Why One Tipping Threshold May Not Be Enough
A one-parameter sensitivity analysis varies an assumption, such as a delta adjustment for missing outcomes, while holding the analysis and decision rule fixed. It is tempting to expect the decision to move in one direction as the parameter changes. That expectation is not guaranteed. The decision may change from favorable to unfavorable and later return to favorable, producing more than one tipping point over the range considered.
The issue is not merely how to locate a zero. It is how to determine whether each candidate value actually changes the decision, whether additional changes have been missed, and how to report the resulting pattern without implying that there is a unique threshold.
The margin \(g\) must be defined in the language of the decision. For example, if a treatment is considered worthwhile when its estimated improvement exceeds a clinical threshold, \(g\) can be the estimated improvement minus that threshold. Positive values then meet the rule. A different rule, such as a confidence-interval criterion, would require a margin or decision indicator that represents that rule. The decision rule should not be changed while searching for thresholds.
For a continuous margin, a crossing of zero is a useful candidate for a decision change, but not every zero is a change. Moreover, if the decision procedure is discontinuous, the decision can change at a jump without \(g(\delta)=0\). In that case, inspect the decision states on either side rather than relying on root-finding alone.
Assumptions and a Reliable Search
The search is meaningful only after specifying the parameter range, the delta convention, the decision rule, and the analysis settings. State what increasing \(\delta\) means for the missing outcomes in the chosen parameterization. Keep the outcome scale, imputation procedure, pooling procedure, and decision criterion fixed at every evaluation. This makes changes in the decision attributable to the sensitivity parameter rather than to a changed analysis.
If \(g\) is continuous on the range, a sign change between two parameter values guarantees at least one zero between them. It does not guarantee exactly one zero. A sign change can contain three or another odd number of crossings. Conversely, equal signs at two grid values do not rule out two crossings between them. A zero with no sign change may be a contact, which ordinary sign-change bracketing will not find.
This implication uses continuity. It says nothing about uniqueness unless additional information, such as monotonicity on that interval, is available. The practical approach is to evaluate the margin across the full range, identify every interval with a sign change, and refine each such interval separately. Also inspect the shape of the margin for near-zero contacts and changes that could fall between grid points. A grid is a diagnostic, not a proof that every transition has been found.
Write the favorable and unfavorable conditions explicitly, and define the margin or decision indicator that implements them.
Use a prespecified grid and record the margin and decision state at every value, not just at the two endpoints.
Bracket each adjacent pair with different decision states and refine each bracket. Investigate same-sign intervals that have near-zero values or suspicious curvature.
Check states on both sides of every candidate, distinguish crossings from contacts, and report all relevant thresholds in the stated range.
A short justification explains why the side check matters. If \(g\) is continuous and has opposite signs at \(\delta_L\) and \(\delta_U\), the intermediate value theorem gives at least one \(\delta^*\) between them with \(g(\delta^*)=0\). If \(g\) is strictly monotone on that interval, it can have at most one such zero, so the crossing is unique there. Without monotonicity, the endpoint signs alone do not establish how many zeros occur or whether a particular zero changes the decision. Thus every candidate must be classified locally, and every bracket must be interpreted in light of possible nonmonotonicity.
Worked Examples
Worked Example: Three Crossings in a Prespecified Range
Suppose a hypothetical analysis reports a decision margin, in mmHg, as a function of a dimensionless sensitivity parameter \(\delta\). Increasing \(\delta\) represents a progressively less favorable assumed shift for missing outcomes. The fixed rule is favorable when \(g(\delta)>0\), and the prespecified range is \(0\leq\delta\leq4\). For illustration, let the evaluated margin be
The zeros in the range are \(\delta=1\) and \(\delta=3\); the third zero, \(\delta=5\), is outside the range. Check the signs in each interval. At \(\delta=0\),
so the decision is unfavorable on the first interval. At \(\delta=2\),
so it is favorable between 1 and 3. At \(\delta=4\),
so it is unfavorable between 3 and 4. The decision therefore changes at both \(\delta=1\) and \(\delta=3\): it goes from unfavorable to favorable and then back to unfavorable. The endpoint values alone would show the same unfavorable state at 0 and 4 and would miss both changes. The reported thresholds are dimensionless values of \(\delta\); the margin itself is in mmHg.
Worked Example: Finding Several Candidate Intervals on a Grid
Consider another hypothetical margin, in mmHg, on \(0\leq\delta\leq5\):
Evaluate it at integer grid points. The resulting values are \(g(0)=-5.625\), \(g(1)=2.625\), \(g(2)=1.875\), \(g(3)=-1.875\), \(g(4)=-2.625\), and \(g(5)=5.625\) mmHg. The sign changes occur in three separate intervals: \((0,1)\), \((2,3)\), and \((4,5)\). Each interval contains at least one crossing because the margin is continuous. The middle crossing is exactly at \(2.5\), as \(g(2.5)=0\); the outer crossings are at \(0.5\) and \(4.5\).
