I conveyed these questions to Cook this week over e-mail, and he responded in detail. If you've been jonesing to wonk out on alcohol consumption measurement methodology, today's your lucky day.
So the most common reaction to the chart was shock. People said "No way any person can average ten drinks per day and still function in society, much less that 24 million adults are averaging that." What do you say to that?
In response to the evident interest I have gone back to the original data (NESARC, 2001-2) and re-estimated everything. I was able to recreate the estimates that I published, but I’ve decided that I should make a couple of changes in the assumptions. First, out of the 42,000 respondents, a handful are represented as drinking more than is humanly possible, perhaps because of data entry errors. I have now capped the high responses at 10 ounces of ethanol per day. Also, I now use the assumption that the average ounces of ethanol per drink is 0.6 ounces. In any event, in my new calculations, the top 10 percent still average 8.7 drinks per day, which is close to the original, and strikingly high.
Working with survey data on drinking is a tricky business, and certainly do not provide precise answers. But the general picture seems correct to me. Read on.
Many readers were interested in the methodology - can you describe how you did the calculations to arrive at those numbers, starting with the raw data? Feel free to get as wonky/detailed as you want!
The statistics are computed from the 2001-2 (Wave 1) of NESARC, a large survey of the US adult population concerning their drinking and other related matters. The data are available to the public. The downloadable file includes an item based on responses to a number of detailed questions concerning whether the respondent has had a drink in the previous year and if so how frequently and in what quantity.
That item gives an estimate of the average number of ounces of ethanol per day that the respondent has consumed in the previous year. That number can be converted in various ways. I converted to a number of drinks under the assumption that a standard drink averages 0.6 ounce of ethanol (e.g., one 12 ounce beer that is 5% ethanol, or 1.5 ounces of liquor that is 40% ethanol, or 5 ounces of wine that is 12% alcohol).
I “corrected” the result by a constant multiple that is based on comparing the average ethanol consumption reported by survey respondents with the per capita ethanol sales in the United States. That average is 2.2 gallons per year, which is more than twice the average computed from the NESARC survey. (NESARC is not unusually bad in that respect – typically national surveys of drinking “capture” only 40-60% of sales, presumably because respondents are understating their true consumption levels).
Equipped with the corrected statistic on the number of drinks consumed by each respondent per month (say), I then order them from lowest to highest and group them into deciles. Since there are 42,000 respondents, the first decile consists of 4,200 respondents – all of whom said that they abstained the previous year. I can then determine the percentile values, and the averages for deciles.
It should be noted that the NESARC sample is by design not perfectly representative of the adult public – it oversamples some groups. For that reason I do all computations of percentiles and averages using the sampling weight that is included in the downloadable dataset. This is standard practice in working with survey data.
As mentioned above, the item I used (ounces of ethanol consumed per day) is computed by the Census Bureau from responses to several items on drinking. Of the 42,000 entries there are a handful that are too high to be humanly possible. If I include these they distort the importance of the top decile. They could be dropped from the calculations, but I chose to keep them in and assign them a somewhat arbitrary value (10 ounces per day) that is at least feasible. If instead I had simply dropped them, the results would not be much changed.
Is there other research out there showing similar levels of consumption?
Every good study of drinking patterns across a population has found the same general result – the distribution is “skewed” to the right with a long tail, and indeed looks something like a log normal distribution (excluding the abstainers). That fact was first discovered by the French sociologist Sully Ledermann in the 1950s. The result is that total alcohol consumption for the population is quite concentrated at the high end. While there are always problems with survey data, I think the evidence in support of these general facts about the distribution is very strong.
I would also point to the 20/80 rule (which is not really a “rule” but is a pretty good guide to a number of distributions). Note that the top 10% of adults constitute the top 15% of drinkers, since one third abstain. In my calculations, these 15% consume 75% of the total. That may be a bit too high, and would drop somewhat under different assumption about the top reported values. The point is that these calculations appear to be in the right ballpark.
Surveys like NSDUH show adult alcohol dependency rates somewhere in the level of 5-6% of the adult population. How do you square that with your numbers showing 10 percent of the population averaging 10 drinks per day?
There is a wide range of alcohol consumption within the top decile. The 90th percentile is higher than recommended by medical experts, but not nearly as high as the usual range of those who are active alcoholics. They typically show up still higher in the distribution, where people may be consuming near to the biological maximum.
The numbers in Paying the Tab are from the 2001-2002 NESARC, correct? Is there any reason to think that alcohol use rates may have changed since then?
The per capita sales of ethanol have changed very little since then, so I would be amazed if other aspects of the distribution have changed much. Incidentally, ethanol sales started to drift down in 2008 with the advent of the Great Recession, but very recently have turned up again – a sign of the improving economy.
Would you want the average member to the general public to take away from that chart? What about a policymaker?
I believe that the interest in the chart comes from a natural interest for drinkers to know how they compare with the rest of the population – partly out of a concern that they might be drinking “too much.” I want to remind everyone that average consumption per day or per month is not a sufficient statistic for determining problem drinking. More relevant is to compare the “dose” with body weight. And even adjusting for weight, there is some evidence that a particular dose may have a greater effect on women than men.
The drinking pattern over time also matters. Certainly the effect of having one drink per day is quite different than abstaining 6 days per week and then bingeing on 7 drinks – even though the average will still work out to one drink per day.
Despite these and many other provisos, I would say that per capita consumption is still a very relevant statistic for understanding alcohol as a population-level problem – in part because per capita consumption is so closely tied to how much the heaviest drinkers are drinking, and thus how many are vulnerable to the organ damage that heavy drinking causes.
Anything else you'd like to add?
I of course would be delighted to have people read my book Paying the Tab, which gives the whole story!