How Truncation Changes the Estimating Equation

From Clipped Weights to a New Score Equation

A weighted marginal structural model (MSM) is fitted by solving a weighted estimating equation. Clipping the weights does not merely make large contributions smaller: unless the affected observations have a special relationship to the score residuals, it changes the equation being solved. The central quantity is not just how much the weights change, but how those changes line up with the estimating function.

Let $O$ denote the observed data for one individual, $W=W(O)\geq 0$ the stabilized weight, and $U(O;\beta)\in\mathbb{R}^p$ the estimating function for an MSM parameter $\beta\in\mathbb{R}^p$. Assume the untruncated weighted score is integrable and identifies the target parameter $\beta_0$ through

E[W U(O;\beta_0)]=0.

This equation is the starting point here: under the usual causal identification assumptions for the MSM and a correctly specified estimating equation, it is the population equation for the untruncated weighted analysis. We do not need to specify a particular treatment history or outcome model to examine what clipping does to it.

Definition: For a fixed clipping rule $C$, define the clipped weight $W_C=C(W)$. The clipped population score is $G_C(\beta)=E[W_C U(O;\beta)]$. A clipped population solution $\beta_C$ satisfies $G_C(\beta_C)=0$. For upper clipping at $c>0$, $C(w)=\min(w,c)$; for lower and upper clipping at $l,u$, $C(w)=\min\{u,\max(l,w)\}$.

The distinction between $\beta_0$ and $\beta_C$ is important. The first solves the score using the original weights. The second solves the score actually used after clipping. If a clipped solution is unique and the estimating equation is correctly implemented, computation can be entirely accurate while $\beta_C\neq\beta_0$.

The Exact Residual-Weight Decomposition

Write the weight change as $\Delta_C=W_C-W$. At the untruncated target parameter, add and subtract the original score:

G_C(\beta_0) =E[W_C U(O;\beta_0)] =E[W U(O;\beta_0)]+E[\Delta_C U(O;\beta_0)] =E[\Delta_C U(O;\beta_0)].

The last equality uses the assumed untruncated score equation. Thus the clipped equation still has $\beta_0$ as a solution exactly when the expected change-weighted score is zero. This is a condition on the score residuals after weighting changes, not simply on the amount of clipping.

Theorem: Suppose $E[W\|U(O;\beta_0)\|]$ and $E[W_C\|U(O;\beta_0)\|]$ are finite, and $E[W U(O;\beta_0)]=0$. Then $$ G_C(\beta_0)=E[(W_C-W)U(O;\beta_0)]. $$ Consequently, $\beta_0$ solves the clipped population equation if and only if $E[(W_C-W)U(O;\beta_0)]=0$. This statement concerns whether $\beta_0$ is a solution; uniqueness of the clipped solution is a separate requirement.

The theorem follows directly from linearity of expectation and the two integrability assumptions, which ensure the expectations in the decomposition exist. It does not say that clipping must change the solution: the residual-weight term can vanish through cancellation, or because clipping leaves the relevant weights unchanged. Nor does a nonzero term by itself guarantee a unique alternative root; it establishes that $\beta_0$ is not a root of the clipped equation.

Worked Examples

Worked Example: Clipping Shifts a Weighted Intercept

Suppose there are two equally likely types, with $(W,Y)=(0.5,0)$ for the first type and $(W,Y)=(1.5,2)$ for the second. Fit the intercept-only score $U(Y;\beta)=Y-\beta$, and upper-clip at $c=1.2$.

Before clipping, $E[W]=\{0.5+1.5\}/2=1$ and $E[WY]=\{0.5(0)+1.5(2)\}/2=1.5$. Therefore the untruncated root is $\beta_0=E[WY]/E[W]=1.5$.

After clipping, the weights are $0.5$ and $1.2$, so $E[W_C]=0.85$ and $E[W_CY]=1.2$. The clipped root is

\beta_C=\frac{E[W_CY]}{E[W_C]}=\frac{1.2}{0.85} =\frac{24}{17}\approx 1.412.

