Reading Tornado Diagrams Without Overclaiming

What a Tornado Diagram Shows

A tornado diagram is a visual summary of a set of one-way sensitivity analyses. Each bar represents the range of a decision outcome obtained by changing one input across specified bounds while holding the other inputs at their reference values. Bars are commonly ordered from widest to narrowest, giving the chart its characteristic shape.

The diagram is useful for seeing which specified input ranges produce the largest changes in an outcome. But its ranking is conditional on the chosen bounds, reference values, model, and outcome measure. A long bar is not, by itself, evidence that the input is more uncertain, more likely to change, or more important in the real world.

Definition: Let the decision outcome be $y=f(x_1,\ldots,x_k)$, where $x_1,\ldots,x_k$ are decision inputs. Let $x^0$ be the reference input vector and $y^0=f(x^0)$ the reference outcome. For input $x_i$, a one-way tornado bar shows outcomes $y_{i,L}$ and $y_{i,U}$ calculated at specified lower and upper input bounds, with all other inputs fixed at their reference values. The bar spans the two outcomes; its width is $W_i=|y_{i,U}-y_{i,L}|$.

A complete interpretation starts with the chart’s construction. Identify the outcome and its units, the reference outcome, the input being varied, and the lower and upper bounds used for that input. Check whether the plotted bar shows endpoint outcomes, changes from the reference, or both. The reference line is especially informative: endpoints on either side indicate a change in direction across the scenarios, while endpoints on the same side mean both scenarios move the outcome in the same direction relative to the reference.

“Lower input” and “lower outcome” are not interchangeable. An input may have a negative or non-monotone relationship with the outcome, so its lower bound can produce the higher outcome. Read endpoint labels or the analysis table rather than inferring direction from the words “low” and “high.”

Why Bar Width Depends on the Bounds

The width is the difference between two model outputs, not an intrinsic property of an input. The following short derivation shows why the selected input range matters. If the outcome is approximately linear in $x_i$ over the varied interval, while all other inputs remain fixed, write its local relationship as $y\approx a+\beta_i x_i$. Then:

W_i =|y_{i,U}-y_{i,L}| \approx |(a+\beta_i U_i)-(a+\beta_i L_i)| =|\beta_i|(U_i-L_i).

Thus, even under a simple linear relationship, bar width combines the outcome’s responsiveness to the input, $|\beta_i|$, with the width of the selected input range, $U_i-L_i$. A comparatively wide bar could reflect a broad scenario range rather than unusually strong responsiveness. With nonlinear relationships, the endpoint difference can also depend on where the interval lies; equal-width input intervals need not produce equal-width bars.

The ranges should therefore have a defensible interpretation. They might represent plausible alternative assumptions, a specified scenario range, or another explicitly justified interval. They are not automatically confidence or credibility intervals, probability distributions, or predictions of what will happen. In a standard one-way tornado analysis, other inputs are fixed at their reference values; this makes the comparisons interpretable as isolated changes, but it does not describe all combinations of changing inputs.

Worked Example: Ranking Bars for a Net-Benefit Outcome

Hypothetical data. A development team uses incremental net benefit, in millions of dollars, as its decision outcome. Its reference inputs give $y^0=\$12$ million. For each row below, the model is evaluated at the listed lower and upper input bounds, with all other inputs at reference values.

Input varied	Outcome at lower bound	Outcome at upper bound
Treatment effect	$4 million	$21 million
Development cost	$8 million	$15 million
Eligible population	$10 million	$14 million

For treatment effect, $W=|\$21-\$4|=\$17$ million. Its endpoints are $-\$8$ million and $+\$9$ million relative to the $\$12$ million reference outcome.

For development cost, $W=|\$15-\$8|=\$7$ million. Its endpoints are $-\$4$ million and $+\$3$ million relative to reference.

For eligible population, $W=|\$14-\$10|=\$4$ million. Its endpoints are $-\$2$ million and $+\$2$ million relative to reference. The descending bar-width order is therefore treatment effect, development cost, eligible population.

This ranking says that, over the particular bounds used in this analysis, varying treatment effect produces the largest endpoint range in incremental net benefit. It does not say treatment effect is the most uncertain input, nor does it give the probability that net benefit will take any value. Because the treatment-effect endpoints fall below and above zero, this one-way sweep crosses the zero-net-benefit reference; it does not establish what happens when other inputs also change.

Read the Reference Case and the Bounds Together

A tornado diagram ranks outcome changes under selected scenarios, so the reader needs to know why each scenario bound was chosen. Before interpreting a long bar, ask whether its input range is credible for the decision being considered, whether its bounds use the same scale or rationale as other inputs, and whether the reference value is appropriate. The earlier material on choosing plausible scenario bounds is relevant here: a mathematically valid sweep can still be decision-irrelevant if the bounds do not represent a defensible question.

The reference value matters because the model may respond differently in different parts of its input range. A bar around one reference case is not a universal ranking. If the reference changes, or the varied bounds shift, the endpoints and even the ordering can change. This is one reason to report the underlying values and ranges alongside the figure rather than treating the graphic as self-explanatory.

Worked Example: A Wider Bar Can Reflect a Broader Scenario Range

Hypothetical data. Consider two inputs in a model of incremental net benefit, measured in millions of dollars, with a reference outcome of $\$6$ million. A broad biomarker-effect scenario range produces outcomes of $-\$3$ million and $\$15$ million. A narrower exposure-scaling range produces outcomes of $\$1$ million and $\$11$ million.

