There has been a growing consensus that moderate consumption of alcohol is associated with a lower risk of mortality and that this association is probably causal. However, a recent review article has raised a serious challenge to this consensus. In short, it determined that most prior research in this area committed serious misclassification errors; furthermore, among those studies that were free of these misclassification errors, no support for a protective role of alcohol consumption was found. This article reexamines the issue using prospective data for more than 124,000 persons interviewed in the U.S. National Health Interview Surveys of 1997 through 2000 with mortality follow-up through 2002 using the Linked Mortality File. The study involves about 488,000 person-years. Controlling for a variety of covariates, this study finds that compared with nondrinkers, those who consume a moderate amount of alcohol have lower all-cause and CHD mortality. The fact that the current study has taken care to avoid the pitfalls of some earlier studies and still finds that those who consume a moderate amount of alcohol have lower all-cause mortality and CHD mortality lends credence to the argument that the relationship is causal.
At least since Pearl’s classic study in the United States in the 1920s (1926), empirical evidence has suggested that moderate alcohol consumption is associated with a lower risk of mortality. As more studies confirmed this basic finding, more researchers have concluded that the relationship is probably causal (e.g., Marmot 2001). Various refinements in methodology have been made since Pearl’s study. As discussed herein, however, recent research by Fillmore et al. (2006) has identified serious methodological shortcomings that remain in most of the previous research on the relationship between moderate alcohol consumption and mortality. The present study reexamines the relationship between moderate alcohol consumption and mortality in the United States, using recent nationally representative data, while attempting to avoid the methodological errors found in some previous studies.
Most researchers credit Raymond Pearl with realizing that we must distinguish between heavy and moderate drinkers. Pearl wrote that “moderate drinking of alcohol beverage did not shorten life. On the contrary moderate steady drinkers exhibited somewhat lower rates of mortality, and greater expectation of life than did abstainers” (1926:226). This pattern—moderate drinkers having lower mortality than both abstainers and heavy drinkers—has often been referred to a U-shaped (or, sometimes, J-shaped) relationship (Emberson et al. 2005; Marmot et al. 1981). Pearl, however, was cautious in his conclusion; he noted that moderate drinking was “not harmful.” Klatsky (2001) suggested that Pearl’s caution might have been influenced in part by the social period in which he wrote—namely, during Prohibition in the United States.
Since 1926, many additional studies have found the U-shaped relationship between alcohol consumption and mortality. Many studies focus not just on all-cause mortality, but specifically on mortality attributable to cardiovascular disease or coronary heart disease. In 1984, Marmot (1984) prepared a major review of the literature on alcohol and mortality risk attributable to coronary heart disease (CHD). By 1984, a large literature based on aggregate international comparisons, aggregate time trends, case-control studies, and (most importantly) panel studies documented that moderate drinkers tend to have a lower risk of CHD mortality than do nondrinkers. Marmot (1984) clearly believed that there was likely a causal basis to the observed association. Marmot (2001:731) later clarified that he believed that although the relationship between alcohol consumption and all-cause mortality was U-shaped, it was never clear that the relationship between alcohol consumption and CHD mortality was U-shaped. That is, while moderate consumption of alcohol was associated with reduced risk of CHD mortality, it was not clear that heavy consumption of alcohol was associated with an increased risk of CHD mortality.
In 1988, Shaper et al. (1988) proposed a critique of previous research. They argued that the higher mortality of nondrinkers might be due to the inclusion of ex-drinkers who might have given up drinking alcohol because of ill health, disability, frailty, and/or medication use. They argued, therefore, that it is important to distinguish between former drinkers and lifelong abstainers. This critique has sometimes been called the “sick quitter” hypothesis (Rimm 2001). As a result of Shaper et al.’s (1988) critique, a common practice has been to either exclude former drinkers or include them as a group separate from lifetime nondrinkers.
Another critique of the early literature (pre-1984) was that not only the amount of alcohol consumed but also the pattern of drinking is important. For example, consuming one alcoholic beverage every day of the week might have a different effect than consuming seven alcoholic beverages on one day. Recent research has therefore typically included information about the pattern of drinking (Fillmore et al. 2006). While controlling for binge drinking does not appear to have become standard practice, several researchers imply that it would be important to do so (Klatsky 2001; Marmot 2001; Thun et al. 1997).
