Why Locate a Tipping Point?
A decision can be sensitive to an assumption without being close to changing direction. Conversely, a small change in one uncertain input can move a decision from go to no-go even when most other assumptions have little effect. Tipping-point analysis makes that distinction explicit: it identifies the change in a specified assumption, or combination of assumptions, that brings a decision rule to its boundary.
The result is conditional on the decision rule, the reference case, the permitted range of assumptions, and the way changes are measured. A tipping point is not by itself evidence that the assumption is wrong, that the decision should be reversed, or that the boundary is likely to be reached. It is a structured way to ask how much an assumption would need to change before the stated conclusion changes.
The score need not be a clinical effect estimate. It might be a net-benefit measure, an expected utility difference, or another prespecified quantity that combines decision-relevant consequences. What matters is that the score and the rule translating it into go or no-go are defined before interpreting the tipping point. Equality also needs a convention: here, \(D=0\) is assigned to go, so a score must become negative to produce no-go.
Finding a One-Input Boundary
For a single uncertain input \(x\), hold all other assumptions fixed at their reference values and solve \(D(x)=0\). If there are multiple roots, check which are within the defined plausible range and which can be reached from the reference value. The nearest reachable root is the one-input tipping point under absolute change in \(x\). If a different distance measure is intended, such as percentage change, state it and use it consistently.
A root alone is not enough. Check the sign of \(D(x)\) on each side: the score may touch zero without crossing it, or it may cross more than once. Also check that the equation has not been altered by multiplying or dividing by a quantity that could be zero or change sign. These checks establish whether the calculated boundary corresponds to the claimed decision change.
Worked Example: Efficacy Score With Two Decision Boundaries
Suppose a hypothetical program uses the score \(D(E)=10E/(E+0.4)-6E-2\), where \(E\) is a dimensionless efficacy assumption, and goes if \(D(E)\geq 0\). The reference value is \(E=0.72\), and the plausible range is \(0.15\) to \(0.85\). Find the nearest tipping point.
At the reference value, \(D(0.72)=10(0.72)/(0.72+0.4)-6(0.72)-2 =7.2/1.12-4.32-2 =6.4286-6.32 =0.1086\). The reference decision is go.
To locate the boundaries, solve \(D(E)=0\). In the stated range, \(E+0.4\) is positive, so multiplying by it preserves the equation's roots: \(10E-(6E+2)(E+0.4)=0\). Expanding gives \(6E^2-5.6E+0.8=0\). Thus \(E=[5.6\pm\sqrt{5.6^2-4(6)(0.8)}]/12 =[5.6\pm\sqrt{12.16}]/12\), which gives roots \(E\approx0.1761\) and \(E\approx0.7573\).
Both roots are in the plausible range, but their distances from \(0.72\) differ: \(0.72-0.1761=0.5439\), whereas \(0.7573-0.72=0.0373\). The upper root is the nearest boundary. The score is decreasing in this neighbourhood because \(D'(E)=4/(E+0.4)^2-6\), which is negative from \(E=0.72\) to \(0.7573\). For example, \(D(0.700)=7/1.1-6.2\approx0.1636\), while \(D(0.800)=8/1.2-6.8\approx-0.1333\). These checks support a crossing from go to no-go as efficacy increases past the upper boundary. With go assigned at equality, \(E\approx0.7573\) is the tipping boundary; a value just above it gives no-go.
The result is a change of about \(0.0373\) efficacy units from the reference value. It does not say that such a change is likely. Its evidentiary context should identify where the reference value and plausible range came from, and whether the range reflects data, expert elicitation, or a prespecified scenario analysis.
Safety Assumptions and Threshold Interpretation
For a safety input, the sign and units matter. A risk entered as a proportion is not the same numerical scale as a risk entered in percentage points. State the unit at the model interface, preserve it during the calculation, and translate the result into a clinically readable change. As with efficacy, the boundary is conditional on the other fixed assumptions and on the specified score.
Worked Example: Safety-Risk Tipping Point
Consider a hypothetical decision score \(D(q)=4.8-20q\), where \(q\) is the probability of a serious adverse event, expressed as a proportion. Go is assigned when \(D(q)\geq0\). At the reference assumption \(q=0.12\), the score is \(4.8-20(0.12)=4.8-2.4=2.4\), so the decision is go.
The boundary solves \(4.8-20q=0\), giving \(q=4.8/20=0.24\). If the plausible range is \(0.08\) to \(0.30\), this boundary is within range. The increase from reference to boundary is \(0.24-0.12=0.12\), or 12 percentage points in event probability. At \(q=0.25\), for example, \(D(0.25)=4.8-5=-0.2\), so the decision is no-go. Under the stated equality convention, \(q=0.24\) is the tipping boundary and a value above it reverses the decision.
The report should say whether the plausible range is supported by observed safety data or is a stress-test range, and whether the event definition and population match the decision being made. A boundary inside a broad, poorly supported range is not equivalent to a boundary inside a well-supported range. In either case, the calculation describes a conditional threshold, not the probability that the true risk exceeds it.
Joint Changes Require a Distance Rule
When several assumptions may change together, there is generally no unique “smallest” joint change until a distance rule is specified. Inputs have different units and ranges; an unscaled Euclidean distance, for example, can be dominated by whichever input happens to use larger numerical units. One option is to scale each change by a prespecified plausible-range width and find the closest point on the decision boundary in those scaled coordinates. That choice is a sensitivity convention, not a statistical law. It must also be checked against dependence and feasibility: a mathematically close point may be an implausible combination of assumptions.
