What did this study find? Using Bayesian inverse probability simulation, this study estimates the population-level risk that a legally permissible interrogation tactic contributes to a false confession wrongful conviction (FCWC). Across a wide range of empirically plausible assumptions, the median estimated risk clusters near 1%.
This stands in contrast to figures of 8–25% often cited in the literature, which represent the proportion of known wrongful convictions that involved false confessions—a fundamentally different quantity that does not account for the vast majority of interrogations that produce no false confession.
Results below are shown for three values of ρ (rho), which represents the share of FCWCs attributable to police conduct—as opposed to false confessions that are voluntary or made to cover for someone else. At ρ = 1.0, all FCWCs are assumed to result from police tactics; at ρ = 0.75 and ρ = 0.5, only 75% and 50% are, respectively.
Posterior Distributions of FCWC Risk
Why Prior Estimates Are Misleading
| Source | Estimate | What It Measures | Key Limitation |
|---|---|---|---|
| Bedau & Radelet | 14% | FC | WC | Outcome-selected cases only |
| Connors et al. | 17.9% | FC | WC | Outcome-selected cases only |
| National Registry of Exonerations | ~13% | FC | WC | Outcome-selected cases only |
| Innocence Project | 25% | FC | WC | Outcome-selected cases only |
| Huff et al. / Cassell | 8.4% | FC | WC | Outcome-selected cases only |
| This Study (ρ=1.0) | 1.30% | FCWC | T | Integrates base rates + full denominator |
| This Study (ρ=0.75) | 0.97% | FCWC | T | Integrates base rates + full denominator |
| This Study (ρ=0.50) | 0.65% | FCWC | T | Integrates base rates + full denominator |
FC | WC = false confessions among known wrongful convictions (outcome-selected). FCWC | T = probability a lawful tactic contributes to a false confession wrongful conviction (population-level).
Explore the model. Adjust the assumptions below to see how the estimated risk of a false confession wrongful conviction changes. The calculation applies Bayes' theorem to combine the base rate of FCWCs with how often a given tactic appears in FCWC versus non-FCWC cases.
Set Your Assumptions
How much risk is “acceptable”? It depends on how you weigh the harm of convicting an innocent person against the harm of letting a guilty person go free. The value λ captures this trade-off.
For example: If you believe wrongly convicting one innocent person is as bad as letting 10 guilty people go free, then your λ = 10. This is the famous “Blackstone ratio.” A higher λ means you place greater weight on protecting the innocent—at λ = 20, you’d accept 20 guilty going free to prevent one wrongful conviction.
The maximum acceptable risk is given by: τ(λ) = 1 / (1 + λ)
When police use an interrogation tactic and obtain a confession, what is the chance that the confession is false and leads to a wrongful conviction?
Most existing research answers a different question: among known wrongful convictions, how many involved false confessions? That figure (8–25%) tells us about the composition of failures, not the probability that a given interrogation tactic produces one. The relevant denominator is not wrongful convictions alone, but all interrogations—including the vast majority that produce accurate confessions, no confession, or no charge.
A medical analogy illustrates the importance of base rates. Consider a diagnostic test with 90% sensitivity and 95% specificity applied to a condition affecting 1% of the population. In a cohort of 10,000 people:
• 100 have the condition; the test correctly identifies 90 of them.
• 9,900 do not; the test falsely flags 495 of them.
• Of 585 total positives, only 90 are true cases—just 15.4%.
Even a highly accurate test produces mostly false positives when the condition is rare. The same arithmetic applies to interrogation tactics: if FCWCs are infrequent, even assumptions positing relatively strong associations between tactics and FCWCs will yield modest posterior probabilities.
Many policy arguments about interrogation tactics rely on case compilations selected because they contain a false confession. This is a classic selection bias: only failures are observed, while the far larger pool of interrogations using the same tactics without producing false confessions remains unseen.
Analogous problems are well documented across disciplines: police narcotics dogs may appear highly successful based on hit rates yet coexist with false-positive rates approaching 85%. Credit scoring models appear accurate because outcomes are observed only for approved applicants. Medical screening accuracy is overstated when only positive results are followed up.
Without incorporating both successes and failures, risk cannot be meaningfully estimated.
The posterior probability that the presence of a tactic is associated with an FCWC is:
Where P(T | FCWC) = sensitivity, P(T | ¬FCWC) = 1 − specificity, P(FCWC) = base rate × ρ
Sensitivity (P(T|FCWC)): How often the tactic appears in FCWC cases.
Specificity (P(¬T|¬FCWC)): How often the tactic is absent in non-FCWC cases.
Base rate: The prevalence of FCWCs, estimated at ~1.29% (posterior median).
ρ: The fraction of FCWCs attributable to police conduct (50%, 75%, or 100%).
In practice, interrogations rarely rely on a single technique. Observational research indicates that interrogators use an average of approximately six tactics per session (Leo, 1996). The multi-tactic model extends the single-tactic estimate using a logarithmic adjustment:
Where k = number of tactics, α is calibrated so that ~6 tactics approximately doubles the single-tactic risk.
A logarithmic formulation allows for diminishing marginal effects: as additional techniques are introduced, each contributes incrementally less to overall risk. This avoids implausible risk accumulation unsupported by existing evidence while still acknowledging the potential for additive effects.