As the basic U-shaped relationship between moderate alcohol consumption and risk of mortality persisted after these methodological critiques were leveled and addressed, more researchers have concluded that this relationship between moderate alcohol consumption and mortality is a causal relationship. Marmot wrote, “My view of the causal nature of the association was strengthened by findings subsequent to 1984” (2001:730). Bovet and Paccaud stated that “current epidemiological research supports virtually irrefutably that light-to-moderate drinking substantially reduces the risk of CHD” (2001:734). Rimm et al. wrote, “The inverse association between moderate alcohol consumption and coronary heart disease is well established. Evidence for a causal interpretation comes from over 60 ecological, case-control, and cohort studies” (1996:731).
Should alcohol be recommended as a means to reduce coronary heart disease, a major killer in the United States and other industrialized countries? Bovet and Paccaud came close to suggesting exactly this: “… the evidence currently available is large enough to guide recommendations on alcohol consumption. A net health benefit can be expected in people aged at least 40 and drinking no more than 2–3 drinks a day” (2001:736). Klatsky (2001) gave his paper the provocative title, “Could abstinence from alcohol be hazardous to your health?” Although he seems to have come close to answering his question in the affirmative, he did point out that virtually all the evidence for the health benefits of moderate consumption of alcohol comes from correlational studies rather than controlled experiments. He further reminded us, as did Bovet and Paccaud (2001) and others, that there may be health risks from even moderate consumption of alcohol, most notably fetal alcohol syndrome and certain types of cancer.
Contribution of This Study
Given the large amount of prior research on moderate alcohol consumption and mortality, it is reasonable to ask, what is the contribution of this study? One contribution is that whereas most prior studies have been on local communities (Fillmore et al. 2006:117–124), the present study is nationally representative. There have been several previous national studies in the United States, but they have some limitations. Inasmuch as both Liao et al. (2000) and Rogers et al. (2000) used National Health Interview Study data, their data are the most similar to my own. The current study uses data from a somewhat more recent time frame. Rogers et al. (2000) used data from the 1990 NHIS interviews, with follow-up through 1995; Liao et al. (2000) used data from the 1988 and 1990 NHIS interviews, with follow-up through 1995; and my study uses data from the NHIS interviews in 1997 through 2000, with follow-up through 2002. The current study is also based on a somewhat larger number of person-years than the study by Rogers et al. (2000) (about 488,000 vs. about 235,000). Liao et al. (2000) did not report the number of person-years studied. They did report that the number of cases analyzed is 43,695, and that there was “an average of 6 years of follow-up,” implying that about 260,000 person-years were involved in the study, which would be fewer than the number of person-years in my study. More significantly, although Rogers et al. (2000) used sex as a control variable, they did not analyze the two sexes separately, which is surprising because there are clearly sex differences in age-specific death rates and patterns of drinking, and likely sex differences in the effects of alcohol. Analyzing men and women separately is the norm for studies in this area, including that by Liao et al. (2000). And while Liao et al. (2000) analyzed both all-cause and CHD mortality, Rogers et al. (2000) studied all-cause mortality and circulatory mortality, which they defined as including “heart and cerebrovascular diseases, as well as hypertension” (2000:21). While Rogers et al. (2000) used both quantity of alcohol consumption and pattern of consumption, Liao et al. (2000) used only quantity of alcohol consumption in their analysis. The fact that Rogers et al. (2000) did not analyze males and females separately is a serious shortcoming. Also, the fact that Liao et al. (2000) did not take into account the pattern of alcohol use is a shortcoming. A more serious shortcoming in the analysis by Liao et al. (2000) is that although they included a variety of basic controls in the analysis of all-cause mortality, they included only age in the analysis of CHD mortality.
Interestingly, Rogers et al. (2000) did not find that moderate drinkers have a significantly lower risk of all-cause mortality. It is not clear whether analyzing the two sexes separately would have affected the results.
Another major national study of moderate alcohol consumption and mortality in the United States was that by Thun et al. (1997), based on the Cancer Prevention Study II. Their study pertained to an older time frame than mine; their initial data collection was in 1982, with follow-up through 1991. Their sample was larger (490,000 vs. 132,000) and had a longer follow-up period (nine years vs. two to six years); although they did not report the number of person-years reflected in their study, it is clearly considerably greater than the number in mine. Thun et al. mentioned some limitations of their study. One is that “[b]ecause the study subjects were recruited by American Cancer Society volunteers, they were also more likely than the general U.S. population to be college-educated, married, middle-class, and white” (1997:1706). Another limitation that they mentioned is “the lack of information on sporadic binge drinking” (p. 1713). The current study does not have these particular limitations.
