Weight Sensitivity Analysis (Robustness of a Ranking to Weight Change)

Weight sensitivity analysis does not produce a ranking on its own. It takes a ranking already produced by some base method (for instance TOPSIS, SAW, or another ranking method) and asks: does this ranking hold even if the criterion weights change a little, or can a small change break it? Its output is a "stability interval" for every criterion (the range within which that weight can move without changing the ranking) and the criterion whose interval is narrowest, that is, the criterion to which the ranking is most sensitive. This gives the decision-maker a direct answer to the question "how reliable is your result."

Most ranking methods produce a single ranking from a single set of weights, and that ranking is then presented as if it were a settled fact. Weight sensitivity analysis takes the opposite view: a ranking is valid only for the weights it was given, and those weights are usually contested or approximate. The method asks "how robust is this first place" before it asks "which alternative is first." Its philosophical consequence is this: the value of a ranking is measured not by the ranking itself but by how easily it can be overturned. A narrow stability interval means that even a small disagreement over weight could change the result.

The method proceeds through four steps.

First, the base ranking. A chosen base method (for instance TOPSIS) computes an initial ranking and score using the existing weights.

Second, searching for the critical tipping point. For every criterion, that criterion's weight is raised and lowered while the other criteria's weights are redistributed proportionally. The method finds, by binary search, the smallest weight change at which the top two alternatives swap places.

Third, the stability interval. For every criterion, the lowest and highest weight value at which the ranking stays intact is identified. How far the current weight sits within this interval is recorded as that criterion's "margin."

Fourth, the most critical criterion. The criterion with the smallest margin is flagged as the one most likely to break the ranking, and every criterion's stability profile is reported as a table.

The formulas behind each step, the intermediate tables and the citation formats are given on the DecisionMind method page; this card carries no formulas.

The stability interval tells you how much a criterion's weight can change without swapping the base ranking's top two alternatives. A wide interval means you can be relaxed about that criterion; a narrow interval means even a small disagreement over it could change the result. The most critical criterion is where the discussion should concentrate most; if agreement cannot be reached on this criterion's weight, the reported ranking should not be treated as reliable. This analysis does not tell you "what is the correct weight"; it tells you only "how fragile is the ranking produced with the current weights."

Thus instead of writing:

"The analysis proved the ranking is correct"

the report should read:

"The ranking produced with these weights is stable only within a narrow interval on this criterion; if this criterion's weight is contested, the ranking is contested too"

Weight sensitivity analysis is built on top of a base method that works with crisp data: a decision matrix, criterion directions and initial weights are required; the base method to be used (such as TOPSIS or SAW) must also be specified, since the stability intervals depend on it. The method does not itself produce weights and does not require an external weight source either; it tests weights that already exist. DecisionMind currently has no completed, selectable implementation of this method: since the system has not yet defined a general contract for which base method is tested against which rule set, it does not offer this method to users, and when a request comes in it states plainly "this method is not currently available"; it does not silently produce a wrong or incomplete result. The computational logic described on this card follows the definition in the literature and was computed separately for illustrative purposes by this card's author.

This analysis is appropriate once you have a result from a ranking method and want to see how robust the weights are before using that result as the basis for a decision or report. It is a particularly critical checkpoint when the weights are the average of several people's opinions, or an estimate.

If the weights are already very clear and uncontested (for instance, coming directly from a legal regulation), this analysis may add no further information. Since the method does not itself produce a ranking, it cannot be used alone when the ranking itself is what is needed; it always works alongside a base method.

A ranking already exists, the robustness of the weights needs testing → weight sensitivity analysis

The weights are clear and uncontested → an additional sensitivity test may not be needed

The real question is "which alternative is first" → first use a base ranking method (TOPSIS, SAW, VIKOR)

The question is how the weights themselves should be determined → AHP, BWM, SWARA (subjective); Entropy, CRITIC (objective)

This analysis's most important contribution is that it stops a ranking from being presented as "the one true result"; it shows the decision-maker which criterion needs more caution. Stability intervals tie a contested weight decision to an objective threshold: the sentence "if the weight falls below this value, the ranking changes" is stronger than an abstract argument. The method is designed to work alongside any base ranking method.

The analysis tracks only the change between the top two alternatives; a swap between the third- and fourth-ranked alternatives may go unseen by this measure. Stability intervals depend on the base method; the same data can give different stability profiles under TOPSIS and SAW, and this difference comes not from the sensitivity method itself but from the base method's mathematics (Wolters and Mareschal, 1995). Redistributing the other weights proportionally while one criterion's weight changes is an assumption; if the decision-maker wants to change weights by another logic, the intervals lose their validity. DecisionMind does not currently offer this method as a product; the computation described on this card is an illustrative demonstration of how the method should work.

The most common mistake is reading a wide stability interval as "the result is certainly correct"; the interval only shows the region in which the top two positions are preserved, and proves nothing about the scientific correctness of the result. A second mistake is identifying the most critical criterion and then closing the report without opening its weight to discussion. A third is assuming that a different base method would produce the same stability intervals; the intervals are specific to the base method. A fourth is assuming that the weight change will occur in only one criterion at a time and never testing realistic scenarios where several criteria change together.

The governing principle is this:

A ranking is valid only for the weights it was given; if a debate over weight continues on a criterion with a narrow stability interval, the ranking is just as contested.