Why Boundary Effects Need Their Own Diagnosis
A delta adjustment changes imputations to represent departures from the assumptions used to predict missing values. When a shift is large relative to the outcome’s possible range, the adjusted values may cross a boundary, such as zero for a nonnegative measurement or 100 for a bounded score. Even if an imputation program prevents invalid values from entering the completed data, the method used to enforce that restriction can alter the sensitivity scenario.
The diagnostic question is not simply whether all completed values are legal. It is whether the shifted distribution still represents the intended departure from the original predictive distribution. A pileup at a minimum or maximum, a severely compressed tail, or many values requiring correction can indicate that the constraint is doing much of the work. These effects matter because conclusions under such a scenario may reflect the chosen boundary rule as much as the delta value.
The earlier material on checking implied distributions across delta values motivates comparing distributions, not just summary estimates. Here the focus is specifically on what happens near the support limits and how to identify constraint-driven sensitivity results.
What Counts as a Boundary Effect?
Let \(Y^{\mathrm{imp}}\) denote an imputed value before a shift, and let \(\delta\) be an additive shift on the outcome scale. The unrestrained adjusted value is \(Y^{\mathrm{raw}}=Y^{\mathrm{imp}}+\delta\). If the outcome must lie between known limits \(L\) and \(U\), one possible rule is clipping:
Clipping is not a neutral cleanup step. Values below \(L\) are replaced by \(L\), and values above \(U\) are replaced by \(U\). This can create repeated boundary values and make the realized change smaller than \(\delta\). A valid range must also be substantively justified: an analysis rule such as excluding extreme measurements is not automatically a true support limit.
Boundary checks are most interpretable when the imputations are compared under the same predictive draws, covariates, and delta value, changing only the boundary-handling rule. In practice, multiple imputations also vary because of stochastic draws. The comparison should therefore be repeated or summarized across imputations rather than inferred from a single completed data set.
Why Clipping Changes the Intended Shift
For a bounded outcome, define \(Z=Y^{\mathrm{imp}}+\delta\). The clipping identity \(Y^{\mathrm{clip}}=Z-(Z-U)_+ +(L-Z)_+\), where \(a_+=\max(a,0)\), gives a direct expression for the realized displacement:
If the shift is upward and no values cross the lower limit, the final term is zero. Every value that crosses the upper limit has some excess removed, so its realized displacement is less than \(\delta\). If the pre-shift distribution is continuous, clipping also creates a probability mass at \(U\) equal to \(P(Y^{\mathrm{imp}}+\delta\ge U)\); similarly, it creates mass at \(L\) equal to \(P(Y^{\mathrm{imp}}+\delta\le L)\). This explains why a boundary pileup can be evidence that the restriction, rather than the unconstrained shift, shapes the result.
Worked Example: Clipping a Bounded Score
Suppose a hypothetical test score is known to lie between 0 and 100. Six imputed scores are 72, 83, 91, 96, 99, and 100. Consider an upward additive shift of 12 points. Assume these six values illustrate the same imputation distribution before and after the shift.
Unrestrained shift. The raw adjusted values are 84, 95, 103, 108, 111, and 112. Four of the six values exceed the maximum of 100, so the unconstrained scenario produces out-of-support imputations.
Clip to the valid range. Applying the clipping rule gives 84, 95, 100, 100, 100, and 100. The mean before the shift is \((72+83+91+96+99+100)/6=541/6=90.17\) points. The clipped mean is \((84+95+100+100+100+100)/6=579/6=96.50\) points. Thus the mean increase is \(96.50-90.17=6.33\) points, not the intended 12 points.
The four scores at 100 are a boundary pileup, and the four raw violations show that clipping is consequential in this example. The clipped values are legal, but that fact alone does not make the result a faithful 12-point shift. A report should show the violation rate and the clipped distribution, and should explain that this scenario represents an attempted shift followed by a ceiling rule.