As I suggest earlier, there seems to be an emerging consensus that moderate alcohol consumption is related to lower mortality and, indeed, the relationship is causative, but Fillmore et al. (2006) raised a serious challenge to this consensus. They tracked down virtually all the prospective studies of the association between alcohol use and all-cause and CHD mortality published from the 1950s until mid-2004. Fifty-four studies in 17 countries met their criteria for inclusion; all these studies examined all-cause mortality, and 35 also studied CHD mortality. Fillmore et al. (2006) found that many of these studies committed one or both of two misclassification errors: (1) failure to separate former drinkers from complete abstainers, and (2) failure to separate occasional drinkers (which I call “near abstainers”) from complete abstainers. These errors are potentially significant in light of the “sick quitters” hypothesis. They found that fully 48% of the studies made both types of errors; 39% made only the first type of error; and only 13% of the studies made neither type of error. Liao et al. (2000) did not make either type of error, but Rogers et al. (2000) made the second type of error, and Thun et al. (1997) made both types of errors. More importantly, when focusing on the small number of studies that contained neither type of error, Fillmore et al. (2006) found no significant differences between long-term abstainers and any of the categories of drinkers (former, occasional, light, moderate, and heavy) in terms of either all-cause mortality or CHD mortality.
In sum, Fillmore et al. (2006) found that the vast majority of studies had made at least one type of misclassification error, and the studies that did not make either type of error failed to find that those with moderate consumption of alcohol had a lower mortality rate than abstainers. This is a serious challenge to the emerging consensus that moderate alcohol consumption has a protective effect for all-cause mortality and/or CHD mortality. The goal of this article is to reexamine the relationship, attempting to avoid the errors that Fillmore et al. (2006) identified as well as errors that others have identified.
Although it is common to analyze men and women separately, an additional minor contribution of my study is that I include sex-interaction effects. This approach is useful because although separate analyses of men and women tell us whether alcohol use has an effect on mortality of men and of women, it does not directly test whether the effect is stronger for men or for women. The sex-interaction effects test that difference.
Studies of the effect of moderate alcohol consumption on mortality do not always analyze both CHD mortality and all-cause mortality, but it is useful to do so. If moderate drinking is protective of health, the effect probably operates primarily through a reduced risk of CHD mortality. However, moderate alcohol consumption is known to be associated with higher risks of mortality from some other causes (e.g., some types of cancer). The effect of moderate alcohol consumption on all-cause mortality might be viewed as a measure of the net effect of moderate alcohol consumption.
Materials and Methods
The data used in this article derive from the National Health Interview Survey (NHIS), an annual cross-sectional survey of the noninstitutionalized population in the United States. The NHIS uses a multistage cluster design to select a nationally representative sample of households (National Center for Health Statistics 2003). The NHIS obtains some information about each member of the sample households as well as more detailed information about one “sample adult” from each sample household. This study relies on data from the “sample adult” portion of the NHIS for 1997 through 2000. The NHIS for these years is supplemented with the Linked Mortality File, which indicates, for each adult interviewed in the NHIS, whether they died at any time on or before December 31, 2002. If a respondent died, the date and cause of death are recorded.
For all-cause mortality, participants were followed until their date of death or December 31, 2002 (the end of the follow-up period), whichever came first; those who were alive at the end of the follow-up were censored on December 31, 2002. CHD mortality was followed up in the same way except that deaths from causes other than CHD were censored at their date of death.
The data set contains a weight variable that provides a post-stratification adjustment based on age, race/ethnicity, and sex, using the census population as the control. The weight variable is used in the analysis.
To maintain subject confidentiality, the NCHS Research Data Center linked the NHIS data to the mortality data and performed the computer analyses using SAS.1
Adults who reported having at least 12 drinks in their lifetime were asked, “In the past year, how often did you drink any type of alcoholic beverage?” The responses were coded in terms of the number of days per week that they had an alcoholic beverage; this could range from 0 to 7. Those who reported drinking in the past year were asked, “In the past year, on those days that you drank alcoholic beverages, on the average, how many drinks did you have?” I used the responses to these two questions to code respondents into one of nine categories. There are three categories of nondrinkers and six categories of drinkers.
The three categories of nondrinkers are (1) “lifetime abstainers”, who had fewer than 12 drinks in their lifetime; (2) “near abstainers”, who had 12 or more drinks in their lifetime, but not 12 in one year; and (3) “former drinkers”, who had 12 or more drinks in one year, but not in the past year.