The residuals at the original root are $-1.5$ and $0.5$. The weight changes are $0$ and $-0.3$, so $E[(W_C-W)(Y-\beta_0)]=\{0+(-0.3)(0.5)\}/2=-0.075$. Since $G_C(\beta)=E[W_C(Y-\beta)]$, its root differs from $\beta_0$ by $-0.075/0.85\approx-0.088$, matching $1.412-1.5$. The high-weight type has a positive residual at $\beta_0$; reducing its weight pulls the clipped intercept down.

Worked Example: A Regression Score Can Shift in Both Coordinates

Consider a weighted linear regression with design vector $z=(1,x)^\mathsf{T}$ and score $U=z(Y-z^\mathsf{T}\beta)$. Three equally likely observations have $x=0,1,2$, weights $W=0.5,1.5,1$, and outcomes $Y=-2,5,3.5$, respectively. Clip above at $1.2$.

At $\beta_0=(1,2)^\mathsf{T}$, the fitted values are $1,3,5$, giving residuals $-3,2,-1.5$. The untruncated weighted residual-score sum is $(0.5)(1,0)(-3)+(1.5)(1,1)(2)+(1)(1,2)(-1.5)=(0,0)$. Thus $\beta_0$ solves the untruncated equation.

Only the second weight changes, by $-0.3$. The change in the expected score at $\beta_0$ is therefore $(-0.3)(1,1)(2)/3=(-0.2,-0.2)^\mathsf{T}$, which is nonzero. To find the clipped root, the common factor $1/3$ in the expected score can be dropped. The clipped weighted design matrix and outcome vector are

\sum_i W_{C,i}z_i z_i^\mathsf{T} = \begin{pmatrix}2.7&3.2\\3.2&5.2\end{pmatrix}, \qquad \sum_i W_{C,i}z_iY_i = \begin{pmatrix}8.5\\13\end{pmatrix}.

The determinant is $2.7(5.2)-3.2^2=3.8$. Solving gives $\beta_{C,0}=\{8.5(5.2)-3.2(13)\}/3.8=0.6842$ and $\beta_{C,1}=\{2.7(13)-3.2(8.5)\}/3.8=2.0789$. The intercept and slope both differ from $(1,2)$. The clipped score changes in both coordinates because the affected observation contributes to both components of the regression score.

Worked Example: Clipping Need Not Move the Root

For an intercept-only score, take three equally likely types with weights $0.5,1,1.5$ and outcomes $Y=2,-1,0$, respectively. Clip above at $1.2$. The original weighted outcome sum is $(0.5)(2)+(1)(-1)+(1.5)(0)=0$, so the untruncated weighted mean is zero. After clipping it remains $(0.5)(2)+(1)(-1)+(1.2)(0)=0$.

The denominator changes from $E[W]=1$ to $E[W_C]=0.9$, but both are positive, so the weighted mean remains zero. Equivalently, at $\beta_0=0$ the only changed weight multiplies a zero residual, making $E[(W_C-W)(Y-\beta_0)]=0$. This example shows why altered weights alone do not establish altered parameter values.

How the Population Solution Moves

The exact decomposition says whether the original parameter remains a root. A local derivative describes the direction of movement when the weights are changed gradually. This is a local result; it does not assert that a large, fixed clipping change has a small effect.

Theorem: Let $W_t=W+t\Delta$ for $t$ in an open interval containing zero, and define $G_t(\beta)=E[W_tU(O;\beta)]$. Suppose $G_t(\beta)$ is continuously differentiable in $(t,\beta)$ near $(0,\beta_0)$, $G_0(\beta_0)=0$, and the matrix $J_0=\partial G_0(\beta)/\partial\beta^\mathsf{T}\vert_{\beta=\beta_0}$ is nonsingular. Then, locally, there is a differentiable root $\beta(t)$ with $\beta(0)=\beta_0$, and $$ \beta'(0)=-J_0^{-1}E[\Delta U(O;\beta_0)]. $$

