From a Grid to a Root
A grid search identifies a tipping point by evaluating a sensitivity curve at selected delta values and locating where the conclusion changes. The crossing may lie between grid points, however, and a finer grid can require many repeated evaluations. Root finding refines a candidate crossing by solving an equation: the pooled estimate or confidence limit equals a specified threshold.
At each value of \(\delta\), the sensitivity quantity is obtained under the chosen delta adjustment and analysis. The root-finding method acts on that quantity as a function of \(\delta\); it does not replace the imputation and analysis steps. For example, the quantity might be a pooled estimate \(\bar{Q}(\delta)\), or a lower confidence limit \(L(\delta)\), computed using the established analysis at that delta value.
The threshold depends on the question. For a clinical or policy threshold, \(T\) could be a pooled estimate and \(c\) the relevant threshold. For a confidence-limit tipping point, \(T\) could be a confidence limit and \(c\) the boundary at which the conclusion changes. The root identifies equality with that threshold; interpretation of the conclusion on either side still depends on the analysis question.
Conditions for a Reliable Search
A root-finding calculation is meaningful only if the function and search range represent the sensitivity question correctly. Before iterating, specify the delta scale and plausible range, the sensitivity quantity, its threshold, and the direction of the conclusion change being investigated. Check that the evaluations of \(T(\delta)\) use the same estimand, model, and analysis procedure throughout the range.
Bisection requires a continuous function on an interval \([a,b]\), with opposite signs at its endpoints. Continuity ensures that a sign change reflects at least one crossing, rather than a jump over the threshold. Under these conditions, the Intermediate Value Theorem guarantees at least one root in the interval. A sign change does not guarantee that the root is unique: a function can cross the threshold more than once.
In practice, sensitivity quantities are evaluated numerically, and repeated imputation can make evaluations vary slightly. A function that appears to change sign only by an amount comparable to its numerical variability does not provide a dependable bracket. Recheck such evaluations, use a consistent analysis, and avoid reporting more precision than the calculations support. A root finder also cannot correct an implausible delta range, an inappropriate threshold, or a sensitivity curve that does not represent the intended estimand.
Bisection: Refine a Verified Bracket
Bisection repeatedly halves an interval known to contain a root. At each iteration, evaluate \(g\) at the midpoint \(m=(a+b)/2\). If \(g(m)=0\), the root has been found. Otherwise, retain the half-interval whose endpoint values have opposite signs. This choice preserves a bracket under the continuity and initial sign-change conditions.
If the midpoint of the final bracket is used as the root estimate, its distance from any root in that bracket is at most half the bracket width. This is a deterministic bound from the interval length, not a probability statement.
Worked Example: Refining a Confidence-Limit Threshold
Suppose a hypothetical analysis evaluates a lower confidence limit \(L(\delta)=1.20-0.15\delta\), in outcome units, over delta values from 0 to 8 in the same units. The specified threshold is \(c=0.50\) outcome units. The tipping-point function is \(g(\delta)=L(\delta)-0.50=0.70-0.15\delta\). This illustrative curve is linear, hence continuous, on the range.
At the endpoints, \(g(0)=0.70\) and \(g(8)=0.70-1.20=-0.50\) outcome units. Their product is negative, so \([0,8]\) is a valid sign-changing bracket. The midpoint steps are:
| Iteration | Midpoint \(\delta\) | \(g(\delta)\), outcome units | Retained bracket |
|---|---|---|---|
| 1 | 4.00 | 0.100 | [4, 8] |
| 2 | 6.00 | -0.200 | [4, 6] |
| 3 | 5.00 | -0.0500 | [4, 5] |
| 4 | 4.50 | 0.0250 | [4.5, 5] |
| 5 | 4.75 | -0.0125 | [4.5, 4.75] |
After five halvings, the bracket width is \(8/2^5=0.25\) delta units. Its midpoint, \(4.625\), is within \(0.125\) delta units of a root in the bracket. Directly, \(g(4.625)=0.70-0.15(4.625)=0.00625\) outcome units. Solving the linear equation gives \(\delta^\ast=0.70/0.15=4.6667\) delta units, which lies in the bracket. The estimate \(4.625\) is therefore a coarse approximation; further halvings would narrow the guaranteed error bound.
Secant Iteration: Faster Proposals, Weaker Guarantees
The secant method uses two current function evaluations to approximate the curve by a straight line and uses that line's zero as the next proposal. For two distinct points \(\delta_{k-1}\) and \(\delta_k\), with unequal function values, the update is
Unlike bisection, the secant method does not require an initial sign-changing bracket and may converge quickly near a well-behaved root. But it does not guarantee that a proposal stays within a plausible delta range or that it converges. If its function-value denominator is zero or very small, the proposal is undefined or unstable. Check proposals against the original function and the scientific search range; where a dependable bracket is important, prefer a bracket-preserving method.
Worked Example: Secant Proposals for a Nonlinear Curve
Consider the hypothetical tipping-point function \(g(\delta)=0.12-0.040\delta+0.001\delta^2\), with \(g\) measured in outcome units and \(\delta\) in the delta units used for the analysis. The function is a continuous quadratic. Its values at \(\delta=2\) and \(\delta=5\) are \(g(2)=0.12-0.080+0.004=0.044\) and \(g(5)=0.12-0.200+0.025=-0.055\) outcome units. These opposite signs provide a bracket as well as two starting points for secant iteration.