“Drinkers”—those who had 12 or more drinks in the past year—are divided into six categories based on both the frequency of consuming alcohol and the number of drinks usually consumed on days when they do drink alcohol. Those who usually drink on zero days in a given week are considered infrequent drinkers; those who usually drink on one or two days in a given week are considered occasional drinkers. Those who drink on three or more days in a given week are considered frequent drinkers. Those who usually consume one or two drinks on days on which they consume alcohol are considered moderate drinkers, and those who usually consume three or more drinks on days on which they consume alcohol are considered heavy drinkers.
Combining frequency and amount of drinking produces six categories of drinkers: (4) “infrequent moderate drinkers” consume an average of one or two drinks on days when they drink, but usually consume alcohol zero days per week (recall that all “drinkers” consumed at least 12 drinks in the past year); (5) “occasional moderate drinkers” consume an average of one or two drinks per day, on one or two days per week (this is the reference category); (6) “frequent moderate drinkers” consume one or two drinks per day, three or more days per week; (7) “infrequent heavy drinkers” consume three or more drinks on days when they drink alcohol, but they usually consume alcohol zero days per week; (8) “occasional heavy drinkers” consume three or more drinks per day, one or two days per week; and (9) “frequent heavy drinkers” consume three or more drinks per day, on three or more days per week.
A separate question, used for some analyses, asked, “In the past year, on how many days did you have five or more drinks of any alcoholic beverage?”
Regarding the validity of self-reports of alcohol consumption, Lang et al. suggested that they “… are generally regarded as valid for classification of individuals into broad consumption bands” (2006:54). Marmot (1984) argued that any inaccuracies in alcohol reporting probably do not substantially affect the correlation between alcohol and CHD mortality.
There has been considerable discussion of whether the type of beverage (wine, beer, or liquor) affects the beneficial effects of alcohol. In an extensive and widely cited article that reviewed 57 studies and focused on precisely this question, Rimm et al. (1996:731) concluded, “Results from observational studies, where alcohol consumption can be linked directly to an individual’s risk of coronary heart disease, provide strong evidence that all alcoholic drinks are linked with lower risk.” Consistent with Rimm et al.’s conclusion, Paganini-Hill et al. (2007:207) “found no difference in the effects of wine, beer, or hard liquor on mortality.” Similarly, Klatsky (2001:741) stated the wine/beer/liquor issue is unresolved, but added, “There is evidence of benefit from each beverage type.” Furthermore, Marmot’s view is that “it is the alcohol in wine that is protective rather than other substances” (2001:731).
A related issue is that the amount of alcohol per ounce obviously varies across beverage types. Also, the size of a typical serving varies across beverage types. The net result is that the total alcohol content is roughly the same, if one consumes one glass of wine, one can or bottle of beer, or one shot of liquor (Bovet and Paccaud 2001).
A discussion of the biochemical explanations of why those who have moderate consumption of alcohol tend to have a lower risk for death, especially from coronary heart disease, is beyond the scope of this paper; for a review, see Agarwal (2002). Genetic factors affect individual variation in the rate of alcohol metabolism; there are three variations of the gene that affects the rate of alcohol metabolism (Hines et al. 2001). Hines et al. (2001) found that moderate alcohol consumption is associated with a decreased risk of myocardial infarction and increased HDL levels (“good cholesterol”), especially for those whose genes cause them to metabolize alcohol relatively slowly. In their discussion, Hines et al. (2001) presented several arguments as to why their findings lend support to the plausibility of a causal interpretation of the association between moderate alcohol consumption and the reduced risk of myocardial infarction.
Covariates include age (14 categories: 18–24; five-year age groups for ages 25–29 through 80–84; and 85 or older), race/ethnicity (non-Hispanic white, non-Hispanic black, Hispanic, and other), education (no high school, some high school, high school graduate, some college, college graduate, and graduate or professional degree), marital status (currently married, cohabiting, widowed/divorced/separated, and never married), work status (worked last week, worked last year but not last week, and did not work last year), smoking status (current daily smoker [fewer than 20 cigarettes], current daily smoker [20 or more cigarettes], current smoker some days, former smoker, and never smoked), and income (11 categories, ranging from less than $5,000 to $75,000 and higher). Except for income, which was analyzed as a continuous variable, each covariate was entered as a series of categorical variables.2
Self-reported health is viewed as an important indicator of health and has been documented to be an important predictor of subsequent mortality, even controlling for known health risk factors, including data from medical records (Benjamins et al. 2004; Idler and Benyamani 1997; Mossey and Shapiro 1982). However, self-reported health is rarely used as a covariate in studies of the effect of moderate alcohol consumption on mortality; Fillmore et al. (2006) reported that only 24% of studies of all-cause mortality and 4% of studies of CHD mortality employ a measure of “general health” as a covariate. In this study, self-reported health is included as a covariate. In the NHIS, the self-reported health of sample adults is categorized as excellent, very good, good, fair, or poor. The responses were recoded so that poor = 1, and excellent = 5.
