$. Their covariance is

A_c^{-1}\operatorname{Var}(Z_i^{(c)})A_c^{-T} = \begin{pmatrix} 1&4\\ 4&16 \end{pmatrix}, \qquad \operatorname{tr}=17.

For example, these contributions are compatible with $c=2$ and score vectors $(1,2)$ or $(-1,-2)$, whose Euclidean norm is $K=\sqrt{5}$. The theorem’s bound is $1^2(2^2)(\sqrt{5})^2=20$, and $17\leq20$, as required. The bound is valid but not exact: it uses the largest possible score norm and the largest stretching factor of $A_c^{-1}$, whether or not both are attained in the same direction.

Why a Clipped Weight Does Not Guarantee Low Variance

The result is a ceiling conditional on its ingredients, not a promise that the ceiling is small. First, $K$ may be large or no finite bound may exist. Weighted estimating functions often include outcomes or covariates, and clipping the weight does not clip those quantities. Second, $\|A_c^{-1}\|_{\mathrm{op}}$ can be large when the estimating equation is weakly sensitive to some parameter direction. In that case even moderate score variation can produce substantial parameter variation. Third, a bound that grows as $c^2$ can itself be large when the chosen threshold is large.

Worked Example: Bounded Weights but Large Estimator Variance

Let $X$ be Bernoulli with $P(X=1)=q=0.01$. Conditional on $X=1$, let $Y$ equal $10$ or $-10$, each with probability $1/2$; when $X=0$, its value does not affect the equation. Use weights $W^{(c)}=1$ for everyone and the scalar estimating equation

\frac{1}{n}\sum_{i=1}^n X_i(Y_i-\beta)=0.

The population target is $\beta_c=0$, since $E[XY]=0$. At this target, the score contribution is $Z=XY$, so $E[Z]=0$ and $\operatorname{Var}(Z)=E[X^2Y^2]=0.01(100)=1$. The derivative of the population equation with respect to $\beta$ is $A_c=-E[X]=-0.01$. Thus the scalar first-order variance is

A_c^{-2}\operatorname{Var}(Z) = \frac{1}{(0.01)^2}(1) = 10{,}000.

The weights are perfectly bounded by $1$, and each score contribution has absolute value at most $10$. Yet only one percent of observations contribute, making the equation insensitive to $\beta$; that weak sensitivity amplifies the variance. For a regular large sample, the corresponding leading variance of $\hat\beta_c$ is $10{,}000/n$. The example shows why checking the weight cap alone cannot establish that the estimator is precise.

Diagnostics, Comparisons, and Limitations

A useful diagnostic sequence separates the pieces of the argument rather than reporting only the threshold:

Inspect the actual contributions.
Calculate the norm or relevant component of each clipped weighted score. A weight cap gives no useful bound if the scores themselves are extreme.

Inspect sensitivity.
Examine the estimated derivative matrix and whether it is close to singular. A large inverse norm signals that bounded score changes can still have a large effect on the fitted parameter.

Check the uncertainty calculation.
Compare the empirical variability of contributions and the resulting sandwich covariance with the crude upper bound. The bound is a worst-case inequality, not a variance estimate.

Check what the contribution omits.
If weights, nuisance models, or other components are estimated, verify that the influence calculation accounts for their estimation. A bound on the weighted score alone is not automatically a bound on the full estimator’s influence.

Compared with raw weights, upper clipping supplies a hard maximum for the weight multiplier; raw weights have no such ceiling. Stabilization changes the scale or numerator of weights but, by itself, does not generally impose an upper bound. Clipping the score or outcome as well could bound a larger part of the contribution, but that changes the estimating equation and potentially its target. These are distinct interventions, not interchangeable ways to obtain the same guarantee.

The trace bound also does not address dependence between observations, such as clustering, unless the variance calculation is adapted to the dependence structure. Nor does it establish that clipping preserves the original causal target: the clipped equation defines $\beta_c$, which may differ from the target under untruncated weights. The bound is most informative when the score is controlled, the sensitivity is well-conditioned, the sampling assumptions are credible, and the estimating equation includes all terms relevant to the estimator.

Common Mistake

Wrong reasoning: “Every clipped weight is at most $c$, so the estimator must have low variance.” This confuses a bound on one factor in a score contribution with a bound on the estimator’s variance. The score may be large, the inverse sensitivity may be large, or the observations may not satisfy the independence assumptions used in the calculation.

Correction: State the complete bound and its assumptions. For the result above, the variance ceiling depends on $c^2K^2\|A_c^{-1}\|_{\mathrm{op}}^2$, not on $c$ alone. Also distinguish the covariance of first-order contributions from finite-sample variance and from uncertainty due to estimated weights or nuisance models.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

If $c=4$ and $\|\psi(O_i;\beta_c)\|_2\leq 3$, what is the largest guaranteed norm of a clipped score contribution?
Why does the first-order variance bound involve $\|A_c^{-1}\|_{\mathrm{op}}$, as well as the weight threshold?
Suppose the score is mean zero and bounded in Euclidean norm by $K$, and the clipped weight is at most $c$. State the bound on the trace of its transformed covariance.
Give one reason why a small weight cap might not yield a precise estimator.

Solutions. (1) The bound is $cK=4(3)=12$. (2) The inverse sensitivity maps score perturbations into parameter perturbations; a nearly singular derivative can magnify them. (3) The trace is at most $\|A_c^{-1}\|_{\mathrm{op}}^2c^2K^2$, under the stated mean-zero and covariance conditions. (4) For example, the score may be unbounded or highly variable, or $A_c$ may be nearly singular. In each case, the weight cap alone does not control the relevant source of uncertainty.