Does a Six-Hour Interrogation Carry a 10% Risk of Wrongful Conviction? A Reply to Smith and Colleagues, Forthcoming in the Journal of Criminal Justice

Status: This paper is forthcoming in the Journal of Criminal Justice. The DOI will be added here once it is published. All data and code are already available at github.com/smourtgos/fcwc-duration-reply.

The exchange

In 2026 we published an analysis in the Journal of Criminal Justice arguing that the false confession literature cannot currently estimate the population risk that a lawful interrogation tactic contributes to a wrongful conviction (Mourtgos & Adams, 2026). The obstacle is not the quality of the research. It is that studies assembled from known false confession cases cannot supply a denominator: the much larger population of interrogations that did not produce a false confession wrongful conviction (FCWC).

Smith and colleagues (2026) responded — and did something unusual and genuinely valuable. They independently reconstructed our analysis and reproduced its results. Replications of this kind remain rare in criminology, and we welcome the effort. Their replication also confirms the central point of the original article: the existing literature does not contain the empirical likelihoods needed to estimate FCWC risk under different interrogation tactics.

They then asked a different question. Rather than conditioning on tactics, they conditioned on interrogation duration, and reported FCWC risk of roughly 1% at one to two hours, 10% at about six hours, and 25% at about twelve. Those are specific, policy-relevant numbers. Our reply asks whether the available evidence supports them.

What we did

We recalculated Smith et al.’s duration model using their own study inputs and model functions, taken from their deposited code, and retaining their lognormal distributions and 0–48-hour truncation. Unless otherwise noted, we fixed the FCWC base rate at .019 so that only the duration inputs changed. We then varied, one at a time, the choices the model requires an analyst to make: which false confession study supplies the FCWC duration distribution, which studies form the comparison group, how those components are weighted, and how much spread the comparison distribution is assumed to have.

What we found

The reported thresholds are not stable estimates of population risk.

• Changing one study nearly triples the twelve-hour estimate. Smith et al. motivate the duration contrast with Drizin and Leo’s (2004) documented false confession cases (mean 16.3 hours), but the model actually uses Redlich et al. (2011): 35 people with serious mental illness who self-reported a false confession, with a mean total questioning time of 3.07 hours across multiple sittings. At 12.06 hours, the published Redlich-only specification gives 26.3%. Including both sources gives 62.1%. Using Drizin and Leo alone gives 72.6%. Nothing else changes. These estimates do not bracket a true value; they show that the answer is determined substantially by inputs that neither study was designed to estimate.
• A comparison within a single study never reaches 10%. Redlich et al. (2011) allow a more comparable check, because both confession groups came from the same population and answered the same questions. Keeping the false confession curve and replacing the pooled comparison group with Redlich’s own 30 true confessors, estimated risk falls from 10.6% to 3.6% at 6.28 hours and from 26.3% to 4.0% at 12.06 hours, never exceeding about 4.2% across the evaluated range. This is not a corrected population estimate — that sample is small, retrospective, and restricted to persons with serious mental illness. What it shows is that the sharp increase in the published model does not come from the difference observed within that study. It appears when that study’s false confessors are compared against the narrower pooled distribution assembled from other sources.
• The two distributions are not measuring the same thing. The comparison group pools twelve sets of summary statistics from eight studies, weighted by sample size — but the sample sizes count different things. About 62.4% of the weight comes from officer surveys, 16.0% from suspects’ retrospective reports, and 21.6% from recorded interrogations. Kassin et al.’s (2007) n = 601 counts officers estimating their typical interrogation length; Kelly et al.’s (2016) n = 29 counts 29 recorded interrogations. Weighting by sample size gives the officer survey roughly 21 times the influence, although the two count and measure different things. The same issue affects the reported variation: Kassin et al.’s standard deviation describes variation across officers’ estimates, not across individual interrogations, yet the model uses it to determine how often individual interrogations run into the right tail.
• Six hours is not a stable feature of the data. The thresholds sit far out in the tails of fitted curves. Under the published specification only about 0.116% of the fitted comparison distribution lies above 12.06 hours, and 0.0024% above 34 hours — outputs of an assumed curve, not observed frequencies. Widening each comparison component’s log-scale spread by 25% moves the 10% crossing from 6.07 to 14.75 hours; widening it by 50% eliminates the crossing entirely; narrowing it by 25% moves it to 3.93 hours. Composition matters too: 7.20 hours with equal weights, 8.69 hours without the officer surveys, 6.62 hours using recorded interrogations only.
• The threshold is uncertain even within the published specification. Applying Smith et al.’s own bootstrap procedure to both duration likelihoods and carrying the draws through Bayes’ theorem, with the base rate fixed, gives a 95% interval of about 5.7%–14.6% at 6.28 hours and 7.1%–42.3% at 12.06 hours. At the point where the published estimate reaches 10%, uncertainty in the fitted likelihoods places the risk on either side of that threshold.