It seems likely that those who have previously been diagnosed with a heart condition would have a greater risk of CHD mortality. Thus, it is useful to control for whether a person has been given such a prior heart-related diagnosis. However, studies examining mortality as related to moderate alcohol consumption usually do not use a prior diagnosis of heart problems as a covariate; Fillmore et al. (2006) reported that only 24% of studies of all-cause mortality and 17% of studies of CHD mortality used “past heart problems” as a covariate. Sample adults in the NHIS were asked, in four separate questions, whether they had ever been told by a doctor or other health professional that they had (1) coronary heart disease, (2) angina (“also called angina pectoris”), (3) a heart attack (“also called myocardial infarction”), or (4) any (other) kind of heart condition or heart disease. A dummy variable was created in which all respondents who answered “yes” to any of these four questions were coded 1, and all others were coded 0. Inasmuch as coronary heart disease is a major cause of death, I included this variable (prior CHD diagnosis) as a covariate in the analyses of both CHD mortality and all-cause mortality.
Inasmuch as it is well established that binge drinking is harmful to health, it is useful to take into account binge drinking. Here, “binge drinking” is defined as consuming five or more alcohol drinks in one day. For the purposes of this article, anyone who engaged in binge drinking on three or fewer days in the 365 days prior to the interview is regarded as “rarely binge drinking.” In some analyses, I restrict the sample to those who rarely binge drink.
For each respondent, I calculated the number of days between the interview and either the date of death or the end of the follow-up period (i.e., December 31, 2002). For the analysis for CHD-related deaths, I used deaths attributable to ischemic heart diseases (codes I20–I25 in ICD-10). This is coded as 058 through 063 in the Underlying Cause of Death Recode: 113 groups.3
I use Cox proportional hazards models (Yamaguchi 1991) to examine the effect of patterns of alcohol consumption on all-cause and CHD mortality. Because the NHIS is based on a complex sample design, I use the Survival procedure in SAS-callable SUDAAN Release 10.0.0; this procedure takes the sample design into account in calculating the standard errors.
I analyze all available cases. However, because relatively few people aged 18–49 died during the observation period, I replicate the analyses restricting the sample to those aged 50 and older. To control for possible confounding effects of binge drinking, I also replicate the analyses for those who rarely binge drink.
Of the 131,731 sample adults interviewed as part of the NHIS in 1997 through 2000, the final mortality status was determined for a total of 124,483 persons. During 178,282,307 person-days of follow-up (488,110 person-years), there were a total of 5,846 deaths, including 1,221 deaths from coronary heart disease.
Table 1 shows the bivariate distribution of individuals across the nine categories of drinking status by various background characteristics.
As is well known, women tend to drink less than men; in particular, women are disproportionately likely to be found in three categories: lifetime abstainer, near abstainer, and infrequent moderate drinker.
Infrequent and occasional heavy drinkers tend to be relatively young (median age of 30 or 31 years). Infrequent and occasional moderate drinkers tend to be older than infrequent and occasional heavy drinkers (median age about 40). Nondrinkers tend to be somewhat older than the four categories just referred to (median age about 45 to 52).
Non-Hispanic whites tend to drink more than individuals in the other three categories of race/ethnicity. In particular, non-Hispanic whites are more likely than those in other race/ethnicity categories to be frequent heavy drinkers, and they are much less likely to be lifetime abstainers. Non-Hispanic whites are also more likely than non-Hispanic blacks and others to be occasional heavy drinkers. It may be advantageous to non-Hispanic whites that they are also more likely to be occasional or frequent moderate drinkers.
Those with less education (i.e., high school graduate or less) are more likely to be lifetime abstainers or near abstainers. Those who did not attend high school are also slightly more likely to be former drinkers. Those who have a college degree or advanced degree are somewhat more likely to be in any of the moderate drinker categories. High school dropouts, high school graduates, and those who completed some college are more likely to be heavy drinkers, usually infrequent or occasional heavy drinkers rather than frequent heavy drinkers.
