Why Grid Resolution Matters
A tipping-point analysis evaluates a prespecified decision rule over plausible values of one or more sensitivity parameters. A grid makes that evaluation manageable: rather than checking every possible parameter value, the analyst checks a finite set. But the grid is only a numerical summary of the parameter space. If its spacing is too wide, a reported boundary can be imprecise, a curved boundary can be poorly represented, or a narrow region with a different decision can be missed entirely.
The goal is not to make the grid arbitrarily dense. A very dense grid costs more computation and can suggest more precision than the analysis supports. The goal is to choose spacing that resolves the boundary at the precision needed for the scientific question, then check that conclusions are not artifacts of the chosen grid.
Let \(\boldsymbol{\delta}\) denote the sensitivity parameter or vector of sensitivity parameters. Write \(D(\boldsymbol{\delta})\) for the classification produced by the prespecified decision rule at that setting, such as whether the conclusion remains on one side of the rule or has tipped to the other. The boundary separates parameter settings with different classifications. When the decision can be represented by a continuous margin \(m(\boldsymbol{\delta})\), the boundary is often described by \(m(\boldsymbol{\delta})=0\), with the sign of the margin indicating the classification.
The grid locates the boundary only to the extent supported by its spacing and the behavior of the decision rule. In particular, a grid point is an evaluated setting, not a boundary location. A boundary estimate based on interpolation is also an approximation unless the margin is known to vary linearly over that interval or cell.
Spacing, Brackets, and Refinement
For a one-parameter analysis over the range \([a,b]\), dividing the range into \(N\) equal intervals gives spacing \(h=(b-a)/N\). If adjacent grid points have different classifications, they bracket at least one crossing when the decision margin is continuous between them. Without further calculation, the crossing is localized only to that interval, whose width is \(h\). Thus, if the desired numerical localization is within 0.05 parameter units, a final bracket wider than 0.05 does not meet that target.
A useful strategy is to first scan the full plausible range, then refine where classifications change. In one dimension, evaluate the midpoint of a bracket and keep the half-interval whose endpoints still have different classifications. Repeating this step halves the bracket width each time. This refinement presumes that the same boundary crossing is being followed and that the evaluated decision is stable. If there may be multiple crossings, each detected bracket must be followed separately, and the initial scan must be fine enough to avoid overlooking crossings between same-class endpoints.
Choose an initial grid spanning the prespecified plausible parameter range. Record the decision at every grid point, rather than beginning with only a small region where a boundary is expected.
In one dimension, flag adjacent points with different classifications. In two dimensions, flag neighboring grid locations or cells where classifications change, while remembering that this will not necessarily reveal a small region enclosed within a same-class cell.
Add grid points around candidate boundaries, re-evaluate the decision rule, and compare the refined map with the initial scan. Retain a sufficiently broad scan to check for other crossings or regions.
Give the grid spacing or final bracket width, identify the parameter units, and distinguish a bracketed location from an interpolated estimate.
Worked Example: Refining a One-Parameter Bracket
Suppose a hypothetical analysis varies a sensitivity parameter \(\delta\), measured in outcome-score points, from 0 to 3. Let the decision margin be \(m(\delta)=1.4-\delta\); the decision tips when \(m(\delta)\) changes from positive to negative. The margin is continuous and strictly decreasing, so it has one crossing in this range.
Start with grid points 0, 1, 2, and 3. At \(\delta=1\), \(m(1)=1.4-1=0.4\), which is positive. At \(\delta=2\), \(m(2)=1.4-2=-0.6\), which is negative. Therefore the crossing is bracketed by \([1,2]\), a bracket one score point wide. The grid has not located the crossing to an accuracy better than that bracket on its own.
Evaluate the midpoint, \(\delta=1.5\): \(m(1.5)=1.4-1.5=-0.1\). The crossing is now bracketed by \([1,1.5]\), width \(0.5\). Evaluate the new midpoint, \(\delta=1.25\): \(m(1.25)=0.15\), so the bracket becomes \([1.25,1.5]\), width \(0.25\). Next, \(m(1.375)=0.025\), giving \([1.375,1.5]\), width \(0.125\). One more midpoint gives \(\delta=1.4375\) and \(m(1.4375)=-0.0375\), so the bracket is \([1.375,1.4375]\), width \(0.0625\) score points.
This establishes a bracket of less than 0.1 score points. Because the example’s margin is exactly linear, solving \(1.4-\delta=0\) gives the exact crossing \(\delta=1.4\). In an application where only grid evaluations are available, the bracket is the warranted numerical result; linearity should not be assumed merely because two grid points were evaluated.
Curvature and Interpolation
In two-parameter analyses, contours and boundaries can curve through grid cells. A coarse grid may show the broad location of a boundary but represent its shape poorly. If a boundary is interpolated between grid locations, its accuracy depends not only on spacing but also on how rapidly its direction or curvature changes across the cell.
The following elementary bound explains why finer spacing helps when the margin is smooth. Suppose a one-dimensional margin \(m(x)\) is twice differentiable on an interval of width \(h\), and \(|m''(x)|\leq M\) throughout that interval. Let \(L(x)\) be the straight line joining the endpoint values of \(m\). The linear interpolation error satisfies \(|m(x)-L(x)|\leq Mh^2/8\). If the true crossing \(r\) and interpolated crossing \(\tilde r\) lie in this interval, and \(|m'(x)|\geq\kappa>0\) throughout it, then the mean value theorem gives \(|r-\tilde r|\leq Mh^2/(8\kappa)\).
