Raw-Scale and Log-Scale Rules for Cumulative Weights

Why the Scale of a Bound Matters

A longitudinal inverse-probability weight is commonly a product of positive contributions across treatment times. Even if each contribution is moderate, repeated multiplication can produce a very large or very small cumulative weight. A fixed upper and lower bound can limit those extremes, but the rule may be described either on the raw weight scale or on the logarithmic scale. The key distinction is how the bounds are chosen and interpreted: once correctly translated, the two scales can express exactly the same clipping operation.

This tutorial concerns fixed bounds on the cumulative weight. It does not replace cumulative-weight truncation with truncation of each time-specific contribution. The latter changes the construction of the product and is a different operation.

Definitions and Exact Correspondence

Let \(W_T=\prod_{t=1}^{T}w_t\) be a cumulative weight through \(T\) treatment times, where every contribution \(w_t\) is finite and strictly positive. Thus \(W_T>0\), and its logarithm is defined. For fixed raw-scale bounds \(0<L\leq U<\infty\), define raw-scale clipping by

C_{[L,U]}(W_T)=\min\{U,\max(L,W_T)\}.

For fixed log-scale bounds \(a\leq b\), define log-scale clipping by first clipping the log weight and then exponentiating:

C^{\log}_{[a,b]}(W_T) =\exp\!\left\{\min\bigl(b,\max(a,\log W_T)\bigr)\right\}.

The lower and upper raw-scale limits corresponding to log limits \(a,b\) are \(\exp(a),\exp(b)\). Conversely, raw-scale limits \(L,U\) correspond to log limits \(\log L,\log U\). The strict positivity of \(L\) is necessary for \(\log L\) to exist.

Theorem: For every \(W_T>0\) and \(0<L\leq U<\infty\), \(C_{[L,U]}(W_T)=C^{\log}_{[\log L,\log U]}(W_T)\). Thus clipping on the raw scale and clipping on the log scale are identical when their bounds are exact transformations of one another.

To prove this, let \(x=\log W_T\). Since the exponential function is strictly increasing, it preserves the order of the three quantities being compared. In particular, \(\exp(\max(c,d))=\max(\exp(c),\exp(d))\) and \(\exp(\min(c,d))=\min(\exp(c),\exp(d))\). Therefore,

\begin{aligned} C^{\log}_{[\log L,\log U]}(W_T) &=\exp\{\min(\log U,\max(\log L,\log W_T))\}\\ &=\min\{U,\max(L,W_T)\}\\ &=C_{[L,U]}(W_T). \end{aligned}

This establishes the correspondence, including weights below the lower bound, inside the interval, and above the upper bound. \(\square\)

The practical difference usually lies in how a rule is specified. For example, symmetric log limits \([-c,c]\) give raw limits \([\exp(-c),\exp(c)]\), a reciprocal pair. They are not generally the same as raw limits symmetric around 1 in an additive sense. A raw interval \([L,U]\) is symmetric around 1 on the log scale only when \(LU=1\).

Worked Example: Translating a Raw-Scale Cap

Suppose six treatment-time contributions are each \(1.4\), and the cumulative weight is clipped to the raw interval \([0.2,5]\). Calculate the raw and log-scale results.

The cumulative weight is

$$ W_6=1.4^6=7.529536. $$

Because \(7.529536>5\), raw-scale clipping returns \(5\). The equivalent log bounds are \(\log(0.2)\) and \(\log(5)\). The cumulative log weight is

\log W_6=6\log(1.4)=2.018833\ldots,

while \(\log(5)=1.609438\ldots\). The log weight therefore exceeds its upper limit, and exponentiating the clipped value gives \(\exp(\log 5)=5\). Both rules return the same weight. The unrounded sum is \(6\log(1.4)=2.0188334196\ldots\); rounding the single log contribution too early can change the last reported digit.

What Accumulation Changes

Products become sums after taking logarithms:

\log W_T=\sum_{t=1}^{T}\log w_t.

This identity is useful because it reveals how contributions accumulate. A sequence of factors slightly above 1 adds positive log contributions; factors below 1 add negative ones. The cumulative weight can therefore grow, shrink, or partly offset earlier changes. A raw-scale upper cap \(U\) is reached precisely when the running sum of log contributions is at least \(\log U\); a lower cap \(L\) is reached when that sum is at most \(\log L\).

Worked Example: Symmetric Log Bounds Are Multiplicative

Suppose the chosen log limits are \([-1,1]\). Their raw-scale equivalents are \([\exp(-1),\exp(1)]\), approximately \([0.3679,2.7183]\). Three consecutive contributions are each \(1.5\).

The cumulative product is \(1.5^3=3.375\), and the log cumulative weight is

3\log(1.5)=1.2163953243\ldots.

This exceeds the upper log limit \(1\), so log-scale clipping returns \(\exp(1)=2.7183\ldots\). The same operation on the raw scale clips \(3.375\) at the upper bound \(e\), also returning \(2.7183\ldots\). In contrast, a raw cap of \(3\) would return \(3\), not \(e\). The scales are equivalent only after the bounds are translated.

The log interval \([-1,1]\) permits weights from \(e^{-1}\) to \(e\), a multiplicative range with reciprocal endpoints. More generally, \([-c,c]\) permits \([e^{-c},e^c]\); the ratio of upper to lower limit is \(e^{2c}\).

Result: If every time-specific contribution equals a constant \(q>1\), then \(W_T=q^T\). An upper cap \(U>1\) is reached at the first integer \(T\) satisfying \(T\geq \log(U)/\log(q)\). The equivalent log-scale cap is \(\log U\).