Compared with those in other marital statuses, cohabiters are less likely to be lifetime abstainers. Cohabiters and the never married are more likely than the married and the previously married to be heavy drinkers (usually infrequent or occasional heavy drinkers, rather than frequent heavy drinkers).
The drinking patterns of those who worked last week are similar to the drinking patterns of those who did not work in the last year; by contrast, those who worked in the last year but not in the last week are less likely to be nondrinkers and are correspondingly more likely to be drinkers, especially infrequent or occasional heavy drinkers.
Compared with nonsmokers, current smokers are more likely to be heavy drinkers. Smokers and nonsmokers are roughly equally likely to be moderate drinkers. Those who never smoked are the most likely to be lifetime abstainers from alcohol.
The median family income for moderate drinkers (all three categories) tends to be higher than that for heavy drinkers (all three categories), which in turn tends to be higher than that of nondrinkers (all three categories).
Among drinkers, patterns of alcohol consumption are only weakly associated with self-reported health. However, nondrinkers tend to have poorer self-reported health than drinkers. Consistent with the “sick quitter” hypothesis, former drinkers have the poorest self-reported health.
Interestingly, compared with drinkers, nondrinkers are more likely to have reported a CHD-related diagnosis at the time of the interview. Again consistent with the “sick quitter” hypothesis, former drinkers are the most likely to have a prior CHD-related diagnosis. Heavy drinkers are least likely to report a prior CHD-related diagnosis.
Mortality Effects for Females
The analyses of mortality related to coronary heart disease (“CHD mortality”) for females are shown in Table 2. As noted earlier, occasional moderate drinkers are defined as those who consume alcoholic drinks on one or two days per week and have one or two drinks on days when they drink.
The focus in this article is the comparison between occasional moderate drinkers and nondrinkers. The first nine coefficients in column 1 of Table 2 report the hazard ratios for women in various alcohol consumption categories, compared with their counterparts who are occasional moderate drinkers. All three of the groups of nondrinkers—lifetime abstainers, near abstainers, and former drinkers—have a significantly higher risk of CHD mortality than occasional moderate drinkers. The hazard ratios are fairly similar for the three categories of nondrinkers: 1.94, 2.15, and 2.09, respectively. In other words, the risk of CHD mortality is roughly twice as high for female nondrinkers as for their female counterparts who are occasional moderate drinkers. The “sick quitters” hypothesis suggests that former drinkers may have stopped drinking because of health problems, and this might explain their higher mortality. However, among women, lifetime abstainers and near abstainers also have a higher risk of CHD mortality, compared with occasional moderate drinkers. By definition, lifetime abstainers and near abstainers never quit drinking—for health reasons or any other reason. The fact that women who are occasional moderate drinkers have a lower risk of CHD mortality than comparable women who are lifetime abstainers or near abstainers is consistent with the idea that moderate drinking is protective of health. When we restrict our attention to those who (a) are aged 50 or older (column 3) or (b) rarely binge drink (column 5), the results are similar: that is, the respective hazard ratios are significant and of similar magnitude.
As expected, males have a significantly higher risk of CHD mortality; the hazard ratio is roughly 2.9 (Table 2, column 1, line 10). However, none of the sex interaction terms are significant, implying that the effect of alcohol consumption on CHD mortality is similar for males and females.
Turning to all-cause mortality among women, all three groups of nondrinkers have a significantly higher risk of all-cause mortality than do their female counterparts who are occasional moderate drinkers (Table 3). The hazard ratios are fairly similar for the three categories of nondrinkers: 1.61, 1.72, and 1.90, respectively. These hazard ratios are not quite as high as those found for CHD mortality. Nevertheless, among women, the net effect of occasional moderate consumption of alcohol seems to be beneficial. Looking again at those who (a) are aged 50 or older or (b) rarely binge drink, the results are similar: that is, the respective hazard ratios are significant and of similar magnitude. As was the case with CHD mortality, men have a higher risk of all-cause mortality; the hazard ratio is about 2.0 (Table 3, column 1, line 10). As I’ll discuss later, some of the sex-interaction terms are significant.
Mortality Effects for Men
Turning to the effects of occasional moderate alcohol consumption on CHD mortality for males, the results are not as clear-cut as they are for females. The hazard ratios for lifetime abstainers (1.56) and for near abstainers (1.42) are significant, but the hazard ratio for former drinkers (1.34) is not significant at the conventional .05 level of significance (see Table 4). Among men who are aged 50 or older, the hazard ratio for lifetime abstainers is significant (1.46), but the hazard ratios for near abstainers (1.34) and for former drinkers (1.28) are not significant at the .05 level. Among those who rarely binge drink, the hazard ratios for lifetime abstainers and near abstainers are significant (1.56 and 1.45, respectively), but the hazard ratio for former drinkers (1.36) is not significant.
