Clipping Bias in a Longitudinal Treatment Effect

Why Clipping Can Change a Longitudinal Contrast

A cumulative treatment weight corrects for the probability of receiving an entire treatment history. Histories that were unlikely under the observed treatment process receive large weights, so they can contribute substantially to an inverse-probability-weighted estimate. Clipping caps those contributions. This may improve numerical stability, but it also changes which histories dominate the weighted outcome mean. In a longitudinal marginal structural model, the change can affect the estimated mean under each treatment regime differently and therefore alter their contrast.

The key point is that clipping is not merely a computational adjustment to an otherwise unchanged target. With a fixed cutoff, it generally changes the population quantity estimated by a normalized weighted mean. The amount and direction of the change depend on how treatment-history probabilities relate to the potential outcomes.

Regime Means and Cumulative Weights

Consider treatment at two times, $A_1$ and $A_2$, baseline covariates $L_0$, an intermediate covariate $L_1$ measured after $A_1$ and before $A_2$, and outcome $Y$. A deterministic regime is $g=(g_1,g_2)$. Write $Y^g$ for the outcome that would be observed under that regime, and $\psi_g=E[Y^g]$ for its marginal mean.

Definition: For a regime $g=(g_1,g_2)$, define its observed treatment-history probability by $$ q_g(L_0,L_1)=P(A_1=g_1\mid L_0)\,P(A_2=g_2\mid L_0,A_1=g_1,L_1). $$ The unstabilized cumulative inverse-probability weight is $$ W_g=\frac{I(A_1=g_1,A_2=g_2)}{q_g(L_0,L_1)}. $$ For a fixed upper cutoff $c>0$, the clipped weight is $W_g^{(c)}=\min(W_g,c)$.

The indicator makes the weight zero for individuals whose observed treatment history does not match $g$. For those whose history does match, the denominator is the product of the two conditional treatment probabilities, each evaluated using the information available just before that treatment. A stabilized numerator would modify the weight; the examples below use unstabilized weights to isolate the effect of clipping the cumulative denominator.

Identification of $\psi_g$ by the untruncated weight requires consistency, sequential exchangeability, and positivity. Consistency means that an individual following regime $g$ has observed outcome $Y=Y^g$. Sequential exchangeability requires treatment assignment at each time to be independent of the relevant future potential outcomes, conditional on the measured history available before that assignment: at time 1, conditional on $L_0$; at time 2, conditional on $L_0,A_1,L_1$. Positivity requires each treatment value specified by $g$ to have positive conditional probability at histories occurring under the regime. Assume also that the outcomes have finite expectations and that the relevant conditional distributions are well-defined.

What the Untruncated Weight Identifies

Result: Under consistency, sequential exchangeability, and positivity, $E[W_g]=1$ and $E[W_gY]=E[Y^g]$. Thus the population normalized weighted mean $E[W_gY]/E[W_g]$ equals the regime mean $\psi_g$.

To justify the result, factor the observed-data law in temporal order. For clarity, write its conditional densities or mass functions as $f(l_0)$, $f(l_1\mid l_0,a_1)$, and $f(y\mid l_0,a_1,l_1,a_2)$, with treatment probabilities $p_1(a_1\mid l_0)$ and $p_2(a_2\mid l_0,a_1,l_1)$. On the part of the joint law where $a_1=g_1$ and $a_2=g_2$, multiplying by $W_g$ cancels the two treatment probabilities:

\begin{aligned} &f(l_0)\,p_1(g_1\mid l_0)\,f(l_1\mid l_0,g_1)\, p_2(g_2\mid l_0,g_1,l_1)\,f(y\mid l_0,g_1,l_1,g_2)\\ &\qquad{}\times \frac{1}{p_1(g_1\mid l_0)p_2(g_2\mid l_0,g_1,l_1)}\\ &=f(l_0)\,f(l_1\mid l_0,g_1)\,f(y\mid l_0,g_1,l_1,g_2). \end{aligned}

Integrating the right-hand side over $y,l_1,l_0$ gives 1, proving $E[W_g]=1$. Integrating it after multiplying by $y$ gives the mean outcome in the distribution generated by setting treatments to $g_1$ and $g_2$. Sequential exchangeability identifies the conditional outcome distributions under that intervention with the displayed observed-data distributions; consistency identifies the observed outcome on the matching history with $Y^g$. The integral is therefore $E[Y^g]$. The cancellation is a joint-law factorization followed by integration in temporal order; it does not condition on variables measured after a treatment while treating that treatment’s assignment probability as unchanged. $\square$