Diagnosing Discrete Outcomes and Distorted Tails
Not all boundary effects appear as impossible values. For a binary outcome, a delta adjustment is often applied on a log-odds scale. If the original imputation probability is \(p\), a shift \(\delta\) gives \(p_\delta=\operatorname{expit}(\operatorname{logit}(p)+\delta)\), where \(\operatorname{logit}(p)=\log\{p/(1-p)\}\) and \(\operatorname{expit}(a)=1/(1+\exp(-a))\). Probabilities remain between zero and one, so there are no invalid values. But a large positive shift may make nearly every imputed outcome equal to 1, while a large negative shift may make nearly every outcome equal to 0. That is a near-degenerate distribution, not an invalid one.
Worked Example: A Large Log-Odds Shift for a Binary Outcome
Consider three hypothetical missing cases with initial imputation probabilities 0.70, 0.85, and 0.92 for an event. Apply a log-odds shift \(\delta=2\). The odds are multiplied by \(\exp(2)=7.389\).
For the first case, the initial odds are \(0.70/0.30=2.333\). The shifted odds are \(2.333(7.389)=17.24\), giving \(p_\delta=17.24/(1+17.24)=0.945\). For the second case, the odds are \(0.85/0.15=5.667\); after shifting they are \(5.667(7.389)=41.87\), so \(p_\delta=41.87/(1+41.87)=0.977\). For the third, the odds are \(0.92/0.08=11.50\); after shifting they are \(11.50(7.389)=84.97\), giving \(p_\delta=84.97/(1+84.97)=0.988\).
The shifted probabilities average \((0.945+0.977+0.988)/3=0.970\), compared with the initial average \((0.70+0.85+0.92)/3=0.823\). Each probability is mathematically valid, but each is now close to one. Bernoulli draws from these probabilities will be dominated by event values. This is a diagnostic warning that the scenario concentrates imputations near the binary outcome’s upper boundary; it is not, by itself, proof that the shift is inappropriate. Its plausibility must be assessed against the intended sensitivity assumption.
For continuous variables, tail distortion can occur even without clipping. A transformation may compress the upper tail, or a constrained predictive distribution may place little probability near a limit. Useful checks include the proportion of raw shifted values outside support, the proportion exactly at each limit after constraint handling, extreme quantiles before and after the shift, and the frequency of repeated identical boundary values. Inspect these by relevant missingness pattern or subgroup when the delta was specified differently across such groups.
Worked Example: A Nonnegative Measurement and a Lower-Bound Pileup
Suppose a hypothetical nonnegative measurement, recorded in hours, has five imputed values: 0.2, 0.5, 1.0, 2.0, and 4.0. Apply an additive shift of \(-1\) hour. Assume zero is a genuine lower support limit and that the values are on the original hour scale.
The raw shifted values are \(-0.8\), \(-0.5\), 0, 1, and 3 hours. Two of five values, or \(2/5=40\%\), are invalid. If negative values are clipped to zero, the adjusted values are 0, 0, 0, 1, and 3 hours. The mean before shifting is \((0.2+0.5+1+2+4)/5=7.7/5=1.54\) hours. The clipped mean is \((0+0+0+1+3)/5=0.80\) hours, a decrease of only 0.74 hours rather than 1 hour. Three of five values now equal zero, including the value that was already zero after the raw shift.
A multiplicative shift on the log scale provides a different scenario: multiplying by \(\exp(-1)=0.3679\) yields approximately 0.074, 0.184, 0.368, 0.736, and 1.472 hours. These remain positive and preserve the ordering, but they represent proportional reduction, not a one-hour decrease. This alternative avoids negative values; it does not establish that the proportional scenario is substantively more credible.
Workflow for Boundary Diagnostics
State the justified lower and upper limits, units, and whether the delta acts additively, multiplicatively, or on a link scale. Confirm that the adjustment corresponds to the intended sensitivity assumption.
Before enforcing limits, calculate how many adjusted imputations fall below or above support. Report counts and proportions, preferably separately for the missing values to which the shift applies.
Check boundary frequencies, quantiles, range, and repeated values. Compare these with the unshifted imputations and with values in the observed data, while recognizing that observed values do not necessarily reveal the missing-value distribution.