The biomarker-effect bar has width $|\$15-(-\$3)|=\$18$ million. The exposure-scaling bar has width $|\$11-\$1|=\$10$ million. By bar width, biomarker effect ranks first for these chosen bounds.

Now suppose the team’s evidence review supports a narrower biomarker-effect range for the decision at hand, and that range gives outcomes of $\$4$ million and $\$8$ million. The corresponding width is $|\$8-\$4|=\$4$ million. Under that alternative scenario range, exposure scaling has the wider bar: $\$10$ million versus $\$4$ million.

The arithmetic does not establish which input is more uncertain; that requires information beyond the bar widths. It demonstrates that the ranking depends on the ranges. A sound interpretation reports the ranges and treats a ranking based on them as conditional, rather than claiming the first ranking is a permanent measure of input importance.

Overlapping Bars Are Separate Sweeps

Two bars can overlap on the outcome axis. This means that the outcome ranges from their separate one-way sweeps share values. It does not mean the inputs were varied together, that one input caused the other to change, or that the model contains an interaction between them. Each endpoint was calculated with the other inputs fixed at reference values.

Overlap also does not make the bars identical. Their widths, positions relative to the reference, and the input bounds behind them may differ. Describe the shared outcome range and the distinct sweeps plainly. Do not infer a joint outcome from the overlap: the chart has not evaluated the combination of, for example, both inputs at their lower bounds.

Worked Example: Interpreting Overlap Without Inferring Joint Effects

Hypothetical data. A model gives a reference incremental net benefit of $\$5$ million. Varying the eligible population alone produces outcomes from $\$2$ million to $\$9$ million. Varying unit cost alone produces outcomes from $\$3$ million to $\$8$ million.

The population bar spans $[\$2,\$9]$ million and has width $|\$9-\$2|=\$7$ million. The unit-cost bar spans $[\$3,\$8]$ million and has width $|\$8-\$3|=\$5$ million. Their ranges overlap from $\$3$ million to $\$8$ million. Each sweep has endpoints both below and above the $\$5$ million reference.

A justified reading is: “In separate one-way analyses, the outcome ranges overlap, and the population sweep is wider under the specified bounds.” It would be incorrect to say that the chart shows the outcome when both population and unit cost vary, or that overlap demonstrates an interaction. No joint scenario was calculated. Also, all four endpoints here remain above zero; this describes the endpoints shown, not every possible combination of inputs.

Diagnostics, Alternatives, and Limitations

A tornado diagram is most useful when its construction can be checked and its limits are visible. Before drawing a conclusion, review the model calculations at the reference case and at every displayed bound, verify the units, and check that the chart’s labels match the underlying scenarios. Then consider these failure modes:

Unclear or incomparable bounds. If bounds have no stated rationale, a ranking can reflect arbitrary range choices. If different inputs use different range definitions, interpret their widths accordingly rather than presenting them as a like-for-like measure.
A misleading endpoint summary. For a nonlinear response, endpoint values do not guarantee the response is monotone between the endpoints. If an interior change or reversal is plausible, evaluate additional points rather than assuming the bar captures every outcome in the interval.
Missing reference context. Without the reference values and outcome, it is difficult to tell whether a bar represents a large change in the decision context or merely a wide range on the plotted scale.
Overinterpreting the order. Close widths may provide little practical basis for declaring one input “the most important.” The ordering is specific to the model, outcome, reference case, and bounds.
Unshown uncertainty or dependence. A one-way chart does not encode how likely an input value is or how inputs may vary together. Its bars should not be read as probability statements or joint predictions.

A table of scenario results is a useful alternative when exact input bounds and endpoint outcomes matter more than a visual ranking. The table makes the calculations easy to inspect, while a tornado diagram makes relative ranges easier to scan. Neither alone supplies probabilities for outcomes. When the question concerns uncertainty across a joint distribution of inputs, probabilistic sensitivity analysis is a different approach; a one-way tornado diagram is not a substitute for it.

The central limitation is scope: the diagram describes model output under chosen one-at-a-time changes. It can help identify which assumptions merit attention within that scope, but it cannot, by bar length alone, establish real-world importance, likelihood, causation, a joint scenario, or a final decision.

Common pitfall: “The longest bar identifies the most uncertain and most important input.” This reasoning confuses a scenario-based output range with uncertainty and decision relevance. The correction is to report the input bounds and reference case, interpret the bar as the outcome span under those bounds, and use additional evidence or analyses for claims about uncertainty or real-world importance.

Check Your Understanding

Answer each question, then check the concise solutions below.

A reference outcome is 20 units. One input’s lower- and upper-bound scenarios give 14 and 27 units. What is the bar width, and what are the endpoint changes from reference?
Input A has a bar width of 12 units over its chosen range; input B has a width of 7 units. What can you conclude, and what can you not conclude, from this ranking?
Two bars span outcome ranges of $[2,8]$ and $[5,10]$ units. What does their overlap show, and what does it not show?
Why might checking an interior input value be useful when a model is nonlinear?
A bar’s two endpoints are both above a decision reference of zero. What does that establish about the displayed sweep, and what does it not establish about joint variation?

Solutions. (1) The width is $|27-14|=13$ units. The lower-bound change is $14-20=-6$ units, and the upper-bound change is $27-20=+7$ units. (2) Over the bounds used, A produces the larger outcome span. The widths alone do not establish which input is more uncertain or more important in the real world. (3) The separate outcome ranges share values from 5 through 8 units. The overlap does not show a joint scenario or an interaction. (4) An interior value can reveal curvature or a reversal that endpoint-only results miss. (5) Both displayed endpoint outcomes are above zero; this does not establish the outcome for combinations in which other inputs vary too.