The Population Target After Clipping

Clipping changes the identity above because the multiplier on a matching history is no longer exactly the reciprocal of its probability. Let $E_g$ denote expectation under the distribution of covariates and outcome generated by regime $g$, and define $r_c=\min(1,cq_g(L_0,L_1))$ along that regime. On the matching history, multiplying its observed probability by its clipped weight leaves the factor $q_g\min(1/q_g,c)=\min(1,cq_g)$. The same temporal factorization therefore gives:

Result: If $E_g[r_c]>0$, the normalized population mean using clipped weights is $$ \frac{E[W_g^{(c)}Y]}{E[W_g^{(c)}]} = \frac{E_g[r_cY^g]}{E_g[r_c]}. $$

This formula shows exactly what changes: under the intervention distribution, histories receive an additional relative weight $r_c$. If $cq_g\geq 1$, then $r_c=1$ and clipping has no effect on that history. If $cq_g<1$, then $r_c=cq_g<1$, so that history is downweighted relative to histories whose inverse probability weight is below the cutoff. The clipped mean is a weighted mean of potential outcomes, not generally the unweighted regime mean $E_g[Y^g]$. Equality can still occur in special cases—for example, if $r_c$ is constant or unassociated with $Y^g$—but it is not guaranteed by clipping itself.

Worked Two-Regime Example

Worked Example: A Clipped Two-Time-Point Contrast

Let $X$ be a binary baseline covariate with $P(X=0)=P(X=1)=0.5$. Suppose $P(A_1=1\mid X=0)=0.5$ and $P(A_1=1\mid X=1)=0.8$. For the second treatment, specify the probabilities needed for both regimes being compared: $P(A_2=1\mid A_1=1,X=0)=0.5$, $P(A_2=1\mid A_1=1,X=1)=0.2$, $P(A_2=0\mid A_1=0,X=0)=0.5$, and $P(A_2=0\mid A_1=0,X=1)=0.5$. Assume the potential outcomes are deterministic given $X$: $Y^{11}=0$ for $X=0$ and $10$ for $X=1$; $Y^{00}=0$ for $X=0$ and $2$ for $X=1$. These probabilities and outcomes fully specify the calculation for regimes $11$ and $00$.

Untruncated target means. Since the two values of $X$ are equally likely, $\psi_{11}=0.5(0)+0.5(10)=5$, and $\psi_{00}=0.5(0)+0.5(2)=1$. The regime contrast $\psi_{11}-\psi_{00}$ is $5-1=4$.

Regime $11$. Its history probabilities are $q_{11}(X=0)=0.5(0.5)=0.25$ and $q_{11}(X=1)=0.8(0.2)=0.16$. The matching-history weights are therefore $4$ and $1/0.16=6.25$. With cutoff $c=5$, the clipped weights are $4$ and $5$. The population weighted masses, including the probability of each baseline stratum and its matching treatment history, are $0.5(0.25)(4)=0.5$ and $0.5(0.16)(5)=0.4$. Thus the clipped mean is $(0.5(0)+0.4(10))/(0.5+0.4)=4/0.9=4.444\ldots$.

Regime $00$. Its history probabilities are $q_{00}(X=0)=0.5(0.5)=0.25$ and $q_{00}(X=1)=0.2(0.5)=0.10$. The corresponding weights are $4$ and $10$, so the clipped weights are $4$ and $5$. The weighted masses are $0.5(0.25)(4)=0.5$ and $0.5(0.10)(5)=0.25$. The clipped mean is $(0.5(0)+0.25(2))/(0.5+0.25)=0.5/0.75=0.666\ldots$.

Clipped contrast. The result is $4.444\ldots-0.666\ldots=3.777\ldots$, rather than the target contrast $4$. The bias is approximately $-0.222$. Both regime means changed, and they changed by different amounts. This is why checking only the clipped mean for one regime does not establish that a causal contrast is unaffected.

