How was this data collected?
Fatal crash coroner's reports. Autopsy results, and blood samples taken from accused drivers and used in criminal prosecutions. Incredibly accurate data. The US actually had compiled the data from ALL crashes, so it wasn't a sample but a total population measurement.
It turns out that the accurate, comprehensive data from fatal crashes is good enough to draw conclusions about the cost of drunk driving in general, since we know the total number of drivers in the US.
The estimated cost is a LOWER bound, because it excludes harm done in cases that did not result in a fatality. So, cases where someone was merely paralyzed were excluded.
Also, undetected drunk surviving drivers were implicitly included in the sober category, which minimizes the difference between drunk and sober making the estimator a lower bound for a second reason.
Thus the estimated cost of driving drunk is very clearly a lower bound, the true cost, and true risk, is certainly higher than estimated.
Are you telling me they know each person's BAC when they had an accident after having one drink?
Yes, in the data they have the exact BAC's.
The study broke drivers down into three categories: drunk over the legal limit, under the legal limit, and sober.
Drunk over the limit drivers were 13 times more likely to cause fatal crashes than sober drivers. Drunk under the limit drivers were 7 times more likely. This is enough information to attribute the total number of fatalities to each category, and with some binomial assumptions, the fraction of drivers on the road who are in each category.
By computing the share of fatalities caused by each type of driver and combining that with standard values on human life used in insurance calculations, they determined that a driver found drunk by police at a roadside stop has done a average 10k damage to society (8k, adjusted for inflation).
That is in the ballpark of the value of the average car, and since it is clearly a lower bound, time to start seizing cars.