When clinical trials evaluate medical interventions, they typically compare two or three options directly. However, clinicians in practice often face a crowded marketplace of competing therapies. For instance, deciding which oral antiseptic best prevents ventilator-associated pneumonia or which drug regimen optimizes advanced bladder cancer outcomes requires synthesizing evidence across dozens of disparate trials.
Traditional meta-analyses rely on direct, pairwise comparisons (such as odds ratios or relative risks). When extended to a network of multiple treatments, relying solely on frequentist p-values to rank interventions can become unwieldy and fail to capture the full picture of uncertainty, as noted by Mingfang Yang in Diagnostic value of ultrasound combined with MRI in cholecystolithiasis: a systematic review protocol. If Treatment A beats Treatment B in one trial, and B beats C in another, we cannot simply declare A the definitive winner without quantifying the probability distribution of their relative ranks across the entire network.
The Core Concept: Cumulative Ranking Probabilities
To solve this, Network Meta-Analyses (NMAs) use simulation techniques—often via Markov chain Monte Carlo (MCMC) methods within a Bayesian framework—to estimate the probability that each treatment occupies every possible rank, a methodology highlighted by Wenhua Liang et al. in Network Meta-Analysis of Erlotinib, Gefitinib, Afatinib and Icotinib in Patients with Advanced Non-Small-Cell Lung Cancer Harboring EGFR Mutations and Cheng Xue et al. in Antihypertensive treatments in adult autosomal dominant polycystic kidney disease: network meta-analysis of the randomized controlled trials.
Imagine a network comparing five treatments. For each treatment, the model outputs a probability distribution: the chance it is the 1st best, 2nd best, 3rd best, and so on. While highly informative, looking at a raw matrix of these rank probabilities is cognitively overwhelming. A treatment might have a 40% chance of being 1st, but a 50% chance of being 5th, making its overall utility ambiguous.
This is where cumulative ranking probabilities come in. Instead of looking at individual rank probabilities in isolation, we calculate the cumulative probability that a treatment achieves at least a certain rank, an approach utilized by Rui Wang et al. in Treatment strategies for women with WHO group II anovulation: systematic review and network meta-analysis. For example, we ask: What is the probability that Treatment A is among the top 2 treatments? By plotting these cumulative probabilities across all possible ranks, we create a step-wise cumulative ranking curve.
The Mechanism: Surface Under the Cumulative Ranking Curve (SUCRA)
To convert this curve into an actionable, single-number summary, researchers calculate the Surface Under the Cumulative Ranking Curve (SUCRA), as discussed by Lei Yu-ping et al. in Chinese herbal medicine injections (CHMIs) for chronic pulmonary heart disease and Runting Li et al. in Comparative efficacy of antiangiogenic treatment for newly diagnosed glioblastoma.
SUCRA translates the entire distribution of ranking probabilities into a single percentage between 0% and 100%:
- A SUCRA value of 100% means the treatment is certain to be the absolute best (its cumulative probability is 1 at rank 1).
- A SUCRA value of 0% means the treatment is certain to be the absolute worst.
As Zhigang Zhang et al. explain in Comparison of the effect of oral care with four different antiseptics to prevent ventilator-associated pneumonia in adults: protocol for a network meta-analysis, this metric integrates the area under the cumulative distribution of rankings. It condenses a complex multi-treatment network into a standardized, linear hierarchy, allowing clinicians to instantly see which therapies are most likely to sit at the top of the options pool.
Why It Works: The Intuition of Ranking Surfaces
Why is SUCRA preferred over simply looking at the "probability of being the best" ($P(\text{Rank} = 1)$)?
Consider a scenario where Treatment A has a 45% chance of being the absolute best, but a 55% chance of being the absolute worst (a highly volatile, high-risk intervention). Meanwhile, Treatment B has only a 10% chance of being 1st, but an 80% chance of being 2nd. If a decision-maker looks only at the probability of being number one, Treatment A wins. However, Treatment B is far more stable and consistently effective across simulated scenarios.
SUCRA captures this nuance because it accounts for the entire surface under the curve. It behaves similarly to the Area Under the ROC Curve (AUC) used in machine learning. As shown by Corinna Cortes et al. in Confidence Intervals for the Area Under the ROC Curve, the AUC can be modeled as a Wilcoxon-Mann-Whitney statistic that calculates the probability of correct relative ranking rather than raw score distributions. Just as AUC evaluates a classifier across all possible thresholds, SUCRA evaluates a treatment across all possible ranks. This foundational logic of ranking surfaces also extends to multi-class problems, where metrics like the Volume Under the ROC Surface (VUS) measure the probability that a sequence of items from ordered categories is correctly sorted, as demonstrated by Sylvain Robbiano et al. in Ranking data with ordinal labels: optimality and pairwise aggregation.
Applications and Evolving Metrics
SUCRA and cumulative ranking metrics have become vital tools for resolving clinical uncertainty across diverse medical fields, from evaluating post-surgery cancer therapies (Li Wang et al. in Evaluating the efficacy of post-surgery adjuvant therapies used for ductal carcinoma in situ patients: a network meta-analysis) to ranking exercise regimens for chronic illness (Zhikai Qin et al. in Effects of various exercise therapies on 6-min walk distance in patients with stable chronic obstructive pulmonary disease: a systematic review and network meta-analysis). Furthermore, as Jianjun Zhang demonstrates in Aspirin plus dipyridamole has the highest surface under the cumulative ranking curves (SUCRA) values in terms of mortality, intracranial hemorrhage, and adverse event rate among 7 drug therapies in the treatment of cerebral infarction, SUCRA allows for simultaneous, multi-outcome ranking, helping identify therapies that optimize both efficacy and safety profiles.
The field continues to refine these probabilistic ranking tools to address specific statistical limitations:
- Frequentist P-scores: In frequentist network meta-analyses where Bayesian MCMC simulations are not used, researchers often utilize P-scores. As explained by Jihad Mallat et al. in Comparative efficacy and safety of prostacyclin therapies for pulmonary arterial hypertension: a systematic review and network meta-analysis, P-scores serve as a direct frequentist analogue to SUCRA, measuring the mean probability that a specific intervention is superior to competing options.
- Predictive P-scores: Traditional SUCRA scores only summarize the evidence within the collected historical trial data. To bridge the gap to real-world clinical application, Kristine J. Rosenberger et al. introduced the predictive P-score in Predictive P-score for treatment ranking in Bayesian network meta-analysis. This metric incorporates between-study heterogeneity to estimate how treatments will rank in a future clinical setting, ensuring that clinical decisions are robust to the real-world variability across patient populations.