Median Ranking

If you hold several rankings already produced for the same alternatives, say from three separate jury members or three different decision methods, median ranking turns them into one common order. Its output is a single number for every alternative: the median of that alternative's positions across all the rankings. The smaller this number, the better the alternative is judged. The method produces no new score and asks for no criterion weights; it only merges rankings that already exist. Its source is Orakçı's (2024) Turkish-language book on aggregation techniques, which places this technique alongside classical average-based aggregation as a more robust alternative against outlying opinions.

The idea comes from a familiar statistical principle. Do you measure a class's performance by its mean mark or its median mark? If one student scores zero on an exam, the mean drops sharply, while the median barely moves. Median ranking carries the same logic into rank aggregation. Even if one source, a jury member or a method, rates an alternative extremely poorly or extremely well, this single extreme opinion does not pull the median; the shared view of the other sources prevails.

This carries a philosophical consequence: median ranking seeks robustness, not consensus. An average-based aggregation, such as Borda, weighs every source's opinion equally and sums them, whereas median ranking finds the "middle" view and disregards the outlying opinions at the edges. If one source consistently thinks differently from the rest, for instance a jury member who always gives outlying scores, the median automatically dampens that anomaly.

The method proceeds through a single step.

One step: finding the median position. For each alternative, the positions it holds across all the rankings are laid out as a list. Suppose an alternative comes first, second, third and third across four separate rankings; the median of these four numbers is taken. Where the count is even, the average of the two middle values is used; where it is odd, the middle value is taken directly. This is done separately for every alternative, giving each a single median value. Alternatives are then ordered from the smallest median to the largest; the smallest median marks the best alternative. If two alternatives' medians tie, the method draws no distinction on its own; a further rule is then needed, such as looking at the mean rank or adding one more source.

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

The median value shows where an alternative typically sits across the different rankings; it says nothing more. A value such as 2.5 is not "2.5 per cent" or a probability; it means only that "the middle of this alternative's positions lies between 2 and 3." The value itself depends on the number of sources and the rank range they use, and cannot be directly compared with a median coming from a different set of rankings.

Two alternatives tied on median are not "identical"; they are simply indistinguishable by this measure. The report should not hide this, and should state clearly which further rule was applied, be it alphabetical order, the mean rank, or the addition of another source. The median also does not show the size of any disagreement between sources: whether two of four sources place an alternative first and two place it last, or two sources consistently place it second, a similar median can result; yet the first case reflects a serious difference of opinion among the sources.

Thus instead of writing:

"Median ranking found the best alternative"

the report should read:

"Given this set of sources, the most robust middle position by majority opinion among the alternatives is this one; how much agreement exists among the sources should be examined separately"

Median ranking works with a crisp rank matrix: alternatives in rows, different ranking sources (experts, votes, or the outputs of different methods) in columns, and an integer rank position from 1 to m in every cell, with 1 marking the best position. Every source must have ranked all the alternatives; empty cells are not accepted. The method produces no weights and asks for none; every source counts equally. DecisionMind currently holds no further extension alongside this base method. A minimum of two alternatives and two separate ranking sources is required, and there is no practical upper limit on the number of sources.

Median ranking is suitable if you already hold several rankings and want to collapse them robustly into a single order that resists outlying opinions. Its advantage shrinks with only two or three sources, because the median behaves almost like the mean with so few values. Where a cyclical contradiction exists among the sources, for instance A beats B, B beats C and C beats A, median ranking cannot see this cycle and gives a one-dimensional answer to a problem that requires pairwise comparison.

Combine several rankings, be robust to outliers → Median Ranking

Few sources, preserve fine differences → Average-based aggregation (such as Borda)

Seek a consistent winner via pairwise comparisons, resolve cycles → Schulze, Nanson, Kemeny-Young

Combining raw scores/data rather than rankings → Weighted-average-based aggregation operators

Median ranking's greatest strength is its simplicity: it can be computed by hand and requires no parameter at all. An outlying source, a jury member who thinks very differently or a method that has malfunctioned, is prevented from distorting the overall result, because the median is unaffected by extreme values. It is easy to interpret: an alternative's median directly shows its typical position across the sources.

Its limitations also follow from its simplicity. First, ties are common; where the number of sources or alternatives is small, several alternatives may share the same median, and the method cannot separate them on its own. Second, the median gives no guarantee of pairwise consistency; an alternative preferred by a majority of the sources (a Condorcet winner) may not come out first under median ranking (Young, 1988). Third, because the median reduces everything to a single figure, it hides the size of any disagreement among the sources; this kind of information loss in rank-aggregation methods generally has been discussed in a separate literature (Dwork, Kumar, Naor and Sivakumar, 2001).

The most common mistake is computing the median without first filling in missing data, when a source has not ranked every alternative; the median then comes out wrong. A second mistake is confusing the median with the mean, reading "median 2.5" as "mean rank 2.5"; the two statistics carry different information. A third is declaring alternatives with equal medians "identical" without applying any further rule; which further rule was applied should be reported. A fourth is hiding a serious disagreement among sources behind the single median figure, when the sources' underlying views differ sharply.

The governing principle is this:

Median ranking finds the sources' shared middle opinion; in doing so it does not show the size of any disagreement among sources, so the report must separately answer how much the sources agreed.