Time-Specific Truncation in Longitudinal Products

Why the Time of Clipping Matters

A longitudinal stabilized weight is built by multiplying treatment-specific factors across time. A large factor at an early treatment time appears in every later cumulative product; a large factor at a late time appears only in products that reach that time. Therefore, a rule that uses different clipping thresholds at different times does more than put an overall ceiling on a final weight: it determines which parts of each treatment history are altered and how those alterations carry forward.

This tutorial focuses on that time-specific structure. It assumes the treatment-specific stabilized weight factors have already been defined. The introductory material on weight truncation distinguishes clipping from deletion; here, the central task is to apply a specified clipping rule at each time and track the resulting products.

Definition and Product Structure

Definition: For individual \(i\) and treatment time \(t=1,\ldots,T\), let \(R_{it}>0\) be the stabilized treatment-weight factor at time \(t\). The cumulative weight through time \(t\) is \(W_{it}=\prod_{s=1}^{t}R_{is}\), with \(W_{i0}=1\). Given time-specific bounds \(0<l_t\leq u_t<\infty\), define the clipped factor \(R_{it}^{c}=\min\{u_t,\max(l_t,R_{it})\}\). The cumulative weight from component clipping is \(W_{it}^{c}=\prod_{s=1}^{t}R_{is}^{c}\). For an upper-only cap, use \(R_{it}^{c}=\min(R_{it},u_t)\).

The limits \(l_t\) and \(u_t\) may differ across treatment times. A small upper bound at one time restricts that component even if the bounds at other times are large. Likewise, a lower bound can raise a small component. Unless stated otherwise, the examples use fixed, positive thresholds and apply the same time-specific rule to every individual.

A separate operation is to clip the cumulative prefix itself: \(W_{it}^{p}=\min\{b_t,\max(a_t,W_{it})\}\), for time-specific cumulative bounds \(a_t,b_t\). This acts on the already-multiplied prefix, not on each factor. The two operations generally produce different answers. In particular, component-clipped prefixes obey \(W_{it}^{c}=W_{i,t-1}^{c}R_{it}^{c}\); directly clipped raw prefixes need not obey that recursion.

Worked Example: Three Different Upper Caps

Suppose one individual has factors \(R_{i1}=0.5\), \(R_{i2}=4\), and \(R_{i3}=3\). Apply upper caps \(u_1=1\), \(u_2=3\), and \(u_3=2\), with no lower clipping.

First calculate the raw cumulative weights: \(W_{i1}=0.5\), \(W_{i2}=0.5(4)=2\), and \(W_{i3}=0.5(4)(3)=6\).

The time-specific clipped factors are \(R_{i1}^{c}=0.5\), \(R_{i2}^{c}=3\), and \(R_{i3}^{c}=2\). Multiplying successively gives \(W_{i1}^{c}=0.5\), \(W_{i2}^{c}=0.5(3)=1.5\), and \(W_{i3}^{c}=0.5(3)(2)=3\).

Thus, the last prefix is reduced from 6 to 3. The first prefix is unchanged, while the second and third reflect the caps encountered up to those times. For comparison, a uniform upper cap of 2 would give clipped factors \(0.5,2,2\) and a final product \(0.5(2)(2)=2\), not 3.

How a Component Clip Propagates

The product definition gives a precise account of the effect at each time. Define the component multiplier \(q_{it}=R_{it}^{c}/R_{it}\). It records how much of the original factor remains after clipping: \(q_{it}=1\) when the factor is inside its bounds, less than 1 when an upper cap reduces it, and greater than 1 when a lower bound raises it.

Result: For every prefix \(t\), \(W_{it}^{c}=W_{it}\prod_{s=1}^{t}q_{is}\). Consequently, clipping a component at time \(s\) changes every later clipped prefix \(t\geq s\) by including the same multiplier \(q_{is}\); it does not change prefixes \(t<s\).

Proof. By definition, \(R_{is}^{c}=R_{is}q_{is}\). Substitute this identity into the clipped product:

W_{it}^{c} =\prod_{s=1}^{t}R_{is}^{c} =\prod_{s=1}^{t}(R_{is}q_{is}) =\left(\prod_{s=1}^{t}R_{is}\right) \left(\prod_{s=1}^{t}q_{is}\right) =W_{it}\prod_{s=1}^{t}q_{is}.

A factor at time \(s\) occurs in precisely those prefixes whose endpoint is \(s\) or later. This proves the propagation statement. \(\square\)

For upper-only caps, \(0<q_{it}\leq1\), so \(W_{it}^{c}\leq W_{it}\) for every prefix. The ratio is especially transparent on the log scale:

\log W_{it}^{c}-\log W_{it} =\sum_{s=1}^{t}\log q_{is}.

Each time-specific cap contributes an additive log change, and the changes accumulate across the product. This identity is useful for locating which time points account for most of the reduction. With two-sided clipping, some \(q_{it}\) can exceed 1, so the cumulative clipped product need not be smaller than the raw product.

Worked Example: A Lower Bound Can Raise a Product

Suppose there are two factors, \(R_{i1}=0.4\) and \(R_{i2}=2\), and the time-specific clipping intervals are \([0.5,1]\) and \([1,3]\), respectively. The raw final product is \(0.4(2)=0.8\).

At time 1, clipping raises \(0.4\) to \(0.5\). At time 2, the factor \(2\) is already in its interval. The clipped final product is therefore \(0.5(2)=1\), which is greater than the raw product \(0.8\). The multipliers are \(q_{i1}=0.5/0.4=1.25\) and \(q_{i2}=2/2=1\), so the product ratio is \(1.25(1)=1.25\). Indeed, \(0.8(1.25)=1\).