A linear interpolation within a bracket gives an approximation, not a general exact answer. In the first bracket it gives
In the last bracket it gives
Both estimates differ from the exact roots because the margin is curved over those intervals. Refining each bracket, for example by repeated evaluation and bisection under continuity, narrows the location of each root. The important first step is retaining all three brackets. Searching only near one previously suspected threshold would fail to identify the other two.
Worked Example: Equal Endpoint Signs Do Not Classify a Candidate
Suppose an investigator reports \(g(1)=-0.5\), \(g(2)=0\), and \(g(3)=-1.5\) mmHg, and asks whether \(\delta=2\) must be only a contact. It does not. The values at 1 and 3 are not necessarily the signs immediately to the left and right of 2.
For a crossing consistent with all three reported values, take the hypothetical margin
At the reported points, \(g_{\text{cross}}(1)=-0.5\), \(g_{\text{cross}}(2)=0\), and \(g_{\text{cross}}(3)=-1.5\) mmHg. But just to the left of 2, at 1.9, the value is
and just to the right, at 2.1, it is
Thus the decision changes at 2. A contact with exactly the same three reported values is also possible:
This margin equals \(-0.5\), \(0\), and \(-1.5\) mmHg at 1, 2, and 3, respectively, but it is negative on both sides sufficiently close to 2. Therefore the three reported values do not determine whether 2 is a crossing or a contact. Evaluate sufficiently close on both sides, or otherwise establish the local behavior, before classifying it.
Diagnostics, Alternatives, and Limitations
A useful report shows the decision state over the full investigated range, not only the first crossing. A table or plot should include enough evaluations to reveal reversals, and a table should list each refined threshold with the states on either side. Check that the direction of increasing \(\delta\) matches the stated delta convention. A reported return to a favorable decision should not be described as a single monotone loss of robustness.
Several warning signs call for more investigation:
- Same endpoint states. Equal decisions at the ends of the range do not imply no changes between them.
- Large grid gaps. Two changes can occur between adjacent grid values and leave the sampled endpoint signs unchanged.
- Near-zero margins without a sign change. These may reflect a contact, a very narrow favorable or unfavorable interval, or numerical error.
- Unstable evaluations. If multiple imputation or another stochastic computation makes \(g(\delta)\) noisy, small apparent reversals may be simulation variation rather than a stable decision pattern.
When computation is stochastic, use sufficiently precise and reproducible evaluations at candidate values; for example, increase the number of imputations or use a consistent random-number strategy when appropriate. Then check whether the apparent transitions persist. A threshold estimated from a noisy margin should not be reported with more precision than the calculation supports.
The main alternative is to assume a single monotone crossing and apply a one-bracket method. That is efficient when monotonicity over the relevant interval is scientifically and computationally justified. It is not a safe default for a nonmonotone margin. A full-range grid followed by separate refinement of every candidate bracket is more transparent, though a grid alone cannot prove that no narrow decision interval was missed. If the margin is available in a form that permits analytic root finding or reliable root isolation, that can strengthen the search. In all cases, the final classification still depends on the decision states on each side.
A one-parameter search has limits. Its thresholds describe the chosen decision rule, analysis, and parameter range; they do not establish that values outside the range are plausible or that one missing-data assumption is more credible than another. A root estimate is also not automatically a confidence interval for a tipping point. The purpose here is to describe how the specified decision changes as the sensitivity parameter varies, not to assign probability to the assumptions.
Common Mistake
Wrong reasoning: “The margin is negative at \(\delta=1\) and \(\delta=3\), and zero at \(\delta=2\), so \(\delta=2\) cannot be a decision-changing tipping point.” This compares values at a distance from 2 and assumes they describe the local signs.
Correction: Those three values are consistent with both a crossing and a contact, as the worked example demonstrates. Check the decision states immediately on either side of 2 or establish the local behavior mathematically. More generally, do not label every zero a tipping point, and do not assume one crossing merely because a method or plot shows one.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A continuous decision margin is positive at \(\delta=0\), negative at 1, positive at 2, and negative at 3. What can be concluded about the number and location of decision changes?
- At a zero of \(g\), the margin is negative on both sides sufficiently close to the zero. Is that zero a decision-changing tipping point under the rule favorable if \(g>0\)?
- A margin is negative at both ends of a parameter range. Does this establish that the decision is unfavorable throughout the range?
- Why does a sign change in a continuous margin bracket establish at least one zero but not necessarily exactly one?
- A stochastic imputation calculation shows a small sign reversal on a fine grid, but repeated runs move or remove it. What should be checked before reporting a threshold?
Solutions. (1) Each adjacent pair has opposite signs, so continuity guarantees at least one zero in each of \((0,1)\), \((1,2)\), and \((2,3)\). There are at least three decision changes; the signs alone do not rule out additional crossings within those intervals. (2) No. The decision is unfavorable on both sides, so this is a contact rather than a decision-changing tipping point. (3) No. The decision could change twice, or another even number of times, between the endpoints. Evaluate the interior of the range. (4) Continuity and opposite endpoint signs guarantee a zero by the intermediate value theorem, but without monotonicity there may be multiple zeros in the bracket. (5) Assess simulation variability by making evaluations more precise and reproducible, then determine whether the reversal persists and whether local side states are stable. Do not report a numerical threshold based on an unstable sign change.