For the derivation, \(q>1\) implies \(\log q>0\), and \(q^T\geq U\) is equivalent, by taking logarithms, to \(T\log q\geq\log U\). Dividing by the positive value \(\log q\) gives \(T\geq\log(U)/\log(q)\). Thus the first integer time is the smallest integer at least that ratio. This result assumes a constant contribution greater than 1; with varying contributions the relevant quantity is the partial sum \(\sum_{t=1}^T\log w_t\), not a single average factor. \(\square\)

Worked Example: How Many Times Until a Cap Is Reached?

If each contribution is \(1.1\) and the raw upper cap is \(2\), find the first treatment time at which the untruncated product reaches or exceeds the cap.

The condition is \(T\geq\log(2)/\log(1.1)\). Numerically, this ratio is approximately \(7.27\), so the first eligible integer is \(T=8\). Checking adjacent integer products confirms the result: \(1.1^7=1.9487171<2\), while \(1.1^8=2.14358881>2\). At \(T=8\), clipping returns \(2\). The corresponding log-scale cap is \(\log 2\), and \(8\log(1.1)=0.762481438\ldots>\log 2=0.693147\ldots\), giving the same conclusion.

Choosing and Checking a Rule

A fixed raw upper limit can be easy to state in units of the weight: values above \(U\) are replaced by \(U\). A fixed log limit can be easier to interpret when cumulative weights span several orders of magnitude, because additive differences in log weights represent multiplicative ratios in raw weights. Neither description is inherently a different estimator when the bounds are equivalent. If the bounds are selected by different conventions—for example, a stated raw cap versus a symmetric log interval—the resulting limits may differ, and so may the clipped data.

The horizon matters. With repeated positive log contributions, a fixed upper limit can be reached more often as the number of treatment times increases. A rule that is moderate for a short follow-up may clip a large fraction of the cumulative weights over a longer follow-up. Conversely, negative contributions can push weights toward a lower bound. The sequence of partial log sums is informative: the final product alone does not show whether weights approached a boundary early, crossed it, or were moderated later by contributions in the opposite direction.

Useful checks include reporting the raw and log bounds together, the proportion of cumulative weights changed at each bound, and summaries of the unclipped and clipped weights. Inspecting the distribution of \(\log W_T\), or its partial sums by treatment time, helps identify steadily accumulating extremes and unusually abrupt changes. A mismatch between a stated raw cap and the exponentiated log cap usually signals a scale-conversion error. These checks diagnose how a rule acts; they do not establish that a particular bound is causally or statistically appropriate.

Common Pitfalls and Limitations

Using the same numerical limits on both scales. Clipping at raw weight 5 is not the same as clipping log weight at 5. The equivalent log upper limit is \(\log 5\); a log limit of 5 corresponds to raw weight \(e^5\).
Assuming symmetric bounds mean the same thing on both scales. Raw bounds \([1-c,1+c]\) are additive around 1, whereas symmetric log bounds \([-a,a]\) give reciprocal raw limits \([e^{-a},e^a]\).
Ignoring positivity. The logarithm requires a strictly positive cumulative weight. Under the stated setup this follows from strictly positive finite factors. If a computed weight is zero or non-finite, the log calculation is not defined as specified and the data or computation must be investigated.
Treating a fixed bound as horizon-neutral. A fixed cap can affect more observations as treatment times accumulate. It does not guarantee that the same fraction of weights is clipped across follow-up lengths or populations.
Confusing a cumulative rule with component-wise clipping. Clipping \(W_T\) after multiplication is not generally equivalent to clipping each \(w_t\) before multiplication. The latter changes the product itself.

Clipping limits extreme numerical contributions, but it changes the weights and therefore can alter the weighted estimating equation and its target behavior. A raw-scale and log-scale description cannot remove that limitation: if they encode the same bounds, they produce the same clipped values and the same consequences. If they encode different bounds, the difference must be assessed as a difference in the clipping rule, not as a benefit of taking logarithms.

Check Your Understanding

Answer each question, then compare with the solutions.

What raw-scale interval corresponds to log bounds \([-0.5,0.5]\)?
A cumulative weight is \(0.1\), and raw bounds are \([0.25,4]\). What is the clipped value, and what equivalent log-bound comparison gives the same answer?
If every contribution is \(1.2\) and the upper cap is \(2\), what is the first integer treatment time at which clipping occurs?
Why can two clipping rules described with different scales produce exactly the same result?
Name one diagnostic that can reveal increasing accumulation toward a bound across treatment times.

Solutions. (1) The raw interval is \([e^{-0.5},e^{0.5}]\), approximately \([0.6065,1.6487]\). (2) The value \(0.1\) is below \(0.25\), so it is clipped to \(0.25\). The equivalent log bounds are \([\log(0.25),\log(4)]\), approximately \([-1.3863,1.3863]\); \(\log(0.1)\approx-2.3026\) is below the lower limit, and exponentiating \(\log(0.25)\) returns \(0.25\). (3) Solve \(T\geq\log(2)/\log(1.2)\approx3.80\), giving \(T=4\). Indeed, \(1.2^3=1.728<2\) and \(1.2^4=2.0736>2\). (4) The exponential is strictly increasing, so log bounds \([\log L,\log U]\) translate exactly to raw bounds \([L,U]\); the clipping operations then agree. (5) Inspect the partial sums \(\sum_{t=1}^k\log w_t\) by treatment time, together with the fraction of cumulative weights reaching each bound.