Using \(\delta_{k-1}=2\) and \(\delta_k=5\), the first secant proposal is
Substitution gives \(g(3.3333)\approx0.12-0.133332+0.011111=-0.002221\) outcome units. Using 5 and 3.3333 as the next pair, the secant proposal is approximately \(3.2632\). Substituting that value gives \(g(3.2632)\approx0.0001205\) outcome units. A further proposal from these two most recent points is approximately \(3.2668\), where the function is approximately zero at the displayed precision.
As a check, the quadratic formula gives the roots:
The smaller root is approximately \(3.2668\), consistent with the secant proposals; the larger root is approximately \(36.73\). The initial bracket \([2,5]\) contains the smaller root, not the larger one. This check illustrates why the root returned by a numerical search must be interpreted relative to its starting values and search range.
Multiple Crossings and Choosing a Search Interval
A sensitivity curve need not be monotone. If it crosses its threshold multiple times, a search can locate one crossing without revealing the others. Therefore, the interval is part of the definition of the reported tipping point: report which crossing is being sought and why that range is relevant to the sensitivity question.
Worked Example: Two Roots in the Same Sensitivity Curve
Suppose a hypothetical tipping-point function is \(g(\delta)=0.010(\delta-2)(\delta-6)\), with function values in outcome units and delta measured on a specified sensitivity scale. This function is continuous, and it equals zero at both \(\delta=2\) and \(\delta=6\).
On \([1,4]\), \(g(1)=0.050\) and \(g(4)=-0.040\) outcome units, so bisection can locate the root at \(\delta=2\). On \([4,7]\), \(g(4)=-0.040\) and \(g(7)=0.050\), so bisection can instead locate the root at \(\delta=6\). Both brackets meet the continuity and sign-change conditions. A search on a wider interval may find a crossing, but the sign change alone does not say which root it will refine. The analyst must inspect the curve and choose a range that addresses the intended conclusion change.
Diagnostics, Alternatives, and Limitations
A root estimate should be checked, not just returned by software. Evaluate the original \(g\) at the reported delta and report the residual \(g(\hat{\delta})\), the final bracket or stopping tolerance, and the delta units. If the relevant conclusion is determined by which side of a threshold the curve lies on, evaluate nearby delta values on both sides when possible. A small residual alone is not enough if the function is nearly flat, evaluations are unstable, or the result lies outside the meaningful range.
Common warning signs include a proposed delta outside the prespecified range, repeated proposals with little progress, a secant denominator close to zero, inconsistent signs on repeated evaluations, and a grid or plot suggesting more than one crossing. A discontinuity or substantial numerical noise can make a sign change a poor indicator of a genuine crossing. In such cases, investigate the function evaluations and assumptions rather than treating an algorithm's output as a valid tipping point.
| Approach | Strength | Limitation |
|---|---|---|
| Grid search | Shows the curve at a set of interpretable delta values and can reveal broad patterns. | Resolution is limited by grid spacing; a narrow crossing can fall between grid points. |
| Bisection | With a continuous function and a verified sign-changing bracket, it reliably narrows an interval. | Requires a bracket and typically uses more function evaluations than a fast local method. |
| Secant iteration | Can refine a root quickly using function values without derivatives. | May fail to converge, leave the range, or behave poorly when function values are close. |
| Newton iteration | Can converge rapidly when the derivative is available and the starting value is suitable. | Requires a derivative and can take unstable or out-of-range steps, especially near a flat slope. |
These methods locate numerical crossings; they do not establish that a delta value is plausible or that the underlying sensitivity assumptions are credible. Nor do they remove uncertainty in the imputation analysis or in evaluating the sensitivity curve. A reported tipping point should be presented with its threshold, direction of change, delta scale, search range, and numerical tolerance so readers can understand what the crossing means.
Common Mistake
Wrong reasoning: “The endpoint values have opposite signs, so bisection will always retain a strict sign change after every midpoint.” If the midpoint is exactly a root, neither half has endpoints with strictly opposite signs. For example, \(g(\delta)=\delta\) on \([-1,1]\) has opposite signs at the endpoints, but its midpoint is an exact root.
Correction: At each iteration, first check whether the midpoint is a root to the chosen numerical tolerance. If so, stop. Otherwise, retain the half whose endpoints have opposite signs. This distinction preserves the validity of the bisection argument without incorrectly excluding exact midpoint roots.
Check Your Understanding
Answer each question, then compare with the solutions below.
- If \(T(\delta)\) is a pooled estimate and \(c\) is a clinical threshold, what function should be set to zero to locate equality with that threshold?
- A continuous function has \(g(a)=0.3\) and \(g(b)=-0.1\). After four bisections, what is the bracket width as a fraction of its original width?
- For \(g(\delta)=0.5-0.1\delta\), find the root and verify it by substitution.
- Why can a sign-changing interval fail to identify a unique tipping point?
Solutions. (1) Set \(g(\delta)=T(\delta)-c\) to zero; then \(T(\delta)=c\) at a root. (2) Each bisection halves the width, so after four steps the width is \(1/2^4=1/16\) of the original. (3) Solve \(0.5-0.1\delta=0\), giving \(\delta=5\); substitution gives \(g(5)=0.5-0.5=0\). (4) A continuous function can cross zero more than once. A sign-changing bracket guarantees at least one root, not a unique root, so the curve and search interval must be considered.