In other words, whether occasional moderate alcohol consumption has a significant effect on the risk of CHD mortality among males depends to some extent on the category of nondrinker and the subpopulation on which we focus. It is worth noting, however, that the hazard ratios for lifetime abstainers are significant for all cases and for both subpopulations, and that the hazard ratios for near abstainers are significant for all cases and for those who rarely binge drink. On the other hand, none of the three hazard ratios for former drinkers are significant. So, using conventional two-tailed tests at the .05 level, five of the nine hazard ratios are significant (the nine hazard ratios are for all three groups of nondrinkers for the full population and for both subpopulations). If one-tailed tests, rather than two-tailed tests, are used, two additional hazard ratios are significant (i.e., all cases for former drinkers [1.34] and former drinkers who rarely binge drink [1.36]).4 One-tailed tests might be justified because the literature clearly specifies the direction of the effect. The two nonsignificant hazard ratios are found among males aged 50 or older, where the risk of CHD mortality for near abstainers and former drinkers is not significantly different from that of occasional moderate drinkers, even with a one-tailed test.
Turning to all-cause mortality among men, shown in Table 5, the hazard ratios for near abstainers (1.30) and for former drinkers (1.33) are significant; the hazard ratio for lifetime abstainers (1.19) reaches borderline significance (p < .07). These three hazard ratios for all-cause mortality are smaller than the corresponding hazard ratios for CHD mortality. It may be surprising that the hazard ratio for all-cause mortality for former drinkers is significant while the hazard ratio for CHD mortality for the same group is not significant at conventional levels. The explanation probably lies in the fact that the number of CHD deaths is much smaller than the number of all-cause deaths, and hence the standard errors are larger in the analysis of CHD mortality.
Among men who (a) are aged 50 or older or (b) rarely binge drink, the hazard ratios for all three groups of nondrinkers are significant at the .05 level.
As noted earlier, none of the sex-interaction terms for CHD mortality are significant, meaning that the effects of membership in any given alcohol consumption category are comparable for both males and females. However, for all-cause mortality, several of the sex-interaction terms are significant. Which terms are significant varies somewhat, depending on whether one considers the whole sample or one of the subsamples. In terms of the whole sample, all three sex-interaction terms for nondrinkers are significant. The magnitude of these sex-interaction terms (about 1.3 in Table 5) indicates that females who are either lifetime abstainers, near abstainers, or former drinkers have a higher risk of all-cause mortality, compared with their male counterparts. Focusing on those who are 50 or older, two of the sex-interaction terms for nondrinkers are significant; compared with their male counterparts, women aged 50 or older who are either lifetime abstainers or former drinkers have a higher risk of all-cause mortality. Focusing on those who rarely binge drink, all three sex-interaction terms for nondrinkers are significant.
The Effects of Heavy Drinking
It is well known that excessive consumption of alcohol is an important health behavior factor related to mortality (Mokdad et al. 2004). Nothing in this article should be construed as challenging the consensus that excessive consumption of alcohol contributes to mortality. The focus of this article, by contrast, is to investigate the effect of moderate alcohol consumption. The attention has been on the contrast between occasional moderate drinkers and nondrinkers. But what about the effect of heavy drinking (i.e., consuming three or more alcoholic drinks on days on which one drinks)? Reviewing the hazard ratios for the three categories of heavy drinkers in Tables 2 through 5 shows two patterns: (1) the hazard ratios are usually greater than 1.0, implying that the risk of death for heavy drinkers is greater than the risk of death for their same-sex counterparts who are occasional moderate drinkers; but (2) these hazard ratios are usually not significant. In general, among drinkers, I find significant hazard ratios only among frequent heavy drinkers (those who usually drink three or more drinks per day, three or more days per week); even then, the significant hazard ratios are found only for all-cause mortality. Compared with occasional moderate drinkers, frequent heavy drinkers do face a higher risk of all-cause mortality; the hazard ratios are higher for women (about 2.0) than for men (about 1.25). This is consistent with the findings of other researchers that excessive consumption of alcohol is detrimental to health. It is noteworthy, however, that the risk of CHD mortality for heavy drinkers is not significantly higher than it is for occasional moderate drinkers (see Marmot 2001).