The Direction of Bias Is Not Fixed

Worked Example: Reversing the Outcome Pattern

Keep the treatment probabilities and the $11$ regime history probabilities from the previous example, but reverse its potential outcomes: let $Y^{11}=10$ when $X=0$ and $Y^{11}=0$ when $X=1$. The untruncated mean remains $\psi_{11}=0.5(10)+0.5(0)=5$. At cutoff $c=5$, the weighted masses remain $0.5$ for $X=0$ and $0.4$ for $X=1$. The clipped mean is now $(0.5(10)+0.4(0))/(0.5+0.4)=5/0.9=5.555\ldots$, which is above the target by about $0.556$. In the first example, the same cutoff pulled the $11$ mean below its target; here it pulls the mean above. Clipping does not, by itself, imply bias toward zero or toward any particular treatment effect.

Worked Example: Changing the Cutoff Changes the Target

Return to the original $11$ outcomes, $0$ for $X=0$ and $10$ for $X=1$. The untruncated weights are $4$ and $6.25$, and the matching-history probabilities are $0.25$ and $0.16$. At cutoff $c=4$, both weights become $4$, so the weighted masses are $0.5(0.25)(4)=0.5$ and $0.5(0.16)(4)=0.32$. The clipped mean is $(0.5(0)+0.32(10))/0.82=3.2/0.82=3.902\ldots$.

At $c=5$, the masses are $0.5$ and $0.4$, giving $4.444\ldots$, as calculated above. At $c=6.25$, neither weight is clipped: the masses are $0.5(0.25)(4)=0.5$ and $0.5(0.16)(6.25)=0.5$, giving $(0.5(0)+0.5(10))/1=5$. Thus the estimated population target changes with the cutoff. This calculation does not imply that every contrast changes monotonically with the cutoff; it shows why the cutoff is substantively relevant.

Diagnostics, Alternatives, and Limitations

A useful diagnostic is to examine which matching histories lose relative weight after clipping. Report the cutoff and, where practical, compare untruncated and clipped estimates for each regime as well as for the contrast. Check contributions by important baseline and time-varying histories: a large change concentrated in particular strata can reveal which part of the covariate distribution drives the shift. Also assess whether the weighted covariate distribution under each regime still represents the intended target population. A small change in the overall contrast can conceal larger, opposing changes in its two regime means.

Clipping is not the same as deleting observations with large weights. Clipping retains their outcomes but caps their contributions; deletion removes them from the weighted calculation. Neither operation automatically preserves the original regime mean. Leaving weights untruncated retains the inverse-probability identity under the assumptions above, but may leave highly variable estimates. Clipping deliberately trades some fidelity to that target for bounded contributions. The appropriate choice depends on the analysis goal and should be accompanied by sensitivity analyses, not presented as a consequence-free correction.

The derivations also rely on the treatment probabilities being correctly specified or otherwise consistently estimated, in addition to the causal assumptions. Clipping cannot repair unmeasured confounding, a positivity violation, or an incorrectly defined intervention. If histories have zero probability under the observed treatment process, their inverse-probability weights are not defined and no finite clipping cutoff restores identification from those data.

Common Mistake

Pitfall: “Clipping only limits extreme values, so the weighted estimate still targets the same regime mean.” The correction is that the cancellation proving the untruncated identity depends on multiplying each matching history by exactly $1/q_g$. With clipping, the remaining factor is $\min(1,cq_g)$, which generally varies across histories. The resulting mean is $E_g[r_cY^g]/E_g[r_c]$, and it equals the original regime mean only under additional conditions.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

For the $11$ regime in the first example, why is the history weight for $X=1$ equal to $6.25$, rather than $1/0.8$?
Using the first example, calculate the clipped $00$ mean when $c=10$.
In the clipped-target formula, what is $r_c$ for a regime history with $q_g=0.1$ and cutoff $c=4$?
Does clipping necessarily move a regime mean toward zero? Give the relevant lesson from the examples.

Solutions. (1) The matching $11$ history requires both treatments: $q_{11}=0.8(0.2)=0.16$, so its weight is $1/0.16=6.25$. (2) The $00$ weights are $4$ and $10$, so at $c=10$ neither is reduced. The masses are $0.5$ and $0.5$, giving mean $(0.5(0)+0.5(2))/1=1$. (3) $r_c=\min(1,4(0.1))=0.4$. (4) No. Reversing the outcome pattern while keeping treatment probabilities fixed changed the clipped $11$ mean from below its target to above it.