Relating a Tipping Threshold to a Plausible Range

Why Compare a Threshold With a Plausible Range?

A tipping threshold summarizes how far a sensitivity parameter must move before a specified conclusion changes. A separate judgment is needed to decide whether that parameter value is plausible for the application. Comparing the two can make sensitivity results easier to interpret: it tells us whether the threshold falls among the sensitivity values the analysis regards as plausible.

That comparison is informative, but limited. A threshold inside a plausible range signals that the conclusion can change at a value in that range, subject to the decision rule and model used. It does not establish that the value is likely, or that the conclusion changes throughout the range. A threshold outside the range does not, by itself, establish that the conclusion is stable throughout the range. Those stronger statements require checking how the decision behaves across the range.

The key is to keep three things distinct: the sensitivity model, the decision rule that defines tipping, and the rationale for the plausible range. The threshold is conditional on the first two; it does not validate the third.

Definition and Conditions for the Comparison

Definition: Let \(\delta\) be a sensitivity parameter, and let \(C(\delta)\) denote the conclusion under a specified sensitivity model and decision rule. A tipping threshold \(\delta^*\) is a value at which that conclusion changes, or reaches the specified decision boundary. A prespecified plausible range is an interval \([L,U]\) of sensitivity values judged relevant for the application before interpreting the threshold. The threshold is inside the range when \(L \leq \delta^* \leq U\), and outside it otherwise.

The comparison is meaningful only when \(\delta^*\), \(L\), and \(U\) refer to the same parameter, scale, and direction. For example, a threshold expressed as a departure in outcome points cannot be compared directly with a range expressed as a standardized departure unless the conversion is made explicit. Sign conventions matter too: a positive value in one parameterization may represent the opposite departure from a positive value in another.

The plausible range should have an application-specific rationale, such as subject-matter knowledge or external evidence, and should be stated independently of the threshold result. It is an analysis input, not a statistical confidence interval and not automatically a probability distribution for \(\delta\). Without a defensible basis for its endpoints, the inside-or-outside label has little substantive meaning.

A threshold also depends on its decision rule. Tipping may mean a change in the sign of an estimated contrast, loss of statistical significance, or crossing a clinical decision boundary. Those are different events and can yield different thresholds. The range comparison must use the threshold for the conclusion the analysis actually intends to assess.

What the Comparison Can Establish

If the threshold lies inside the plausible range, at least one boundary of the specified conclusion lies within the values considered plausible. Under continuity, if the decision is determined by whether a numerical quantity \(g(\delta)\) is above or below zero, a sign change across the range implies a zero somewhere in the range. This does not imply that all values in the range lead to the same conclusion, or that values near the threshold are more probable than others.

If a threshold lies outside the range, then the location of that particular boundary is beyond the range. To conclude that the decision does not change anywhere inside the range, more is needed: for example, that the decision function is continuous and monotone over the range, that the relevant threshold is the only crossing, and that the range lies wholly on one side of it. Alternatively, the decision can be evaluated throughout the range directly. If there are multiple crossings, one threshold alone cannot summarize the behavior.

Useful implication: Suppose \(g(\delta)\) is continuous and strictly monotone on \([L,U]\), and the decision changes only where \(g(\delta)=0\). If \(g(L)\) and \(g(U)\) have opposite signs, the intermediate value theorem guarantees a threshold in \([L,U]\). If both endpoint values have the same nonzero sign, strict monotonicity ensures that \(g\) keeps that sign throughout the interval, so there is no threshold inside it.

For a linear decision quantity \(g(\delta)=a+b\delta\), with \(b\ne 0\), the threshold is \(\delta^*=-a/b\). Substitution into the endpoints gives \(g(L)=b(L-\delta^*)\) and \(g(U)=b(U-\delta^*)\). When \(L \leq \delta^* \leq U\), the endpoint values are on opposite sides of zero or one is zero. When the threshold is outside the interval, the two endpoint values have the same sign. Since a linear function has no additional crossing between its endpoints, these checks establish the decision behavior across the interval. This justification depends on linearity; endpoint checks alone need not suffice for a nonmonotone function.

Worked Example: Threshold Outside a Plausible Range

Suppose a hypothetical sensitivity analysis uses a dimensionless parameter \(\delta\), and the estimated treatment contrast, in outcome points, is \(\theta(\delta)=2.4-1.2\delta\). The specified conclusion is positive contrast when \(\theta(\delta)>0\). Subject-matter judgment, set before examining the threshold, gives the plausible range \([-0.5,1.5]\). Does the threshold lie in that range, and what can be concluded?

Find the threshold. Set the contrast equal to zero: \(2.4-1.2\delta=0\), so \(\delta^*=2.4/1.2=2.0\). The threshold is outside the plausible range because \(2.0>1.5\).

Check the range. At its lower endpoint, \(\theta(-0.5)=2.4-1.2(-0.5)=2.4+0.6=3.0\) outcome points. At its upper endpoint, \(\theta(1.5)=2.4-1.2(1.5)=2.4-1.8=0.6\) outcome points.

The contrast is linear and decreasing, so its smallest value on this interval is \(0.6\), which is still positive. Under this model and decision rule, the positive-contrast conclusion holds throughout the stated range. This conclusion follows from the linearity and endpoint check, not from the outside label alone. It does not show that the range is correct or that other sensitivity models would give the same result.