Worked Example: A Joint Efficacy-and-Safety Tipping Point
Suppose a hypothetical score is \(D(E,q)=8E-25q-1.5\), with go when \(D\geq0\). The reference assumptions are \(E=0.70\) and \(q=0.10\), giving \(D=8(0.70)-25(0.10)-1.5=5.6-2.5-1.5=1.6\), a go decision. Assume plausible ranges \(0.60\leq E\leq0.80\) and \(0.08\leq q\leq0.16\).
A one-input check gives an efficacy boundary at fixed \(q=0.10\): \(8E-2.5-1.5=0\), so \(E=0.50\), outside its plausible range. At fixed \(E=0.70\), the safety boundary is \(5.6-25q-1.5=0\), so \(q=4.1/25=0.164\), slightly above its plausible upper bound. Neither one-input change reaches the boundary within its own range.
Now define scaled changes \(u=(E-0.70)/0.20\) and \(v=(q-0.10)/0.08\), using the full plausible-range widths. The score becomes \(D=1.6+1.6u-2v\), so the boundary is \(1.6u-2v=-1.6\). Under ordinary Euclidean distance in the \((u,v)\) plane, the nearest point on this line to \((0,0)\) is the projection \((u,v)=-1.6(1.6,-2)/(1.6^2+(-2)^2)\approx(-0.3902,0.4878)\).
Returning to the original inputs gives \(E=0.70+0.20(-0.3902)\approx0.6220\) and \(q=0.10+0.08(0.4878)\approx0.1390\). Both lie in their stated ranges. Substitution checks the boundary: \(D=8(0.6220)-25(0.1390)-1.5\approx4.976-3.475-1.5\approx0.001\), with the small residual due to rounding; the unrounded projection gives zero. This joint result depends on the selected scaling and Euclidean metric. If efficacy and safety assumptions are dependent, or this combination is not considered feasible, the point is not an admissible tipping point for that uncertainty structure.
Reporting, Diagnostics, and Alternatives
A useful report gives enough information for a reader to reconstruct what the threshold means. Include the reference decision and score, the decision rule including its equality convention, the tipping value and direction of change, the plausible range, the change metric, and the source or rationale for the assumptions. For joint changes, name the scaling and explain why the combination is feasible. Present the score or decision boundary alongside the input values so that the conclusion is not detached from its evidentiary context.
Several checks help detect a misleading result. Recalculate the score at the reference case and at values just to either side of the proposed boundary. Check every algebraic root against the original equation, the input domain, and the plausible range. For numerical searches, check the search interval and resolution; a coarse grid can miss a narrow crossing. If the score is non-monotone, search for multiple roots rather than assuming one boundary. If no root occurs in the plausible range, report that the decision does not tip within that range under the specified analysis; do not extrapolate a distant root as if it were plausible.
Tipping-point analysis answers a different question from several related approaches. One-way sensitivity analysis shows how a decision score changes over a range for one input, while a tipping-point calculation extracts the boundary from that relationship. Threshold plots can display that boundary across values of a second input. Probabilistic sensitivity analysis can describe how often sampled assumptions yield go under specified distributions; a tipping point instead identifies a boundary and does not assign it a probability. Robustness criteria summarize decision stability under defined uncertainty sets, whereas a tipping point identifies where the boundary lies relative to a reference case. These approaches can complement one another, but none makes the others' assumptions unnecessary.
The main limitations are consequential. A tipping point may be sensitive to the score's construction, the chosen reference case, the treatment of interactions, the range definitions, or the distance metric. It can also conceal uncertainty in the decision rule itself. A nearby boundary signals fragility conditional on the model and assumptions; it does not establish which side is correct. A distant or unreachable boundary does not prove that a decision is safe from uncertainty outside the assessed range.
Common Mistake: Calling the Boundary “the Likely Reversal”
A common error is to find a boundary inside the plausible range and conclude that the decision is likely to reverse. That reasoning confuses possibility under a sensitivity analysis with probability. A range may be a set of deliberately selected scenarios and may have no probability distribution. The correction is to report that the boundary is reachable within the stated range, describe the evidence supporting that range, and avoid probability language unless a separate probabilistic analysis justifies it.
Check Your Understanding
Answer each question, then compare with the concise solutions below.
- A score is \(D(x)=3-5x\), with go when \(D\geq0\). At \(x=0.40\), is the decision go, and what is the boundary?
- For \(D(q)=6-30q\), with \(q\) a probability proportion, find the boundary and express the change from \(q=0.10\) in percentage points.
- Why is a two-input tipping point not fully specified by saying “find the smallest change”?
- A calculated boundary is outside the stated plausible range. What conclusion is warranted, and what conclusion is not?
Solutions. (1) \(D(0.40)=3-2=1\), so the decision is go. Setting \(3-5x=0\) gives \(x=0.60\); values above it produce no-go. (2) \(6-30q=0\) gives \(q=0.20\). The change from \(0.10\) is \(0.10\), or 10 percentage points. (3) Inputs can have different units and scales, so “smallest” requires a specified distance or scaling rule, plus a feasible set for joint changes. (4) The warranted conclusion is that the decision does not tip within the assessed range under that analysis. It does not follow that the decision cannot tip under assumptions outside the range.