This is why the statement “truncation can only reduce weights” requires an upper-only rule. It is true for upper caps, but not for two-sided clipping that includes a lower bound.

Time-Specific Rules and History Contributions

Each observed treatment history receives a weight determined by its own sequence of factors. Applying the same time-specific rule across individuals does not mean every history is affected equally: factors that exceed the cap at different times produce different products of multipliers. The following calculation describes changes in numerical weight contributions, not a treatment-effect estimate.

Worked Example: Comparing Two Treatment Histories

Consider two individuals with three factors each. Individual A has factors \((2,0.5,4)\), and individual B has factors \((1,2,1)\). Apply upper caps \((1.5,1,3)\) at times 1, 2, and 3.

For A, the raw prefixes are \(2\), \(2(0.5)=1\), and \(1(4)=4\). The clipped factors are \((1.5,0.5,3)\), giving clipped prefixes \(1.5\), \(1.5(0.5)=0.75\), and \(0.75(3)=2.25\).

For B, the raw prefixes are \(1\), \(1(2)=2\), and \(2(1)=2\). The clipped factors are \((1,1,1)\), giving clipped prefixes \(1\), \(1(1)=1\), and \(1(1)=1\).

At the final time, the raw weights sum to \(4+2=6\), of which A contributes \(4/6=2/3\). The clipped weights sum to \(2.25+1=3.25=13/4\). A's share of that clipped total is \(2.25/3.25=9/13\), approximately \(0.692\). Although both weights fell under upper caps, A's relative share increased because the two histories were reduced by different amounts. The result depends on the factor sequences and the time-specific thresholds, not just on the final raw weights.

Component Clipping and Cumulative-Prefix Clipping

Different time-specific limits can be applied to components or to cumulative prefixes. These rules answer different operational questions: component clipping limits the size of each factor before multiplication; cumulative clipping limits the size of each raw prefix after multiplication. They should not be treated as interchangeable.

Worked Example: Two Rules Give Different Prefixes

Let the factors be \((2,0.4,5)\). For component clipping, use upper caps \((1.5,1,3)\). The clipped factors are \((1.5,0.4,3)\), so the clipped prefixes are \(1.5\), \(1.5(0.4)=0.6\), and \(0.6(3)=1.8\).

The raw prefixes, by contrast, are \(2\), \(2(0.4)=0.8\), and \(0.8(5)=4\). Now directly clip those raw prefixes using time-specific upper bounds \((1.5,1,2)\). The resulting prefixes are \(1.5\), \(0.8\), and \(2\).

The second prefix is \(0.6\) under component clipping but \(0.8\) under cumulative clipping; the final prefix is \(1.8\) versus \(2\). Directly clipped raw prefixes are computed separately at each endpoint. They are not obtained by multiplying together the earlier directly clipped prefix and a newly clipped component.

A practical diagnostic is to retain the raw factor, the threshold at that time, the clipped factor, and the raw and clipped prefixes for each observation. This makes it possible to identify which time points trigger clipping and to verify that each prefix has been formed by the intended rule. If an upper-only component rule produces a clipped prefix larger than its raw counterpart, check for an arithmetic error, a mismatched threshold, or an unintended lower bound. For two-sided rules, a larger prefix can be legitimate.

Common Mistakes and Limitations

Using a single cap by habit. A time-specific rule requires applying \(u_t\) at time \(t\), not reusing one threshold for every factor. Check the threshold sequence alongside the factor sequence.
Clipping only the final product when the rule is component-based. That does not reproduce the component-clipped product, because each factor has its own time-specific operation.
Assuming the first cap affects only the first prefix. A clipped factor at time 1 appears in every later cumulative product.
Assuming all clipping reduces weights. This conclusion follows for upper-only caps, but a lower bound can increase a factor and the product.
Interpreting weight changes as a treatment-effect conclusion. These calculations describe reweighting. They do not, by themselves, establish the consequences for an estimator or its covariate balance.

Time-specific clipping is a deliberate modification of the observed weight products. Its consequences depend on the chosen thresholds and on where unusually large or small factors occur in a treatment history. The arithmetic identifies those changes, but it does not establish that a particular set of thresholds is appropriate for a given analysis.

Check Your Understanding

For each question, calculate the requested quantities before reading the solution.

Factors are \((1,3,2)\), with upper caps \((2,2,1.5)\). Find the raw and component-clipped cumulative prefixes.
For upper-only component caps, explain why clipping a factor at time 2 cannot alter the prefix at time 1, but can alter every later prefix.
Factors are \((0.5,4)\), with component intervals \([0.8,2]\) and \([1,3]\). Find the raw and clipped final products, and say whether clipping reduced the product.
Raw prefixes are \((2,0.8,4)\). Directly clip them using upper bounds \((1.5,1,2)\). Explain why the resulting sequence need not equal prefixes formed by component clipping.

Solutions. (1) The raw prefixes are \(1\), \(1(3)=3\), and \(3(2)=6\). The clipped factors are \((1,2,1.5)\), so the component-clipped prefixes are \(1\), \(1(2)=2\), and \(2(1.5)=3\). (2) A prefix through time 1 is the product of factors with indices \(s\leq1\), so it contains no time-2 factor. Every prefix through \(t\geq2\) includes the time-2 factor, so its clipped product includes the time-2 multiplier. (3) The raw product is \(0.5(4)=2\). The clipped factors are \(0.8\) and \(3\), giving \(0.8(3)=2.4\). Clipping increased the product because the first factor was raised from \(0.5\) to \(0.8\), while the second was reduced from 4 to 3. (4) The directly clipped prefixes are \(1.5\), \(0.8\), and \(2\). This rule clips each raw cumulative value at its own time-specific bound; component clipping instead clips the factors before multiplying, so the two operations generally yield different prefixes.