Worked Example: A Curved Two-Parameter Boundary
Consider a hypothetical two-parameter analysis with \(\delta_1\) and \(\delta_2\) both measured in score points. Suppose the margin is \(m(\delta_1,\delta_2)=\delta_2-[0.40+0.10\delta_1+0.08\delta_1^2]\), so the boundary is \(\delta_2=0.40+0.10\delta_1+0.08\delta_1^2\). Assume the plotted parameter range includes these boundary values.
On a coarse grid with \(\delta_1\) values 0 and 1, the boundary heights are \(0.40+0.10(0)+0.08(0)^2=0.40\) and \(0.40+0.10(1)+0.08(1)^2=0.58\) score points. Straight-line interpolation halfway across, at \(\delta_1=0.5\), gives \((0.40+0.58)/2=0.49\) score points.
But the actual boundary at \(\delta_1=0.5\) is \(0.40+0.10(0.5)+0.08(0.5)^2=0.40+0.05+0.02=0.47\) score points. The coarse linear representation is 0.02 score points too high at that location. A refined grid that evaluates \(\delta_1=0.5\) directly records the boundary there as 0.47. This example shows how refinement can improve the representation of a curved boundary; the error in a particular application is not known unless the margin or boundary behavior is assessed.
When a Coarse Grid Misses a Region
A more serious problem arises when a region with one decision is narrow and lies between grid points. Checking only whether neighboring points have different classifications will not detect such a region if the points on both sides have the same classification. The same issue occurs in two dimensions: a small region can be enclosed within a grid cell whose corners all have the same classification. No refinement rule triggered only by observed changes at grid nodes can guarantee detection of an unobserved region.
Worked Example: A Narrow Tipping Island
For illustration, let a hypothetical two-parameter decision margin be \(m(\delta_1,\delta_2)=(\delta_1-1.5)^2+(\delta_2-1.5)^2-0.04\), with the tipped classification where \(m<0\). This decision region is a circle centered at \((1.5,1.5)\) with radius \(0.2\) parameter units.
Use the coarse grid with each parameter taking values 0, 1, 2, and 3. The closest grid points to the center are \((1,1)\), \((1,2)\), \((2,1)\), and \((2,2)\). At each, the squared distance from the center is \(0.5^2+0.5^2=0.5\), so \(m=0.5-0.04=0.46>0\). All four corners of the cell around the center have the untipped classification, and the other grid points are farther away. Thus the coarse grid reports no tipping region.
Now add the point \((1.5,1.5)\). There, \(m=0^2+0^2-0.04=-0.04<0\), so the tipping region is detected. A fine grid with increments of 0.1 would also include this center point. The example makes an important limitation explicit: a coarse grid can miss a narrow region even when every evaluated value is calculated correctly. Refinement prompted by changes at grid points alone would not catch this island.
Diagnostics, Alternatives, and Limitations
Useful checks include repeating the scan with a finer grid, comparing the locations and shapes of the resulting boundaries, and checking whether a finer scan reveals additional crossings or disconnected regions. In one dimension, record the width of each final bracket. In two dimensions, report both spacings and inspect whether the displayed boundary shifts materially when the grid is refined. If the decision is computed by simulation or another numerically variable procedure, check that repeated evaluations at the same parameter setting give a stable classification before attributing apparent changes to the grid.
Uniform grids are straightforward and provide broad, predictable coverage, but can spend many evaluations far from a boundary. Adaptive refinement allocates more evaluations near suspected crossings and is often more efficient. Its weakness is that local refinement usually follows what the initial scan has already revealed. A coarse initial scan can hide a narrow island, so targeted refinement should supplement—not replace—a scan of the relevant range. Where scientifically justified, known smoothness or monotonicity can guide the grid; those properties should not be presumed solely for computational convenience.
Grid resolution concerns numerical localization in the sensitivity-parameter space. It does not by itself establish that the sensitivity-parameter range is clinically plausible, that the decision rule is appropriate, or that the substantive conclusion is robust. Nor does a fine grid remove uncertainty in the underlying analysis. Report grid spacing and numerical precision in the parameter’s units, and avoid giving an interpolated boundary more digits than the computation and assumptions warrant.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A one-dimensional grid covers 0 to 4 score points in 8 equal intervals. What is its spacing? If a boundary is bracketed by adjacent points, what is the maximum bracket width before refinement?
- A bracket has width 0.8 parameter units. How many midpoint refinements are needed to make its width no more than 0.1 units?
- Why can a two-dimensional grid with identical decisions at all four corners of a cell fail to show every tipping region?
- What should be reported when a boundary is interpolated between grid points rather than evaluated exactly?
Solutions. (1) The spacing is \(h=(4-0)/8=0.5\) score points. An adjacent-point bracket is at most 0.5 score points wide. (2) Each midpoint refinement halves the width: \(0.8/2=0.4\), \(0.4/2=0.2\), \(0.2/2=0.1\). Three refinements suffice. (3) A narrow region can lie wholly inside the cell, with no corner inside it; the grid evaluates corners, not every point between them. (4) Report the grid spacings, the parameter units, and that the location is interpolated; do not present it as an exact evaluated boundary or imply precision unsupported by the spacing and assumptions.