Summary and Discussion
There has been a growing consensus that moderate consumption of alcohol is associated with a lower risk of mortality and that this association is probably causal. Fillmore et al. (2006) identified serious misclassification errors in most of the existing prospective studies of alcohol use and mortality, thereby casting doubt on this growing consensus. Using nationally representative data for the United States, the present study has attempted to avoid the methodological errors that the literature has identified. In particular, the present study avoids the misclassification errors identified by Fillmore et al. (2006); examines both all-cause and CHD mortality; incorporates a number of important covariates in the analysis of both all-cause and CHD mortality; takes into account not only the quantity of alcohol consumption but also the pattern of use; and examines the effect of alcohol use on mortality separately for men and women.
This study finds that compared with nondrinkers, women who consume a moderate amount of alcohol have lower all-cause and CHD mortality. Among men, the pattern is not quite as consistent. In most comparisons, however, the risk of both types of mortality is lower for males who are occasional moderate drinkers than it is for their male counterparts who are nondrinkers.
The sex-interaction terms for CHD mortality indicate that for nondrinkers, the risks of mortality are elevated by a similar amount for both males and females. The sex-interaction terms for all-cause mortality indicate that for nondrinkers, the risks of mortality are elevated to a greater extent for females than for males.
It seems likely that, in general, the health behaviors of lifetime abstainers, and even of near abstainers, may be better than those of drinkers (Fraser 1999; Kent and Worsley 2008; Minugh 1998). Although I control for smoking behavior, it would clearly be better to control for additional health behaviors as well. However, this suggests that my test of the effect of moderate alcohol consumption is a conservative test. That is, if I had controlled for more health behaviors, the likely effect is that the mortality difference between occasional moderate drinkers and abstainers would have been greater than that observed.
Another limitation is the relatively short follow-up period, only two to six years. A longer follow-up period would obviously yield a larger number of person-years and would likely yield a larger number of deaths, which in turn would likely lead to more precise estimates.
Although most prior studies of the effect of moderate alcohol consumption on mortality have used community samples, the analysis in this article is based on a large, nationally representative data set. Whereas not all previous studies in this area have examined both CHD mortality and all-cause mortality, this study analyzes both. Although some studies have categorized drinkers according to only the frequency or quantity of drinks, this study categorizes drinkers according to both the frequency and quantity of their drinking. And whereas most previous studies in this area have made classification errors with respect to nondrinkers, this study avoids such classification errors. In addition, a larger than usual number of covariates is included in the analyses. In particular, self-reported health and having a prior CHD-related diagnosis are included as covariates here, whereas neither one is typically included. Finally, in calculating the standard errors, I take into account the complex sample design.
This article presents strong evidence that occasional moderate alcohol consumption has a protective role in health. The fact that the current study has taken care to avoid the pitfalls of some earlier studies and still finds that those who consume a moderate amount of alcohol have lower all-cause mortality and CHD mortality lends credence to the argument that the relationship is causal.
I would like to thank Margaret Gaines for computational assistance on this paper.
I use the restricted-use data set. There is a corresponding public-use data set. With the public-use data set, an analyst can obtain the data set to be analyzed at his or her own institution. Owing to confidentiality concerns, however, there are some disadvantages to using the public-use data set. First, the NCHS used “standard data perturbation techniques that introduce statistical noise into the data set” (Lochner et al. 2007:2). Second, “synthetic data were substituted for the actual date and underlying cause-of-death data for selected decedent records” (Lochner et al. 2007:2). Third, while the restricted-use data set reports date of death, the public-use data set reports quarter of death. Given the limitations of the public-use data set, I opted to use the restricted-use data set. For more details, see Lochner et al. (2007).
Not all families reported their income. To deal with missing data, the National Center for Health Statistics (NCHS) used multiple-imputation methodology to create five imputed estimates of family income for each family. As recommended in NCHS documentation, each of the analyses reported in this article was replicated five times, once using each imputation of family income, and the results were then combined. More details are provided by NCHS at http://www.cdc.gov/nchs/about/major/nhis/2000imputedincome.htm and the Technical Documentation linked to that site.
These are found in the document titled “Underlying and Multiple Causes of Death September 2006” located at http://www.cdc.gov/nchs/data/datalinkage/h1_underlying_and_multiple_causes_of_death.pdf.
Former drinkers who rarely binge drink are the same group as all former drinkers; however, the comparison groups are different (i.e., occasional moderate who rarely binge drink versus all occasional moderate binge drinkers).