Risk-Limiting Audits for Parliamentary Majorities

Instead of checking every single local election result to confirm a winner, this new method checks if a party won enough seats to form a government. By focusing on the overall majority rather than every individual seat, auditors can confirm election results while inspecting thousands of fewer ballots.

In parliamentary democracies, the political stakes are rarely tied to a single local race. Instead, the critical question is whether a party or coalition has secured enough seats to command a majority. Current risk-limiting audits (RLAs)—statistical procedures that sample paper ballots to verify election outcomes—typically treat each seat as an isolated contest. An RLA ensures that if a reported winner actually lost, the audit will detect the error with a high degree of mathematical certainty. This certainty is called the risk limit.

However, applying this seat-by-seat logic to an entire parliament is incredibly inefficient. If an auditor insists on certifying every single winning seat to confirm a national majority, they end up inspecting a massive volume of ballots. Many of these belong to "safe" seats where the margin of victory is overwhelming. This creates a massive logistical bottleneck. The researchers argue that we can achieve the same level of confidence regarding the government's formation without the exhaustive overhead of a full-scale recount for every constituency.

The inefficiency of seat-by-seat certification

The status quo in election auditing assumes that to trust a collective outcome, you must trust every constituent part. In a parliamentary system, this means an auditor would ideally certify every reported winning seat to ensure the total count is correct. As the authors demonstrate, this approach is computationally and logistically expensive.

If an auditor follows the "All seats" benchmark, they are essentially performing a series of independent tests. Even if a party wins a seat by a landslide, the "All seats" method requires continued sampling to meet the statistical threshold for that specific seat. This becomes a significant drain on resources in large-scale elections. The authors illustrate this inefficiency by comparing their method to traditional approaches in simulated environments .

When the goal is simply to verify a majority, treating every seat as a mandatory checkpoint is highly inefficient.

Verifying the majority through partial conjunction

The authors reformulate the problem as a "partial conjunction" hypothesis test. In statistics, a conjunction is a logical "AND" statement. A partial conjunction asks if at least a certain number of conditions are met. Instead of testing if every seat in the winning set $W$ is correct, the authors test the hypothesis that the winning party truly won at least $r$ seats.

The mechanism relies on a specialized test statistic, $E_{r/W,t}$, which the authors build from seat-level "e-processes." An e-process is a mathematical tool used to accumulate evidence sequentially. Think of it like a betting game where you increase your stake as the evidence for a claim grows. To determine if a majority exists, the authors take the product of the smallest $|W| - r + 1$ seat-level statistics [Equation 11]. This mathematical structure ensures that the audit only stops when there is sufficient evidence that enough seats have been won to satisfy the majority requirement.

To make this practical, the paper introduces three adaptive sampling schemes designed to decide which ballots to pull next :

1. Greedy: This strategy identifies the "weakest" seats—those providing the least evidence for a win—and concentrates sampling effort there.
2. Filtered: This uses Bayesian inference to calculate a posterior probability ($\pi_s$) for each seat. If a seat appears to have been falsely reported, the filter redirects effort away from that seat. This avoids wasting ballots on a "lost cause."
3. Greedy Filtered: This combines both. It uses a moving window of the weakest seats while simultaneously filtering out those that look mathematically improbable.

Massive reductions in ballot inspection

The results of the simulations suggest that shifting the focus from individual seats to the parliamentary majority yields dramatic savings in labor. In controlled simulations with 100 seats, the authors report that their majority-based strategies were markedly more efficient than the "All seats" benchmark .

The impact is even more pronounced when applied to real-world data. The authors simulated audits using the 2014 Indian Lok Sabha election. This election involved approximately 282 million ballots across 543 seats. Using the "All seats" baseline, an auditor would need to inspect roughly 250 million ballots to certify the result. In contrast, the authors report that their adaptive schemes required only a few hundred thousand ballots. This represents a reduction of several orders of magnitude .

The researchers found that while the "Greedy" scheme performs well when all reported winners are legitimate, it struggles if some seats are falsely reported. In those cases, it wastes time trying to "fix" a seat that was never truly won. The "Filtered" and "Greedy Filtered" variants proved much more robust. They maintained high efficiency even when the number of incorrectly reported seats ($n_{false}$) increased .

Constraints on deployment

While the mathematical framework is robust, the authors identify several practical hurdles. First, this method is specifically designed for elections where the governing majority is determined by a fixed number of seats. If a country uses a system where governing coalitions are negotiated after the election, the "majority" to be audited is not known upfront. This makes the specific partial conjunction approach inapplicable.

Second, the implementation requires significant cross-seat coordination. Unlike a single-contest audit, this method requires a central authority to aggregate data from many constituencies in real-time. This authority must update the global $E_{r/W,t}$ statistic. This adds a layer of administrative complexity and potential latency to the auditing process. Finally, the study focused on "ballot-polling" audits (manually checking paper slips). The authors note that if digital voter records are available for "comparison audits," the sample sizes could potentially be reduced even further.

Verdict: A scalable solution for large parliaments

The findings suggest that transitioning from seat-level to majority-level auditing is a practical necessity for large-scale verification. By moving away from the "all-or-nothing" requirement of certifying every single seat, the authors provide a way to maintain rigorous statistical guarantees. This approach significantly reduces the physical workload.

For practitioners, the "Greedy Filtered" approach appears to be the most reliable choice. This is especially true in environments where reporting errors might exist. The methodology is ready for integration into existing audit workflows. Authorities could potentially certify a government "midstream" while continuing to perform detailed seat-by-seat recounts in the background. Code for the implementation is reportedly available; see the paper for the canonical link.

Figures from the paper