Where appropriate, compare clipping with a support-respecting predictive model or a scientifically suitable transformation. Make clear that these choices define different assumptions, not interchangeable repairs.
If many imputations are invalid or pile up at a limit, explain how the constraint changes the realized shift. Do not describe the result as the unmodified delta scenario.
Choosing and Comparing Constraint Strategies
| Approach | What it does | Boundary diagnostic to emphasize |
|---|---|---|
| Clip or truncate after shifting | Replaces out-of-range values with a boundary or otherwise excludes them | Raw violation rate and the new mass or ties at the limit |
| Transform the outcome scale | Applies a shift on a scale that can map back into the valid range | Implied change in original units and tail compression |
| Use a support-respecting predictive model | Generates values from a model whose distribution respects the outcome’s support | Whether the shifted model still yields plausible tail probabilities |
| Use a two-part or mixture model when warranted | Separates a boundary value, such as a structural zero, from positive values | Whether the shift affects boundary occurrence, positive values, or both as intended |
These approaches are not equivalent. For example, truncation can change a continuous distribution by removing probability beyond a limit, whereas clipping converts that probability into an atom at the limit. A transformation changes the meaning of a fixed delta: an additive shift on the log scale is a multiplicative change on the original scale. A constrained predictive model can avoid impossible values, but its tail behavior depends on its model assumptions. Select the approach that matches the measurement process and the scientific interpretation of the sensitivity scenario.
A particularly important distinction is between a true support limit and a practical or administrative boundary. If a value is merely unusual, censoring it at a convenient threshold can suppress genuine tail behavior. Conversely, if a measurement is physically impossible below zero, retaining negative imputations is not a defensible representation of that variable. The reason for each constraint should be stated rather than inferred from the data range alone.
Common Mistakes and Limitations
- Calling clipped imputations “the delta-shifted values.” Clipping modifies the intended shift. Report both the raw shift and what the constraint did to the resulting distribution.
- Checking only whether values are legal. A completed data set can contain only valid values and still have a severe pileup or compressed tail. Check frequencies and distributional summaries as well as range.
- Treating all boundary values as artifacts. Some outcomes genuinely have many zeros or values at a ceiling. Compare the pattern with how the variable is generated and measured before attributing it to the shift.
- Assuming a transformation solves the sensitivity question. It may prevent invalid values while changing the scale and meaning of delta. State the implied change in original units.
- Using observed extremes as proof of the missing-value support. The observed sample may not include the full range. Support should come from the variable’s definition or measurement process where possible.
Boundary diagnostics can show that a scenario is numerically strained or that its implementation is constraint-dominated. They cannot establish which missing-data assumption is true. A low violation rate also does not guarantee that the unobserved tail is correctly represented. Interpret diagnostics alongside the substantive justification for the delta and for the chosen predictive model.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A bounded outcome has upper limit 50. Four raw shifted imputations are 42, 51, 55, and 60. What is the raw violation proportion, and what boundary pileup results from clipping?
- Why can a binary outcome show a boundary effect even when every shifted probability is between zero and one?
- A nonnegative variable is shifted downward by 2 units and many raw values become negative. Name two diagnostics and one alternative modeling strategy.
- True or false: if the completed data contain no invalid values after clipping, the intended delta shift has been preserved. Explain.
Solutions. (1) Three of four values exceed 50, so the violation proportion is \(3/4=75\%\). Clipping yields 42, 50, 50, and 50, creating a three-value pileup at the upper limit. (2) A large log-odds shift can push event probabilities close to one, so Bernoulli imputations are almost all events; the distribution becomes nearly degenerate without invalid probabilities. (3) Report the fraction below zero and the number or proportion at zero after any correction; also compare quantiles or the lower tail before and after adjustment. A support-respecting predictive model is one alternative, provided its assumptions fit the variable. (4) False. Clipping can make every value legal while attenuating the realized shift and creating a boundary mass. The raw violation rate and post-clipping distribution are needed to assess that change.