We also address three further issues the response raises: that completed interrogation duration is partly an outcome rather than pure exposure (a suspect who confesses at hour two and is questioned four hours more is coded as a six-hour interrogation); that shifting the estimand from P(FCWC) to P(FC | WC) answers a different question rather than supplying the missing denominator; and that the problem we identified is not collider bias, which requires a causal structure neither analysis specifies.

What this does not show. These analyses do not establish that prolonged interrogation is safe, or that duration is unrelated to FCWC. Prolonged interrogations may well increase risk. The finding is narrower and about measurement: the available studies do not identify the magnitude of the relationship, and the six-hour threshold is a product of selected inputs and fitted tails rather than a risk boundary established by observed interrogations.

Where we agree

More of this exchange is agreement than disagreement. Smith et al.’s replication is a real contribution, and it confirms our central result. Both papers agree on what the field should do next: code actual interrogations for tactic presence, duration, and outcome as recordings become more widely available. Both agree that even a 1% FCWC rate would be problematic — as the original article put it, a society may judge even low probabilities unacceptable, and such judgments are ethical and political rather than empirical.

The disagreement is narrower than it may appear. It concerns whether the available evidence establishes the magnitude of the risk. The seriousness of an outcome informs how much risk society should tolerate; it cannot establish how frequently that outcome occurs.

Recording alone, though, will not supply the denominator. Researchers would also need to define the population from which interrogations are sampled, distinguish individual sessions from cumulative questioning, record the time of the first admission, link interrogation characteristics to case outcomes, and distinguish a false confession from one that contributed to a wrongful conviction.

Explore the results

Every figure in the dashboard below is computed live in your browser from the published model. Move the duration slider and watch the three source choices diverge; scale the comparison curve’s spread and watch the six-hour threshold move between roughly four and fifteen hours; swap the comparison group and watch it disappear entirely.

Methods summary

• Model: Smith et al.’s (2026) duration model, reusing their study inputs and model functions from their deposited code — mixtures of 0–48-hour truncated lognormal densities for FCWC and non-FCWC interrogations, combined with a base rate through Bayes’ theorem
• Base rate: fixed at .019 throughout, so that only the duration inputs vary. Smith et al.’s published results use base-rate draws and report posterior medians, so their values and ours are not computed the same way
• Sensitivity analyses: alternative FCWC sources (Redlich only, both sources sample-size weighted, Drizin & Leo only); alternative comparison groups (published pooled, equal weights, officer surveys excluded, recorded interrogations only, within-study); log-scale spread scaled by ×0.75 to ×1.5
• Uncertainty: Smith et al.’s parametric-bootstrap likelihood procedure applied to both duration likelihoods, 400 replications, carried through Bayes’ theorem at a fixed base rate
• Reproducibility: one script reproduces every number and figure in the reply; github.com/smourtgos/fcwc-duration-reply

References

Mourtgos, S. M., & Adams, I. T. (forthcoming). Interrogation duration and the estimation of false confession wrongful conviction risk: A reply to Smith and colleagues. Journal of Criminal Justice.

Mourtgos, S. M., & Adams, I. T. (2026). Recalibrating the risk of false confession wrongful convictions: Interrogation tactics and inverse probability. Journal of Criminal Justice, 103, 102600. https://doi.org/10.1016/j.jcrimjus.2026.102600

Smith, T. B., Catlin, M., May, B., Redlich, A. D., Meissner, C. A., & Kelly, C. E. (2026). Simulating the probability of false confession-wrongful conviction conditional on features of police interrogation: A response to Mourtgos and Adams (2026). Journal of Criminal Justice, 107, 102742. https://doi.org/10.1016/j.jcrimjus.2026.102742