EVIDENCE LIBRARY
Can a Bayesian BAC Reconstruction Overcome an Uncertain Drinking History?
A forensic case report used population pharmacokinetics and bootstrap predictions, but uncertain inputs and extreme parameter estimates limited what its probabilities could establish.
No. A Bayesian alcohol model can express uncertainty within its assumptions, but it cannot make an uncertain drinking history true. A 2017 forensic case report combined a later breath result, a reported drinking pattern and a population model to estimate earlier concentrations. The authors also reported extreme fitted parameters and a mismatch between the observed result and the model’s prediction interval.
The case illustrates why a numerical probability needs an explanation of the evidence and assumptions that generated it. It is not a direct measurement of BAC while driving.
What evidence entered the model
Lim, Soung and Bae analyzed one defendant’s case at a prosecutor’s request. The reported history was four cups of soju over approximately 30 minutes before driving. A breath test approximately one hour and 24 minutes after drinking ended produced a reported blood-equivalent alcohol value of 0.173%.
The model had previously been developed from 178 breath-derived observations in 24 healthy Korean men. Those were repeated observations from 24 people, not 178 independent participants. The case result was also breath-derived; the paper did not present a directly assayed blood concentration for the relevant measurement.
The authors used a one-compartment model with absorption and saturable elimination. They updated its population information using the defendant’s reported history, body weight and later test result.
What the bootstrap added
The investigators generated 1,000 resampled versions of the original dataset and repeated the model-fitting and individual-estimation process. They then summarized predicted concentrations and prediction intervals over time.
At 30 minutes after the onset of drinking, for example, the table reported a median predicted concentration of approximately 0.088%, with a 95% prediction interval of approximately 0.036%–0.103%. Ninety-three percent of the modeled replicates were at or above the paper’s comparison value of 0.05% at that time.
That percentage describes the outputs of this modeling procedure. It is not an independently validated 93% probability of guilt, proof of impairment or a universal error rate for Bayesian BAC reconstruction. The paper’s historical comparison value is not a statement of current law.
Important warning signs remained
The authors reported that some model fits terminated or produced unrealistic individual parameter estimates; those results were retained. The individual elimination parameters also differed markedly from the population values, with extremely broad intervals.
They further stated that the observed breath-derived concentration exceeded the upper limit of the model’s 95% prediction interval. That mismatch matters: a probability calculation does not become conclusive simply because many simulations were run.
The authors suggested that underreporting of alcohol could explain the unusual estimates. That was a possible explanation, not independent proof that the defendant’s history was false. Model inadequacy, limited calibration information and uncertainty in inputs also need consideration.
A source-unit problem cannot be repaired silently
The paper describes its assumed dose as 32 mg while also describing four 40 mL cups of 20% soju. Those descriptions are not compatible as written. A reader should not quietly replace the printed unit and then present the reconstruction as fully reproducible.
The issue does not show that the entire modeling approach is unusable. It does mean that the actual dose implementation and units require clarification before reproducing or adopting its numerical case conclusions.
What a DUI review should ask
Identify which inputs were measured, which were reported and which were assumed. Determine whether uncertainty in drinking quantity, timing, absorption, instrument conversion and model structure was actually included. Resampling the development dataset does not automatically encompass every one of those uncertainties.
The site’s retrograde BAC guide explains the distinction between a later measurement and an earlier estimate. The calculation-uncertainty analysis addresses why uncertainty depends on the scenario rather than one fixed percentage.
This single case report did not validate prediction accuracy across independent defendants with known earlier concentrations. It supports careful examination of model inputs, fit and sensitivity; it does not allow a modeled distribution to replace missing historical facts.
Primary source
Lim HS, Soung JH, Bae KS. Forensic science meets clinical pharmacology: pharmacokinetic model based estimation of alcohol concentration of a defendant as requested by a local prosecutor’s office. Translational and Clinical Pharmacology. 2017;25:5–9. https://doi.org/10.12793/tcp.2017.25.1.5.