Bayesian BWM (Mohammadi and Rezaei, 2020)

Three things change; the best-to-others and others-to-worst comparison logic does not.

Cells and the number of decision-makers. Crisp BWM expects a single best/worst choice and two comparison vectors from one decision-maker. Here multiple decision-makers (k of them) evaluate the same criteria, each with their own best/worst choice and their own comparison vectors. Every decision-maker's figures are, as in crisp BWM, whole numbers on a 1–9 scale; what changes is how many people they come from.

Combining the weights. Where crisp BWM has more than one decision-maker, the simplest route is to solve each decision-maker's weights separately and then average them. Bayesian BWM does not do this. Instead, it builds a three-layer hierarchy that assumes the decision-makers' weights are drawn from a shared "group" distribution. Each decision-maker's weight vector is distributed around a group mean, and that group mean is in turn distributed around a higher-level reference distribution. The best/worst comparisons the decision-makers give are treated as observations feeding this hierarchy. The model finds the probability distribution (the posterior) that best fits these observations by sampling (adaptive block Metropolis). DecisionMind runs this sampling with a fixed starting seed and convergence checks (R-hat, effective sample size, acceptance rate); if any of these checks fails, it stops before producing a result.

Outcome and defuzzification. Crisp BWM derives a single weight vector and one consistency indicator (ξ*) from a single linear model. Here, three things are derived from the thousands of possible weight vectors the sampling produces: each criterion's average weight (the posterior mean), a 95 per cent confidence interval for each criterion, and the probability that one criterion's weight exceeds another's (a credal, or probabilistic, ranking matrix). Crisp BWM's ξ* consistency indicator does not exist here; its place is taken by the sampling's convergence checks. These answer a different question: ξ* asks "are the comparisons consistent with one another," while the convergence checks ask "is this numerical solution reliable."

DecisionMind fixes, for this extension, the mathematical form of the hierarchy (the Dirichlet-Gamma hierarchy published by Mohammadi and Rezaei) and the confidence threshold used for the probabilistic ranking (95 per cent, adjustable by the user).

The average weight, as in crisp BWM, shows a criterion's relative importance against the others within this particular set of comparisons. Here, though, it is not a single number but the summary of a distribution. The confidence interval shows how wide that distribution is; the probabilistic ranking table shows how robust the superiority between two criteria actually is.

This distinction matters: even where two criteria's average weights appear in a given order, the probability of that order can fall below the declared threshold (95 per cent, say). In that case the averages may suggest an order, but that order cannot be treated as statistically settled.

Thus instead of writing:

"The posterior means put C2 ahead of C3, so C2 is more important"

the report should read:

"The posterior means put C2 ahead, but if the probability of this superiority falls below the declared 95 per cent threshold, the order cannot be treated as statistically settled"

Use Bayesian BWM when several decision-makers will weight the criteria, and you want any disagreement between them reported openly rather than masked by a simple average. With a single decision-maker this hierarchy contributes nothing; crisp BWM is sufficient and runs faster.

Crisp BWM's exit condition applies here too: decision-makers must be able to choose the best and worst criterion with confidence. In addition, if the sampling fails to converge (R-hat above threshold, insufficient effective sample size), DecisionMind stops before producing a result; the number of iterations should then be increased, or the decision-makers' comparisons reviewed.

Confusing the order of posterior means with the probabilistic ranking. Even where two criteria's average weights appear in a given order, the probability of that order can fall below the threshold; declaring a "definite" order from the averages alone is wrong.

Leaving the sample size too small. Too few iterations produce an unstable posterior; DecisionMind catches this with its R-hat and effective-sample-size checks, but that protection weakens if the threshold is set too loosely.

Letting one severely inconsistent decision-maker dominate the group. If one decision-maker's comparisons conflict sharply with the others (the kind that would give a high ξ* in crisp BWM), that decision-maker can single-handedly pull the group posterior; decision-maker reliability should be assessed separately.

Presenting only the average without reporting the confidence interval at all. This hides the method's main contribution, which is keeping uncertainty visible.

The governing principle is this:

Bayesian BWM weights summarise a distribution that arises from multiple decision-makers' comparisons. The average may suggest an order; how robust that order actually is can only be read from the probabilistic ranking table and the declared confidence threshold.