Worked Example: Threshold Inside the Range

Consider a second hypothetical analysis with \(\theta(\delta)=2.75-2\delta\) outcome points and the same rule: the conclusion is positive when \(\theta(\delta)>0\). Suppose the prespecified plausible range is \([1,2]\). Determine whether the threshold is inside the range and interpret the result.

Find the threshold. Solving \(2.75-2\delta=0\) gives \(\delta^*=2.75/2=1.375\), which is inside \([1,2]\).

Verify the behavior. At \(\delta=1\), the contrast is \(2.75-2(1)=0.75\) outcome points. At \(\delta=2\), it is \(2.75-2(2)=-1.25\) outcome points. The linear contrast decreases from a positive value to a negative one, and it equals zero at \(\delta=1.375\).

Thus, under this model, the positive-contrast conclusion changes within the plausible range. The comparison does not say that the treatment contrast is more likely to be negative: the range is not a probability distribution. Nor does it determine which sensitivity value is most credible. It identifies where the decision boundary falls relative to the range that was specified.

When an Outside Threshold Can Mislead

An outside threshold can be falsely reassuring if the decision function has more than one crossing. The following hypothetical calculation illustrates why the full decision behavior matters. Let \(g(\delta)=(\delta-0.5)(\delta-2)\), with a positive conclusion when \(g(\delta)>0\), and consider the plausible range \([1,2.5]\). The thresholds are \(0.5\) and \(2\). The first is outside the range, but the second is inside it.

Worked Example: Multiple Crossings and an Incomplete Threshold

At the lower endpoint, \(g(1)=(1-0.5)(1-2)=(0.5)(-1)=-0.5\). At \(\delta=1.5\), \(g(1.5)=(1)(-0.5)=-0.5\). At the upper endpoint, \(g(2.5)=(2)(0.5)=1.0\). The decision changes at \(\delta=2\), which lies inside the plausible range.

If a report mentioned only the threshold \(0.5\), it could incorrectly suggest that the plausible range contains no threshold. The quadratic function is not monotone on this interval: it decreases to a minimum and then increases. Here the second root, not the first, lies in the range. The calculation shows why a threshold comparison must account for all relevant crossings or be supplemented by evaluating the decision over the range.

Diagnostics, Alternatives, and Limitations

Before interpreting an inside-or-outside comparison, check the following:

Parameter alignment. Confirm that the range and threshold use the same delta definition, scale, sign convention, and arm-specific or contrast parameterization.
Decision alignment. Confirm that the threshold corresponds to the intended conclusion, rather than a different rule such as statistical significance versus effect direction.
Range rationale. Explain why the endpoints are credible for the application. A range chosen after seeing the threshold can make an apparently reassuring comparison circular.
Behavior across the interval. Check monotonicity, possible multiple crossings, and the decision at relevant points. Endpoint checks are decisive only with suitable shape assumptions, such as linearity or monotonicity.
Numerical stability. Verify a computed threshold by substituting it back into the decision function. If the threshold was found numerically, evaluate on both sides to confirm the expected decision change and check that the search did not miss other crossings.

Comparing a threshold with a plausible range is a compact summary. Directly evaluating the decision rule at a sufficiently informative set of values, or tracing the decision boundary, can reveal behavior that one threshold conceals. Those checks are especially important when the function may be nonlinear or nonmonotone. They do not replace a justified range: they answer how the specified model behaves at selected sensitivity values, not whether those values are credible.

The range itself is also a limitation. A narrow range may exclude important departures; a wide range may include values that are not substantively meaningful. Different experts may justify different endpoints. The comparison is therefore conditional: it reports what follows from a chosen range, sensitivity model, and decision rule. It is not a proof that missing-data assumptions are correct, a guarantee about unexamined models, or a probability statement about the conclusion.

A useful report gives the threshold, the range and its rationale, the direction of the decision change, and any checks for nonmonotonicity or additional crossings. State whether the conclusion changes anywhere in the range only when that behavior has actually been checked or justified by the function's shape.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A threshold is \(\delta^*=1.2\), and the prespecified plausible range is \([0,1]\). What does the outside-range comparison establish, and what does it not establish by itself?
For \(g(\delta)=3-\delta\), with a positive conclusion when \(g(\delta)>0\), find the threshold and determine the conclusion throughout the plausible range \([0.5,2]\).
For \(g(\delta)=(\delta-1)(\delta-3)\), the reported threshold is \(1\), and the plausible range is \([1.5,2.5]\). Is it valid to infer from that reported threshold alone that no decision change occurs in the range?
Why is a plausible range not automatically a confidence interval or a probability distribution for the sensitivity parameter?

Solutions. (1) The particular threshold \(1.2\) lies above the range. That locates this boundary, but does not alone establish that the decision is constant throughout \([0,1]\); other crossings or nonmonotone behavior could matter. (2) Solve \(3-\delta=0\), giving \(\delta^*=3\), outside the range. Since \(g\) is decreasing and \(g(2)=1>0\), it remains positive throughout \([0.5,2]\), so the positive conclusion holds there. (3) No. The function has roots at \(1\) and \(3\), and neither is in \([1.5,2.5]\); however, \(g(2)=(1)(-1)=-1\), while \(g(1.5)=(0.5)(-1.5)=-0.75\) and \(g(2.5)=(1.5)(-0.5)=-0.75\). The conclusion is negative throughout this interval, but that requires checking the function's behavior, not relying only on the reported root at \(1\). (4) The range represents selected sensitivity values considered relevant, not sampling uncertainty in an estimated parameter. A probability distribution requires an explicit probabilistic model; the interval alone supplies none.