Alcohol as a control variable

[Omer Gersten]

Hi. I've been a long-time user of Stata and was hoping to get some feedback on a problem I'm having. The question is more methodological, than technical. If there is a better place to post the following question, can somebody please tell me?
Alcohol as a control variable When I was in graduate school I was taught that a moderate level of alcohol consumption is good for you. But now we know that this may not be the case and that any level of alcohol may be toxic. Thus, if alcohol were to be used as a control variable, I would recommend making it a dummy of never/barely drinks and excessive alcohol consumption. According to the NIH, a drink counts as a can of beer, a glass of wine, and a shot of spirits. Also, moderate alcohol counts as up to 1 drink a day for women and 2 drinks a day for men. If I were to design a survey that gathers enough relevant information to be able to control for alcohol consumption adequately I think I would need to ask roughly the following questions: X. Would say that you never or barely drink any alcohol? X. If you drink beer regularly, do you make sure to limit yourself to no more than 1 drink a day if you're a woman and 2 drinks a day if a man? X. If you drink wine regularly, do you make sure to limit yourself to 1 glass a day if a woman and 2 glasses a day if a man? X. If you drink spirits regularly, do you make sure to limit yourself to half a shot a day if a woman and 1 shot a day if a man? I am not certain how to deal with the phenomenon of infrequent, but binge drinking. According to the NIH, binge drinking generally occurs after 4 drinks for a woman, and 5 drinks for a man. So, I might come up with questions like this: X. If you seldom drink beer, but drink a lot when you do, do you make sure to limit yourself to no more than 4 drinks a session if you're a woman and 5 drinks a session if a man? X. If you seldom drink wine, but drink a lot when you do, do you make sure to limit yourself to 4 glasses a session if a woman and 5 glasses a session if a man? X. If you seldom drink spirits, but drink a lot when you do, do you make sure to limit yourself to 4 shots a session if a woman and 5 shots a session if a man? These questions do not get at the phenomenon of drinking a mix of different types of alcohol in the same session. I'm thinking of when someone might go out and drink a mix of shots and beer. It seems hard though to come up with a good question that captures this. Maybe one could ask, "If you drink, do you tend to drink a mix of spirits and beer?" Maybe one could assume that if the respondent answers yes they are probably drinking in excessive amounts. But, let's say, a male goes out and has two shots and two beers, this still wouldn't be considered excessive drinking. X. Instead of the previous 4 questions I mentioned, one could simply ask "When you drink, do you often drink to get drunk?" In the paper I am currently a reviewer for for the Journal of Biosocial Science (JBS), the authors only say that alcohol consumption is coded as "never, sometimes, and frequent". I am perplexed about what to think about this coding. As indicated above, one could be a frequent drinker, but still be considered a moderate drinker. Is it safe to assume that roughly speaking, if someone considers themselves a "frequent" drinker that they drink to excess? My intuition is that this question in this paper is simply not granular enough to be useful. I would recommend that it is simply is not used. Does this make sense? Thanks,
Omer

***************************************

Omer Gersten, Ph.D.

[Brad Anderson]

There is a ton of extant literature using different measures of alcohol use. At one time people were often asked separate questions about consumption of beer, wine, and spirits separately. I don't see that much anymore. Instead, I think these days we more often just see a generic statements such as a drink is defined as .... " I'd say the current gold standard measure is the Time-Line Follow-Back. I believe Sobel and Sobel were the originators. Basically it's a calendar based questionnaire asking persons report the number of drinks they've had on each day for the past X (often 30) days. That would allow you to define frequency of alcohol use, frequency of binge alcohol use, and measures of total alcohol consumption. Except in the context of assessing alcohol abstinence (such as for a treatment study where the desired outcome is abstinence), I'm of the opinion that measures of frequency that don't quantity into consideration are generally of little value. A person who has a glass of wine with dinner every day has a very different risk profile than a person who binges everyday. And when you start looking at people entering treatment, the daily quantities of alcohol consumed can be truly staggering. You could also use something as simple as the first 3 items of the AUDIT.

[Clyde Schechter]

I agree with everything Brad Anderson says. I'll just add that the changing opinions about whether a little alcohol is good for you may represent both methodological issues in different studies and also differences in the outcome looked at. It may be that low-level alcohol intake may be protective for certain things but still toxic for others. I don't think a simple "good/bad" characterization is applicable.

[Omer Gersten]

Thank you Brad and Clyde. In the paper I'm reviewing the only information I'm given is that respondents report "never, sometimes, or frequent" alcohol use. In my mind this question is simply not subtle enough to be useful and so I would recommend to the authors that they don't include it in their models. As I alluded to before, "sometimes" drinking alcohol might include binge drinking (unhealthy), and "frequent" alcohol consumption might include a glass of wine a day at dinner time (either helpful or mildly unhealthy -- depending on what the science reveals)... Would you agree with my decision?

[Clyde Schechter]

Yes, I would agree with your recommendation.

[Phil Bromiley]

Clyde - I know nothing about alcohol measurement, but wondered if you could help me understand the general issue at hand here? As I understand it, Omer has a control variable that is measured with substantial error.

If the measure is all error, then we have no problem dropping it. But if it is somewhat correlated with the true value, how should we choose?

As I understand the general results, omitting a control variable that should be included gives bias. However, including a variable that shouldn't be included can increase error variance. If measurement error is uncorrelated with the variables, including a noisy control should give lower error variance than omission, but is clearly not as good as an accurately measured control. And, if measurement error is not random (e.g., could be correlated with included variables), then we have bias again. Please correct me if I have this wrong.

Would it be fair to say that your recommendation to drop the variable rests on an evaluation that the measure is almost complete or complete noise?

Phil

[Clyde Schechter]

Would it be fair to say that your recommendation to drop the variable rests on an evaluation that the measure is almost complete or complete noise?

Yes. That has been my experience with items that ask about alcohol frequency and don't delve into quantity.

Added: Let me elaborate on that a bit more. People who drink infrequently and irregularly, say at social occasions only, tend not to have accurate recall of their frequency, so their responses are something like a random guess. For the people who drink on a regular and frequent basis, the distribution of quantity consumed turns out to be quite wide. There are people who have a glass of wine once a day. There are people who will binge 10 drinks once a week. And the opposite combinations are about equally likely. So the frequency response in these people is nearly independent of the quantity consumed. It's just a strange thing about the patterns of alcohol consumption that frequency alone is almost entirely uninformative. (And I won't even get into the issues of socially desirable response bias--which is a different matter entirely.)