On the consequences of taxing alcohol

The Open Philanthropy Project just released another big literature review of mine, this one on whether taxing alcohol save lives. I conclude that it probably does.

That’s hardly shocking. Making stuff more expensive generally leads to people to buy less. And alcohol in excess is bad for you. Perhaps the more significant finding is about the number of lives that could be saved, which is not so great next to other things that “Open Phil” might fund. E.g., we dream of financing the invention of a new research technique that leads to a cure for Alzheimer’s. The grant chasing that dream would be a longshot…but then so might be funding advocacy for raising taxes.

Coroners in the US attribute 23,000 deaths/year to alcohol-caused diseases, according to my calculation (see the report for more). The most rigorous studies I found produced a rather wide range of elasticities of death rates with respect to alcohol prices: 1–3. That means each 1% price rise reduces deaths 1–3%. And, if you do the math carefully, tax hikes sufficient to raise alcohol prices 10% would cut the alcohol death rate 9–25%, or 2,000–6,000 lives/year. This math leaves out any reduction in deaths from drunk driving, which currently amount to 10,000/year. The benefit there would presumably be of the same order of magnitude.

A few interesting things I learned and did along the way:

  • The United States has made great progress since the 1980s reducing drinking and its side effects, including deaths from alcohol-related diseases and drunk driving. This progress appears to stem from a movement spearheaded by groups such as Mothers Against Drunk Driving, which led to many policy changes: a higher drinking age, criminalization of drunk driving per se (regardless of whether harm is done), and more. Cultural and demographic changes may have mattered at least as much. Meanwhile, ironically, alcohol taxes fell steadily because of inflation.
  • Alcohol prices and consumption are perhaps most clearly linked, among bad consequences, to cirrhosis of the liver. At the bottom I’ve pasted a 55-year-old graph that speaks eloquently to this point. Although cirrhosis is a chronic disease, because it is progressive, at any given time a pool of cirrhotics is, coarsely speaking, one binge away from death. As a result, the death rate from this long-term disease is surprisingly sensitive in the short-term to changes in taxes and probably other alcohol control policies.
  • Alcohol taxes change rarely, which is good for research. Discontinuities help support credible analysis of short term impacts. If I tell you taxes rose in April and deaths fell in May, that can demonstrate causality much more convincingly than if I tell you taxes rose in 2000 and deaths fell in 2010.
  • Sudden policy changes such as tax hikes produce what statisticians call interrupted time series. To study them, you produce a model for the normal evolution of the time series and then add a factor that allow for a “structural break” at the moment of discontinuity in trend such as drunk driving fatalities.
  • A popular modeling family for (interrupted) time series is ARIMA, which has a lot of moving parts, as I explain in the report. (Perhaps too many: 30 years ago a couple of professors had 12 trained grad students fit ARIMA models to computer-generated data sets for which the professors knew the true model. The students got it right 28% of the time.)
  • I worried that interrupted time series studies I read looked for sharp death rate drops only where they wanted to find them: right after a tax increase. Taking inspiration from the pedagogic guidance of Imbens and Lemieux on regression discontinuity analysis, I thought one ought to look for “jumps at points where there should be no jumps” as a falsification test. If a study’s method for sensing a sudden change in an outcome right after the tax also picks up a significant change a year before and a year after, it becomes less certain that tax impacts are driving the study’s results.
  • Partly for this reason, I replicated many of the studies I reviewed—more than I ever have before, and more than I’ve seen done in a review. I toyed with calling it a “replication review.” Data, code, and spreadsheets are here.
  • I came to favor studies that look at jumps in many states at once, and model the time series in plainer ways—panel studies.
  • I learned about an imperfect but intriguing research method called Mendelian randomization. As a result, I came to question the proposition that moderate drinking is good for you. That at least partially rebuts one argument against the health benefits of alcohol taxes.
  • I learned that you can download data on every death in the US from 1959 to 2004, which I did, importing it all into a 20 gigabyte MariaDB database. Line 1808433 of the 1963 file reads:


    Translation: Nov. 22, 1963, white male, 46, died in Dallas County, TX, resided in District of Columbia. Cause of death: assault by firearm/explosive.

The graph I promised:

Alcohol price and consumption and liver cirrhosis death rate, Ontario, 1929-58, Seeley 1960