Proof. By the implicit function theorem, the differentiability and nonsingularity assumptions give a local differentiable solution to $G_t(\beta(t))=0$. Differentiate this identity at $t=0$. The product and chain rules give $0=E[\Delta U(O;\beta_0)]+J_0\beta'(0)$. Solving this linear equation yields the stated derivative. The differentiation under the expectation is justified by the assumed continuous differentiability of $G_t(\beta)$, which includes the required integrability and interchange conditions. $\square$

For the fixed clipping rule, $\Delta=W_C-W$ describes the full change, but the theorem’s derivative is about a path of weights near $t=0$. It should not be treated as an exact finite-change formula. For the actual fixed rule, the exact score decomposition remains valid, and a clipped root must solve $G_C(\beta_C)=0$.

Assumptions, Diagnostics, and Alternatives

The score argument requires finite expectations for the weighted estimating functions and a well-defined untruncated target score. To interpret that score causally, the usual identification conditions for the MSM must also hold, including consistency, exchangeability conditional on the relevant observed history, and positivity. Clipping does not repair a failure of those conditions. For a unique, stable clipped solution, one also needs adequate identification by the clipped score—for example, a nonsingular score derivative near the root.

A useful diagnostic is to evaluate the empirical change-weighted score at the untruncated estimate: $\sum_i(W_{C,i}-W_i)U(O_i;\hat\beta)$. Its coordinates indicate which score components are directly affected by clipping. Also check which observations change weight, whether those observations have unusually large residual-score contributions, and whether the fitted root moves substantially across reasonable prespecified cutoffs. These checks describe sensitivity; a small empirical term is not proof that the population term is zero.

Clipping differs from leaving the weights unchanged, and also from deletion. With no clipping, the original score is retained. Deletion replaces affected contributions by zero, which is another weight modification and therefore has its own residual-weight term. A common positive rescaling of every weight leaves the zero of a homogeneous weighted score unchanged, but clipping is generally not a common rescaling: it changes relative contributions. Consequently, normalizing the clipped weights to have a chosen sample mean does not, by itself, undo the score change.

Pitfall: “The clipped weights are still positive, so the original MSM parameter remains the solution.” Positivity alone is insufficient: the required condition is $E[(W_C-W)U(O;\beta_0)]=0$.
Pitfall: “A nonzero empirical residual-weight sum proves the population parameter is biased by that exact amount.” The empirical sum is a diagnostic, not an exact population shift; the parameter change also depends on the score’s response to $\beta$.
Limitation: The decomposition identifies the source of a changed equation, not the size of the final parameter change without solving that equation or adding conditions for an approximation.

Key takeaway: Clipping replaces the original population score by a new one. At the untruncated target, the discrepancy is exactly the expected product of the weight change and the original score. That term—not clipping alone—determines whether the target remains a solution.

Check Your Understanding

Answer each question, then compare with the brief solutions.

For a fixed clipping rule, write the clipped population score at $\beta_0$ in terms of the untruncated score and the weight change.
In an intercept-only weighted mean equation, what residual pattern makes upper clipping leave the target mean unchanged?
Why does multiplying every clipped weight by the same positive constant not generally restore the untruncated root?
In the local perturbation result, what two quantities determine the first-order parameter movement?

Solutions. (1) $G_C(\beta_0)=E[(W_C-W)U(O;\beta_0)]$, since the untruncated score is zero. (2) The expected product of the weight change and $Y-\beta_0$ must be zero; it can vanish, for example, if changed weights occur only where the residual is zero. (3) A common positive factor multiplies the entire clipped score without changing its zero, but the relative weights remain clipped and generally differ from the original relative weights. (4) The expected change-weighted score $E[\Delta U(O;\beta_0)]$ and the inverse score derivative $J_0^{-1}$ determine $\beta'(0)=-J_0^{-1}E[\Delta U(O;\beta_0)]$.