A narrow confidence interval for a vitreous-to-blood ethanol regression slope does not mean that blood ethanol can be predicted precisely for one person. Pounder and Kuroda demonstrated this distinction in 345 autopsies: the fitted relationship was statistically well defined, but individual prediction intervals remained wide.
How the dataset was assembled
The investigators retrieved 349 medicolegal cases with paired vitreous and blood ethanol measurements from July 1988 through December 1992. Blood was typically, but not universally, femoral. Vitreous was obtained by direct needle puncture. Blood tubes contained potassium fluoride and potassium oxalate; vitreous did not. Samples were held at 4°C and analyzed within one week.
Headspace GC–FID measurements were performed in duplicate, and their means entered the analysis. The article describes the instrument and operating conditions but does not supply a complete numerical validation dataset for analytical limits, precision and recovery. Case-specific death-to-collection intervals were not provided.
Outlier review changed the final sample
Eight cases initially lay outside the 95% prediction interval, all with blood ethanol above vitreous. Four trauma cases were excluded because blood contamination was considered probable. The other four, considered potentially associated with death during absorption, remained in the final 345-case model.
This distinction matters: 345 was the fitted dataset, while 349 was the retrieved series. The exclusions were interpretive decisions after record review, not direct experimental proof of contamination. The model should not be transferred without qualification to a new specimen already suspected of contamination.
A precise slope and a wide prediction can coexist
The fitted equation was blood ethanol = 3.03 + 0.852 × vitreous ethanol, with concentrations expressed in mg%, meaning mg/100 mL. The slope’s standard error was 0.0098 and its 95% confidence interval 0.833–0.871. The intercept’s confidence interval, −1.57 to 7.63, included zero.
Those intervals concern the fitted relationship. They do not describe the scatter of individual blood concentrations around it. Table 2 makes the practical difference explicit: vitreous ethanol of 90 mg/100 mL gave a predicted mean blood concentration of 80 mg/100 mL, but a 95% individual prediction interval of 29–131 mg/100 mL and a 99% interval of 13–147 mg/100 mL.
At vitreous 150 mg/100 mL, the predicted mean was 131 mg/100 mL, with a 95% prediction interval of 80–182 mg/100 mL. Reporting only 131 would omit the much wider uncertainty relevant to an individual.
What the statistical limits assume
The prediction intervals belong to this fitted model and its sampling population. They are not guarantees for every decomposed, contaminated or differently collected specimen. Residual normality testing detected departures from normality; the authors considered the departure small enough for their analysis. That judgment should remain part of the interpretation.
The researchers also reanalyzed raw data from seven previous publications and found that equations were often different. Their central conclusion was that individual prediction uncertainty, rather than merely the precision of a conversion factor, limits quantitative usefulness.
Corroboration and reconstruction are different tasks
Paired vitreous testing can help examine an already measured blood result. Predicting a missing blood concentration requires an additional model and its individual prediction interval. Neither procedure independently reconstructs the concentration at an earlier driving event or proves the origin of ethanol.
This study did not establish impairment or cause of death from vitreous alone. Our discussion of individual blood–vitreous disagreement addresses the separate observation that either matrix can have the higher result; the present study explains how uncertainty must accompany a numerical prediction.
Primary source
Pounder DJ, Kuroda N. Vitreous alcohol is of limited value in predicting blood alcohol. Forensic Science International. 1994;65:73–80. doi:10.1016/0379-0738(94)90262-3.
Leave a Reply